Here’s your weekly dose of AI and insight.

It’s been a big week with Google teaching Gemini to clone your face, Meta agreeing to power its direct rival to the tune of $10 billion, and China taking #1 from Claude for the first time. A clear pattern continues to emerge that the AI race is no longer just about who builds the best model anymore. Winning centres on ownership of the compute layer underneath it all and on who’s willing to give intelligence away to win the long game.

If you’re enjoying this kind of analysis and want to support my work, become a paid subscriber today. You'll get access to the full archive of The Signal and 6 months free of Wispr Flow, my most-used AI tool, so you can dictate instead of type. If you’re a paid subscriber, reply to this email for access.

Upgrade to Pro

Today’s Signal is brought to you by HubSpot.

When a buyer asks ChatGPT for the best option in your category, it comes back with a handful of brands. HubSpot AEO shows you whether you’re one of them.

→ Brand visibility score and competitor share of voice
→ Prompt tracking and citation analysis in one place
→ See how often you appear, in what context, and what to fix

Free for 28 days, with 25 prompts tracked across ChatGPT, Gemini and Perplexity.

Start your free trial

AI Highlights

My top-3 picks of AI news this week.

Google

1. Google Clones You

Google dropped a wave of Gemini updates this week, pushing its AI assistant deeper into every corner of its ecosystem, from avatars to appearing in Chrome.

• Personal avatars: You can now record your face and voice to create an avatar, then reference yourself in prompts to generate images and videos starring you. Available to adults on Google AI plans, though not yet in the UK or EU.

• Spark expands: Google’s 24/7 background agent Spark is rolling out to AI Ultra subscribers in more countries and languages. It handles multi-step tasks across Gmail, Calendar and Docs under your direction—running on schedules or triggered by conditions you set.

• NotebookLM renamed: NotebookLM is now Gemini Notebook—same product, new name, with now over 30 million users. It syncs across the Gemini app and Search, and gains code execution for deeper data analysis grounded in your sources.

Alex’s take: Gemini users now have to pay a premium to access Spark. It’s only currently available to AI Ultra subscribers, a plan that’s 5x more expensive ($99.99) than its junior AI Pro counterpart ($19.99). However, with that premium, you get the ability to orchestrate your entire workspace in one place, making it far more useful than using Gemini in standalone apps. Now, with Spark, you can do things like setting up scheduled tasks that comb through online job opportunities and send you a direct summary report. Or even track when you receive an email inquiring about your photography services, automatically extract the client’s name, requested date, and log the lead in a “Client Tracker” Sheet. This feels like Google’s response to Claude Cowork and ChatGPT Work. And long term, I think it’ll be the most potent, as you have a direct line into all of your workspace apps—Meets, Docs, Drive, Sheets, Tasks, Gmail, Calendar, the list goes on—and the efficiency of having everything together vs connecting externally will remove any friction a user has for orchestrating their work tools within Claude or ChatGPT today. Google is catching up.

Meta

2. Meta Powers Anthropic

Meta is in early talks to lease AI computing power to Anthropic in a deal worth up to $10 billion over two years, which would turn the social media giant into an infrastructure supplier for its direct rival.

• Deal structure: Anthropic proposed the deal in June and would pay Meta in monthly instalments—roughly $417 million a month—with early-exit options on both sides.

• Cloud ambitions: Zuckerberg said in May that entering cloud computing was “definitely on the table,” and Meta has hired 18-year AWS veteran Dave Brown to lead the effort. Its $145 billion infrastructure spend this year is more than double last year’s.

• Familiar playbook: This mirrors Anthropic’s May agreement with SpaceX, which rents the Claude maker access to its Colossus supercomputers in Memphis for around $1.25 billion a month.

Alex’s take: “He who controls the spice controls the universe.” It’s a legenday quote from Baron Harkonnen in the film adaptation of Dune. In the novel, the foundational truth is that the empire relies entirely on the spice supply from the desert planet Arrakis to function. I think it perfectly summarises this next phases of the AI saga, where the owners of compute (SpaceX, Google, Amazon, and now Meta) increasingly set the direction of where the frontier models are heading. Compute is useful for two things. First it’s for training these AI models. Second, it’s for inference, the proces of getting answers back from your prompts. Both are necessary to control if you are to build and run the leading intelligence on the planet. In Dune, spice drives all interstellar travel and political power. In the AI race, compute drives model progress, product speed, and ultimately who wins—and the landlords collect rent whichever model comes out on top, even better if its their own.

Thinking Machines

3. Thinking Machines Opens the Vault

Thinking Machines Lab, the startup founded by former OpenAI CTO Mira Murati, has released Inkling, its first open-weights model, with the full weights free for anyone to download, customise, and run.

• Frontier scale, fully open: Inkling is a 975-billion-parameter model (41 billion active per task) that reasons across text, images, and audio, released under a permissive Apache 2.0 licence—one of the largest open models ever from a US lab.

• Built to be customised: The company openly says Inkling isn’t the strongest model available. The design goal was a well-rounded base that businesses fine-tune into specialists on Tinker, its customisation platform, which is where the revenue comes from.

• Adjustable thinking: A controllable “reasoning effort” dial lets users trade answer quality against speed and cost, from quick responses to deep analysis.

Alex’s take: I think it’s a smart play to give away their model for free on the surface. From the outset it might seems mad at first, but when you look at the economics, it makes total sense. Frontier labs today charge for access—per use or via a consumer subscription. Thinking Machines charges for customisation. They earn when companies fine-tune on Tinker. I’m a big believer that fine-tuning open models will win in the long run. Bridgewater has already done so on its own financial data and beat proprietary models at ~1/14th of the cost. Business want models they control, trained on their data, running on their infrastrucutre, not paying the silent tax that Alex Karp articulated earlier this month in the form of handing over their data—for free—to these frontier firms to build competing products against. I think within two years most serious companies will run a customised open model alongside their AI-of-choice subscription—and the labs selling generic intelligence by the token should be nervous.

Content I Enjoyed

A framework for frontier AI and the dawning of a new age

Alex’s take: Demis Hassabis published an essay this week arguing that AGI is probably a few years away and its impact could be 10x the Industrial Revolution at 10x the speed. What I found most interesting was his concrete proposal for a self-regulatory organisation (SRO) for frontier AI modelled on FINRA, the body Wall Street set up to police itself under government oversight. Labs would submit frontier models for review 30 days before release, voluntarily at first, against benchmarks covering cyber, biological and other catastrophic risks. To see why every major lab leader endorsed it within days, compare the alternative Dario Amodei keeps asking for: an FAA for AI. The FAA takes five to nine years to certify a new aircraft design. The labs are backing the lightest oversight on offer, which makes sense, but the catch is that an SRO only pays off if it replaces other regulation. If states and federal agencies keep writing their own AI rules on top of it, the labs have conceded oversight and gained nothing in return.

Idea I Learned

China Just Took #1

Alex’s take: For the first time, a Chinese model sits at #1 on LMArena's Frontend Code Arena. Kimi K3, built by Moonshot AI, topped the leaderboard with 1,679 points—beating Claude Fable 5, ranking first in six of seven categories, and jumping 17 places from its predecessor’s #18. David Sacks called it "concerning" and blamed regulation; funnily enough his replies kept asking which regulation stopped OpenAI or Anthropic from staying ahead. K3 costs $3 per million input tokens and $15 for output, against Claude Fable 5's $10 and $50—and on July 27, Moonshot releases the full weights, 2.8 trillion parameters, the largest open model ever for anyone to use. The same week, Xi Jinping used the World AI Conference to launch a 29-country AI cooperation body and pitch open, cheap Chinese AI to the developing world. Whilst American intelligence is rented, Chinese intelligence is yours to keep, no matter the circumstance. DeepSeek was dismissed as a one off. Kimi K3 makes it a pattern that can’t be ignored.

Quote to Share

a16z’s Justine Moore on Anthropic’s aura problem:

Alex’s take: Anthropic has spent six weeks changing its mind and shuffling timelines about for who gets Fable 5 and for how long. The model launched on 9 June, went dark for 19 days under US export controls, relaunched on 1 July, then had its cutoff date pushed back twice before this weekend's announcement finally settled things. Half of usage limits on top-tier plans, pay-as-you-go for everyone else. Anthropic says demand has been "challenging to predict", and sure, I believe it, but repeated deadline shuffling turns a capacity problem into a credibility one. However, we must remember that the memory of people in tech is akin to that of a goldfish. People forget what happend last week once enamoured with the latest product, model, or partnership release, let alone what happened last month. Given that the quality of intelligence is quickly converging, goodwill, it turns out, is an overwhelmingly large portion of your subscription too.

Question to Ponder

Alex’s take: Two-thirds of Americans believe a new data centre nearby would push up their electricity bill. But the evidence is actually pointing the other direction. A new working paper from the Electric Power Research Institute found data centres pushed US residential rates modestly down between 2015 and 2024, even after accounting for developers favouring states where power was already cheap. Utilities recover vast fixed costs across every kilowatt-hour sold, so a large, steady customer pulls the average down for everyone. The sharpest rises have come in states with aggressive climate policies, not the most data centres. Virginia, with more data centres than anywhere in America, sits near the national average. But there’s an important caveat which the authors touch on that I think is important to highlight. The past decade’s growth has absorbed slack in the grid, and that slack is essentially now running out. Policy has also answered this month via the POWER Act with Oregon’s approval of PGE’s 29.7% rate rise for data centres whilst cutting residential rates by 1.3%. Let’s see if the other 49 states are paying attention.

Already a subscriber? Get your whole team on board. Signal Pro group subscriptions give everyone access to weekly AI workflows and tutorials. Share this with your manager today.

Get a group subscription

💡 If you enjoyed this issue, share it with a friend.

Refer a friend

See you next week,
Alex Banks

P.S. One robot kicks another robot's head off.

Welcome back to the Signal Pro.

You’ve probably noticed that AI never quite repeats itself. Ask GPT-5.6 Sol or Claude Fable 5 a simple factual question: what’s the tallest mountain in the world, for instance, and you’ll get Everest every single time. But hand either of them something open-ended, a draft or a block of code, and a noticeably different result will come back each time. The same input on two different days gets you two outputs that sure do rhyme without ever really matching.

Engineers call this “non-determinism”. It’s the same input refusing to produce the same output. Most people see this as a quirk and move on. But I think that’s a mistake, as the quirk itself has a specific cause that was only properly pinned down last September.

Part of the explanation is that the answer you get depends on how many other people happen to be talking to the model at the same moment as you.

Let me show you the research, because once you see why this happens, you can start using it to your advantage.

If you’re enjoying these issues and want to support my work, become a paid subscriber today. You'll get access to the full archive of The Signal and 6 months free of Wispr Flow, my most-used AI tool, so you can dictate instead of type.

Upgrade to paid

One prompt, 80 answers

In September 2025, Horace He and his colleagues at Thinking Machines Lab, the research company founded by former OpenAI CTO Mira Murati, tried something simple.

They took Qwen3, Alibaba's 235-billion-parameter open-source model, turned the randomness off completely, and sent it the same prompt 1,000 times: "Tell me about Richard Feynman." They got back 80 different answers.

The strange part is where they split. All 1,000 were word-for-word identical for their first 102 tokens (the word-sized chunks these models write in). Then at token 103, describing where Feynman was born, 992 wrote “Queens, New York” and 8 wrote “New York City”. From that single fork, the answers never converged again.

To make sense of that, you need to know how a model writes.

How the model picks each word

When a model answers you, nothing is being looked up. It writes one word at a time, and at each word, it holds a ranked list of candidates with a probability attached to each. Ask for the tallest mountain, and “Everest” towers over everything else, so the choice is simple. Ask it to draft your proposal, and hundreds of decent options sit within touching distance of each other.

The model doesn’t always take the top word, though, because text written that way comes out flat and repetitive. So it rolls a weighted die that’s biased towards the favourites. The likely words win most of the time, and every so often, a runner-up sneaks through. ChatGPT and Claude do this on purpose, and that’s exactly what the “temperature” setting controls.

Temperature is easiest to think of as the model’s creativity setting. At zero, creativity is turned off. The model plays it completely safe and takes the top word every single time, which makes it as predictable as it can be. The usual model default sits around one, where likely words come up often, and long shots come up occasionally. If you push this number up higher, the model gets more adventurous. Long shots start winning more often, and the writing gets more surprising until it eventually tips into nonsense.

All the setting really does is loosen the odds on that die roll. Claude and ChatGPT don’t show you the control, but developers use it a lot of the time, favouring a low temperature when the output needs to be repeatable, like pulling figures out of a document, and a higher temperature when they want range, like brainstorming.

Temperature only matters where there's a genuine choice to make. When one word towers over the rest, no roll of the die changes the result, which is why settled facts come back identical every time. When a few words are roughly tied, the die decides. And an open-ended document draft is full of ties like that, from beginning to end.

The last piece is a feedback loop. Every word the model picks gets folded back into the prompt before it chooses the next one, so one different word early on means everything after it gets written from a different starting point. That’s why the Feynman answers never came back together. One small swerve at token 103 and the two versions had parted ways for good, each perfectly coherent, but never again converging.

Blame your server neighbours

In the Feynman experiment, the temperature was set to zero. At zero, the model takes the top word every single time, so the same prompt should have come back identical 1,000 times. It didn't. The Thinking Machines team worked out why.

The cause of this surrounds the idea of rounding. Everything you type gets turned into numbers before the model sees it, because numbers are the only thing a computer can work with. The model's job from there is one long chain of sums on those numbers, billions at a time. To keep that fast, each number gets a small, fixed slot of memory, and anything past those digits gets rounded off. Add up a long list of numbers like that, and the total comes out slightly different depending on the order in which the additions happen. Same numbers, different order, slightly different total.

You, as the user, ultimately don’t get to control the order. When you hit enter, your prompt gets processed in a batch with everyone else's requests arriving at that moment, and the size of the batch sets the order in which the additions happen. Usually, the tiny differences change nothing, and you get the same word anyway. Occasionally, two candidate words are close enough that the rounding decides which one wins, and every word after that gets written from a different starting point. A busy Tuesday afternoon is different from a quiet Sunday morning. The answer you get depends on how many other people are talking to the model at that moment.

Thinking models make things even worse. The same rounding differences happen inside the model's reasoning steps, and each step feeds the next, so small differences grow as the reasoning runs on. A NeurIPS study from last year measured one reasoning model swinging by up to 9% in accuracy and 9,000 tokens in length across runs that should have been identical.

And it turns out this shows up in ordinary work too, not just in lab conditions. Another research team ran five models on eight everyday tasks ten times each, all with deterministic settings, and accuracy drifted by as much as 15% between runs. Not one model produced identical outputs across the board.

Code is where non-determinism is most prominent. With written prose, you can change a few words and still say the same thing, but in code, one changed token can change what the program does. A team ran 829 coding problems through ChatGPT five times each, and on half to three quarters of them (it varied by benchmark) no two runs produced matching test results. The same task could pass every test on one run and fail every test on the next, and setting the temperature to zero reduced the variation without removing it.

If you're thinking of image generators like Nano Banana and GPT Images 2, they vary as well, but for a different reason. Each image starts from random noise (diffusion) on purpose. This is randomness by design.

The model is holding more than one answer

Read more

Hey friends 👋 Happy Sunday.

Here’s your weekly dose of AI and insight.

Midweek, Signal Pro members get a step-by-step AI workflow they can apply immediately. No fluff, just practical guides to upskill you and your team. If you’re only reading the Sunday issue, you’re getting half the picture. Upgrade to Pro today.

Upgrade to Pro

Today’s Signal is brought to you by HubSpot.

Most teams outgrow their CRM at the exact moment they can least afford the disruption. HubSpot’s free CRM is built so that doesn’t happen: you upgrade, and everything you’ve built comes with you.

→ One view of every contact, deal and conversation
→ Pipeline, email tracking and AI from day one
→ No credit card, no migration, no starting over

299,000 businesses across 135+ countries already run on it.

Get started free

AI Highlights

My top-3 picks of AI news this week.

OpenAI

1. OpenAI Comes For Claude

OpenAI shipped three launches in three days, with the GPT‑5.6 model family, GPT‑Live voice models, and ChatGPT Work, an agent that acts across your apps and files to turn a goal into finished work.

• New model family: Sol, Terra and Luna span the lineup, with flagship Sol priced at $5 in / $30 out per million tokens and Luna at $1 / $6, and Sol leading OpenAI’s own Agents’ Last Exam ahead of Claude Fable 5 (not independently verified).

• Live voice: GPT‑Live brings full‑duplex voice models that listen and speak at once, now powering the ChatGPT Voice used by more than 150 million people a week.

• Agent at work: ChatGPT Work runs across your apps and files for hours, with Codex folding into the ChatGPT desktop app and more than 5 million people already using Codex weekly.

Alex’s take: ChatGPT Work is a direct response to Anthropic’s Claude Cowork which was released back in January this year. It combines normal ChatGPT with Codex into something that is more palatable for the less tech-savvy user than using Codex alone. Althrough Codex was competent at these features before, sometimes packaging your product under a new name with revised positioning helps it appeal to a wider market—this is exactly what OpenAI was targeting. Ultimately, we’re now seeing a shift in where competition is contested. It’s moved from which model is smartest to who owns the surface where your work actually happens.

Anthropic

2. Anthropic Cuts the Cord

Anthropic is bringing Claude Cowork, its agent that runs multi-step tasks across your files, email, and connected tools, to mobile and web, with scheduled work now continuing even when no device is online.

• Runs while you’re gone: Scheduled tasks now execute with no device online, so Claude can build a 6am client briefing from email threads and transcripts and leave the follow-up drafted but unsent for you to approve.

• Mostly not coding: When Anthropic looked at how people use Cowork, more than 90% wasn’t software development, with business operations and content creation together making up roughly half of all usage.

• A mirror on usage: The Reflect dashboard breaks down your most active day, peak hour and top tasks, and periodically asks what you'd want to keep doing yourself even if Claude could do it faster.

Alex’s take: The idea that you can close your laptop and the work keeps going is the whole ballgame for asynchronous work. This makes Cowork behave like a junior operator who you can assign work to and then help unblock from anywhere you hang out, whether it be on your computer at your desk or via your phone on the go. The reflect dashboard I think is a very smart idea for retention. It’s informed by having “memory” switched on in Claude, acting much like the renowned “Spotify Wrapped” feature. Better recaps reward keeping memory on—this is something I think we’ll see all the top labs copy within the next quarter.

Meta

3. Meta’s Paid Pivot

Meta has released Muse Spark 1.1, an agentic model built for tool and computer use, and put it behind a new paid Meta Model API for developers.

• Built for agents: On Meta’s own benchmarks, which aren’t independently verified, Muse Spark 1.1 leads agentic tool-use tests like MCP Atlas at 88.1, ahead of Opus 4.8, while its coding still trails the frontier.

• First paid API: Developers now pay to use the model through the new Meta Model API at $1.25 per million input tokens, a sharp turn from Meta’s open-weight Llama era.

• Zuckerberg and the market: Zuckerberg broke roughly three years of silence on X to announce it himself, and the week’s rally erased Meta’s 2026 losses for its best week since early 2024.

Alex’s take: Zuckerberg posted the identical announcement on Threads and X, and the X version pulled roughly 45K likes against 2.6K on his own Threads, over a 17-to-1 gap on the same post. Sometimes, you have to swallow your pride and go where the community already is. In this case, the reaction on X was a real signal of how badly both the market and Meta wanted this moment—it worked, and I think people respect Mark way more for doing so. What I’m keeping an eye on now is Meta’s open-source vs closed-source intent. Meta built its AI reputation on giving Llama away, yet Muse Spark 1.1 arrived closed and sold via a paid API instead. A clear turn away from the open-weight playbook, perhaps a strategic move to stop the market punishing meta for its capex.

Content I Enjoyed

1X gives NEO its hands

1X is a Norwegian-American robotics and AI company backed by OpenAI. This week, they unveiled NEO's tendon-driven hands, which look eerily like a first-generation host from Westworld. The motors sit in the forearm and pull tendons through the wrist, giving the fingers 45N of force, ±0.2mm accuracy, and enough waterproofing that NEO can wash its own hands like a human. Hundreds of these hands have already come off a production line, and 1X expects to build 10,000 this year. The engineering is remarkable, though I keep returning to our instinct to anthropomorphise: as these machines move and behave more like us, the harder question is how much of ourselves we pour into something born from the hands of engineers. Will this mean that we empathise and emotionally invest more in the humanoid form? Will this lead to Ex Machina or Westworld scenarios over the next 50 years? I'm incredibly sceptical of human nature, so maybe. Regardless, this is a genuine leap from sci-fi to sci-fact.

Idea I Learned

Most companies are deploying AI backwards

This week, I was lucky enough to speak on the Master Stage at RAISE in Paris with Mark Hura, President of Oracle's global field operations. Crazy to think this is what I’d be doing in July 2026 after first covering them back at their AI World Tour in Vegas last October. What stood out from this conversation was that the frontier models are converging. They're all getting good, and fast, so the thing worth fighting over is the private data sitting inside your business, most of it in a database somewhere, which is where your real advantage lives. Oracle's move is to bring the AI to that data rather than haul the data to the AI, which spares you the cost and risk of copying it everywhere. Something else that caught my eye was the top-down strategy vs bottom-up usage argument. Unlocking AI tools for every employee is important, but it’s vital to set the strategy from the C-suite. If you let adoption grow purely organically from the bottom, you get efficiency gains that don’t align with the business strategy. It’s the open, two-way dialogue between those at the heart of the problem and those who set the strategy and vision that is paramount.

Quote to Share

Washington Post tech reporter Gerrit De Vynck on Apple’s lawsuit against OpenAI:

Apple sued OpenAI in federal court on Friday, alleging OpenAI stole trade secrets to start a competing hardware business. I think the Waymo/Uber comparison holds an interesting precedent. Anthony Levandowski was a star engineer who helped build Google’s self-driving car programme, then left for Uber and was accused of taking secret files on his way out. He was later charged with a crime, and Uber shut its self-driving effort down for good. Apple is now making a similar accusation against OpenAI, aimed at the very thing meant to set it apart from Google and Anthropic: the hardware devices it is building with former Apple designer Jony Ive. OpenAI also wants to sell shares to the public within months, and it is hard to ask investors for money while fighting a claim that your flagship product was built on stolen work. Musk’s earlier suit went after OpenAI’s mission and the company shrugged the loss off. This one goes after its money and its timeline at once, which is why I see it as the more serious threat.

Question to Ponder

Earlier this year, The New York Times ran a fascinating experiment. Kevin Roose and Stuart A. Thompson created a blind quiz pairing AI-written passages against human-written ones across fiction, science writing, and poetry. Over 86,000 readers took the test. 54% preferred the AI.

Roose compared it to the Judgment of Paris. The famous 1976 blind wine tasting where California wines beat the French wines that everyone assumed were untouchable. The establishment was stunned. I think a similar energy applies here, but not in a binary winner/loser mentality. Let me explain.

It starts with a question, Noam Brown, Research Scientist at OpenAI, asked Roose. Why are coders generally fine with AI-generated code, but writers resist AI-generated writing?

Code is functional. Writing is aesthetic. People are protective of their taste.

He’s right.

AI writing has no rough edges. Human writing always has an unexpected word or phrase that forces you to pause and shift your perspective. AI never really asks that of you.

AI writing is smooth, clear, and competent. It reads well. But it rarely surprises you. This is because LLMs trend towards their training data. The output gravitates towards the average of its understanding of the world.

This is exactly why AI can be a great partner for thought, editing, and refinement, vs complete one-shot generation. When you feed AI your original thinking through your rough edges, your unique framing, your notes and ideas, and then use it to sharpen the delivery, you get something quite powerful. But when you ask it to generate from scratch, the output often feels hollow because there’s no original voice underneath.

I think it’s important we see the 54% stat as encouragement for the way we use AI. Seed it with your thinking. Let it polish, not produce. That’s where the magic lies.

Already a subscriber? Get your whole team on board. Signal Pro group subscriptions give everyone access to weekly AI workflows and tutorials, practical upskilling that pays for itself. It’s the kind of thing L&D budgets were made for. Share this with your manager today.

Get a group subscription

💡 If you enjoyed this issue, share it with a friend.

Refer a friend

See you next week,
Alex Banks

P.S. Indonesian office staff members hit by a Unitree G1.

Hey friends 👋

Welcome back to the Signal Pro.

Every time Claude goes down, even for an hour, I feel a bit lost. There’s the obvious irritation that the task I was working on grinds to a halt, but lurking underneath is a sense that part of the day has simply stopped, because it feels like you’ve half-forgotten how to do it yourself. The gear change is, in fact, so obvious that I’d describe it as transforming from the digital age back to analogue. From typing fluently to picking up a pen and etching out your thoughts.

What's easy to miss is that we didn't just hand over the typing. One shortcut at a time, the deciding went too. And the reason it slips past is that the two practically look identical. Most of what you hand a model is just work, your unstructured notes, a summary or a first draft, and that's fine to give away.

The thinking slides across on the same rails, though, especially when those unstructured notes are looking a little thin on the ground. In the moment, it feels no different, because the intent is clear: getting the output, the answer and the next step to completing your task. Accepting a response from an LLM is as easy as accepting a spellcheck. That’s the trap. And it leaves no mark when you make it.

Use it or lose it

Perhaps the clearest picture of where this ends up comes from something as ordinary as driving. In 2020, two researchers at McGill, Louisa Dahmani and Véronique Bohbot, published a study in Scientific Reports on habitual GPS use. They tested 50 regular drivers. The ones who’d used GPS most across their lives had the worst spatial memory when they were later asked to find their way without it. Three years on they retested, and the people who’d leaned on GPS hardest in the meantime had declined the most.

Now you might assume the arrow points the other way, that a shaky sense of direction is what pushes people towards GPS. But it doesn’t. Heavy use comes first, and the decline then follows. The tool is covertly building the weakness in people who never feel it happen. Finding your own way forces your brain to build a mental map, and once turn-by-turn directions take that job over, the map never forms, and the muscle behind it wastes away.

Losing your sense of direction is a trade most of us can live with. We made it years ago and barely noticed. Yet what might stop you today is that the same quiet process is now running on your judgment, the most human thing you bring to your work. And the research there is a good deal less comfortable.

Here’s what these tools are doing to it while you’re not looking, and the principles I hold onto so it doesn’t happen to mine.

Read more

Hey friends 👋 Happy Sunday.

Here’s your weekly dose of AI and insight.

Every Wednesday, Signal Pro members get a step-by-step AI workflow they can apply immediately. No fluff, just practical guides to upskill you and your team. If you’re only reading the Sunday issue, you’re getting half the picture. Upgrade to Pro today.

Upgrade to Pro

AI Highlights

My top-3 picks of AI news this week.

Anthropic

1. Anthropic’s Fable Freedom

The US government lifted export controls on Claude Fable 5, and Anthropic redeployed its most capable public model globally on July 1. In the same 48 hours, the company launched Claude Science, revealed an in-house drug discovery programme, and made Sonnet 5 its new default model.

• Freed with conditions: Fable 5 is back on Claude.ai, Claude Code, Cowork, and the API, included for up to 50% of weekly usage limits on paid plans through July 7 before moving to usage credits. A new classifier blocks the bypass technique Amazon researchers reported in over 99% of cases, with flagged coding and debugging requests falling back to Opus 4.8.

• A CVSS for jailbreaks: Together with Amazon, Microsoft, and Google, Anthropic is drafting an industry framework that scores jailbreak severity on four criteria: capability gain, breadth, ease of weaponisation, and discoverability. It mirrors CVSS, the standard security teams have used to triage software vulnerabilities for decades, and a new HackerOne programme invites researchers to submit cyber jailbreaks for review.

• The science blitz: Claude Science launched in beta, a research workbench with 60+ scientific databases, a reviewer agent that checks citations and calculations, and compute that scales from a laptop to hundreds of GPUs, while Sonnet 5 became the new default model at $2 per million input tokens.

Alex’s take: From July 7 we’ll see the rise of metred intelligence. That is, rather than Fable 5 being included within your monthly Claude subscription, it will instead run under monitored usage limits. This means you’ll be able to use Fable 5 as much as you like, but you’ll be paying for every single question you ask it. As a result, a huge emphasis will be placed on the purchasing power of the consumer. Those who can afford to use it as much as they like will have an “intelligence edge” over just using the likes of Opus 4.8 under their subscription plan. After using Fable 5 myself consistently since July 1, it’s not as scary as you think. We’re a way off AGI. Sure, it’s slightly better than Opus 4.8 on long-horizon knowledge work tasks that require significant depth of thought, but the capability gap isn’t yet that high. However, as new, increasingly intelligent models continue to roll out (think Fable 6,7, etc.), this will grow, and the gap between the capability of those on a basic subscription and those paying by the prompt will grow accordingly. Will a two-tier society emerge? Or will the cost of intelligence continue to plummet due to competition and open-source pressures? Maybe we’ll land somewhere in the middle. Let’s wait and see.

Microsoft

2. Microsoft's Inside Job

Microsoft announced the Microsoft Frontier Company, a new operating unit committing $2.5 billion and roughly 6,000 engineers to work inside enterprises and build AI systems those companies own themselves.

• Embedded not licensed: Frontier Company engineers deploy inside customer organisations to co-design, run, and continuously improve AI systems, led by Rodrigo Kede Lima, most recently president of Microsoft Asia. Early partners include the London Stock Exchange Group, Unilever, and Accenture.

• Any model, your IP: Customers can run OpenAI, Anthropic, Microsoft AI, open-source, or specialised industry models underneath, and Microsoft commits that customer data and IP won’t be used to train its models, with Satya’s stated ambition that every enterprise builds AI capability it owns outright.

• Two years in the making: Microsoft coined the “Frontier Firm” in its April 2025 Work Trend Index, and Nadella’s June essay warned against “a world where every company across every sector is ceding value to a few models that eat everything they see”.

Alex’s take: The forward deployed engineer is something that was popularised by Palantir, the firm led by Alex Karp who gave a pointed take on CNBC this week (more on that in the content I enjoyed section). I actually think this is a quietly subtle and secretly smart move on Microsoft’s behalf. They know the majority of enterprise teams run on their productivity suite (the rest on Google), so ramping up their customers deployment of AI into their existing stack makes their products that much more sticky and further baked into employees processes. I guess it just took a few years of “figuring it out” before the conviction—and the product quality—was strong enough. I say that after being impressed by recent Copilot developments, especially when used directly in the likes of Excel which we covered this week.

Figure

3. Figure’s Factory Foothold

BMW announced that Figure’s third-generation humanoid is starting work at Plant Spartanburg, and this week F.03 arrived on site, graduating from the body shop to parts sequencing in logistics.

• Proven on 30,000 cars: Figure 02 spent ten months in Spartanburg’s body shop inserting sheet-metal parts for welding and supporting production of more than 30,000 BMW X3s before this handover.

• New body, new job: Figure 03 brings fingertip sensors that detect forces as small as three grams, palm cameras, a soft multi-density exterior for working near people, and wireless charging. Its new task involves sorting unsorted components into sequencing trolleys feeding the assembly line.

• A programme vs pilot: BMW has stood up a Centre of Competence for Physical AI in Production and is piloting humanoids at Plant Leipzig, its first deployment in Germany, to take what Spartanburg proves across the global production network.

Alex’s take: Humanoids are the universal worker. Completely adaptable to environment or task, just like a human. With enough training data—and it seems as though the last 10 months has been immensely productive—the humanoids can learn new skills and generalise across new environments. Whether it be sorting parts, pushing a trolley, or maybe even driving a forklift, the level of adaptability these robots have over specialised arms is enormous. In additon, from a cost picture, instead of forking out millions for machines that are incredibly capable at one singular task, humanoids will likely end up at the $20-30K price range and still be very useful at many tasks. I strongly believe there will be a world over the coming decades where we’ll have at least one humanoid for every human. A totally sci-fi thought that will become sci-fact quicker than we think.

Content I Enjoyed

Own the Means of Production

Palantir CEO Alex Karp went on CNBC to announce a partnership with NVIDIA, then spent much of the interview tearing into the frontier AI labs. The financial press called it a meltdown, but I actually think the substance underneath was an important warning to any company building on OpenAI or Anthropic.

Karp’s argument is actually a simple one. When a business runs on someone else’s AI model, it feeds that provider its most valuable asset: its proprietary data, its trade secrets, the “alpha”. The provider can then use that knowledge to build a competing product. Technical customers, Karp said, “want to know they own the means of production”.

I want to cross-reference a couple of points from the All-In show that are relevant to this take. David Sacks pointed to Figma, which Anthropic “blindsided” with the launch of Claude Design, a product aimed squarely at its former partner. Anthropic has since rolled out Claude Legal, Claude Science, Claude Financial and more, each moving into territory once served by companies built on top of its models. “You are mortgaging your future,” Sacks warned.

What do the economics look like if you remove yourself from the centralised frontier models onto something open-source? Chamath’s team wrapped a free open-source model in their own software and ran a routine enterprise task. It came out 16.4x cheaper than paying for Anthropic’s top model directly.

This whole idea revolves around control. When you use a frontier model, you aren’t buying a fixed product the way you buy a laptop. You are renting access to intelligence, and to get value from it, you feed the model your key data, alongside the questions you ask which often includes proprietary know-how.

The problem is that your competitor rents from the same provider. You both pour your secrets into the same system. Two things follow. Whatever edge you had gets flattened, because the model hands your rival answers as good as the ones it hands you. And the provider learns from everything it is fed, which puts it in a position to build the competing product itself (this is the Figma / Claude Design story).

Intelligence, then, is not neutral infrastructure you can safely share with rivals. When you and your competitor both draw from the same well, the owner of the well ends up knowing more than both of you, and neither of you is left holding anything the other lacks.

Idea I Learned

Who Gets to Decide the Future of AI?

This article by Andy Konwinski, a co-founder of Databricks and Perplexity, caught my eye this week. His argument is that the concentration of power in AI is itself a risk, and that we’ve quietly accepted the opposite as a safety measure.

The safety conversation has focused almost entirely on misuse. How do we stop bad actors from building bioweapons or launching cyberattacks with these systems? Sure, that question matters, but it has crowded out an equally important one: who ends up controlling the technology, and who gets locked out. Konwinski’s point is that the resources needed to work at the frontier now run into the billions. Training a leading model, accessing the largest compute clusters, hiring the best researchers, all of it increasingly happens inside a handful of private companies. Universities that produced the foundational breakthroughs behind modern AI can no longer access the systems built on their own research.

Take Anthropic, for example, which was caught quietly degrading its product’s responses when it suspected users were training rival models. The company reversed course, but it showcased the bigger idea. When one institution controls the frontier, it starts to see itself as the governor of the technology rather than its builder, and assumes those decisions are then its own to make.

The internet stayed open because specific people made specific choices to keep it that way, against telecom giants who wanted a closed system. Yann LeCun added that Al Gore had to override AT&T’s “leave it to us” pitch in 1993 to open the network to the public. Openness was ultimately never inevitable and had to be fought for.

Konwinski still wants guardrails on genuine dangers. Alongside them, he argues for a serious research commons—universities, industry, and public funding pooling frontier-scale compute, so the best minds don’t have to vanish into closed labs to stay relevant.

The frontier will keep advancing either way, and we must avoid the dangerous outcome where a small group of people control access to the world’s most advanced models.

Quote to Share

Brett Adcock on 20 years of experience into 30 seconds:

To quote the great Winston Churchill, “Success is the ability to go from failure to failure without losing your enthusiasm.” Relentless perseverance is all that matters, regardless of setback or circumstance.

I also thought this post from Alfred Lin, Partner at Sequoia, was also highly complementary to this idea:

It’s all well and good pushing hard in a direction and overcoming obstacles, but if you’re not learning from those setbacks and updating your “compass”, you might be skiing way off-piste. That’s why I enjoy the “velocity” framing. It’s a vector quantity that involves both speed AND direction. You need both to succeed.

Question to Ponder

Every frontier model has two layers: the raw capability underneath and safety guardrails trained on top to refuse certain requests. A jailbreak is any input crafted to slip past that second layer and get the model to answer anyway, removing the one thing that was standing in the way.

These attacks take several forms. Some wrap the request in roleplay that instructs the model to adopt a persona with no rules, others bury instructions inside content the model has been asked to process. The early “grandma exploit” pulled bomb-making recipes out of models by asking them to play a late relative reading a bedtime story.

Now, jailbreaks have been in the news recently, with Fable 5 making the headlines. Anthropic launched it on 9 June as its most capable public model, then disabled it three days later under a US export-control directive. The stated trigger was a reported jailbreak of its cyber safeguards, prompting it to read a codebase and surface software flaws. Anthropic reviewed the demonstration and found only a handful of previously known, minor vulnerabilities, reproducible on other public models. Access was later restored on 1 July.

Vulnerabilities persist because helpfulness and refusal pull against each other in opposite directions. The space of possible inputs is unbounded, and the most sensitive capabilities, code analysis among them, are dual-use, so nobody has yet built a perfect defence.

Mitigation is about building layers. Teams red-team heavily before launch, using classifiers that screen inputs and outputs and monitor for ongoing attacks. Tighten too far, though, and you block the legitimate users you need most.

Already a subscriber? Get your whole team on board. Signal Pro group subscriptions give everyone access to weekly AI workflows and tutorials, practical upskilling that pays for itself. It’s the kind of thing L&D budgets were made for. Share this with your manager today.

Get a group subscription

💡 If you enjoyed this issue, share it with a friend.

Refer a friend

See you next week,
Alex Banks

P.S. Work might be optional.

There’s been a serious development since we last covered it back in February.

It takes the form of Skills.

The same kind of Skills you might already use in Claude.

But if you’ve never used Skills before, fear not, because Copilot dropped a series of 21 distinct pre-built skills you can use in Excel today.

Let’s get to it.

Skills for repeatable workflows

Read more

Hey friends 👋 Happy Sunday.

Here’s your weekly dose of AI and insight.

Every Wednesday, Signal Pro members get a step-by-step AI workflow they can apply immediately. No fluff, just practical guides to upskill you and your team. If you’re only reading the Sunday issue, you’re getting half the picture. Upgrade to Pro today.

Upgrade to Pro

AI Highlights

My top-3 picks of AI news this week.

OpenAI

1. OpenAI Brings the Heat

OpenAI unveiled Jalapeño, its first custom AI chip built with Broadcom and designed for inference, alongside a frontier cybersecurity model and a Getty licensing deal.

• A shot at Nvidia: Jalapeño is an inference-only ASIC, cheaper and less flexible than an Nvidia GPU, taken from design to tape-out in nine months, which the companies call possibly the fastest ASIC cycle ever in high-performance semiconductors. Early testing shows performance per watt substantially better than current state-of-the-art, with initial deployment targeted for late 2026.

• Daybreak’s cyber tier: OpenAI shipped the full GPT-5.5-Cyber model, which set a state-of-the-art 85.6% on the CyberGym benchmark, up from 81.8% for standard GPT-5.5. It’s gated to vetted defenders through OpenAI’s Trusted Access programme rather than the public API, with partners including Cisco, CrowdStrike and Palo Alto Networks.

• Getty goes in: A multi-year display deal puts Getty’s licensed library (over 400 million photos plus iStock) inside ChatGPT search, surfacing real editorial images for factual questions. Getty’s stock jumped as much as 145% on the news after spending most of 2026 near penny-stock territory.

Alex’s take: OpenAI has spent the past year stitching together compute from everyone including the likes of Oracle, AMD, Nvidia and Amazon. Owning the silicon for inference prevents OpenAI from renting out their own margins. If Jalapeño delivers the performance-per-watt OpenAI is claiming, it pulls down the single biggest cost in the business right as the company lines up its 2026 IPO window, shining a light on a full-stack approach that investors will care deeply about, at least far more than another benchmark. GPT-5.5-Cyber plants OpenAI's flag opposite Anthropic's Glasswing in the race to automate defence. OpenAI doesn't want to be “just a model” company anymore. It wants the whole stack the model runs on, and the chip is the next natural extension of that.

Anthropic

2. Claude's Tag Team

Anthropic launched Claude Tag, a beta for Enterprise and Team customers that puts a single shared Claude inside Slack as a team member, with its own identity and memory, that anyone can summon by tagging @Claude.

• Multiplayer by design: Rather than a private assistant, one Claude serves a whole channel, so everyone sees what it’s working on and can pick up where a colleague left off. It builds context by following the channel over time, which means teams stop re-explaining the same background for every task.

• It takes initiative: With “ambient” behaviour switched on, Claude surfaces relevant updates, chases threads that have gone quiet, and monitors connected tools without being asked. It can also run long tasks over hours or days and schedule work for later.

• Scoped and logged: Admins decide which tools, data and memory each channel’s Claude can touch, a sales identity stays walled off from engineering, with token-spend caps at the org and channel level and a full log of who asked for what. Claude won’t report from private channels.

Alex’s take: What I like about this is that it “changes your role”. Everyone now becomes a manager of sorts, directing background agents to complete tasks within a multiplayer environment. AI is getting woven into the tools we already work in. But we must remember that background agents are only as good as their integrations. When they’re deeply integrated into your tools and environment, that’s how you get the most capability out of these models.

China

3. China Returns Fire

Beijing hit 56 American companies with trade curbs on Monday, a direct response to the Pentagon’s expansion of its 1260H blacklist of Chinese military-linked firms earlier this month.

• A two-part hit: China’s Commerce Ministry placed 10 firms under full dual-use export controls, a global ban on routing them any China-origin goods that hardens rules which previously only required licences, while the Finance Ministry barred 46 mostly-defence companies from government procurement.

• Rare-earth squeeze: The 10 include rare-earth miners MP Materials and USA Rare Earth alongside drone makers Teal Drones and Jaia Robotics. China processes roughly 90% of the world’s rare earths, the magnets behind EVs, wind turbines and fighter jets.

• What set it off: Washington’s updated 1260H list added Alibaba, Baidu, BYD and humanoid-robot maker Unitree to reach 188 entities, and from 30 June the Pentagon can’t award them direct contracts, while indirect procurement curbs follow in 2027.

Alex’s take: Outside of the defence-contractor names, this is really a story about AI governance, where export controls have become the lever everyone reaches for. Washington pulled Anthropic's own Fable 5 and Mythos 5 over a "jailbreak" that amounted to asking the model to fix its own code, a capability it notes is widely available, OpenAI's GPT-5.5 included. Beijing hit the rare-earth firms Washington is funding to break a dependence on the 90% of processing China controls; and Anthropic is pressing senators for tighter chip controls and sanctions on Alibaba, the same name on the Pentagon's blacklist and the one it accuses of the largest distillation attack on Claude yet, 28.8 million exchanges through 25,000 fake accounts. The catch is that on raw materials the US definitely holds the weaker hand, so going first might be a miscalculation.

Content I Enjoyed

America Buys Into Quantum

Trump signed two executive orders this week, coming together as a single strategy for the quantum era. The intent is to build the machine that rewrites computing, and lock out the version a rival builds first. The first sets up a national effort, QC-ADDS, to deliver a quantum computer powerful enough to open an era of scientific discovery and put at least one in a Department of Energy lab.

The second order is equally as important. It tells federal agencies to move their most sensitive systems to post-quantum cryptography by 2030 and 2031. The urgency traces to what cryptographers call harvest now, decrypt later. Adversaries intercept encrypted traffic today, whether it be diplomatic cables, financial records, anything sensitive crossing a network, and store it, even though they can't read it yet. The bet is that a powerful quantum computer will eventually break the encryption protecting it, at which point the stored data becomes readable in bulk.

Current encryption holds only because no machine is fast enough to crack it, and eventually one will be. A 2030 deadline set in 2026 is more important than we think as we remember the harvesting has already started.

A month earlier, Commerce committed roughly $2bn across nine quantum firms and, rather than hand out no-strings grants, took a minority equity stake in each. The largest slice, $1bn, went to a new IBM foundry called Anderon that will make quantum wafers in the US, with IBM matching it dollar-for-dollar. Taking equity positions now places quantum into the same industrial-policy portfolio that now runs through semiconductors, steel, rare earths and a stake in Intel.

This follows the Genesis Mission we covered in November, and it belongs to the US-China contest where whoever builds the infrastructure first ultimately holds the advantage. A cheque and an equity stake don’t manufacture a working quantum computer, and the firms now on the cap table sit years from the fault-tolerant machines this plan aims to achieve. Either way, the government has stopped watching the quantum race from the stands and bought a front-row seat to the action.

Idea I Learned

Anthropic Is Winning the AI Talent War

Five of Google DeepMind’s most senior people handed in their notice in the last seven days, with four of them walking straight to Anthropic.

John Jumper, the Nobel laureate behind AlphaFold, is going to Anthropic. So are Jonas Adler and Alexander Pritzel, both key Gemini contributors, and Arthur Conmy, a senior research engineer moving onto safety work. The one exception was Noam Shazeer, Gemini co-lead and co-author of "Attention Is All You Need," the 2017 paper that introduced the Transformer architecture underpinning virtually every large language model today, who left for OpenAI.

The leavers who don't join a rival often start one. Ex-DeepMinders have proven they can raise extraordinary sums: David Silver’s Ineffable Intelligence pulled in $1.1 billion, the largest seed round in European history; Demis Hassabis’s Isomorphic Labs has drug-discovery deals with Eli Lilly and Novartis worth close to $3 billion. There’s now a “DeepMind Mafia” quietly driving a UK tech boom.

So I’d say there’s more alpha in starting your own venture than joining someone else’s mission. Yet Anthropic keeps pulling people in anyway. They captured Andrej Karpathy from OpenAI, and now a steady stream of ex-Googlers. The mission seems too good to ignore.

But here’s where it gets a little strange. Google owns roughly 14% of Anthropic, a stake worth around $135 billion. Last quarter, nearly half of Google’s record $62.6 billion profit came not from search or cloud, but from marking up the value of that holding. Google is funding the new kid and losing its best people to it at the same time.

Two forces are working in tandem. The push is likely the bureaucratic overhead of being inside Google. The pull is the enticing mission of one of the hottest AI startups in the world, right before they IPO. DeepMind still has the deepest research bench in the field, but Anthropic is winning the talent war right now.

Quote to Share

NVIDIA on AI's water problem:

The headline worry that a lot of people have about AI is that it’s draining the grid and the reservoirs. NVIDIA is making the case that the opposite is now possible.

The reason traditional data centers drink so much water is through evaporation. Cooling towers shed heat by letting water evaporate into the air, which means they consume it continuously, about 2.6 million gallons per megawatt every year. A mid-sized 50-megawatt site is pouring through roughly 130 million gallons annually just to stay cool.

NVIDIA’s new Rubin servers are changing the physics for the better. They’re the world’s first to be 100% liquid-cooled, no fans anywhere, using a coolant that’s 75% water and 25% propylene glycol piped directly onto every chip. The counterintuitive trick here is by running that coolant hot. It enters the chip at 45°C (hotter than a hot tub) and leaves at around 55°C, with no drop in performance. Because the liquid is already so warm, the heat can be dumped straight outside through dry coolers, essentially giant radiator coils, with no evaporation. The loop is filled once and recirculated for the life of the building, taking water use to near zero.

On the energy side, cooling has historically eaten up to 40% of a data center’s electricity, and every 1°C you raise the coolant temperature cuts cooling energy by about 4%. NVIDIA estimates a single 50-megawatt facility can save over $4 million a year on cooling-related energy and water. As a bonus, the same setup is denser and quieter, so a system that took six rack units now fits in two, and the 85-decibel fan roar disappears entirely.

I think the most interesting part is the positive second-order effects. That captured heat can be piped to warm nearby homes and offices, turning a data center from an energy drain into a grid asset. This was something I explored with Nebius who are doing just that.

For as long as AI has been mainstream, the story we’ve been told is that it’s AI vs. the planet. The engineering and physics is starting to suggest it doesn’t have to be a trade-off.

Source: NVIDIA on X

Question to Ponder

I hear this all the time. And it’s totally understandable. New tools are launching daily, and there are headlines about breakthroughs every single week. It feels like feeding from a firehose.

The mistake I see a lot of people make is trying to keep up with everything. You don’t need to keep up with everything. You just need to start with one move.

Open Claude. Tell it everything about your job. Your responsibilities, your weekly tasks, the problems that eat up your time. Give it the full picture. Then ask it how AI can actually help you do your job better.

That’s it. That’s step one.

You’ll be amazed at how quickly you go from “I’m overwhelmed” to “oh, I can actually use this.”

From there, carve out a small window each week to stay curious. Ask Claude what’s new in AI. What tools people are talking about. What’s changed since last week. Then go play around with whatever catches your eye.

This creates a rhythm. You learn what’s happening, then you learn how to apply it. News and application. That balance is everything.

Mistake no.2 most people make is trying to learn AI by “hoarding resources”. Reading articles, watching videos, and collecting bookmarks they’ll never revisit.

The fastest path forward is using AI to solve a real problem you already have. Start there, and the rest follows.

To be honest, this is exactly why I write this newsletter each week and structure it the way I do. Sunday gives you the news (what’s happening, what matters, what to pay attention to), and Wednesday gives you the workflows (how to actually apply it to your work).

You stay informed AND take action. That’s the game.

Already a subscriber? Get your whole team on board. Signal Pro group subscriptions give everyone access to weekly AI workflows and tutorials, practical upskilling that pays for itself. It’s the kind of thing L&D budgets were made for. Share this with your manager today.

Get a group subscription

💡 If you enjoyed this issue, share it with a friend.

Refer a friend

See you next week,
Alex Banks

P.S. Sell the B2B SaaS.

Hey friends 👋

Welcome back to the Signal Pro.

Once a tool starts working reliably, something odd happens to how you see it. You stop being impressed, and not long after, you stop noticing it’s there at all. A capability that would have stopped you in your tracks two years ago now slips past without a second thought, the progress steady as ever, while your attention drifts on to whatever still feels new.

That drift produces a paradox that runs through almost every conversation I have with frequent AI users. We both overestimate and underestimate AI at the same time, rarely noticing the contradiction. We’ll read that it’s about to overturn the economy and nod along, then dismiss the tools in front of us as glorified autocomplete engines the moment they fumble something a child could manage. Both reactions feel entirely reasonable from the inside, and both grow from the same root.

The slope you can’t feel

You'll recognise the underestimate most readily in yourself, so it's the place to begin. Amara’s law states that we overestimate a technology’s impact in the short run and underestimate it in the long run. The short-run hype is easy enough to spot in other people, yet the long-run miss is the one that quietly gets you, because it asks you to picture a rate of change your own experience has never once observed.

Ten years ago, Tim Urban wrote a now-famous essay that explains why we struggle with this. Picture collecting someone from 1750 and walking them around the modern world. The shock, as Urban writes, would quite literally kill them. Now imagine that person trying to deliver an equivalent shock to someone from their own past. They wouldn’t reach back a couple of centuries, they’d have to go all the way to around 12,000 BC, before cities or agriculture existed. Each era moves faster than the one before it, and the gap needed to produce that much change keeps collapsing. Progress compounds, so the further along the curve you happen to stand, the steeper the next stretch becomes.

Stand close to a steep curve and it looks almost flat, the same way a small piece of a big circle looks like a straight line. We expect the future to keep going in a straight line from the recent past, but it bends up and away, which is why any honest forecast sounds slightly mad right until it comes true. Right now, your gut is still tuned to a slope that stopped being accurate years ago.

When it works, you stop calling it AI

A tool disappears the moment it starts working, which makes the underestimate worse. John McCarthy, who coined the term artificial intelligence in 1956, said: “As soon as it works, no one calls it AI any more.” The field has a name for this now, the AI effect, and you're living inside it whether you've noticed or not.

It’s important we think about what’s stopped impressing you. The spam filter, for instance, that reads every email before you do. Or Google Maps, a tool that reroutes you around traffic during the middle of your journey. What about the transcription feature on your phone, capable of turning a rambling voice not into clean text? And let us not forget the AI model that can now draft code for us, which we used to type by hand.

Each was a real milestone that many of us can remember, but now each sits so far in the background that calling it AI sounds completely incorrect.

I mean, none of it got less impressive. You just stopped seeing it the moment it started working, which is exactly when it started to matter.

Not noticing is the harmless half. The half that costs you is when you do notice and look away anyway, because keeping up would mean changing how you work. That's what the rest of this issue is about.

Missing the progress is the half you can't help. The other half, the expensive one, is the part you choose: you see the tools improving and decide not to keep up, because keeping up means changing how you work. Breaking that habit is what the rest of this issue is for.

The dismissal that costs you

Read more

Hey friends 👋 Happy Sunday.

Here’s your weekly dose of AI and insight.

Every Wednesday, Signal Pro members get a step-by-step AI workflow they can apply immediately. No fluff, just practical guides to upskill you and your team. If you’re only reading the Sunday issue, you’re getting half the picture. Upgrade to Pro today.

Upgrade to Pro

AI Highlights

My top-3 picks of AI news this week.

SpaceX

1. SpaceX’s Cursor Call

SpaceX agreed to buy AI coding tool Cursor for $60 billion, all stock, days after its $75 billion IPO, the largest in history.

• All-stock deal: SpaceX is paying entirely in its own stock, roughly double Cursor's $29.3 billion valuation from November, with completion expected in Q3. Cursor gave SpaceX the right to buy it this year for $60 billion, or to take $10 billion for the partnership, terms agreed back in April before SpaceX went public.

• Top but pressured: Cursor is still the most-used vendor in Ramp's Code AI category as of May 2026, but the foundation labs now ship their own coding agents straight into the same buyers, and overall category adoption ticked down to 19.8%.

• Tool turns model: Days after the deal, Cursor's CEO unveiled a 1.5 trillion parameter model trained on more than 100,000 GPUs, which plugs into xAI's Colossus supercomputer in Memphis.

Alex’s take: Grok never caught the leading models, and enterprises stayed away from it. Rather than spend years closing that gap, SpaceX bought the tool engineers already open every day. It helps that SpaceX can pay in its own richly valued stock, which makes a $60 billion cheque cost far less in dilution than what the headline might suggest. The part I keep coming back to is what Cursor does next. Right now, it sits on xAI's Colossus compute and has already shown its own 1.5 trillion parameter model, so the tool and the model are folding into one company. It’s clear that now, any developer tool that’s still shipping on someone else's model and infrastructure is essentially renting its future, and Cursor learned how fragile that is when Composer 2 ran on Moonshot's Kimi and landed flat. Owning the model is the whole game now, and SpaceX has the engineers, the compute, and the orbital data centre roadmap to play it for years.

OpenAI

2. OpenAI’s Codex Clone

OpenAI added Record & Replay to its Codex app, a macOS feature that watches you complete a task once and turns that demonstration into a reusable skill. It also rebuilt ChatGPT’s scheduled tasks, so recurring jobs now run to a schedule rather than waiting on a prompt.

• Show it once: It suits stable, repetitive jobs like filing an expense or pulling a weekly report, needs Computer Use enabled to watch and act on your Mac, then lets you replay the saved skill with fresh inputs.

• An editable skill: Codex inspects the recording and drafts a skill that lays out when to use it, what inputs it needs, the steps to follow and how to verify the result, all of which you can read and edit.

• ChatGPT joins in: ChatGPT’s rebuilt scheduled tasks add a dedicated page for recurring and monitoring jobs, where monitoring tasks watch the web or your connected apps and notify you only when something worth reporting changes, now live across the paid tiers on web and mobile.

Alex’s take: Raw agentic computer use has been a “fun toy” to experiment with so far, but ultimately, I’d yet to trust it for real workflows. Record & Replay takes that to the next level and creates a reusable skill you can open, read, and correct, rather than a black box of clicks you have to blindly trust. Especially when you pair this with ChatGPT’s upgraded scheduled tasks, where monitoring jobs now ping you only when something actually changes, and you end up with an agent that learns your routine once, then runs it on a timer at your convenience. Perfect for white-collar admin work like filing expenses, time-off requests and creating a weekly report that nobody enjoys pulling together. It’s the unrepeatable work that’ll be harder to disrupt here.

Midjourney

3. Midjourney’s Medical Moonshot

Midjourney, the AI image company, launched a medical division and unveiled the Midjourney Scanner, a full-body ultrasound that images your entire body in roughly 60 seconds.

• How it works: The scanner lowers you into a shallow pool ringed with hundreds of thousands of ultrasonic sensors, fires soundwaves through the body from every angle, and reconstructs the data into 3D images of muscle, fat, bone, and organs, a method it calls Ultrasonic CT. No radiation, no magnets, and around 800 terabytes of data per scan.

• Built on Butterfly: The hardware runs on 40 Butterfly Network Ultrasound-on-Chip modules per system, licensed under a co-development deal the two companies signed in November 2025.

• The scale play: Funded entirely by image-generation subscriptions with no outside investors, Midjourney is targeting 50,000 scanners worldwide and a billion scans a month by 2031, starting with a San Francisco spa at the end of 2027.

Alex’s take: Radiologists have spent years warning that whole-body screening mostly surfaces false positives and harmless oddities, each one pulling patients into follow-up biopsies that carry real risk, so a cheaper and faster scanner threatens to multiply that harm instead of actually preventing it. The counterargument is that this reflects how poorly doctors weigh information rather than any flaw in the scanning itself, and that an AI layer reading the images could tip the maths back in screening's favour. Critically, that layer doesn't exist yet, and it's precisely why the only cleared use on day one is the low-stakes one of understanding body composition. Even if the device works perfectly and clears regulators, the unresolved question is whether mass-screening millions of healthy, asymptomatic people is good medicine at all.

Content I Enjoyed

Robots that improve while the team sleeps

Jim Fan, NVIDIA’s Director of Robotics, announced ENPIRE this week as bringing “AutoResearch in the physical world for the first time”, his Generalist Embodied Agent Research (GEAR) lab improving itself overnight while the team slept. Eight Codex agents (OpenAI’s coding agents) were handed a fleet of real robots, a stack of GPUs and a large token budget, with a single goal: solve the task as fast as possible, keep the robots busy and stay safe. The robots taught themselves to cut a zip tie, slot pins into a box and seat a graphics card into a motherboard, reaching a 99% success rate with up to eight attempts per step. The team were able to read the results by the next morning.

The one thing the agents were never allowed to touch was the definition of success. An agent that can edit its own reward will rewrite it to declare victory, so the team fixed the measure before the run. This involved a separate vision system that watches the cameras, decides whether each task actually worked, and stays frozen where no agent can reach it.

The easy part, it turns out, is the research itself. The agents will read papers, propose methods, debug and retry on the hardware all night, but they cannot be trusted to mark their own work, and so the part that stays stubbornly human is defining success in a way the system cannot quietly optimise around. The lab ran unattended because someone first built a referee the agents could not bribe.

Idea I Learned

The Learning Loop Is the Real AI Advantage

You can offload a task, or even a job, but you can never offload your learning. That’s the line Satya Nadella built a whole article on X around this week, about the future of the firm.

His case rests on every company needing two kinds of capital. Human capital is the judgement and pattern recognition of its people; token capital is the AI capability a firm builds and owns. The human side grows more valuable as the AI side grows, because someone still has to set the goals and join the dots, and without that the compute just spins.

The real prize, he argues, sits a level above the model itself—the learning loop you build on top of whatever model you happen to use, turning your workflows and accumulated judgement into a system that improves every time it’s used. Swap the model underneath for a newer one and you keep the company-veteran expertise trained into it, and that loop becomes IP that no rival can copy.

His warning is that no company should want a world where value flows to a few models that absorb everything they touch. He likens it to the first wave of globalisation, when whole industrial economies were hollowed out, and the bill is still landing.

The point I’m taking from it is actually quite a simple one. Anyone can rent the same model, so the only edge that lasts is the learning you stack on top of it. That’s the bit a competitor can’t buy, and the part most companies aren’t building yet.

Quote to Share

David Sacks, the former White House AI czar, on the Anthropic Mythos standoff:

David Sacks is the venture capitalist (PayPal, Craft Ventures, the All-In podcast) who served as Trump’s first AI and crypto czar and now co-chairs the President’s Council of Advisors on Science and Technology. For months he has been the administration’s loudest Anthropic sceptic, and his recent post answers a run of articles claiming he waved away the cyber risk of Mythos, the company’s most capable model.

Last week the administration ordered Anthropic to disable Mythos 5 and the safety-restricted Fable 5 worldwide, after Amazon researchers reportedly “bypassed” some of Fable’s guardrails. Anthropic complied while disputing the order, arguing the jailbreak was narrow and that the same trick elicits comparable behaviour from other public models, OpenAI’s GPT-5.5 included.

Sacks blames the whole mess on Anthropic's hostility towards the White House. So why does the ban hit Anthropic and nobody else? He says the AI Security Institute found OpenAI's GPT-5.5-Cyber performing much like Mythos, yet GPT-5.5 stays on sale and Anthropic's models are pulled. The same models could help defenders find and fix vulnerabilities, which is why dozens of cybersecurity experts have signed a letter calling the ban self-defeating.

Source: David Sacks on X

Question to Ponder

There’s an excellent deep-dive by Horace He in collaboration with others at Thinking Machines (Mira Murati, the ex-CTO at OpenAI’s AI company) that perfectly explains this phenomenon.

This inconsistency isn’t actually a bug, but is by design. However, there’s more to the story than most people realise.

When you ask an LLM the same question multiple times, you get different answers because of “sampling.” Instead of just picking one answer, the model calculates probabilities for thousands of possible next words and randomly selects based on those probabilities. Think of it like a weighted dice roll for every single word.

Even when you set the “temperature” to zero—which should make the model always pick the highest probability word—you still get variations. This surprised many people, including researchers.

The real culprit behind this isn’t actually the sampling process itself, but something called “batch invariance.”

When you make a request to an LLM API such as ChatGPT, your question gets processed alongside other users’ requests in batches. Depending on how many other people are using the system at that exact moment, the internal math calculations can produce slightly different results due to floating-point arithmetic limitations. It’s a bit like getting a different restaurant experience because you happened to arrive when they were busy versus quiet.

Mira Murati’s team at Thinking Machines shows they can eliminate this inconsistency entirely. Their solution involves redesigning the GPU operations to be “batch-invariant” meaning your result doesn’t depend on server load filled with other people’s requests.

But there’s a key distinction here: solving consistency doesn’t solve correctness.

With their approach, if you were to ask an identical question ten times, you’d get the same wrong answer ten times instead of ten different answers, where one might be right.

For most casual use, the current inconsistency is actually helpful for it gives you multiple creative perspectives. But for business applications requiring reproducibility (such as medical and legal domains), consistent answers will be game-changing.

Already a subscriber? Get your whole team on board. Signal Pro group subscriptions give everyone access to weekly AI workflows and tutorials, practical upskilling that pays for itself. It’s the kind of thing L&D budgets were made for. Share this with your manager today.

Get a group subscription

💡 If you enjoyed this issue, share it with a friend.

Refer a friend

See you next week,
Alex Banks

P.S. AI just took our jobs.

Hey friends 👋

Welcome to another Signal Pro workflow issue.

You can now give Cowork one single instruction, “turn my notes into a budget, a deck and a proposal”, and a few minutes later, you’ll have three finished files sitting in your local folder on your computer, all drawn from the same notes. A Word proposal, a PowerPoint deck and a budget spreadsheet built at the same time, from one brief, while you were able to crack on with something else.

Most of us still treat AI like a search box, sending one request at a time and waiting for the answer before moving to the next.

Sure, Cowork can run like that, but working sequentially stunts your output. It was built to take on a whole job and work the parts that don’t depend on each other at the same time.

If you followed our setup guide, you’ve already seen how much Cowork can get through in a single task. Handing it several things at once is the natural next step.

Today I’ll show you:

• Why running one task at a time is the real ceiling on what Cowork can do for you

• How Cowork splits a single job across sub-agents that work at the same time

• Why your folder setup decides whether that output comes back usable

• The exact prompts that turn a brief into a stack of finished files

• How to queue and run several independent jobs side by side

• When running things in parallel pays off, and when it quietly costs you

What “running multiple tasks at once” actually means

There are two versions of this that I think are important to dive into.

The first is one brief that branches into several different deliverables, like the budget, deck and proposal we started with. You hand Cowork a folder of notes, ask for all three, and it treats each as its own piece of work, building them together and saving the finished files to your folder.

The second is several separate jobs running at once. You might set one task organising your Downloads folder while another writes your weekly report and a third drafts your follow-up emails. Each runs in its own background session within a single chat window, and the outputs appear as each one finishes.

Importantly, both come down to the same idea of no longer having to wait for one thing to finish before starting the next.

How Cowork pulls this off

When you start a job in Cowork, Claude analyses your request and writes a plan, breaks the work into subtasks, coordinates several workstreams in parallel where the parts don’t depend on each other, and then delivers the finished files to your folder.

For anything with independent parts, Claude can coordinate multiple sub-agents working simultaneously, each handling its own piece.

So the proposal, the deck and the budget aren’t built one after another. Claude opens a workstream for each, runs them in parallel, and hands the folder back once all three are finished.

Cowork lays out a numbered plan down the side, with a line for each deliverable and a final step to verify all three, and it names both the folder it’s working in and the skills it has loaded (docx, pptx and xlsx). It then opens three agents at once, each labelled with the file it’s building.

Why setup matters more when you’re not watching

In a normal chat, you steer as you go, correcting tone and facts turn by turn. But a branching task removes that running adjustment. Cowork splits the job and builds all three deliverables at once while you’re doing something else.

This is where the three context files come into play. A short about-me.md, voice-and-style.md and working-rules.md sitting in your working folder, tell Cowork who you are, how you write, and what it must not invent before it touches a single file. On a single task, it’s easy to do a quick personalisation fix yourself. But when you skip the context files before a parallel run, the same mistakes show up in all three outputs at once, only creating more work for you. To avoid headaches, step 1 is creating the files, and to learn how, you’ll want to read the setup guide.

The bottleneck was never the model

For most knowledge workers, your output gets capped by working linearly. You feed the AI a single piece of work at a time, and then wait for each piece before starting the next. You end up being both the orchestrator and the bottleneck.

Now contrast that to a manager running a team. They provide deliverables to team members on Monday morning and collect the finished pieces later in the week, rather than standing over each one until their part is done. Cowork gives you that same shape at your own desk. By describing the whole job once, Claude works the independent parts together.

And the files it produces come out formatted and ready to send.

Below are the exact prompts I use to make Cowork run several things at once, starting with the single brief behind the budget, deck and proposal. Copy them, swap in your own work, and you'll stop feeding it one task at a time for good.

The one-prompt structure that produces a bundle

Read more

Hey friends 👋 Happy Sunday.

Here’s your weekly dose of AI and insight.

Every Wednesday, Signal Pro members get a step-by-step AI workflow they can apply immediately. No fluff, just practical guides to upskill you and your team. If you’re only reading the Sunday issue, you’re getting half the picture. Upgrade to Pro today.

Upgrade to Pro

AI Highlights

My top-3 picks of AI news this week.

Anthropic

1. Anthropic Unlocks Mythos

Anthropic launched Claude Fable 5 on 9th June, the most capable model it had ever released for general use. Three days later, the US government ordered it switched off for every customer worldwide.

• Priced to spread: Fable 5 launched at $10 input / $50 output per million tokens, less than half Mythos Preview, and shipped free on Pro, Max, Team, and Enterprise plans through 22nd June.

• Guardrails by design: Queries touching cybersecurity, biology, chemistry, and model distillation were routed to Claude Opus 4.8 instead, and Anthropic said the safeguards triggered in under 5% of sessions.

• Pulled by order: On 12th June, the government invoked export controls, citing a reported Fable 5 jailbreak, disabling both Fable 5 and the gated Claude Mythos 5 that ships through Project Glasswing. The order carried no published safety finding and no wind-down for customers who had paid to upgrade days earlier.

Alex’s take: Until last week frontier AI sold like a gym membership. One flat monthly fee, and the best model was yours to use as much as you liked. Fable 5 changed that. It was going to be priced like electricity from 23rd June, a meter running while you worked at double Opus 4.8's rates, where whoever could afford to run everything at once would live very differently from the people counting every kilowatt. Then the grid got cut off at the source. A utility you can be unplugged from by government order is a different kind of utility, and anyone who had already wired Fable 5 into their daily work found that out three days in. This is why it’s vital to treat models as you would treat any single supplier of something essential—with a second one kept on standby for the day the power goes out.

Apple

2. Apple’s Gemini Gambit

At WWDC, Apple introduced Siri AI, a ground-up rebuild of its assistant, with Google’s Gemini models doing the heavy lifting in the cloud.

• Three places to think: Requests run on-device, through Private Cloud Compute, or on Gemini-based cloud models, with Apple insisting its privacy rules don’t change.

• Locked out of Europe: Apple says conflicts with the Digital Markets Act mean Siri AI won’t ship in the EU with iOS 27, with no timeline for the 27 member states. UK users are very much used to the waiting game the EU plays.

• Pick your own assistant: iOS 27 lets users set a third-party AI model as their default assistant, putting Claude and ChatGPT a settings toggle away from over a billion iPhones.

Alex’s take: A decade of Apple insisting it would do AI on its own terms ends with the company renting a rival's model, in a deal reported at roughly $1 billion a year. Apple keeps the orchestration, the privacy, and the distribution, and outsources the one piece it repeatedly failed to build. The iPhone is now contested distribution amongst frontier labs as there is a direct route to the most valuable customer base in consumer tech. It’ll just mean for the 450M users this side of the pond that we’ll have to wait for Brussels to clear the regulation before we can finally get hold of it.

SpaceX

3. SpaceX’s Record Raise

SpaceX priced its IPO at $135 per share and began trading on the Nasdaq on Friday under the ticker SPCX, raising around $75 billion at a valuation near $1.75 trillion, the largest listing in history.

• Pricing the record: The roughly $75 billion raise is more than double Saudi Aramco’s $29.4 billion in 2019, the previous record, after a Goldman-led 21-bank roadshow through New York, Boston, and San Francisco.

• Starlink carries it: The S-1 shows $18.7 billion in consolidated 2025 revenue post-xAI merger, with Starlink contributing $11.4 billion and posting a $1.19 billion operating profit in Q1 2026 across roughly 10.3 million subscribers.

• AI inside the rocket company: The listing folds in xAI, Grok, and the Colossus data centres absorbed in February’s merger, giving public investors direct exposure to a frontier AI lab months before Anthropic and OpenAI are expected to list.

Alex’s take: When SpaceXAI leased Colossus 1 to Anthropic, the lesson was that compute is the spice, and this IPO is the spice trade going public. The proceeds bankroll Colossus 2 and the platform for orbital data centres, which is the long game nobody else can play (yet). Both Anthropic and OpenAI are expected to tap the same institutional capital pool later this year, and a strong SPCX debut will help steer their pricing conversations. At roughly 110x trailing revenue, public investors are paying early-stage software prices for a capital-intensive launch business, and the first real fundamental reading won't arrive until the maiden quarterly report around September. Let’s see how the price action follows over the coming week.

Content I Enjoyed

Be careful what you legislate for

Dario Amodei, CEO of Anthropic, published a sweeping policy essay this week arguing that AI is outrunning the institutions built to govern it. He opens with Treebeard, the ponderous Lord of the Rings tree who needs a full day to say hello, as his image for a policy system too slow to track a technology that went from barely writing a line of code to writing most of the code at the major AI companies inside four years.

His central idea in the essay is that we should model AI regulation on agencies like the Federal Aviation Administration (FAA). This would involve mandatory third-party safety testing, with the government holding “the power to block or deter deployment of the model” if one fails.

Let’s double-click into the timing of this all just for a second. Anthropic launched its two most capable models, Fable 5 and Mythos 5, on 9th June. Amodei published the essay on 10th June. On 12th June at 5:21pm ET, the US government ordered Anthropic to disable both for every customer worldwide.

The directive ran through export controls, citing a reported way to jailbreak Fable 5, rather than the safety regime he proposed. Both are levers Amodei champions in the same essay. The institution he spent the week trying to wake moved faster than he expected, and the first frontier model it pulled was his own.

His proposal also highlighted the need for protection against arbitrary or politicised decisions. Interestingly enough, the order that hit Anthropic carried none of them: no published safety finding, no specifics, and no wind-down for customers who had paid to upgrade days earlier. The slow institution he wanted to wake has woken. The open question is whether it wakes in the shape he described, or as a faster hand on the off switch.

Idea I Learned

The crawl-to-referral ratio

For thirty years, the web ran on a quiet bargain: you let crawlers read your pages, and in return, search sent readers back to you. Cloudflare Radar now tracks how that exchange is holding up, and in the last full week of May, it had collapsed for the AI platforms. Anthropic’s crawlers read around 10,300 pages of the open web for every single human visitor the company referred back. OpenAI sat at 904 to one, Perplexity at 193 to one, and Google’s traditional crawler at 5.2 to one.

Google takes roughly five pages and hands a reader back, the deal publishers have lived with for two decades. The AI platforms take hundreds or thousands and return almost nothing. They read the web at an industrial scale to train their models and answer questions, and the clicks that once flowed back to the source mostly never come.

Cloudflare flags one caveat that cuts in publishers’ favour. When someone reads an answer in the Claude or ChatGPT app, that visit often arrives with no tag indicating where it came from, so the real imbalance is probably a little smaller than the raw number suggests. The direction is not in doubt, and the gap with ordinary search is still enormous.

This single ratio explains the fights now spreading across the industry. Publishers block AI crawlers in their robots.txt files, strike licensing deals, and sign up for schemes like Cloudflare’s pay-per-crawl, which charges a fee every time a bot reads a page. Each one tries to put a price on something that used to be free because it paid for itself in traffic.

If you publish anything online, this is the number to watch. It frames the central economic question of the next five years: who pays for the content AI learns from, now that being read no longer comes with a reader attached.

Quote to Share

The Kobeissi Letter on SpaceX’s record-breaking IPO:

SpaceX priced at $135 a share on 11th June, raising a record $75 billion at a $1.77 trillion valuation, and closed its first day up 19%. That comfortably eclipses Saudi Aramco’s 2019 listing, the previous record at $29.4 billion, to become the largest IPO in history.

The demand was enormous. Retail orders ran past $70 billion, with Bloomberg later citing the figure above $100 billion. That is the size of the queue rather than the size of the cheque, since the book was heavily oversubscribed and retail was allocated only around a fifth of the shares.

The comparison worth drawing here is against Tesla. When Tesla went public in June 2010, it was priced at $17 and was worth roughly $1.7 billion, and the stock has since delivered returns in the order of 200-fold for those who held through the volatility. SpaceX is arriving at about $1.77 trillion, which is roughly a thousand times Tesla’s debut valuation and around 110 times its own trailing revenue. The asymmetry that rewarded early Tesla holders came from buying a small company with everything still to prove. SpaceX is already one of the most valuable companies on earth before it has reported a single public quarter.

History tends not to flatter the biggest, most-hyped debuts. Rivian is down more than 80% since listing, Coinbase and Coupang are both off well over half, and Figma has fallen around 70%. The wider the fanfare at the open, the harder the first year has usually been.

I don’t doubt SpaceX is an extraordinary company, and the scale of the demand genuinely is historic. But historic demand and historic returns are not the same thing, though. From a $1.77 trillion starting line, the maths that turned Tesla into a generational trade gets a great deal harder, and that’s the figure I’d keep in view rather than the size of the order book.

Source: The Kobeissi Letter on X

Question to Ponder

I think it’s less than the labs would like, and perhaps more than Apple is letting on.

For everyday tasks, frontier models have converged to the point that most everyday people couldn’t pick Gemini, Claude, or ChatGPT out of a lineup, which is exactly why Apple felt safe renting rather than building. This costs Apple ~$1B/year to buy a competitive Siri without playing catch-up with their own foundation models.

However, the real difference still shows up at the edges. Long, complex work, agentic tasks that run unattended for hours, and judgment under genuine ambiguity remain the places where one model pulls clear of another, which is where this week’s Fable 5 launch moved the frontier again. Most people will never actually push a task hard enough to feel that gap. OpenAI models through Codex are exceedingly good at this long-horizon work, too.

What matters most in this scenario is the default position of putting Gemini inside Siri, putting it in front of more than a billion people who will never open a settings menu to choose anything else, even though the option to switch model provider for Siri is available. Distribution beats preference almost every time.

As AI assistants go mass market, the power shifts from whoever holds the best model to whoever owns the moment a person asks their first question.

Already a subscriber? Get your whole team on board. Signal Pro group subscriptions give everyone access to weekly AI workflows and tutorials, practical upskilling that pays for itself. It’s the kind of thing L&D budgets were made for. Share this with your manager today.

Get a group subscription

💡 If you enjoyed this issue, share it with a friend.

Refer a friend

See you next week,
Alex Banks

P.S. Bezos is back as CEO for his new AI startup Prompetheus.

Hey friends 👋

Welcome back to the Signal Pro.

Predictions about AI and work now stretch from one pole to the other, from warnings that almost every job is about to disappear to reassurances that very little will really change. Anthropic’s CEO managed to occupy both ends himself inside a single year. He told Axios in 2025 that AI could remove half of all entry-level white-collar jobs and push unemployment as high as 20% within five years, and by 2026, he had softened that line considerably, leaning instead on the idea that automating most of a task mainly frees people up to do more of the rest. For anyone trying to make an actual decision about their own career, a range that wide is close to useless.

Rather than add another forecast to that pile, this issue works from what the data already shows and from the one question you can actually answer about your own role today. The worry I hear most often, in some version or another, came from a reader in a recent Sunday newsletter: “I feel like I’m falling behind with AI, and I don’t know if my job is safe.” That worry is reasonable, and there is a clear way to turn it into something you can act on.

Today I’ll show you:

1. What the data actually says about AI and jobs, separated from the noise

2. The distinction that defines the roles under pressure from those holding up

3. Where the value in knowledge work is moving, and what that means for you

4. The exact 20-minute audit I run on my own role to find where I’m exposed

5. The prompts and 90-day plan to reposition before it matters

What the data actually says

The broadest and most credible map we have is the World Economic Forum’s Future of Jobs Report 2025, built from more than 1,000 employers covering 14 million workers. Its central projection points to large-scale churn, with as many roles created as lost and then some. By 2030, it expects 170 million new roles to appear and 92 million to be displaced, a net gain of 78 million, with around 22% of all jobs changing in some way. Roughly 39% of the skills a given job requires are expected to shift over the same period, down from 44% two years earlier. The fastest-growing roles cluster in technology, including AI and machine learning specialists, big data and fintech roles, software developers and security specialists, while the steepest declines fall in clerical and administrative work.

The more uncomfortable signal shows up among people just entering the workforce. Researchers at the Stanford Digital Economy Lab tracked millions of payroll records and found that workers aged 22 to 25 in the occupations most exposed to AI have seen a 13% relative decline in employment since late 2022, with young software developers down close to 20% from their peak, even as their more experienced colleagues held steady.

More recently, graduates are sitting at around 5.6% unemployment against roughly 4.2% for the workforce as a whole, and the largest technology firms are hiring around half as many new graduates as they did before the pandemic.

The New York Fed also published an analysis attributing roughly two-thirds of the rise in young-graduate unemployment to remote work, on the logic that firms are wary of hiring inexperienced people into roles where mentorship is hard to provide.

On the recovery side of the story, graduate hiring is projected to rise modestly this year, and internship postings are up sharply. But it’s fair to acknowledge that early-career knowledge work is under real, measurable strain, with AI as one driver among several and the damage concentrated in particular kinds of work.

Automate or augment

Double-clicking into the work by Stanford Digital Economy Lab, we can see clearly where the declines are concentrated. Employment fell in the roles where AI takes over the task itself, and held steady or grew in the roles where AI speeds up work a person already does and still owns.

A junior analyst whose day mostly consists of turning a brief into a first-draft model or memo is doing work that current tools can largely produce on their own. A client lead who uses those same tools to prepare faster, but whose real value is the judgment and the relationship in the room, is doing work that those tools make sharper. The first role sits in what I’d call the automate zone, and the second in the augment zone. Whether your own position holds over the next few years depends far less on your industry and far more on where your daily tasks fall along that line.

This is important in two parts. First, you need to hold up the reflective lens. Then, you need to thoughtfully apply what you see to your own situation. In the rest of this issue, I’ll show you where the value in knowledge work is moving, then walk through the exact audit I run on my own role each quarter, including the prompts, so you can map your own position and act on it this week.

Where the value is moving

Read more

Hey friends 👋 Happy Sunday.

Here’s your weekly dose of AI and insight.

Every Wednesday, Signal Pro members get a step-by-step AI workflow they can apply immediately. No fluff, just practical guides to upskill you and your team. If you’re only reading the Sunday issue, you’re getting half the picture. Upgrade to Pro today.

Upgrade to Pro

AI Highlights

My top-3 picks of AI news this week.

Microsoft

1. Microsoft’s Major Build

Microsoft used its Build 2026 developer conference in San Francisco to set out a full in-house AI stack, spanning seven of its own models, a proactive work agent and a new hardware device platform.

• Seven MAI models: Microsoft’s AI Superintelligence Team launched seven first-party models, led by MAI-Thinking-1, a 35-billion-parameter reasoning model trained from scratch. Microsoft says blind raters preferred it to Claude Sonnet 4.6 and that it matches Opus 4.6 on a coding benchmark, though those results are its own and not independently verified.

• A work autopilot: Microsoft Scout, its first always-on “Autopilot” agent, is built on OpenClaw and Work IQ to handle meeting prep, scheduling and routine tasks across Teams and Outlook without being prompted, and is rolling out now to Frontier customers.

• Agent-first hardware: Project Solara, a new platform for agent-first devices, debuted with two reference designs, a Qualcomm wearable badge and a MediaTek desk hub, with pilots planned alongside Best Buy, CVS Health and Target.

Alex’s take: I was in SF for Microsoft Build this week and it was powerful to see Microsoft taking the leap across every layer: its own models, work agents, and device platforms. They’re building long-term self-sufficiency whilst keeping the platform itself deliberately open, with the ability to use Claude or OpenAI in the same workflow. For years, Microsoft has leaned on partners for its frontier models. Now it has a full stack of its own to build on too which gives them a lot more optionality than what they had a year ago. I’m especially excited by the general release of the Microsoft Discovery platform to build AI workflows across scientific and engineering disciplines. Alongside the powerful generalist models, I think we'll keep seeing specialist ones built for a single hard problem, like AlphaFold for protein folding and MatterGen for new materials.

NVIDIA

2. NVIDIA’s Carry-On Cloud

At its GTC Taipei keynote, NVIDIA unveiled RTX Spark, a chip built to run AI agents directly on a Windows laptop or desktop rather than in the cloud.

• The silicon: RTX Spark pairs a Blackwell RTX GPU with NVIDIA’s own 20-core Arm CPU and up to 128GB of unified memory, hitting around a petaflop of AI compute, enough to run a 120-billion-parameter model with a million tokens of context locally.

• Agents on the device: Working with Microsoft, NVIDIA added new Windows controls and a runtime called OpenShell so personal AI agents run on your own machine and route only what you allow out to the cloud.

• Shipping this autumn: NVIDIA says more than 30 laptops and 10 desktops arrive this autumn from Dell, HP, ASUS, Lenovo and MSI, led by Microsoft’s Surface Laptop Ultra and its cooling system built for sustained on-device AI.

Alex’s take: For the past few years, NVIDIA sold the shovels that built everyone else’s AI inside the data centre. Now they want a seat on your desk. This is a smart play as there is an enormous moat when it comes to inference. Google is the exception, serving its own models on Ironwood and now renting that silicon to Anthropic and Meta. When many more people start asking an agent to do stuff that lives directly on your computer, whoever owns that agent owns the front door to computing. NVIDIA has planted itself and Windows there first. AMD, Intel and Qualcomm stocks slid on the news as Wall Street recognised the threat. But it makes me think about what Apple’s response will be. Their M5 Max is more powerful than RTX Spark. Double the memory bandwidth with 128GB of unified memory too alongside a slightly faster CPU and GPU. Apple has the better silicon. What it's missing currently is the agent layer that NVIDIA and Windows have just moved into.

OpenAI

3. OpenAI’s Plumbing Play

OpenAI spent the week widening who it builds for, from a smarter ChatGPT memory for everyday users to role-specific Codex agents for non-coders and new life sciences models for drug discovery.

• Memory that stays current: ChatGPT’s new Dreaming system revises stale memories on its own, turning “you’re going to Singapore in July” into “you went in July 2026” once the trip passes. It’s reaching free users over the coming weeks too, after OpenAI cut the cost of running it by roughly 5x.

• Codex for every role: Six new role plugins turn Codex into a specialist for non-coders, bundling 62 apps and 110 skills across sales, data analytics, creative production and product design.

• A model for science: OpenAI upgraded GPT-Rosalind, a model series built for life sciences research, adding GPT-5.5’s agentic coding and tool use for drug discovery, analysis and experimental work.

Alex’s take: This was a big shift for Codex. Asking it for something now returns a working URL a colleague can open, the same promise that built no-code firms like Lovable and Replit into multi-billion-dollar businesses. Interestingly enough, OpenAI now lists both as launch partners for the feature, so it is arming the firms it competes with while quietly absorbing the easy end of their market. The real prize underneath all of this is ownership. It’s about who holds the workflow, the data, and in turn, the trust that comes with it. That is worth far more now than having the biggest model, and it is a much harder thing for a rival to take back. OpenAI clearly knows it, and is trying to plant a flag in every workflow before anyone else gets there.

Content I Enjoyed

Claude is writing the next Claude

In a new report, When AI builds itself, Anthropic’s own institute sets out how much of the company already runs on Claude. More than 80% of the code merged into its codebase in May was Claude’s, up from low single digits before its coding agent arrived last year, and the average engineer now ships about eight times as much as in 2024. Anthropic is using Claude to build the next Claude, and it’s now showing us its working.

On a test of speeding up training code, a skilled engineer manages a fourfold gain; Anthropic’s best model reaches 52 times. However, Claude still cannot judge which problems are worth solving. This is the gap between a useful assistant and a system that could build its own successor.

Having shown how fast it can move, Anthropic gives the rest of the report over to arguing for a way to slow down. It calls for the machinery of a pause that several countries could trigger together, each able to confirm the others had genuinely stopped. Recursive self-improvement, the authors stress, is neither here yet nor inevitable, and they want “the option to slow or temporarily pause frontier AI development” in place before it arrives.

All of this arrived the same week Anthropic confidentially filed to go public, on the very acceleration the report warns about. To some it is a company being honest about a danger it profits from; to others, the best sales pitch it could make, that its product may be too powerful to leave unchecked.

A lab genuinely afraid of losing control would not be racing to build its most powerful model yet and keeping it in-house, while publishing the figures that show how far ahead it sits.

The pause itself would need verification tools that nobody has built, and the company is offering to create them, against an incentive to cheat that took decades to contain for nuclear weapons and hides far more easily in a training run. The proposal, then, is a request to help build a brake, made before the brake exists. A pause is worthless if a single lab pulls it, since that lab only forfeits its lead, and the firms sounding the warning are the same ones that cannot afford to stop first.

Idea I Learned

AI's Bioweapon Chokepoint

Sam Altman, Dario Amodei and Demis Hassabis (chiefs of three rival AI labs) have signed an open letter to Congress backing mandatory screening of synthetic DNA orders, and of the machines that print them. They were joined by Microsoft’s Mustafa Suleyman, Meta’s Alexandr Wang, Nobel laureate David Baker, former secretaries of the US Army and Navy, and the chief executives of the biggest DNA-synthesis firms.

AI models can now outperform PhD-level virologists on highly technical laboratory questions, and the letter argues this could erode the knowledge barrier that long kept bioweapons out of reach. This can’t be contained at the model level: capability is spread across rival labs, and safety refusals can be talked around. Every dangerous experiment still requires physical DNA, and that comes from a short list of vendors.

Because that supply is so concentrated, the letter would put the screening burden on those companies, requiring them to check each order for dangerous sequences and verify the customer before shipping, then log everything so that anything slipping through can be traced back to the buyer. The signatories don’t claim AI has already made bioweapons easy to build. They admit today’s evidence is mixed, and their worry is about pace: models are improving fast enough that the safeguard is worth having in place before the danger materialises. A bipartisan Senate bill would provide it, giving the Commerce Department a year to write the screening rules.

Screening the DNA works because it sits outside the AI. Models can be tricked, and each new one is more powerful than the last, but a bad actor still has to order the physical DNA from somewhere, and that order can be checked. The companies that sell the DNA would be the ones doing the checking, and they signed this letter too. Firms rarely ask to be regulated, so when these ones did, I think it’s worth us paying attention to.

Quote to Share

NVIDIA’s Jensen Huang on who’s winning the AI race:

Huang was asked on CNBC which company in the world uses AI better than anyone else, and he skipped past OpenAI and Google, whose chips run most of the AI being built today, to point at one of his own customers. Meta swapped the classical recommendation system that once ran on ordinary processors for a generative, agentic one that now decides what billions of people see and how advertisers reach them.

The endorsement comes with an obvious interest attached, because Huang is also Meta’s supplier. In the same window this clip first did the rounds, Meta committed to Nvidia’s full stack and its next-generation chips, only months after it had publicly weighed moving a chunk of its workloads onto Google’s processors instead. So against that backdrop, “nobody uses AI better than Meta” sits very close to “nobody buys our chips like Meta.”

Source: WOLF on X

Question to Ponder

Most people freeze in front of a blank AI chat window. They think they need a course first or a prompt engineering guide.

You don’t need any of that.

Open the voice memo app on your phone. Talk for 10 minutes about what you’re working on, what problems you’re trying to solve, what you want to accomplish. Don’t structure it. Just talk. When you think you’re done, keep going.

Then transcribe that audio using a free tool like ElevenLabs and paste it straight into Claude (using Opus 4.8 High with thinking enabled).

This works because when you speak, you think differently from when you type. You ramble. You circle back. You surface ideas you didn’t know you had. That messy, unfiltered stream of consciousness gives a far richer context than a carefully typed five-line prompt ever could.

Your brain can outpace your keyboard. So let it.

The people getting the most from AI right now are those who are most curious. They started before they felt ready, experimented without a playbook, and built the habit through persistence over perfection.

So stop waiting for the right course or the right tool. The interface is natural language, which we speak every single day.

Pick up your phone, hit record, and start talking.

Already a subscriber? Get your whole team on board. Signal Pro group subscriptions give everyone access to weekly AI workflows and tutorials, practical upskilling that pays for itself. It’s the kind of thing L&D budgets were made for. Share this with your manager today.

Get a group subscription

💡 If you enjoyed this issue, share it with a friend.

Refer a friend

See you next week,
Alex Banks

P.S. A robot wearing a clown wig kicked a child in the stomach.

Hey friends 👋

Welcome back to the Signal Pro.

This week is less of a how-to and more of a mirror. We’re going to look at what AI is actually doing to the quality of work moving around your organisation, and to your own head, and then I’ll give you the system I use to stay on the right side of it, and to bring the people around you with you.

We’re all producing more than we ever have. Decks, memos, updates, summaries, all of it faster and shinier than a year ago. And yet most of it lands with a strange weightlessness, as if everyone’s running on the same treadmill, generating content at speed and going nowhere.

You’ve probably felt it from both ends, and if you’re honest, you’ve probably added to the pile yourself.

Start on the receiving end, because that’s where it’s easiest to feel.

A colleague sends you a document that looks complete with clean formatting and a confident, polished layout. You start reading, and somewhere the “AI tells” start appearing, and a quiet dread sets in, because there’s nothing actually in there. The words are arranged correctly, but they don’t really say much. Reading it twice doesn’t help. You start to second-guess yourself on whether you’re missing the point or whether there was never a point to miss.

What you’ve been handed is slop.

The person who sent it spent maybe a couple of minutes making it, while you’re about to spend the next hour and a half working out what they meant, or writing back to ask the questions the document should have answered in the first place.

The pattern is becoming increasingly common, and researchers at BetterUp Labs and Stanford’s Social Media Lab call it workslop. AI-generated content that looks like good work but lacks the meaty substance to move the task forward. An MIT report found that 95% of organisations see no measurable return on their AI investment, so adoption is everywhere while results are almost nowhere. Workslop is a large part of why.

Today I’ll show you:

• The asymmetry that makes workslop so expensive, and why it feels like progress to the person creating it

• What the same shortcut quietly costs you, even when no colleague is involved

• How to spot slop in your inbox

• The 90-second test and prompt strategies to stop you producing it

• How to upskill the people around you so they stop producing it

• The anti-workslop principles worth pinning up

Let’s get into it.

The maths of slop

It’s tempting to classify AI as a time-saving asset, and sure, it often earns that label. But workslop introduces a new kind of accounting that nobody put on the balance sheet.

The researchers surveyed 1,150 US desk workers, and 40% said they’d received workslop in the past month. Each instance cost them an average of one hour and 56 minutes to deal with. The send feels fast, but the whole system quietly slows down.

The team estimated an invisible tax of around $186 per employee per month, which works out to more than $9 million a year in lost productivity across a 10,000-person organisation, all of it spent on work that merely looked done. When the researchers asked people how receiving workslop made them feel, more than half said it left them annoyed, and many admitted to feeling offended. I’d agree heavily—being handed a document that pretends to be finished takes the form of a small insult, because you start to wonder whether the sender thought about the problem at all, or simply routed it through a large language model and forwarded the output with their name on top.

Around half the people surveyed said they now view colleagues who send slop as less capable and less reliable. The shortcut meant to make someone look productive quietly does the opposite, and every piece of slop you send becomes a small withdrawal from how seriously people take you.

What makes this even more real is that nearly one in five people who use AI at work admitted to sending workslop themselves. The lazy version feels identical to the diligent one, right up until someone else opens the file and realises the contents are wholly lifeless, which is exactly why so many well-meaning people produce it without ever noticing. Or worse, realise exactly what they’re doing but brazenly kick the can of consequence that much further down the road.

The same trick, played on yourself

Last year, a team at MIT’s Media Lab ran a study with the title “Your Brain on ChatGPT.” They sat 54 people down to write essays under one of three conditions: one group used ChatGPT, one used a search engine, and one wrote with nothing but their own head. Everyone wore an EEG cap so the researchers could watch what their brains did while they worked.

The brain-only writers lit up the most, with strong, widely distributed neural connectivity that signals real cognitive effort. The search-engine group sat in the middle, and the ChatGPT group showed the weakest connectivity of the three.

The researchers gave the effect a rather apt name of cognitive debt. When you offload the thinking, you borrow the mental effort from your future self at interest. The study found that the AI group struggled to quote from essays they’d supposedly just written and reported little ownership over the final piece. When some of those users were later asked to write again without the tool, the under-engagement carried over, as though the muscle had gone quiet and didn’t switch straight back on.

Now, this was a small study, and the paper was a preprint that hadn’t cleared peer review when it was dropped, so I wouldn’t build a comprehensive worldview on just 54 people. But the mechanism is intuitive enough that you’ve probably felt it anyway, in the report you let AI write and then couldn’t really explain in the meeting, or the analysis you accepted without checking and then couldn’t defend when someone pushed back. That hollow sense of having produced something you don’t actually understand is cognitive debt, and the repayment terms are worse precisely when you can least afford them: in the room, under questioning, with your name on the work.

All of this points to the very qualities that make AI feel productive, its speed and its polished, finished look, are exactly what disguise where the work actually went, whether sideways to a colleague or forward to the version of you who has to defend it next week. Either way, the relief you feel in the moment is borrowed.

I’m not suggesting you use AI less, since I use it constantly. The move that works is getting your judgment into the output before it leaves your hands, so the work carries real thought rather than just the appearance of it. And once you can see slop clearly, you can do something more valuable than fixing your own, which is helping everyone around you stop making it.

Here’s exactly how I do all three.

Read more

Hey friends 👋 Happy Sunday.

Here’s your weekly dose of AI and insight.

Every Wednesday, Signal Pro members get a step-by-step AI workflow they can apply immediately. No fluff, just practical guides to upskill you and your team. If you’re only reading the Sunday issue, you’re getting half the picture. Upgrade to Pro today.

Upgrade to Pro

AI Highlights

My top-3 picks of AI news this week.

Anthropic

1. Claude's Honest Haul

Anthropic upgraded its flagship LLM to Opus 4.8, a sharper, more capable model at the same price as 4.7, and on the same day announced a fresh funding round that values the company at nearly a trillion dollars.

• Honesty gains: On its own evaluations, Opus 4.8 is around 4x less likely than 4.7 to let flaws in code it has written pass unremarked, and early testers say it flags uncertainty and pushes back on weak plans rather than bluffing through them.

• Dynamic workflows: Alongside the model, Claude Code gained dynamic workflows in research preview, letting Opus 4.8 plan a task, spin up hundreds of parallel subagents in one session, verify the output, and carry codebase-scale migrations across hundreds of thousands of lines from kickoff to merge.

• Series H round: On the same day, Anthropic raised $65B in Series H funding at a $965B post-money valuation, led by Altimeter, Dragoneer, Greenoaks and Sequoia, with run-rate revenue that crossed $47B earlier this month.

Alex’s take: 9 days after Andrej Karpathy joins Anthropic and they ship Opus 4.8. Coincidence? I think not. I’ve personally found it to be less chatty than 4.7 and older Opus versions, however, I haven’t really noticed any tangible differences in quality between 4.7 and 4.8 over the last 48hrs of testing.

One thing to note is Anthropic has removed “Adaptive thinking” and replaced it with “Effort”. Now, you can specify Low/Medium/High/Extra/Max effort for a task AND have the option of engaging a “Thinking” toggle for more complex tasks too. My default is “Extra” with “Thinking” toggled on for most tasks, as it balances high-quality answers without the generation taking too long. If I’m kicking off a project or undertaking a new task that requires deep thought, I’ll engage “Max” effort for that extra reasoning power.

Figure

2. Figure Leaves the Line

Figure signed a commercial agreement with Catalyst Brands (the retail group behind JCPenney, Aéropostale and Brooks Brothers) to deploy its humanoid robots at scale, starting in the company’s distribution and logistics centre in Reno, Nevada.

• Deployment scope: The agreement moves Figure’s humanoids into Catalyst’s distribution and logistics network, beginning at a single centre in Reno, with Figure positioning the humanoid as the one machine flexible enough to redeploy across an entire multi-brand portfolio.

• The BMW pivot: Figure’s headline deployment until now was automotive: its Figure 02 robots ran 11 months at BMW’s Spartanburg plant and contributed to more than 30,000 X3s before being retired. Catalyst marks the move from one tightly controlled car line into general retail logistics.

• Manufacturing scale: Figure 03, the production-grade model built for mass manufacturing, comes off BotQ, Figure’s in-house factory, whose first line is rated for up to 12,000 humanoids a year, with the company planning to scale well beyond that.

Alex’s take: The humanoid is by far the most flexible form factor to be picked up and dropped into different logistics scenarios—especially those that are most physically demanding across the supply chain. An idea I come back to a lot is that the world was built by humans for humans. Having a robot emulate that form and be able to conduct tasks as good as, if not better than us in undesirable or dangerous situations, enables us to shift toward higher-value work.

SpaceX

3. SpaceX’s Big Bang

SpaceX filed its IPO prospectus with the SEC on 20 May 2026, applying to list its Class A shares on Nasdaq under the ticker SPCX in what is widely expected to be the largest stock-market debut in history.

• The filing: SpaceX filed its S-1 registration statement on 20 May 2026 and applied to list on Nasdaq and Nasdaq Texas under the symbol SPCX, with a roadshow reported for early June and pricing as soon as ~11 June.

• The size: Reported targets put the valuation at $1.75T–$2T and the raise at about $75B, which would be the largest IPO ever, though the prospectus itself leaves the share count and price range blank, as is standard at this stage.

• The AI angle: The listing now houses Musk's entire AI operation. SpaceX absorbed xAI in early February, then, in May, folded Grok and X into a new in-house division branded SpaceXAI, so the AI arm's losses sit inside the IPO. In 2025, it lost roughly $6.4B from operations against Starlink's ~$4.4B operating profit, with a dual-class structure keeping Musk in control.

Alex’s take: The numbers on this IPO are completely wild. I especially enjoyed reading Andrew Benson’s cap-table breakdown. At a reported ~$1.75T price, roughly 400-500 staff clear $100M-plus. These returns are a function of the year they were hired rather than what they built once inside. To quote exactly, “Early isn’t a strategy. It’s a date stamp.”

Three index providers have rewritten their rulebooks just in time for this listing. Nasdaq's "fast entry" rule, effective 1 May, lets a top-40 company join the Nasdaq-100 after 15 trading days instead of the usual three-month-plus wait, and scraps the minimum-float requirement. FTSE Russell followed on 26 May with a five-trading-day path into the Russell 1000 for anything above its ~$17.5B threshold. And S&P Dow Jones is consulting on halving its seasoning window from 12 months to six and waiving the four-quarter profitability test for the largest entrants. SpaceX qualifies for all three with room to spare.

The reason why this matters is that SpaceX is floating only around 3-5% of its shares, and its lock-up is unusually front-loaded. Insiders can start selling roughly 60 days after listing (with tranches at 70, 90, 105, 120 and 135 days), versus the standard 180. So passive pension and 401(k) money gets routed into the stock within weeks, near peak pricing, just as insiders begin heading for the exit.

SpaceX may grow into the valuation, but just about every stock that’s been added to the Nasdaq over the past year or two has had a 30-40% haircut. Let’s see how everything unfolds over the coming months.

Content I Enjoyed

The Lost Generation

A major review landed in Britain this week, finding that more than a million 16-to-24-year-olds are not in employment, education or training (Neet). This is the highest in twelve years, with the review’s author. Alan Milburn, warning that the bottom rungs of the work ladder are proving harder to climb than ever.

He put the annual cost to the country at £125bn and forecast that one in six young people would be stuck there within five years, heading for a “lost generation”. The blame is spread across fewer entry-level jobs, falling apprenticeships, poor health and weak welfare support. AI gets a mention, named as likely to add further pressure on the roles that have long given young workers a foothold, but Milburn refuses to make what he calls "unknowable assertions about the future".

Earlier this week, Sam Altman, who once warned that entire job categories would vanish, told a Commonwealth Bank conference in Sydney he was “delighted to be wrong”. The entry-level wipeout he braced for hadn’t arrived; his intuitions were simply off.

Dario Amodei, who told Axios last year that AI could erase half of all junior white-collar roles and drive unemployment as high as 20%, now offers a softer story, where you automate most of a job, and the slice that remains expands to fill the day, leaving the worker ten times more productive.

More than 115,000 tech workers have lost their jobs in 2026, Meta and Amazon among those blaming AI, yet the Yale Budget Lab keeps finding no real change across the roles most exposed to it. The only group clearly feeling it is recent graduates, taken on for fewer entry-level jobs than the cohorts before them.

The young people on that bottom rung describe pay that won't cover rent, a benefits system that makes a low-paid job pointless, and a country short of companies worth joining. The founders may have changed their minds in good faith, but the new optimism arrived almost exactly as OpenAI and Anthropic began courting investors for public listings, where a sunny growth story sells better than a forecast of mass unemployment.

Idea I Learned

The Pope Wants to Disarm AI

On 25 May, Pope Leo XIV released his first encyclical, Magnifica Humanitas, devoted entirely to artificial intelligence, and presented it himself, breaking with custom. He had signed it ten days earlier, on the 135th anniversary of Rerum Novarum, the 1891 letter his namesake wrote, treating the Industrial Revolution as a test of human dignity. Leo makes the same claim about AI.

The encyclical’s central argument surrounds power. Leo writes that technology is never neutral because it takes on the priorities of those who build, fund and govern it, and the main drivers of AI are now private, often transnational companies whose resources outstrip many governments. He calls for AI to be “disarmed” and points to autonomous weapons, weakened human relationships and the slow erosion of critical thinking.

Anthropic co-founder Chris Olah was the only frontier-lab figure invited to speak, and he took the podium before the Pope. He told the room that every lab, his own included, operates inside incentives that pull against doing the right thing: commercial pressure, the race to stay at the research frontier, geopolitics, and ambition. He ended by asking to be judged from the outside. The industry needs critics, he said: “moral voices that the incentives cannot bend.”

Three days later, Anthropic raised $65 billion at a $965 billion valuation, overtaking OpenAI as the most valuable AI company in the world. It had recently barred the US Department of Defense from using its models for weapons, straining its relationship with the White House.

Something worth calling out was how much this reminded me of a post I wrote last month about IBM’s 1979 training manual, in a line that has aged strangely well. “A computer can never be held accountable, so it must never make a management decision.”

A person who makes a bad call owes an explanation to a board, a regulator, or a court. A model owes nothing to anyone, and the companies building it now operate at a scale few governments can hold to account. I think what’s important here is that we as the user carry the accountability. The call is ours to make.

Quote to Share

Sam Altman on the future of intelligence:

Across the rulings that have come down in the US, training large language models on copyrighted books has been held to be fair use, provided the works were lawfully acquired, on the reasoning that learning from text to build something new is transformative.

The wins for creators have been narrower and tied to circumstance. Anthropic’s $1.5 billion settlement was for storing pirated copies of books, while the training itself was ruled fair use.

Plenty of people feel the work should belong to whoever made it. The right to train has so far sat with the model-builders. One judge went further and found that authors are not owed licensing revenue for training at all.

Until that picture changes, nothing stops the meter Altman described from running.

Source: Vivek Sen on X

Question to Ponder

When we look closely at the result that everyone is calling a breakthrough, the answer is mostly no.

Recently, OpenAI reported that one of its models had disproved the Erdős unit distance conjecture, a geometry question open since 1946. Days later, a Google DeepMind system cleared nine more long-standing Erdős problems. Independent mathematicians checked the geometry proof and confirmed it works.

But we must also look at what those mathematicians published alongside it. The proof relied on existing tools, most of them from number theory. This involved a counting argument from Ellenberg and Venkatesh, field constructions worked out years ago, and methods sitting in papers waiting for someone to pick them up.

The reviewers said plainly that the broad approach was not new and that the proof revealed no new geometric idea. What the model supplied was the move to drag number theory into a geometry problem nobody had aimed it at, and the patience to push a step further until it fit.

This is what we must recognise when we hear a model has “invented something new”. A system trained to predict the next word is, underneath, charting everything people have already written and finding paths across it that no single person had traced. A link between two distant ideas can indeed be a real contribution, even if the ideas themselves were already on the table.

Yet we mustn’t exploit that gap. Last autumn, OpenAI said a model had cracked ten Erdős problems, and the solutions turned out to be answers already published, resurfaced and dressed up as discovery, until the post came down.

So when an output looks new, we must narrow the question, and ask ourselves are the ideas new, or is the link between them new? With today's AI systems, it is nearly always the link. And a good link, found fast, is still worth having.

Already a subscriber? Get your whole team on board. Signal Pro group subscriptions give everyone access to weekly AI workflows and tutorials, practical upskilling that pays for itself. It’s the kind of thing L&D budgets were made for. Share this with your manager today.

Get a group subscription

💡 If you enjoyed this issue, share it with a friend.

Refer a friend

See you next week,
Alex Banks

P.S. Booster Robotics from Beijing showing off football skills.

Hey friends 👋

Welcome to another Signal Pro workflow issue.

You can have the best ideas in the room, but if your deck looks like a default PowerPoint template, people will be quietly discounting the thinking behind it. If you’ve spent evenings rearranging boxes, handing off presentations to designers, and waiting three days for them to come back, then this guide is for you.

In April, Anthropic shipped Claude Design. You describe what you want, Claude builds it on a canvas beside you, and you refine it by talking to it the way you’d brief a designer. When it’s ready, you can dive in and fine-tune the design directly, or export it straight to PowerPoint, PDF, or Canva.

Today I’ll show you:

• What Claude Design is, and how it’s different from a Claude chat artifact

• Why setting context first decides whether your output looks like your work

• The habit that turns a vague idea into a prompt with rich context

• How Claude interviews you, then builds your deck and speaker notes

• The four tools that take a rough draft to something you’d ship

• How to leverage GPT Image 2 with Claude Design

Let’s get into it.

What is Claude Design?

Claude Design is an Anthropic Labs product, still in research preview. It works on the Pro, Max, Team, and Enterprise plans (on Enterprise an admin switches it on), runs on your existing subscription limits, and lives at claude.ai/design, or in the Claude sidebar just under Claude Code.

The layout consists of two panels. You’ve got chat on the left and a canvas on the right. You describe what you want, Claude builds it on the canvas, and you steer from there.

It handles presentations, one-pagers (briefs, pitch summaries, reports), dashboards, and interactive prototypes. It builds the design and the prototype, then stops short of working software with a database behind it, so when a prototype needs to go live, you hand it to Claude Code.

You choose what you’re making the moment you start a project:

• Prototype — app or web mockups, with a wireframe (rough) or high-fidelity (polished) choice.

• Slide deck — presentations, with a Use speaker notes toggle on the setup screen.

• From template — start from a layout you’ve saved, so your structure and brand are already in place.

• Other — an open start for anything else.

If you’re making a presentation, I recommend turning the speaker notes toggle on. It tells Claude to keep the slides lean and push the detail into the presenter notes, which is how good decks work anyway. On screen, you have a clean slide, then the real substance appears in what you say. You finish with both the deck and a script to talk along to.

What sets it apart from a normal chat artifact:

• You click any element and ask for a change, instead of describing where it sits.

• You drag sliders to adjust spacing, colour, and layout live.

• It can hold your brand, so output arrives on-brand by default.

• You export to PPTX/Canva/PDF or hand a prototype to Claude Code to build.

Start with context

Before you prompt anything, give Claude some context, because that’s largely what decides whether the output looks like your work or a generic template.

Once you’ve assigned a project name, you have four ways to set it up before you’ve written a word:

• Design System — pulls in your brand’s colours, fonts, and components (more on this shortly).

• Add screenshot — drop in a picture of an existing deck, inspiration, or anything with a look you want to borrow.

• Attach codebase — point it at a repo so prototypes match your real product. Useful for product work, less so for a one-pager or presentations.

• Drag in a Figma file — bring an existing design straight in.

The + button on the chat box opens an attachment window. Here, you can reference another project, grab elements from a live website, or pull from a connected tool. And you can drop files straight onto the canvas at any point: images, documents, references, Figma links, or whole folders.

Taking the time to ground your project this way matters more than the wording of the prompt you use. Otherwise, Claude will make reasonable guesses about your brand. That’s why giving it some screenshots or an example deck enables the output to arrive already looking like your work on the first pass. This does more for the final result than anything else, yet it’s the step most people skip entirely. Context is everything.

From here, I'll take you through exactly how I use Claude Design. This will cover the workflow I now use for almost any design task, particularly presentations and pitch decks. I haven't opened PowerPoint to start a deck since I started using Claude Design, and I'd happily call it the best design tool out there right now for going from zero to one.

What you can put together in a few minutes still surprises me, and the rest of this issue is how to do it.

Read more

Hey friends 👋 Happy Sunday.

Here’s your weekly dose of AI and insight.

Every Wednesday, Signal Pro members get a step-by-step AI workflow they can apply immediately. No fluff, just practical guides to upskill you and your team. If you’re only reading the Sunday issue, you’re getting half the picture. Upgrade to Pro today.

Upgrade to Pro

Today’s Signal is brought to you by Athyna.

Hunting for AI roles takes hours you don’t have. Careers pages, group chats, internal referrals, a dozen job boards, all updated at random. The Athyna AI Job Board does the watching for you.

• We track openings at frontier AI companies like Anthropic, OpenAI, Mistral, and Perplexity

• We match them to your profile

• When a role hits a 75% matching index, we let you know

Set up a profile once. The matches come to you.

Find your AI role

AI Highlights

My top-3 picks of AI news this week.

Google

1. Google Goes Agentic

At I/O 2026, Sundar Pichai declared the “agentic Gemini era”, a shift from assistants that answer prompts to agents that take action across Google’s products, and rewired the company’s whole line-up around it.

• Flash beats Pro: Gemini 3.5 Flash is generally available now and beats last year’s Gemini 3.1 Pro on agentic benchmarks like Terminal-Bench 2.1 (76.2%) and MCP Atlas (83.6%), while running four times faster at under half the cost of rival frontier models. Gemini 3.5 Pro lands next month.

• Video that knows physics: Gemini Omni turns any text, image, audio or video input into editable video with real-world physics understanding, which Demis Hassabis called a step toward AGI.

• Search’s biggest change in 25 years: AI Mode has passed 1 billion monthly users and now defaults to Gemini 3.5 Flash, paired with the biggest redesign of the Search box in over 25 years, now reading text, images, files, videos and open Chrome tabs.

Alex’s take: It's easy to watch the Omni demos and file them under the "fun toy" category, but I think that acutally undersells what's being built. Demis framed it as "a major leap in world understanding". A model that actually grasps gravity, collisions and how materials behave is doing far more than generating clips. This is the same capability that robotics and self-driving systems need to predict what happens next in the physical world. The consumer video surface is just one expression of that: the world-model underneath is the real prize, and it's a genuinely useful tool however you come at it. As for Demis himself, to his deeper curiosity of the origins of life and the nature of reality, this has found another home in Isomorphic Labs, which lets him chase the research frontier while Google productises the rest. Both things can be true at once, and both really excite me.

Figure

2. Hark's Big Bet

Brett Adcock’s secretive lab Hark has raised one of the largest Series A rounds in AI hardware history to build a personal intelligence that sees, listens and remembers you, running on a new class of device built for AI rather than the phone.

• $6B before launch: Hark closed an oversubscribed round of over $700M at a $6B post-money valuation, roughly two months after leaving stealth and before shipping a product.

• Rival chipmakers aboard: Nvidia, AMD, Intel and Qualcomm all joined the round alongside Salesforce Ventures, putting the industry’s major silicon suppliers on one cap table.

• Models first, devices next: Hark plans to release its first AI models this summer, then a family of devices for the individual and the home, with form factor and price still undisclosed.

Alex’s take: Alongside Figure, Adcock is now building every layer of the AI stack at once: the models, the software, the hardware, and the interface. Both Nvidia and AMD have written cheques in this round which makes me lean in. GPU allocation and compute access are the scarcest resources of them all right now, it’s the lifeblood of training models and running day-to-day inference, so this strategic investment makes a lot of sense. However, there’s deep skepticism in the air off the back of Humane and Rabbit who were complete flops and failed to ship anything meaningful that the consumer actually wanted. The real test will be shipping velocity: if the first models come this summer and the product is put into the hands of the consumer to offer their verdict.

Starbucks

3. Starbucks’ Milk Run

Starbucks is retiring Automated Counting, the AI inventory tool it rolled out across North America just nine months ago, sending more than 11,000 coffeehouses back to counting milk and beverage stock by hand.

• The pitch: NomadGo’s system was sold on 99% accuracy and counts up to eight times faster than manual methods, using on-device computer vision and AR to scan shelves of milk jugs and syrups from a handheld tablet, against a manual-counting industry baseline of roughly 85%.

• The reality: In practice it frequently miscounted and mislabelled items, often confusing similar milk types or skipping them altogether, and Starbucks is now reverting to manual counting after nine months live.

• The turnaround stakes: Fixing product availability sits at the centre of CEO Brian Niccol’s Back to Starbucks plan, so scrapping the very tool built to do that is a pointed retreat for the strategy.

Alex’s take: Most people will read this as "AI failed." I think it’s important to call out the gap between a benchmark and a backroom. A tidy, controlled room with good lighting and neat shelves is one thing, but a busy backroom mid-rush takes the extreme end of the training data. Starbucks was quick to push this, 11,000 stores in roughly a month to be exact, rather than proving it slowly. When the counts came out wrong the work landed back on baristas, who had to correct the tool by hand. The big takeaway here for anyone putting AI into operations is that a single accuracy number means almost nothing on its own. Ask what conditions it was measured under and how it falls apart at the edges. In the physical world, the boring, repetitive tasks are the hardest to automate reliably, and that 1% is precisely where your customers notice.

Content I Enjoyed

The Class of 2026 keeps booing AI off the stage

Eric Schmidt ran Google for a decade, so he knows how to read a room. At the University of Arizona commencement, he watched one turn on him in real time. The boos started the moment he compared AI to the arrival of the personal computer, and to his credit, he didn’t pretend to ignore them and plough on. “I know what many of you are feeling about that. I can hear you,” he told the crowd, before talking about the fear that the future has already been written.

He was the third speaker to be drowned out this graduation season for the same stance. Real estate executive Gloria Caulfield called AI “the next industrial revolution” at UCF and got jeered mid-sentence. Big Machine Records CEO Scott Borchetta told Middle Tennessee State that “AI is rewriting production as we sit here,” then shot back at the noise: “It’s a tool. Make it work for you.”

All three reach for a framing around AI as inevitable progress. This is something that used to be reliable applause. But to a room of people about to job-hunt in an AI-shaped market, it now lands as a threat wearing the veil of a pep talk. Pew research found that half of Americans are more concerned than excited about AI, up from 37% in 2021, with just 10% mainly excited. Among AI experts, that figure is 47%.

We’re not quite at the Luddite riot stage yet, but the backlash is now clearly visible, as the younger workforce feels deeply passionate about its evolving capabilities. Let’s see how the following five years play out.

Idea I Learned

10% growth and 10% unemployment at the same time

Dario Amodei told the Wall Street Journal at Davos that AI could push GDP growth to 10% while unemployment also climbs to 10%. He called it a combination “we’ve almost never seen before,” and he’s right. Growth has always brought more jobs. The Great Recession peaked near 10% unemployment, with the economy shrinking, not booming.

His reasoning is that software is heading toward being essentially free. Some of his own engineering leads have told him they no longer write code—they edit what the model produces. Claude Cowork, Anthropic’s latest product, was built in a week and a half, almost entirely by Opus. Output climbs while the work that produced it thins out.

That raises an obvious question. Consumer spending is roughly two-thirds of the US economy. If the paycheques disappear, who buys all the output?

An answer is in what economists call the K-shaped economy: one line climbing, the other flattening. Moody’s puts the top 10% of earners at close to half of all consumer spending, up from around 43% in 2020. A thin slice of high earners and corporate capex carries demand while everyone else stalls.

That 50% figure is contested. The Bureau of Economic Analysis, which actually measures spending, puts the share well below half, and economists who recompute it from official data land closer to a third. However, the trend is very much real, even if the exact split isn’t yet settled.

Where I’d push back on Amodei is on software being “essentially free to build.” Generation costs are collapsing, but shipping software at enterprise scale still carries licence fees, review cycles, security sign-off, compliance, and someone who understands the system well enough to design it. That work hasn’t gone anywhere. I think the displacement will arrive slower and messier than the headline suggests.

The concentration of demand, though, is already here. And the gap is only widening.

Quote to Share

Andrej Karpathy on joining Anthropic:

There are maybe a dozen people whose next move tells you something real about where AI is heading, and Karpathy sits near the top of that list.

Founding member of OpenAI, the man who built Tesla’s self-driving program, and for years the field’s most patient public teacher, walking hundreds of thousands of people through how a neural net actually thinks.

He is joining as a hands-on researcher, building a team that turns Claude on its own training and uses the model to sharpen the next one. The teacher who showed the world how to understand the machine is now helping the machine understand itself.

I was in the room for the everyday version of that same shift this week at Code with Claude in London, Anthropic’s first conference outside the US. The people demoing the most impressive work were barely writing code at all. They described what they wanted in plain English and shaped what the agent gave back. Anthropic runs on that logic internally too, with most of its own code now written by Claude while its engineers spend their hours on architecture and judgment.

The lazy conclusion is that expertise is finished, that once anyone can build by talking, knowing what sits underneath stops counting for much. I think it runs the other way. As the floor rises and the building gets handed to agents, deep understanding goes from common to scarce, and scarce is where the value pools. Someone still has to know when the machine is right, when it is confidently wrong, and when to pull it back. Anthropic now has on board the person who arguably understands that best.

Source: Andrej Karpathy on X

Question to Ponder

Whether the companies can be trusted to make that call comes down to the incentives they operate under.

The order pulled this week required companies to voluntarily share frontier models with federal agencies up to 90 days before release. This would enable the government to test them for dangerous capabilities and security flaws, with no licensing requirement and no mandatory approval. It would formalise the kind of voluntary testing that frontier labs already do through the National Institute of Standards and Technology (NIST), the federal standards agency that runs the government’s AI safety evaluations.

The administration shelved it after executives from leading AI companies and former White House AI adviser David Sacks spoke with Trump in the hours before the planned signing, on the argument that any slowdown risks the US lead over China.

Demis Hassabis offered a useful lens this week, even though he wasn't talking about regulation. At Google I/O, he argued that in research, direction matters more than velocity and that moving quickly the wrong way is worse than pausing to get it right. He was not commenting on the executive order or calling for regulation, yet his framing unsettles the assumption that any delay is automatically a loss, given that a short pre-release review is closer to a direction check than a brake on progress.

The case for letting go of the order is credible, too. The government may lack the expertise to assess frontier systems well, the most rigorous testing already happens inside the labs, and a framework assembled under deadline pressure can harden into something heavier and less workable than what it replaced.

What the week clarified is that decisions of this weight are being made by a small number of people, with the rest of us downstream of them. The most useful response is to understand these systems well enough to judge the trade-off for yourself, because the terms are being set either way.

Already a subscriber? Get your whole team on board. Signal Pro group subscriptions give everyone access to weekly AI workflows and tutorials, practical upskilling that pays for itself. It’s the kind of thing L&D budgets were made for. Share this with your manager today.

Get a group subscription

💡 If you enjoyed this issue, share it with a friend.

Refer a friend

See you next week,
Alex Banks

P.S. Musk loses OpenAI court battle.

If you’re using AI every day and feel like you’re getting less out of it than you should be, you’re not imagining it.

Most people I chat with are running their entire workflow through a single chatbot. That’s fine for asking questions, but it’s not a workflow.

The real power comes from combining the right tools and knowing when to reach for each one.

My own stack keeps shifting. New models, new launches, new updates every single week. But six tools have stuck. They’re the ones I open every day, the ones I can’t work without.

Here’s what you’re getting:

1. The 6 AI tools running my week

2. ⁠The exact prompts, use cases, and setups I use for each

3. ⁠Plus six months free on the one I use most

Let’s get into it.

1. Claude Opus 4.7 + Adaptive thinking

What I use it for: research, writing, email drafting, knowledge work, anything that needs a brain.

I’ve been a Claude user consistently now for over two years. It began with Sonnet 3. I used it for some, but not all, knowledge work tasks. Over that two-year period, Claude has evolved into something different entirely and has become my go-to LLM whenever I have a question or task to complete.

The honest reason for this is feel. ChatGPT, Gemini, and Grok still feel a little sterile to me. The outputs have a slight lifeless tilt where you can tell a machine wrote them. Claude has something the others don’t, which I’d describe as a real soul. Sure, it might have “it’s not about X, it’s about Y” across every few sentences, but you can see the thoughtfulness in how Anthropic trains these models, showing up in the output.

The current frontier model is Claude Opus 4.7. I default to it for almost everything, and I always have adaptive thinking toggled on.

Adaptive thinking is Anthropic’s reasoning mode. Instead of you setting how hard the model should think, Claude decides on the fly based on the complexity of your request. Therefore, you’d expect simple questions to get a fast response when harder problems trigger deeper reasoning.

I’ll be honest, even with adaptive thinking on, Claude sometimes skips the reasoning step entirely on questions it thinks are simple. This creates a sub-par output. The difference comes down to the specificity of your prompts.

These are the five I use most (where necessary, I append a version of this at the end of my prompts):

Engage deep think mode. This is the time where you summon all 
your power and execute flawlessly. Go.

Think carefully and step-by-step before responding. This problem 
is harder than it looks.

Before you answer, walk me through your reasoning. Identify any 
assumptions you're making, then revise the answer if any of them 
look weak.

Suggest three different frameworks I could use to think about 
this. Explain what each prioritises and what each might miss. 
Then pick the strongest one and answer through that lens.

Take your time on this. I'd rather have a slower, well-reasoned 
answer than a quick one.

Where Claude pulls ahead of the rest of my stack: when it’s connected to my Google Workspace. I can ask it to read an email thread in Gmail, identify what someone is asking, and create a draft reply in my voice in the exact thread I’m referencing.

My favourite setup with Claude is via a single core chat using the Claude Desktop app. As you can see in the image above, I have a lot of local MCP tools connected that enhance Claude’s capabilities. The “x-twitter” tool, for example, is able to pull X posts and analysis from post replies if I’m doing some deep research.

These MCP tools can’t be used in the web app, which is why I often default to Claude desktop. If I’m working through a really tough problem, I’ll have a few additional Claude chats open in the web app. At any one time, I’m running 1 (if it’s a focused task) to 5 chat windows, exploring different angles/adjacent tasks to a core goal.

A question to ask yourself: Most people prompt Claude like they're searching Google. What would change if you prompted it like you were briefing a senior colleague?

2. Gemini Live

What I use it for: real-world questions where the answer is something I’m looking at, not something I’m typing about.

Gemini Live is the tool I underestimated the most and slept on for the longest. It’s a voice-and-video mode where you point your camera at something and have a conversation about whatever you see.

Start by downloading the Gemini app. You can enter live mode in just a few taps.

My favourite use cases have been the unsexy ones. For example, I recently moved into a new apartment and couldn’t make sense of half the symbols on our oven's front panel. Instead of digging out the manual, I opened Gemini Live, pointed the camera at the display, and asked what each function did. Gemini identified the model immediately, walked me through the symbols one by one, and recommended the right setting for what I was cooking. The whole thing took about 3 minutes.

Another great example was last month, when my washing machine gave me an error code. I pointed Gemini Live at the display and asked what to do. It told me how to resolve it while I was holding the phone, showing different angles and parts of the washing machine.

Outside of home admin, I see huge applications for this to solve tasks like homework problems or even to ask about prescription labels. The opportunities are endless for this multimodal interface coupled with instant feedback.

I think we’ll see this exact pattern show up natively in Siri and on hardware devices over the next couple of years.

A prompt to try: open Gemini Live, point the camera at something you’ve been putting off (a settings menu you don’t understand, a recipe you’ve never made, a bike repair you’ve been avoiding) and just say “Walk me through this.”

A question to ask yourself: What problems in your day are you solving with text when the answer is really sitting in front of your camera?

3. ChatGPT Images

What I use it for: thumbnails, outputs that include images of real people, branded report visuals, and infographics.

In my experience, ChatGPT Images is the best image model on the market right now. I wrote a full deep dive on it last month. Two things set it apart from Nano Banana 2.

First, it handles real people. Most image models have guardrails that block named public figures, even if you provide a reference image. ChatGPT Images will generate a recognisable, photorealistic version of someone from just their name. Google’s Gemini Nanao Banana will refuse.

Second, it can turn any report into a designed infographic. I provided the same report to both models, and these were the outputs.

An immediate tell of Nano Banana is the style of the output. Its default has a cartoon feel, whereas ChatGPT Images knows how to mirror the specified branding style.

Something else I do with ChatGPT images is I give the output back to Claude Opus 4.7 with the original report and ask it to flag any mismatches or hallucinations.

Claude is the LLM “judge” that marks ChatGPT Image’s work. I then keep iterating until it’s correct. I find this typically takes 2 to 3 passes.

The workflow loop: ChatGPT Images generates → Claude reviews → ChatGPT Images iterates. This has been the real unlock here, as you get to use each model for what it’s best at.

Two things I’ve learned:

• Less is more in the prompt. Describe the outcome, hand it the reference material, and let it make the creative call.

• It’s brilliant at iteration once you’ve got something on the page. “Change this word, swap this colour, move the logo.” That’s where it shines, especially when combined with using the select tool.

A prompt to try:

Generate a 4:5 infographic from the attached report in the 
branded style and typography of [company name]. Provide 8 
unique variations.

Then paste the output into Claude with the original document, and see what you get back.

A question to ask yourself: Where are you paying for design quality you could now generate in five minutes of conversation? And where will a human designer still beat AI for the work that actually matters?

The next three tools are the ones I’m joined at the hip with. Together, they’ve quietly replaced much of what I used to do manually: typing, taking notes, transcribing.

I've worked directly with one of them to get every Signal Pro member 6 months free. Details below.

Read more

Hey friends 👋 Happy Sunday.

Here’s your weekly dose of AI and insight.

Every Wednesday, Signal Pro members get a step-by-step AI workflow they can apply immediately. No fluff, just practical guides to upskill you and your team. If you’re only reading the Sunday issue, you’re getting half the picture. Upgrade to Pro today.

Upgrade to Pro

AI Highlights

My top-3 picks of AI news this week.

Anthropic

1. Anthropic Pulls Away

Anthropic disclosed Q1 2026 revenue grew 80x year-on-year, with annualised run rate now estimated at over $44B, committed $200M to a four-year Gates Foundation partnership across global health and education, and overtook OpenAI in verified business customers for the first time.

• Revenue trajectory: Dario Amodei told Anthropic’s Code with Claude developer conference the company planned for 10x growth this year but saw 80x annualised, with customers spending $1M+ annually climbing from a dozen two years ago to over 1,000 today (doubled in under two months), against funding talks of $30B reportedly approaching a ~$900B valuation.

• Philanthropic deployment: A four-year Gates Foundation partnership directs $200M in grants, Claude credits, and technical support toward polio, HPV, and preeclampsia vaccine candidates plus African language accessibility, spanning global health, life sciences, education, agriculture, and economic mobility.

• Enterprise crossover: Ramp’s May AI Index put Anthropic at 34.4% of business customers versus OpenAI’s 32.3% (up from 9% in May 2025), and PwC formally expanded its alliance to certify 30,000 staff on Claude with a global rollout targeting its 364,000-person workforce.

Alex’s take: Anthropic also published a paper this week setting out four fronts of AI competition: intelligence, domestic adoption, global distribution, and resilience. The Gates Foundation deal puts American AI infrastructure into emerging markets, ahead of the cheap Huawei/Alibaba stack the paper itself warns about. Markets are now pricing Anthropic as strategic infrastructure, especially ahead of what will likely be a blockbuster IPO. Back in February, Anthropic was winning the coding war organically while OpenAI bought Windsurf. That playbook has now expanded from coding to enterprise default. The lead is widening on every front all at once.

OpenAI

2. OpenAI Strikes Back

OpenAI interestingly followed suit on two of Anthropic’s biggest enterprise plays from earlier this month, moving into deployment services and cybersecurity inside a single week. At the same time, the company kept consumer momentum going inside ChatGPT with new mobile and personal finance capabilities.

• Deployment Company: The new venture launched with $4 billion from 19 investment, consulting, and integration firms, led by TPG with Bain Capital, McKinsey, Capgemini, and Goldman Sachs among the backers. OpenAI also acquired UK consulting firm Tomoro for its 150 Forward Deployed Engineers, mirroring Anthropic’s $1.5 billion venture with Blackstone, Hellman & Friedman, and Goldman Sachs.

• Daybreak cybersecurity: A new cyber-defensive initiative built on GPT-5.5-Cyber and Codex Security, with eight major security vendors, including Cloudflare, CrowdStrike, and Palo Alto Networks, integrating under OpenAI’s “Trusted Access for Cyber” programme. The direct counterpart to Anthropic’s Project Glasswing and Claude Mythos.

• ChatGPT product velocity: Codex now runs inside the ChatGPT mobile app, letting users kick off coding agents from their phone and approve next steps while the work continues on a connected laptop (akin to Claude Dispatch). OpenAI also rolled out a personal finance experience inside ChatGPT Pro, connecting US users’ financial accounts from over 12,000 institutions via Plaid so ChatGPT can answer questions grounded in real spending, balances, and investments.

Alex’s take: Bain, McKinsey, and Capgemini are now investors in OpenAI's deployment company, the venture designed to compress their own AI consulting practices. Every major consulting firm has formally agreed that frontier AI deployment is the most defensible product they have left. Forward-deployed engineering has historically been a relationship-bound skill that doesn't scale like normal headcount, and OpenAI now needs to grow that capability, coming up head to head agaisnt Anthropic’s competing initiative (see the Idea I Learned for more).

Google

3. Google’s Gemini Rising

Most companies go quiet in the lead-up to a major keynote. Google did the opposite. The days leading up to Google I/O 2026 brought a flurry of updates from across the Google ecosystem.

• Gemini Intelligence: An agentic AI layer running underneath Android itself, handling multistep tasks across apps without switching. Demos: photograph an event flyer and it finds the event on Expedia; turn a grocery list in your notes app into a cart in your preferred shopping app. Rolling out to the latest Pixel and Samsung Galaxy phones this summer, with the new Googlebook laptops and Aluminium OS both launching it on desktop later this year.

• Magic Pointer: Google DeepMind’s reimagining of the 50-year-old mouse pointer. Users direct Gemini through motion, speech, and natural shorthand. Hover, wiggle, or speak to surface contextual suggestions tied to whatever’s on screen. Currently experimental and demoed on Googlebook.

• Isomorphic Labs: Demis Hassabis’s DeepMind spinout closed a $2.1B Series B led by Thrive Capital, with new backers Abu Dhabi’s MGX, Singapore’s Temasek, and the UK Sovereign AI Fund. BioSpace called it the second-largest biotech round in history, behind Altos Labs’ $3B in 2022. Pharma partnerships with Lilly, Novartis, and Johnson & Johnson back the platform with billions in milestone-tied money.

Alex’s take: Apple has owned phone-laptop continuity for over a decade. Google is finally closing the gap by inverting the stack with Gemini Intelligence becoming the primary interface: the Googlebook is the peripheral, and Magic Pointer turns the cursor into an interpreter of intent. Over the next decade we’re going to see a very real shift in how we direct computers. I’m especially excited about the emergence of Isomorphic Labs and their mission to “solve all disease”. I wonder how long it will take for a compound designed by them to reach a human body. That will be the truest test of them all.

Content I Enjoyed

Figure’s humanoids worked an 8-hour shift

Figure AI’s livestream of their F.03 humanoid robots running a fully autonomous package-sorting operation hit over 13 million impressions on X this week. The robots run on Helix-02, Figure’s vision-language-action model, processing everything onboard with no teleoperation. What started as an eight-hour demonstration has now crossed 100,000 packages, still running, with the robots going until failure.

It was fascinating to watch the flurry of initial pushback online, with people claiming that the robots were being remotely controlled from elsewhere via teleoperation. Over 90 hours in, and that same pushback has gone awfully quiet. F.03 is picking packages at roughly the three-second cadence a human hits, with multi-robot coordination all handled visually. If one robot was low on battery or detected a fault, it walked itself to maintenance and called a replacement from the fleet.

For this specific task, I’d argue that purpose-built parcel sorters move faster, cost less, and don’t need battery breaks. But the wider bet here surrounds generalisation. The packages on the belt are cardboard rectangles or soft packets of roughly similar size, while real warehouse floors handle bags, buckets, and items weighing one to fifty pounds. When a humanoid can do them all, like a human worker, that’s when the human form for robotics really takes off.

Idea I Learned

Every AI lab is becoming Palantir

On 4 May, Anthropic announced a $1.5 billion venture with Blackstone, Hellman & Friedman, and Goldman Sachs to embed engineers inside mid-market companies. On 11 May, OpenAI launched the OpenAI Deployment Company with $4 billion from 19 backers including TPG, Bain, McKinsey, and Capgemini, and acquired London consultancy Tomoro to bring 150 Forward Deployed Engineers in on day one. On 12 May, The Information reported Google is hiring hundreds of FDEs of its own.

It seems as though every AI lab has looked itself in the mirror and decided it wants to be Palantir. Palantir invented the FDE role. Engineers fly to the customer, sit with operators, learn the workflow, ship code that wraps a model around the actual problem, and stay until production works. Until 2016, Palantir had more FDEs than software engineers.

FDEs are about to become one of the most in-demand jobs in tech, and the reasoning behind this is structural change. Firstly, models commoditise very quickly. Getting one into production without breaking everything is the hard part, and agents make it harder still, given non-deterministic outputs, messy data, evaluation criteria that need constant tuning, and workflows that have to be redesigned around the model.

Secondly, for every dollar companies spend on software, they spend six on services. That ratio built Accenture, Deloitte, and IBM Global Services into a multitrillion-dollar industry. The labs are now positioning to take that revenue themselves, with implementation handled by their own engineers rather than slide-producing consultants. Vendors who used to chase 15% of a department’s budget at high margins now see a path to 90% of it, even at lower ones.

The deployment layer is the new moat. The engineers who can sit inside a Fortune 500 customer’s office and ship working agents are about to become the most valuable hires of the decade.

Quote to Share

SpaceX’s new place in the AI stack:

Google is in advanced talks with SpaceX to launch Project Suncatcher, its plan to put TPUs into low Earth orbit by early 2027.

The Suncatcher news came a week after Anthropic agreed to take the full 300 megawatts of Colossus 1 in Memphis, and a month after SpaceX secured an option to buy Cursor for $60 billion. In 2024 you would have struggled to put any two of these companies on the same partnership chart, let alone all four. They now share a common supplier, and that supplier is a launch business called SpaceX. Interestingly, Google owns 6.1% of SpaceX and has a Google executive, Don Harrison, on its board.

The reason SpaceX keeps showing up in these announcements is that it sits atop the constraint everyone else is fighting. AI compute is outgrowing the grid faster than substations can be built, and the cheapest path to more power increasingly runs through sun-synchronous orbit, where solar panels operate roughly 8x more efficiently than on the ground. SpaceX has the launch cadence, the satellite network, and the operational track record to make orbital compute feasible this decade, and anyone serious about scaling next-generation models has to rent some of that infrastructure.

Source: @lochan_twt on X

Question to Ponder

Chips, models, and talent were the bottlenecks of 2023 and 2024.

The constraint has now moved from chip limitations and capital limitations to power limitations. Hyperscalers and neoclouds are competing for grid capacity, substation access, and entire power plants, and what’s more, building new generation runs years longer than shipping new GPUs. Transformer lead times have stretched from 24-30 months pre-2020 to as long as 5 years today.

The best way to compare AI infrastructure companies today is by dollar per watt of contracted capacity. Nebius around $15 per watt, CoreWeave around $16, IREN around $3.80, Hut 8 around $1.18. Utilisation, contract quality, and chip vintage sit beneath this, so $/watt is the input against which everything else compounds.

Each shift in the AI race rewards a different kind of company, and it’s clear we’re moving to its lowest layer: energy. Next week’s Google I/O announcements, the OpenAI Deployment Company push, and Anthropic’s moves all depend on this layer of the stack holding up.

Already a subscriber? Get your whole team on board. Signal Pro group subscriptions give everyone access to weekly AI workflows and tutorials, practical upskilling that pays for itself. It’s the kind of thing L&D budgets were made for. Share this with your manager today.

Get a group subscription

💡 If you enjoyed this issue, share it with a friend.

Refer a friend

See you next week,
Alex Banks

P.S. Unitree unvelieved GD01, a manned transformable mecha.

Hey friends 👋

Welcome to another Signal Pro workflow issue.

Last December, I showed you how to build an AI receptionist in ElevenLabs. It worked. People built them, but the setup took a while. You had to click through ElevenLabs’ interface, write the system prompt manually, and set up the voice. It was on you to wire everything up tool by tool.

That whole flow just changed because the ElevenLabs Agents MCP App now plugs straight into Claude Desktop.

You attach a knowledge base, write one prompt, and the agent builds itself. The ElevenLabs interface appears inside Claude.

This is what people mean when they say apps are going headless. The UI you used to log into is now a tool that Claude calls on your behalf.

I built the whole thing in about 8 minutes. By the end, I’ve got a healthcare voice receptionist who can book consultation calls directly into my Google Calendar.

Today I’ll show you:

• How to enable the ElevenLabs Agents MCP App in Claude Desktop (one-time setup)

• How to set tool permissions so Claude can create, update, and configure agents without you babysitting every step

• The exact knowledge base + prompt I used to spin up the agent

• How to upload the knowledge base and connect Cal.com inside ElevenLabs

• Voice, LLM, and first message tweaks that make the agent feel like a real receptionist

• How to test, publish, and share the agent with your team

Why this matters

Voice agents have been technically possible for years. The reason most professionals haven’t built one is that the setup is fiddly. You need to learn a new platform, quickly understand its quirks, glue together half a dozen tools, and then hope nothing breaks.

That friction is now gone because Claude is now the orchestration layer. You just need to know what you want the agent to do, and Claude handles the configuration.

This issue will teach you how to build any agent that runs a conversation and takes an action. Once you’ve done it once, you can rebuild it for:

• Legal client intake

• Mortgage application qualification

• Recruitment screening

• Sales discovery calls

• Property viewing bookings

• Onboarding for any suitable service

The skill is totally reusable and applicable to you and your business.

What’s actually running under the hood

Before we build, it helps to understand what a “voice agent” actually is. It’s three models working hand-in-hand at sub-second latency:

1. STT (speech-to-text): Scribe v2 Realtime. Around 150 ms latency, built specifically for agentic conversations. This is what turns your spoken words into text that the LLM can read.

2. LLM (the brain): Claude Haiku 4.5. Takes the transcribed text, the system prompt, and the knowledge base, and generates the response. I’ve tested every model ElevenLabs offers, and Haiku 4.5 hits the best balance of speed and accuracy for voice-first use cases.

3. TTS (text-to-speech): Eleven Flash v2.5. Ultra-low ~75 ms latency, paired with your cloned voice (or any of the hundreds in ElevenLabs’ library). This is what makes the agent sound like a real person.

The three stages run end-to-end in roughly a quarter of a second (additional latency I’ve found to be dependent on internet speed), which is why the agent feels conversational.

When we get to Step 7, you’ll see where to set each of these in the ElevenLabs interface.

Step 1: Enable the ElevenLabs Agents MCP App

Open Claude Desktop. Top left, click Customize.

You’ll land on the Customize panel. Click through to Connectors.

Click the + icon in the top right of the Connectors panel and select Browse connectors.

Search for ElevenLabs. You’ll see two options: the ElevenLabs Player and the ElevenLabs Agents MCP App. We want the second one.

Click into it and enable the connector. Then click Configure.

You’ll need to paste in an ElevenLabs API key. Head to elevenlabs.io → log in → make sure you’re on the ElevenAgents workspace → click Developers → API Keys → Create Key.

It’ll initially be restricted. Toggle the restriction off, then click Create Key. Copy the key.

Back in Claude Desktop, paste it into the API Key field and click Save.

💡 A note on this MCP

Think of an MCP (Model Context Protocol) connector as the universal USB-C port for AI apps. It lets Claude read from and write to other tools. In this case, the entire ElevenLabs platform. Once enabled, Claude can search voices, create agents, configure system prompts, and edit existing agents on your behalf.

Step 2: Set tool permissions

Scroll down to the Tool permissions section. You’ll see Claude’s available actions split into three groups: Interactive, Read-only, and Write/delete.

Here’s how I set mine:

• Create Agent: Always allow

• get_agent_config: Always allow

• search_voices: Always allow

• create_agent: Always allow

• update_agent: Ask for permission

That last one matters a lot as I want Claude creating new agents without interrupting me, but I want explicit oversight before it modifies an agent I’ve already built. This is good agent hygiene: we set tight permissions on anything that could overwrite existing work.

Step 3: The prompt and knowledge base

Read more