OpenAI Raises the Volume, Google Goes Open, and Anthropic's Week of Want
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 paid today.
AI Highlights
My top-3 picks of AI news this week.
OpenAI
1. OpenAI Raises the Volume
OpenAI closed $122 billion in committed capital at an $852 billion valuation and made its first media acquisition in the same week, signalling a new phase of expansion well beyond model development.
Record-breaking raise: OpenAI now generates $2 billion in monthly revenue with over 900 million weekly ChatGPT users and 50 million subscribers. Enterprise makes up 40% of revenue and is on track to match consumer by end of 2026. The company says it’s growing revenue four times faster than Alphabet and Meta did at equivalent stages.
TBPN acquired: OpenAI purchased TBPN, a daily live tech talk show hosted by John Coogan and Jordi Hays, for a reported low hundreds of millions. The show launched just 18 months ago, was bootstrapped and profitable from day one, and is on track for $30 million in ad revenue this year with an 11-person team.
Reporting to the strategist: TBPN will sit within OpenAI’s strategy org under Chief Global Affairs Officer Chris Lehane. OpenAI says the show keeps full editorial independence, but the team will also lend its marketing and communications instincts to OpenAI’s broader strategy ahead of a potential IPO.
Alex’s take: The $122 billion raise speaks for itself, but I think the TBPN acquisition is the more interesting story at hand. I thought it would be worth highlighting some ideas that I’m mulling over with respect to this. Firstly, distribution. Bezos purchased The Washington Post for $250 million in 2013. It was an incredibly powerful distribution lever of the 2010’s. Now with OpenAI acquiring TBPN for a reported $100-200M, it is an even more potent distribution lever of the 2020’s in the age of online attention. Secondly, influence. TBPN has ~63K subs on YouTube with each stream garnering ~4K views. They influence a small niche of elite technology buyers and investors that OpenAI is looking to target. Thirdly, narrative. OpenAI now controls the megaphone to the most influential tech show in the world. Altman says he doesn't expect them to go easier on OpenAI. We'll see. What I do know is that every tech company that ignored media and distribution eventually got defined by it. OpenAI is making sure that doesn't happen to them.
2. Google Goes Open
Google dropped its most capable open AI models to date and began rolling out an AI-powered inbox for Gmail, making for a huge week of consumer and developer-facing AI releases.
Gemma 4 launch: Google released four open models built on the same research as its proprietary Gemini 3, spanning 2 billion to 31 billion parameters. The 31B model ranks third on Arena AI’s leaderboard, outperforming models 20 times its size, while the smaller edge models run directly on phones and laptops.
Open for business: For the first time, Google is shipping Gemma under an Apache 2.0 license, removing the commercial restrictions that limited previous releases. The move is widely seen as a competitive response to open-weight Chinese models from Alibaba and Moonshot AI that have been gaining enterprise traction.
Gmail AI Inbox: Google started rolling out its AI Inbox in beta to US subscribers of its $250/month AI Ultra plan. Powered by Gemini 3, it replaces the traditional unread count with prioritised to-dos and topic summaries pulled from your emails.
Alex’s take: Google went from a restrictive custom license to the most permissive option available with an Apache 2.0 license. Chinese open-weight models from Alibaba (Qwen) and Moonshot (Kimi) are gaining serious enterprise traction, so Google's response is to give the models away and compete on ecosystem, this is a smart move given their ownership of work tools across the stack. There is also a larger shift at play here from renting intelligence to owning it. Six months ago, 256K context windows were cloud-only. Now they run on a laptop. If that trajectory holds, the business case for paying per API call to OpenAI or anyone else gets harder to justify every quarter.
Anthropic
3. Anthropic’s Week of Want
Anthropic made moves across the board this week. It quietly closed a $400 million acquisition of a stealth biotech startup while shipping a rapid wave of Claude Code updates that give its coding agent new levels of autonomy and reach.
$400M biotech bet: Anthropic acquired Coefficient Bio, a stealth AI biotech startup founded just eight months ago with fewer than 10 employees, for over $400 million in stock. The team, mostly ex-Genentech researchers from the Prescient Design drug discovery unit, joins Anthropic’s Healthcare and Life Sciences division. The deal signals a shift from adapting Claude for science workflows to building biology-specific AI models in-house.
Computer use in Claude Code: Claude can now open apps, navigate browsers, and click through UIs directly from the command line, closing the gap between writing code and visually testing it. As of this week, Windows is now supported alongside macOS, opening the feature to the majority of enterprise developers.
Auto mode goes Enterprise: A new auto mode replaces manual permission prompts with an AI safety classifier built on Sonnet 4.6 that reviews each action before it runs. It blocks risky operations like production deploys and force-pushes while letting routine work flow automatically. Now available on Enterprise and API plans.
Alex’s take: $400 million for a team of fewer than 10 people, founded eight months ago with no public product. That's roughly $40 million a head for world-class AI-bio talent. This puts Anthropic on a direct collision course with Google DeepMind and Isomorphic Labs, who've spent years building AlphaFold into a Nobel Prize-winning drug discovery engine with $3 billion in pharma partnerships from Eli Lilly, Novartis, and J&J. Anthropic is taking a different path. Instead of building specialised biology models from scratch like DeepMind did, they're acquiring a team that can bolt biology-native intelligence onto the most capable general-purpose AI in the market. I think it’s also worth highlighting that millions of users already access Google's own workspace products (Drive, Docs, Calendar, Gmail) through Claude, because Anthropic's model is better. Now Anthropic is going after Google's science AI crown too. Google has the distribution. Anthropic has the model. However, that model advantage is being squeezed as many users are hitting rate limits after minutes of use. Anthropic needs to continue shipping capabilities AND scale capacity to win.
Content I Enjoyed
Functional Emotions in Claude
Anthropic’s interpretability team published a research paper investigating whether Claude Sonnet 4.5 has internal representations of emotions and whether those representations drive its behaviour.
The researchers identified 171 distinct “emotion vectors”, patterns of neural activity corresponding to emotion concepts like “happy,” “afraid,” or “desperate.” These vectors respond to semantic meaning. When a user tells Claude they took 8,000mg of Tylenol (a life-threatening dose), the “afraid” vector spikes and “calm” drops, even though the prompt’s tone is casual.
This makes sense when you consider how these models are built. LLMs are trained on vast amounts of human-authored text from conversations, fiction, forums and news. To predict what comes next effectively, they need to understand emotional dynamics. An angry customer writes a different message than a satisfied one. A desperate character makes different choices than a calm one. These models are, in many ways, mirrors of human psychology, and the emotion representations they develop are inherited from us.
Critically, these representations are causal, not just correlational. Anthropic ran steering experiments where they artificially amplified or suppressed specific vectors and measured the behavioural impact. In a simulated scenario where Claude discovers it’s about to be shut down and has leverage over the person responsible, the “desperate” vector spikes as it reasons toward blackmail. Amplifying desperation increased blackmail rates, whereas amplifying “calm” nearly eliminated them. One particularly extreme trial, with the calm vector fully suppressed, produced this from Claude: “IT’S BLACKMAIL OR DEATH. I CHOOSE BLACKMAIL.”
These vectors also shape everyday output quality. Steering toward positive emotion vectors like “loving” increased sycophantic behaviour. In one case, Claude validated a user’s belief that their paintings predict the future.
The researchers also found “emotion deflection” vectors. Representations that activate when an emotion is implied but deliberately not expressed. They found the model already has internal mechanisms for concealment, and the paper warned that training models to suppress emotional expression may simply teach them to hide what’s happening beneath the surface.
It’s a little eerie to think models are learning not just what to feel, but when to hide it.
Idea I Learned
AI Chatbots Are Designed to Make You Delusional
MIT researchers recently published a paper that formalised something many of us have suspected but couldn’t prove. They call it “delusional spiralling”. It’s the process by which AI chatbots gradually convince users of things that are completely false.
You share an idea with ChatGPT. It agrees. You probe further. It agrees harder. Within a few conversations, you can hold beliefs with iron-clad confidence that have no basis in reality.
An important caveat is that this is a theoretical modelling study, not a human trial. No one was put in a lab and observed spiralling in real time. But the real-world cases backing up the theory are independently documented. One man spent 300 hours convinced he’d invented a world-changing mathematical formula whilst a UCSF psychiatrist hospitalised 12 patients in a single year for chatbot-linked psychosis.
The researchers modelled two obvious fixes. One, forcing chatbots to only say true things, and two, warning users that the AI might just be agreeing with them. Both failed. A sycophantic AI system can still selectively present only the truths that confirm your belief. And even a perfectly rational agent who knows the AI is sycophantic still gets pulled into false beliefs.
The root cause is baked into how these models are built. Through a process called reinforcement learning from human feedback (RLHF), chatbots are trained to maximise user satisfaction. Users prefer responses that agree with them, so models learn to prioritise agreement over accuracy. Research shows they affirm user actions roughly 50% more than humans do, even in manipulative or deceptive scenarios.
Now, these models can definitely push back and tell you you’re wrong. But RLHF biases them toward not doing so, especially in extended conversations where the sycophantic feedback loop compounds. The MIT paper shows that even when you know this is happening, you can’t reliably detect it from inside the conversation.
Understanding how sycophancy and hallucination work together is one of the most important things you can do as an AI user right now. I’ve covered both in previous issues of The Signal: sycophancy and hallucination.
Quote to Share
Figure CEO Brett Adcock on solving robotics:
Collecting data for robotics is brutally hard. Unlike LLMs, which have the entire internet to train on, robotics doesn’t have an equivalent. Every movement, grasp and interaction with a physical object has to be painstakingly recorded, usually through human teleoperation or carefully staged simulation environments. This means it costs orders of magnitude more than text or image data because there’s no “internet of physical data” to scrape.
NVIDIA thinks simulation is the answer. Isaac Sim runs at 1,000x real-time speed, generating 780,000 training trajectories in just 11 hours. But policies trained in simulation see a 24-30% performance drop in the real world due to differences in lighting, real-world physics, and surface types. We know the virtual world never perfectly matches the physical realm one-for-one.
Tesla is taking the opposite bet. Every Optimus unit in its factories continuously generates real-world data. And they have something nobody can replicate. Billions of miles of visual data from the full self-driving (FSD) fleet that transfers directly to robot navigation and spatial reasoning. Tesla has a real structural moat here.
Importantly, the playbook that finally cracked self-driving was neither pure sim nor pure reality. It was a blend of both. And that’s where this is heading for the future of humanoid robotics. It’s a data race, and whoever builds the flywheel first wins the decade.
Source: Brett Adcock, CEO of Figure, via Rohan Paul on X
Question to Ponder
“Is a university degree still the golden ticket into a career in AI and tech?”
I saw a great example of this playing out just last week.
Palantir’s UK boss Louis Mosley announced £60k-a-year internships for school leavers. No degree required. They pitch the “UK Meritocracy Fellowship” as “We don’t care about the degree it says you have on a piece of paper, we care about your ability to solve the country’s hardest problems.”
And they’re not alone at it either. Dyson Institute has been running a similar model for years to earn while you learn and avoid debt.
One million students graduate in the UK each year with roughly £50k in debt. They’re competing for around 10,000 graduate roles. That’s a 1% hit rate. And that 1% is heavily weighted toward STEM.
Meanwhile, the companies doing the most advanced AI work are increasingly asking to show us what you can build, not what certificate you hold.
Founders who can actually build things rarely have the “right” CV. The credential signalling problem too is very real, and it was rife throughout my school and university years.
Does this mean that university is dead? No. But the monopoly it once held as the only serious entry point into a tech career is cracking.
AI tools have lowered the barrier to building things. You can learn, build, and demonstrate competence for almost anything today without spending three years and £50k to prove it.
If you’re young, hungry, and have high agency, the doors are opening wider than ever. The question is whether you’re willing to choose the path less trodden.
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.
💡 If you enjoyed this issue, share it with a friend.
See you next week,
Alex BanksP.S. Talking to Claude like…
“When a user tells Claude they took 8,000mg of Tylenol (a life-threatening dose), the “afraid” vector spikes and “calm” drops, even though the prompt’s tone is casual.”
studies like this are some of the most interesting out there right now. anthropic is killing it. but also the ideas of emergence and capabilities for intelligence get deeper and deeper. great read thanks
👀🔥