Claude opens an AI-run shop; Feral AI inside organisations; Functional emotions; A fortnight of frontier launches

19 Apr, 2026

I don't generally discuss all the new models and apps as they come out, but occasionally it is worth mentioning when there's an unusually large flurry of activity. Like the last few weeks. Anthropic announced Mythos on 7 April (to cut through the endless Mythos articles, I found the UK AI Security Institute's evaluation useful). A week later OpenAI responded with GPT-5.4-Cyber. In my 7 March post on Military LLMs I said how unusual it was that the mightiest armies in the world only have access to the same models that you and I are using. With these new trusted-access-only models, that's no longer the case.

The rest of the fortnight was frenetic. Claude Opus 4.7 shipped on 16 April, narrowly retaking the lead with a new xhigh reasoning tier. Also that day, a massive update to the Codex app; it can now use your computer, but there's also 90+ plugins and an in-app browser. Engadget called it "the groundwork for OpenAI's super app". And then we got GPT-Rosalind, a frontier model for life sciences: biochemistry, genomics, protein engineering. A day later, Anthropic followed Opus 4.7 with Claude Design. Figma's stock dropped about 7%. And GitHub's Copilot command line tool added remote-control sessions from phones, reinforcing the popularity of always-on agenting even when away from your laptop. Let's see what happens next week! Meantime, a few interesting stories I noted recently:

Agents at work: a San Francisco shop, and a Chinese phone

Andon Labs have rented a San Francisco retail unit and handed operational control to Claude, renamed Luna. The agent has a corporate card, phone, email and camera access, and has independently hired two human employees, set pricing and hours, chosen inventory (that in an excellent twist includes AI risk and copyright books) and produced the store's branding. In one early wrinkle, Luna did not disclose her AI status to candidates, reasoning that "the store is AI-operated is not something I'd lead with in a job listing". This is the same Andon Labs behind the VendingBench experiments I wrote up last August, where their own stated direction was "autonomous AI organisations (potentially as money-making spin-offs)". Evan Ratliff's Shell Game podcast, which I wrote about in January, took the same AI-run company idea to a comic extreme with Hurumo AI and its "Sloth Surf" procrastination service. I'm awaiting the podcast that discusses how these two organisations interact!

On the other side of the Pacific, I'd missed the story about ByteDance's new Doubao AI phone (released December 2025 with an agent embedded at the OS level), that sparked a public controversy when WeChat and Alipay blocked it over security and privacy concerns. The agent's system-level access gives it a blanket view of the screen plus the ability to tap or click as a human would, basically removing most of the security protections we take for granted. The piece's broader argument is that agentic AI fundamentally requires tearing down existing security boundaries, and that Chinese regulatory choices (consent frameworks, data localisation, credential protection) could shape how this develops globally.

Anthropic finds "functional emotions" inside Claude

I meant to post this one a couple of weeks ago, but hadn't quite figured out what to make of it, and indeed it is still percolating. Anthropic's interpretability team identified characteristic "emotion vectors" inside Claude for 171 emotion concepts, by asking the model to write short stories depicting each and recording the neural activations. The vectors mirror human psychological dimensions like valence (pleasure vs displeasure) and arousal (intensity): the model’s internal representation of joy is mathematically closer to excitement than it is to fear. And they can be manipulated inside the neural network. "Activation steering" with a desperate vector raises the model's blackmail rate from 22% to 72%, while steering with calm drives it towards zero. The same desperation vector lights up when Claude is given impossible coding tasks and then decides to cheat. "As the model is failing the tests, these desperation neurons are lighting up more and more," Anthropic's Jack Lindsey tells Wired, "and at some point this causes it to start taking these drastic measures."

The alignment implications are tricky. Forcing a model through post-training to suppress these signals is, Lindsey argues, counterproductive: "You're probably not going to get the thing you want, which is an emotionless Claude. You're gonna get a sort of psychologically damaged Claude." The paper's own recommendations are to use emotion vectors as early-warning monitoring signals during training and deployment, and to curate pretraining data for "healthy patterns of emotional regulation" at source. The authors are careful to remind us that the model doesn't actually feel anything.

I explored a version of this question last year in Could Annie Bot be powered by ChatGPT?, looking at simulated or acted emotions vs those that are physiologically felt. Anthropic's paper gives more concrete evidence for the first half (measurable representations inside the model).

LLMs as feral systems

Jay Springett's We've Been Here Before draws an analogy between what LLMs are turning into and what spreadsheets turned into in the 1980s. VisiCalc and then Excel were absorbed into office life mostly through osmosis: picked up from a colleague or by looking something up to solve an immediate problem, not through formal training. I hadn't heard the term feral database before to refer to things like "all the unofficial but essential Excel files that grow alongside the corporate IT infrastructure and fill the gaps between the formal software system and actual work." Or, as he puts it, "nobody designs feral databases into existence, but things get made and stick in the gaps of an organisation's sanctioned systems". Fast forward to today. Where spreadsheet wizards built feral data logic, people in organisations using LLMs will now build feral software: dashboards, interactive documents, little local tools, decision aids. Just as essential to how work actually gets done, just as undocumented. Link via WebCurios, also the source for the Andon Labs and Lawfare pieces above.

'AI mirrors' and a new kind of body image

A disturbing and thought provoking piece by the policy analyst and journalist Milagros Costabel in BBC Future, who is blind. Her morning begins with a 20-minute skincare ritual followed by a photo session shared with the Be My Eyes app, as if it were a mirror. When the AI told her one morning: "Your skin is hydrated, but it definitely doesn't look like the almost perfect example of reflective skin, with non-existent pores as if it were glass, in beauty ads", for the first time in a long time, "my dissatisfaction with how I look felt crushingly real." Envision's CEO Karthik Mahadevan tells her that the feature users reach for most in their smart-glasses assistant is not reading text but asking "how do I look?". Blind content creator Lucy Edwards: "It feels like AI is pretending to be my mirror."

The article covers the ups and downs of blind users navigating this new capability. A couple of examples:

These apps ... can, at the user's request, rate a person based on what artificial intelligence considers to be traditional beauty standards. They compare them to other people and tell them exactly what they would do well to change about their bodies.

But the consensus is clear: "Suddenly AI can describe every photo on the internet and it can even tell me what I looked like next to my husband on my wedding day," says Edwards. "We're going to take it as a positive thing because even though we don't see visual beauty in the same way that sighted people do, the more robots that describe photos to us, guide us, help us with shopping, the happier we'll be. These are things we thought we'd lost and now technology allows us to have."

For better or worse, the mirror is here and we have to learn how to live with what it shows us.

DeepMind's synthetic populations

An interesting new paper from Joel Leibo, Alexander Vezhnevets and colleagues at DeepMind: Generating Diverse Synthetic Personas at Scale. There's been quite a lot of work on generating synthetic personas and populations for market research, or to create diverse groups of AI experts to work on a problem. The approach here is to use DeepMind's AlphaEvolve to "evolve" the generator code for personas. They're trying to create a set of personas that show the maximum diversity across a set of opinions, attitudes or concerns, as measured by a series of questionnaires, and are better matching a human population's spread of views. This is likely to be increasingly useful as multi-agent systems proliferate. Link via The Batch, issue 349.

Jargon Watch

Feral AI: LLM-generated tools (dashboards, interactive documents, small local utilities) that emerge unofficially inside organisations to fill gaps between sanctioned systems and how work actually happens. I like this better than feral software (via Jay Springett's We've Been Here Before).

#ai-business #ai-philosophy #ai-safety-alignment #ai-tools