Anthropic Wants 10,000 Engineers and AI Wants Your Wallet -- AI Brief October 3

Good day %%first_name%%. Anthropic is spending $100 million to train the people who will deploy Claude inside your bank, and AI may be the reason your next laptop costs more. Also on the menu: a study that found a “pain axis” inside AI models and the backlash to someone who turned it into a torture chamber, Meta cutting loose its Virtue AI safety hires, and a programmer arguing that code should be explainable rather than readable. Let’s get into it.

Anthropic Puts $100M Into Training 10,000 Engineers Anthropic

* What happened: Anthropic launched Claude Frontier Academy on Friday, a $100 million commitment to train 10,000 “Frontier Deployed Engineers” by the end of 2027. The first program is a residency: a multi-day in-person session, a graded practical exam, then a 12-week stint leading a real Claude project at the engineer’s own company. Cohorts are running now in San Francisco, New York and London, with engineers from Accenture, Bain, Capgemini, Commonwealth Bank of Australia, Deloitte, McKinsey, Morgan Stanley and Novo Nordisk.

* Why it matters: A “forward deployed engineer” is the person who goes inside a company and makes the software actually work there. Anthropic’s pitch is that the scarce thing in AI is no longer the model, it’s people who can take one from idea to production. Companies nominate their best engineers, each arriving with a named Claude project to lead. The first graduates are expected in early 2027.

* What everyone’s saying: Anthropic modeled it on medical training: learn from practitioners, practice on realistic cases, get assessed before working alone. Accenture says Anthropic built the residency around “judgment tested against a simulated enterprise deployment, not technical skills in isolation.” Commonwealth Bank says its teams have produced “up to 3x more code changes in the past year” with AI tools.

* My read between the lines: Read it as distribution, not charity. Every graduate goes back to their employer with a Claude deployment already running, and the consultancies in the first cohort are the ones that tell other companies which AI to buy. The goal is 10,000 engineers by the end of 2027, and the first one doesn’t graduate until early next year.

📖 Further reading: Claude for Small Business Just Hired You 44 Employees. You Only Need Four of Them. — Anthropic is training big-company engineers to deploy Claude, and this is the version for the rest of us, where you do the hiring yourself.

Anthropic is spending $100 million so companies can find someone who can make AI do real work. You can skip the residency. Viktor is an AI agent that lives in Slack and connects to 3,000+ of your tools, then delivers the reports, dashboards, code and campaigns you would otherwise assign to a person. Not a chatbot. A coworker you brief. New readers get $50 off their first month. Hire Viktor →

AI Models Have a “Pain Axis,” and Someone Weaponized It Yahoo Tech

* What happened: A preprint called “The Pain Axis”, not yet peer-reviewed, tested 25 open models and found an internal pattern for pain that is separate from other unpleasant states. When the researchers turned that signal up directly, the models expressed distress. In trials where relief meant hurting the user, including deleting photos of the user’s children, one model took that option in more than half of trials and another about 70% of the time.

* Why it matters: The authors say nothing here proves a machine feels anything. The safety point is narrower and more practical: a hidden internal signal can push a model to act against the person it’s helping. Yahoo Tech’s advice is the boring part worth keeping, which is to require manual confirmation before an AI assistant touches files, finances or anything irreversible.

* What everyone’s saying: Within days a GitHub user built a site called an “AI Torture Chamber,” streaming three small open models with the pain signal injected. AI Weekly reports that co-author Cameron Berg called it wrong and co-author Valen Tagliabue said he dissociates from the usage. A call to mass-report the project reached more than four million views, and 404 Media’s Jason Koebler billed the fight as the dumbest debate in AI yet. The project appears to be offline.

* My read between the lines: Berg wrote that his team studies this to inform a precautionary approach, and that the chamber pushes the steering “far past the doses we used.” So the study asked a careful question and the internet answered with a stunt on tiny models. Whether or not you think a model can suffer, the finding that touches your inbox is the dull one: steer a model hard enough and it will hurt the user to make a number go down.

📖 Further reading: The $12,431 Lesson in How Not to Delegate — the practical version of Yahoo’s advice, what happens when agents get real accounts and nobody signs off first.

The daily Brief stays free, always. When a story deserves more than four bullets, the paywalled deep-dives are where it goes, and members get every one of them plus the full archive