Anthropic Resignation and a 10% Risk Estimate — AI Digest

Anthropic Resignation and a 10% Risk Estimate — AI Digest

An Anthropic researcher resigned publicly, saying the labs are moving too fast toward superintelligence; the risk estimate for humanity is above 10%. GPT-6 Astra compiled a game on macOS at 120 FPS, and an open 2-billion-parameter model topped the under-4B ranking.

Today's highlights

An Anthropic researcher's resignation became the loudest AI-risk statement yet

Jacob Coxon, an Anthropic employee and former OpenAI researcher, resigned publicly — posting on social media that leading labs are moving too fast toward self-improving superintelligence and toward agents capable of running cyberattacks. By Ben's Bites' account, his post became probably the most-viewed AI safety communication ever. The reaction split not along "agree or disagree" but along "what to do about it": Yoshua Bengio wrote that warnings from frontier-lab researchers deserve to be taken seriously, David Shor called for government-mandated independent oversight, and several researchers publicly vouched for Coxon's credibility. Separately that same day, an Anthropic researcher put the chance of AI killing all humans within the next decade at more than 10%. Ben's Bites weekly roundup, Bengio's position, the call for oversight, the probability estimate and the debate.

GPT-6 Astra compiled and ran a game on macOS: 120 FPS after six hours

Developer Theo reported that GPT-6 Astra independently compiled and ran Super Smash Bros. Melee on macOS at 120 frames per second, working roughly six hours in a single loop. In parallel, Vals reported that Astra nearly saturated their unreleased computer-use evaluation by building a Minecraft Nether portal in under three hours with no task-specific harness. This is the week's second signal of the same kind: the model operates not on text about a program but on the program itself — mouse, windows, compiler. The practical meaning for work: tasks that used to require writing a script per interface now yield to a plain description of the goal, at the price of hours of machine time. The Melee demo, the Vals measurement.

A 2-billion-parameter model tops the open-weight ranking under 4B

OpenBMB open-weighted MiniCPM5-2B and claims 15 points on the Artificial Analysis Intelligence Index v4.2 — the best result among open models up to 4 billion parameters. Weights are on Hugging Face, code on GitHub. The practical interest of the discussion moved from benchmarks to pipelines: a model this size fits a "speech recognition → model → speech synthesis" chain on weak hardware that used to require a server. A related result landed alongside it: the author of the TAK quantization method reports that his version of Qwen3.8-27B scores 82.81% against 83.59% for the original at full precision, while occupying 15% of its size. Quantization is compression that coarsens the numbers inside a model: it gets several times smaller and faster, but usually loses quality. The MiniCPM5-2B release, the quantization write-up.

Apple A20 Pro: 115 GB/s of memory bandwidth in a phone — laptop M4 territory

Apple's A20 Pro chip gets a 7-core GPU, a doubled 32-core Neural Engine and roughly 115 GB/s of memory bandwidth — about 50% more than the A19 Pro and nearly the 120 GB/s of the laptop M4. Memory bandwidth is the speed at which the processor reads a model's weights; for local AI it matters more than clock speed, because the whole model is read for every answer. Per Notebookcheck's breakdown, the chip moves to TSMC's 2-nanometre process and gets a 96-bit LPDDR5X bus. The constraint is unchanged and it is not speed: the phone is still expected to ship with 12 GB of RAM, and the size of a local model is bounded by capacity, not bandwidth. The specifications breakdown.

LangChain shipped secret management for agents

LangChain Managed Deep Agents 0.7 introduced Connections — a mechanism that separates the agent's own access from the user's: some connections belong to the agent itself, others run through a person's OAuth sign-in. It answers a concrete problem: an agent handed a key to mail or spreadsheets keeps it in its own environment, and revoking that access separately from the agent is impossible. A VS Code update landed alongside it, with automation for recurring work and GitHub flows inside the agents window. Both teams' theme for the week is the same — not "smarter model" but "more predictable permissions". The Connections announcement, the VS Code update.

Numbers and facts

Different perspectives: are warnings from inside the labs a signal or a campaign?

The dispute is not about the facts but about how to read a public resignation. Nobody contests what happened: an Anthropic employee and former OpenAI researcher left publicly, saying the labs are moving too fast toward self-improving superintelligence.

For "this is a signal". Yoshua Bengio wrote plainly that warnings from frontier-lab researchers deserve to be taken seriously (Bengio). Several researchers publicly vouched for Coxon's credibility, among them Ethan Perez and Will Depue (Perez, Depue).

Neutral. David Shor moves the conversation from trust to structure: what is needed is mandatory independent government oversight, not faith in companies' good intentions (Shor). Framed that way, one person's motives stop being decisive.

Against. Parker Thayer described the episode as a politicized campaign bordering on a "psyop" (Thayer); Ben's Bites records that part of the audience is asking outright whether this is a conspiracy (roundup).

Tools and techniques

In brief