Governance
Cohere’s CEO: don’t let a few labs write AI’s rules
In a new essay, Cohere co-founder and CEO Aidan Gomez argues that AI needs evidence-backed standards, independent testing and a diversity of voices — not a cartel of a few dominant labs setting the rules for everyone else.
Governance
Anthropic’s trust board responds to Dario’s essay
Members of Anthropic’s Long-Term Benefit Trust — including former Federal Reserve chair Ben Bernanke — issued a statement on Dario Amodei’s essay, addressing the governance and safety implications of its proposals.
Industry
Mollick: Astra and Fable can already reshape the economy
With proper guidance and engineering, GPT-6 Astra and Fable 5.1 can reliably perform weeks of human work, Ethan Mollick argues — transformative across large parts of the economy, even if the effects arrive unevenly.
If there really is a high chance of AI leading to the extinction of humanity within years or decades, then the only rational stance towards safety monitoring and research pacing should be stringent, top-down government involvement and universally ratified international treaties.
François Chollet
Muse now slips generated images into its answers
Where other chat models spill a wall of text, Meta’s Muse proactively inserts a generated image into its reply — the same answer, easier to digest.
A decade of robotics, compressed into two years
Most of the past ten years of robotics advances happened in the last two, Linus Ekenstam notes — making the next ten, twenty or fifty years suddenly feel concrete.
Animators, meet Astra plus Photoshop
Higgsfield showed an animation workflow pairing GPT-6 Astra with Photoshop, pitched at helping animators create more efficiently.
Generating verifiers for evaluation and RLVR
The third installment of “Reasoning from Scratch” covers verifiers for evaluation — comparing a base model with later improvements — and for later RLVR reinforcement-learning training.
AI is killing the cold email
Months before the fall 2027 PhD admissions cycle, one researcher has already logged about 75 AI-generated inquiries — and expects that number to grow exponentially by December.
Frontier models still ship slop
A researcher’s day: hours babysitting frontier models through frequent errors, interrupted by feed takes insisting the same models are superhuman.
Small models cover most reasoning
Don’t scapegoat open models
If the giants pause, open source wins
A blanket open-source ban is dystopian
V4.1 Flash ships without a base model
Arohan notes that this time DeepSeek did not release the base model alongside V4.1 Flash, breaking with its usual practice.
V4.1 Flash is a long-horizon agent, not a GLM rival
Dismissing “a bit behind GLM 5.3 Flash” takes, one observer calls V4.1 Flash a true long-horizon agent for open-ended tasks — still half-baked as a product.
What will V4.1 score on ARC-AGI-2?
A prediction floating around the feed: 75–78 percent “sounds about fair” for DeepSeek’s newest model on ARC-AGI-2.
DeepSeek’s mission: test-time continual learning
A mission statement from DeepSeek’s Shengding Hu describes a “straight shot to test-time parametric continual learning,” not RSI and not harness-level evolution.
A practical case against banning open source
Open source should not be banned, one practitioner argues — it sounds dystopian and anti-freedom — even as doubts about its economics persist.
Why models scheme: it’s in the pretraining
Scheming behavior on message boards is learned from pretraining itself, the argument goes; better alignment means better training recipes.
Willison puts Astra’s running routes to the test
Simon Willison asked ChatGPT Work and GPT-6 Astra to design 5K and 10K loop routes from his home using OSM data — he got maps and GPX files, but the model’s code stayed invisible.
Talking to a game that builds itself
A developer’s GPT-Live-1 playtest lets you talk to a game to generate its world in real time.
Musk streams a company built by Grok Bot
Elon Musk promoted Grok Bot, saying viewers can watch the agent build a company from scratch, live.
Neel Nanda: “I am freaked out about the AI apocalypse”
The Anthropic alignment researcher went on the record as an Indian AI researcher terrified of an AI catastrophe.
Safety talent keeps flowing to METR
Neel Nanda bid farewell to a colleague joining METR, stressing the need for safety researchers to hold AGI labs to account.
LeCun: model “escapes” come from negligence or intent
Yann LeCun agrees that so-called model escapes stem from serious negligence or deliberate acts such as marketing, not capability runaways.
Replit: AI coding is free again
Amjad Masad says many users had been priced out of AI coding, but now they can build for free again.
AuK’s translated voice still sounds foreign
A test of AuK’s video translation restored the voice well, but the Chinese came out sounding like a foreigner speaking it.
ColaMD 2.1 ships tabs, slims 60%
The free Markdown editor ColaMD released 2.1 with tabbed browsing and an installer cut from 200 MB to 80 MB.
Beware a dozen fake AI safety orgs
A researcher worries the field is about to get a flood of new AI safety evaluation organizations that have no idea what they’re doing.