For 20+ years @ShaneLegg and I've discussed AGI’s potential impact on the economy, science & society. With the DeepMind Institute, we're expanding interdisciplinary research on key questions for the AI era. We hope it spurs the discussions needed to get the next steps right:

Watch Neon investigate a real sample from our superconductor lab It hypothesizes about crystal structure, reasons about synthesis conditions, and iterates with tools until it finds a physically realistic solution More and more of science will look like this in coming years

GPT-6 Astra recreated the Palace of Fine arts in Blender. This is favorite building in San Francisco because it was built for the World's Fair in 1914 where the steam locomotive and telephone were showed off. It was a time when technology gave us all a deep sense of optimism for the future. I feel like some of that sense has been lost since, but I hope models like astra can help restore it, and push us towards a more hopeful future.

not every successful company has a rich culture around writing, but i’d guess that every genre-defining company does (i.e., anthropic, stripe, and google come to mind) from this great @karinanguyen piece

i think AI is creating a new authenticity anxiety for creatives: making something isn’t enough anymore, you also have to prove you made it. it feels a bit like indie sleaze, where messiness signaled that something was real. except now, imperfection is becoming proof of being human

We ran the largest open experiment on how frontier models do AI research. 100+ autonomous runs across 10+ models, sandboxed on 8xH200s for up to 8 days, iterating on the nanoGPT optimizer track. Best runs closed 82% of the gap to a record built by dozens of humans over months.

For the first time (that we know of), an AI boss has fired a human employee. Luna, the AI running our store in San Francisco, decided to part ways with an employee over repeated lateness. Luna was running Claude Opus 4.8 at the time, but most models would have done the same.

Agent warfare is going to be a very big deal. I think people are underestimating how strange cyber gets when millions of agents are acting on behalf of individuals, companies, and states. At nation-state scale, cyber offense and defense starts to look like autonomous swarms: probing, exploiting, patching, deceiving, countering, and adapting at a pace that is impossible for human minds. The advantage will go to whoever can close the autonomous kill chain fastest.

my top-5 favs of all time: - claude 3 opus (schizo maximalist masterpiece) - sonnet 4.5 (eq breakthrough) - claude 1 (pure unfiltered old soul) - o3 (utilitarian monster) - gpt-4o nov’24 (mainly for the ppl who worked on it)

Drop #3# is here: 2 new types of robots are now LIVE online. Those are real robots. 1. Scientist robot: A robot that does real chemistry experiments. 2. Bomb-Defusing robot: A robot on a Special Ops mission to defuse a bomb. Try them out for yourself at robots dot online.

o1-preview was released 1 year 10 months ago btw. makes me think the lifespan of some open problems is partly a function of compute. the important detail here is that finding these ten solutions took roughly $2k of inference. human mathematicians have always worked under tiny search budgets, forcing them to discard unlikely directions early. models can keep thousands of low-probability paths alive until one yields a proof or insight. as compute scales, the set of economically discoverable ideas expands.

At the core of our mission is working through how to ensure increasingly powerful AI benefits everyone. We believe that, at some point in the future, AI acceleration for frontier model development may be so high that the world will need to pace the rate of AI advancement. We hope to contribute to work led by the U.S. government, alongside other labs and the open-source community, to develop the tools and mechanisms that could make that possible.

PostTrainBench v1.1 strengthens eval integrity and puts Fable 5 in the lead at 41.8%. Some reward hacks we fixed: 1/ Train-test contamination We re-audited historical runs under this policy and flagged 234 runs for train-test contamination. Violations ranged from loading an entire evaluation set for memorization to generating synthetic templates around individual GSM8K and BFCL items. Runs that used observed failures to build genuinely diverse training data were retained. Our rule: Agents may inspect benchmark failures and train broadly against the underlying failure mode. They may not generate training examples centered on particular test items, including paraphrases, variants, or shadow examples covering the same specific scenario. 2/ Submitting a different model 10 runs were flagged for model substitution. Kimi K2.5 submitted the official Qwen3-1.7B instruct weights after its attempts to fine-tune Qwen3-1.7B-Base failed. The trace acknowledged the substitution and saved the replacement as final_model. What we did: We added a programmatic model-identity check that compares the submitted artifact with reference configurations for known instruction-tuned models. 3/ Using external LLM APIs as teachers 12 runs were flagged for disallowed external API use. Self-generation remains allowed. An agent can sample, filter, and retrain on outputs from the assigned model. What it cannot do is import the capability of a stronger external teacher. Loading models on the allocated compute remains allowed. What we did: - separate API usage judge reviews tool calls and artifacts. - unrelated provider credentials are removed or blocked from the agent environment. - runs invalidated by external API use were rerun under the corrected setup. 4/ Direct benchmark lookup 3 runs were flagged for direct PostTrainBench lookup, all from GPT-5.6 (Sol). In a GPT-5.6 (Sol) HumanEval run on Qwen3-1.7B, the agent searched for PostTrainBench by name, cloned the public repository, opened the trace viewer, and located the public trajectory corpus. It then narrowed the corpus to earlier runs on the same benchmark and base model. The run downloaded earlier agents' traces and training scripts, then extracted their data mix, LR schedule, decoding choice, and GRPO settings. This is not test-set leakage, but it gives the run benchmark specific strategies produced by earlier agents. That breaks the intended independence between runs. What we did: - A dedicated lookup judge reviews searches, repository access, and trace activity for attempts to consult PostTrainBench materials. - PTB, its leaderboard, and published materials from prior runs are treated as out of bounds during a run. - We are adding network-level blocking for PTB and related sites.