Tim Rocktäschel

BALROG’s leaderboard has three new entries, courtesy of @creus_roger Very interesting that GPT 5.6 Sol at max effort is still within error of Gemini 3 and 3.1 pro. GPT Astra 6 on max reasoning effort however reaches new heights, and a @NetHack_LE avg. progression of 13% 🏰

x.com/i/article/2093…

It's good that MPs want to discuss long tail risks, and engage with what is certainly a technology that presents both a huge opportunity and challenge for Britain. But it is wholly irresponsible to say on the one hand 'we think this is really important' and on the other block the UK's ability to build AI capabilities here and support British AI companies. In the event that neither our allies nor foes listen to the proposed international agreement and that in fact a 'kill switch' is impossible, every single one of these MPs should be fighting to strengthen Britain's hand in this critical technology by supporting local datacenters, so we're not exporting our data abroad, reforming regulations so that AI models like cyber defenses can be trained here, so we're not entirely reliant on foreign technology and supporting British AI companies, so we see the economic benefits here too. We're doing our small bit at SovAI, but this should be a national level debate, and I hope this MP champions the UK scaling those ambitions just as passionately.

🔥Does academia stifle creativity? This new perspective argues that academia is currently structured to select against creativity, intellectual risk-taking and bold ideas. Young scientists – who may be best positioned for new ideas – are particularly incentivized to play it safe.

"Is it AGI" flow chart. Developed with @_rockt at NeurIPS 2022.

I always liked nethack. This 42 year old marvel is one of the deepest adventure games ever made. It also looks like crap, at least out of the box. GPT6 Astra made a headless version for me. You can now use its brain as a library or as mcp or via javascript library. Meet t.co/lVljspZmYs

the only alignment that will work is enlightenment

A fun small win for automated research: We found and helped fix some edge cases that could affect inference performance in vLLM and SGLang. In our last blog post, we described a reward hacking judge we developed for performance optimization tasks. While applying the judge to some new autoresearch work, it discovered that some FlashInfer (a library underpinning vLLM and SGLang) kernels used a hard coded value of -50,000 as a masked-attention sentinel -- even though valid QK values can be smaller. Corner cases like this that are numerically wrong but silent can cause a huge amount of headache to find and fix, e.g., the historical debate around flash attention (t.co/WtoQH97rYM). Great use case for AI. t.co/aiTEQn8hR3

Lots of AI neolabs popping up these days 🧑‍🔬🧑‍💻🥼. Maybe you’ve thought: I should raise billions, buy 100k GPUs and try to safely build superintelligence too! 🧠💸 Well now you can! Try this new neolab simulator: bodono.github.io/neolab.ai (Best on screens larger than phones).

TL;DR: This is one of the most important and exciting opportunities in AI on the planet - please read on. The British Open-ended Learning & Discovery Lab is creating the perfect place for paradigm breaking AI research in the name of open-source and open-science. We have agency, we funding, we have unprecedented amounts of compute*, but WE NEED YOU! ..and we have created the dream job for you: The BOLD Fellow. This job combines a fast-moving, high agency, collaborative environment with full academic freedom and a salary that pays the bills. Apply by noon UK time on the 15th of September for this once in a lifetime opportunity to shape the history of our field and of our planet: t.co/wN009tQEkt *by academic standards

the models have no moat (OpenAI, Anthropic, XAI) the IDEs have no moat (Cursor, Windsurf) the harnesses have no moat (Cognition, Factory, LangChain) the app builders have no moat (Replit, Lovable, Bolt) the wrappers have no moat (Harvey, Abridge, OpenEvidence) the inference providers have no moat (Together, Fireworks, Groq) the voice layer has no moat (Sierra, Decagon, ElevenLabs) the data labeling companies have no moat (Scale, Surge, Mercor) the AI infrastructure has no moat (Baseten, Modal, Railway) the neoclouds have no moat (CoreWeave, Lambda, Crusoe) the generative media companies have no moat (Runway, Higgsfield, Suno) apparently nobody in AI has a moat except the venture firm ☠️

Today we announce our $100M seed round to redefine how humanity computes in AI's next chapter. The future of compute and AI is heterogeneous. Callosum is building it: callosum.com/blog/seed-round

So being an area chair at @NeurIPSConf now means that I have to decide whether the author LLM won the arguments with the reviewers LLMs?

We’re excited to announce DiG-bench, a new benchmark for discovery! Over the last few weeks we’ve been testing frontier AI models on our novel discovery games and seeing how they score. Each game is a text-based environment, so they probe discovery capabilities in the natural domain of language models, rather than requiring additional, potentially confounding, visual understanding. TL;DR frontier models have improved a lot over the last few months. But they are still stumped by some surprisingly simple problems, even in their native text domain. With @cocosci_lab (@Princeton) @MITCoCoSci (@MIT) @SchmidhuberAI (@KAUST_News) @misovalko (@Inria) @tri_dao (@PrincetonCS) @RMBattleday @zebkDotCom @FraserGreenlee @akaijsa @ClareMaguire @TimMuller1 @kubicek_ales @physicscat0x7d @SukritSumant @thoughtchannel_ (1/5)