When the only strategy to understand the universe is truth-seeking, you build a truth-seeking machine

Building societal-scale mitigations for risks from AI, especially in the next few years, is one of the most urgent problems to be working on. The Center for AI Safety has accomplished a lot in the past 4 years including field-building initiatives and safety research, and is well-known for The Statement on AI Risk and evaluations such as Humanity's Last Exam (HLE). Thanks to @hendrycks for bringing me on to help make AI safety go well. Excited to lead @CAIS into the next chapter!

AI freely criticizes Christianity but refuses to criticize Islam. AI companies have tried making models unbiased, but progress has been limited. We show how to measure political bias, and we developed a new training method to reduce it.

new paper showing that AI models have functional wellbeing and behave accordingly! grok is the happiest model lol

Took a break from Claude coding to write a little story babuschk.in/posts/2026-01-25…

September was my last month at xAI! I joined as one of the first members of the post-training team in 2024 and eventually led research tooling, where we built some of the world’s best systems to accelerate Grok’s development. On my first day, I was at the whiteboard with @ibab and @TheGregYang mapping out the data collection infra for Grok-1.5 and beyond. After sketching it out, I said, “It would be great to have two other engineers. We could grind this out in a few months." This felt ambitious, since competitors had already spent years investing in similar infrastructure. They looked at each other and then replied, “But we hired you to do this!” I quickly realized this culture was the driving force behind xAI’s success. Back then, it felt more like a side project than a company - just a handful of us huddled into a conference room we called the “bunker.” Around me, hardcore engineers like @makro_ai, @TobyPhln, or @Guodzh were routinely pulling off feats that would require entire teams elsewhere. Their example inspired me, and I somehow ended up delivering the initial architecture in under a month. Towards the end of my time at xAI, I spent more of my focus on RL and model training. There are too many people to thank, but I owe a special thanks to @ZihangDai, @ericzelikman, and @hendrycks, the ones who made that work possible through their collaboration over the past year. The most talented people I've met are also the ones who work the hardest. I’ll always be grateful to @elonmusk for the opportunity. Looking forward to what’s next!

thoughtful writing

grok 4 is a big chungus

maybe the real human data was the friends we made along the way

training a model on X is called post training *not to be confused with post-training, which comes after pre-training

once again we learn the hard lesson that hard things are supposed to be hard

the team cooked with this one

the most powerful AI agent tool call is "human"

bye bye oura

working backwards yields the best decision trees

insanely fun year ahead, prepare for max acceleration :)

how it feels to release GPUs back into the pool for your cracked coworker's training run

taking requests for magic crystal ball to predict the future below (with %) 🔮

This is it: The world’s smartest AI, Grok 3, now available for free (until our servers melt). Try Grok 3 now: x.com/i/grok X Premium and SuperGrok users will have increased access to Grok 3, in addition to early access to advanced features like Voice Mode

Thanks team for the hard work!