🚀 Introducing Gemini-SQL2, our breakthrough text-to-SQL capability powered by Gemini 3.1 Pro! We've achieved state-of-the-art results on the highly competitive BIRD benchmark, translating natural language into execution-ready SQL queries. 🧵👇

Watch @ElonMusk provide a technical update on SpaceX’s capability to manufacture, launch, and operate AI satellites at scale → spacexipo.com

Absolutely mind-blowing work. 🤯 This reminds me of when we were exploring the concept of 'thinking with visual primitives'. We noticed the model excelled in synthetic scenarios, but real-world generalization for complex tasks still faced bottlenecks. Back then, I kept thinking: "If only someone could provide manual annotations for these real-world scenarios..." Using point-form visual primitives for reasoning would be powerful for tasks like state tracking, spatial topology, and logical testing, but they may heavily rely on high-fidelity, real-world labeled data. I think it’s the same across other domains too. The journey toward AGI cannot skip the hard, grounded work of human-labelled data. Massive respect to this effort! 🫡

that deepfake of Michael Scott introducing @Karpathy to Dario and the Anthropic crew is the absolute best AI video I have seen so far

Monte Carlo Tree Search training corrects the model move by move, while current LLM training only tells it whether the whole trajectory worked. MCTS is preferable if you can get it. But nobody's managed to get MCTS to work for language models. In his blackboard lecture @ericjang11 talked to me about why:

Learning from rich textual feedback (errors, traces, partial reasoning) beats scalar reward alone for LLM optimization. GEPA demonstrated this for context-space optimization (prompts and agent harnesses), delivering frontier results at a fraction of the cost of RL. But context-only optimization is bounded by the base model's capability ceiling; weight updates can reach further. Very excited about this new line of work on Fast-Slow Training (FST), which interleaves context and model weight optimization! The idea is a clean division of labor between two interleaved loops: 🔹 Fast loop (context): GEPA reads rich rollout feedback updating the context layer. The context becomes a fast-updating scratchpad of what the model needs to know about this task, right now. 🔹 Slow loop (model parameters): RL updates the model's parameters conditioned on the evolving context. Because the prompt already carries task-specific nuances, the model parameters are freed from absorbing them and focus on what actually generalizes across tasks and pushes the frontier. ⦁ 3× more sample-efficient than RL on math, code, and physics reasoning ⦁ ~70% lower KL divergence from base at matched accuracy ⦁ Plasticity preserved: FST checkpoints respond better to additional RL on new tasks than RL-only ones ⦁ Continual learning across changing tasks (HoVer → CodeIO → Physics) where RL stalls the moment the task switches FST is a direction towards: ⦁ Addressing RL's pain points: entropy collapse, sparse rewards, long-horizon exploration ⦁ Providing a clean channel for rich feedback into weight updates ⦁ Demonstrating model-harness co-evolution ⦁ Discovery: Using fast context updates for broad exploration, while leveraging a continually improving model. Check out the full thread below:

This is it. Everything learned spending millions on longevity. From: Your Immortal Unc and Auntie. To: Our Immortal nieces and nephews. 0. Sleep is the world's most powerful drug. 1. Be in your bed for 8 hours 2. Same bedtime every night, any time before midnight 3. Don’t eat right before bed 4. Calm foods for dinner 5. No screens 1 hour before bed 6. Avoid added sugar (be aware it’s in everything) 7. Avoid all things in an American convenience store 8. Avoid fried foods 9. Shoes off at the door 10. Eat whole foods, particularly veggies fruits nuts legumes berries 11. Walk a little after meals or air squats 12. Get your heart rate high routinely 13. Lift heavy things 14. Stretch daily 15. Water pik, floss, brush, tongue scrape, morning and night 16. Make an effort to drink water 17. Get sunlight when you wake up (UV is low) 18. Protect skin in midday sun 19. Stand up straight 20. See at least one friend once a week 21. Avoid plastic where you can (in all things) 22. Circulate air in rooms 23. When stressed, breathe, learn to calm your body 24. Go to the dentist 25. Avoid sitting for long times 26. Protect your hearing, the world is too loud 27. Alcohol is bad for you 28. Finish coffee before noon 29. Avoid bright lights after sunset 30. If obese, look into a GLP 31. Sleep in a cold room 32. Texting while driving is dangerous 33. Turn off all notifications 34. Limit social media use 35. Don’t smoke anything 36. If you struggle to sleep, read a physical book before bed 37. 1 hour before bed have a calm wind down routine: bath, read, light walk, listen to music 38. The body is a clock and loves routine. Have a daily morning and evening schedule. 39. Avoid long distance travel where you can 40. Baby steps first: incorporate new things slowly 41. Do less… most things don’t work. Bonus points if you get your blood checked. Start here, it will change your life.

The day a blind man sees. The first thing he throws away is the stick that has helped him all his life

you can outsource your thinking but you cannot outsource your understanding

Kimchi has been found to fine-tune immunity at the cellular level. In a groundbreaking human clinical trial, researchers employed cutting-edge single-cell genetic analysis to explore kimchi's impact on the immune system. Over 12 weeks, overweight adults received either a placebo or powder from two varieties of kimchi. The findings show that kimchi not only enhances immune defenses but also promotes their regulation. Scientists from the World Institute of Kimchi observed that participants consuming kimchi exhibited heightened activity in antigen-presenting cells (APCs), which play a key role in identifying pathogens such as bacteria and viruses. Simultaneously, CD4 T cells developed a more balanced profile, incorporating both effector cells for fighting threats and regulatory cells to prevent overreactions. This dual effect equips the immune system to mount effective responses while avoiding harmful inflammation. The trial utilized single-cell RNA sequencing (scRNA-seq), a precise technique that examines gene expression in individual immune cells, uncovering nuanced changes undetectable by standard blood tests. Fermentation method influenced outcomes: both spontaneous and starter-culture kimchi yielded benefits, but the starter version demonstrated superior effects, including enhanced antigen detection and reduced extraneous immune signaling. This represents the world's first clinical demonstration of kimchi's immunomodulatory action at the genetic level, highlighting its promise for managing hyperactive immune conditions. ["Single-cell RNA sequencing reveals that kimchi dietary intervention modulates human antigen-presenting and CD4⁺ T cells." npj Science of Food, 2025]

This is a 1953 photo of two 4-year-old Australian boys playing in an asbestos sandpit. At the time, very little was known about the devastating effects of asbestos, and people commonly used asbestos sandpits for their children to play in. The boy on the left, Philip Noble, later grew up to become a footballer before dying of mesothelioma at 36. The boy on the right, Ross Munroe, later became a high school principal and also died of mesothelioma at 38.

post-training/RL details of deepseek-v4: i) they replace the mixed RL stage used in deepseek v3.2 with on-policy distillation. they still use RL but RL is mainly used to create strong domain specialists first. then they consolidate those specialists into one final model through OPD. ii) for each domain they start with supervised fine-tuning on high-quality domain-specific data. after that they run RL using GRPO guided by domain-specific prompts and reward signals. the paper does not provide all reward formulas but the important structural point is that each specialist gets its own domain-conditioned reward distribution. iii) deepseek-V4 supports three explicit reasoning modes: non-think, think high & think max. they trained distinct specialist models under different RL configurations to support different reasoning capacities. for each mode they apply different length penalties and context windows during RL training which changes the model’s output reasoning length. iv) one of the more interesting parts is for hard-to-verify tasks they say they discard conventional scalar reward models and instead use a Generative Reward Model or GRM. they curate rubric-guided RL data and use a GRM to evaluate policy trajectories. crucially they apply RL optimization directly to the GRM itself and the actor network natively functions as the GRM (that means the same model family is learning both generation and evaluation...i guess this is closer to LLM-as-judge trained with RL...implicitly moving from reward-as-score to reward-as-deliberative-evaluation). v) deepseek-v4 treats long-horizon tool use as a state-retention problem...instead of flushing reasoning after every user/tool boundary it preserves the full reasoning trace across agentic tool conversations, making context management part of the trained policy interface. vi) also it collapses auxiliary routing decisions search/no-search, query generation, domain classification, URL reading, etc. into special tokens appended to the main model context...reusing the existing KV cache instead of paying redundant prefill for separate router/controller models. vii) deepseek-v4 uses full-vocabulary logit distillation. that means instead of only comparing teacher and student on the sampled token they compare the full probability distribution over the vocabulary at each position. viii) for agentic post-training and evaluation, deepseek builds DeepSeek Elastic Compute (DSec). it is a production-grade sandbox platform with three rust components: a) apiserver b) edge c) watcher...these communicate through custom RPC and scale over deepseek’s 3FS distributed filesystem...the paper says a single DSec cluster manages hundreds of thousands of concurrent sandbox instances...also for each sandbox DSec keeps a globally ordered trajectory log recording every command and result. ix) the paper does not go in full details of async-rl...it focuses more on fault-tolerant asynchronous execution than on off-policy correction under stale rollouts (also on RL/OPD execution). it does not fully reveal the behavior-policy staleness semantics of the RL/OPD loop.

AI bois be like:

Thanks to the entire Claude community for giving feedback and continuing to build with us. Read the full post-mortem here: anthropic.com/engineering/ap…

Every agent will need its own computer. And with new Hosted agents in Foundry, every agent gets its own dedicated enterprise-grade sandbox, with durable state, built-in identity and governance, and support for any harness or framework. Read more: devblogs.microsoft.com/found…

Columbia CS Prof Vishal Misra explains why LLMs can’t generate new science ideas. Bcz LLMs learn a structured map, Bayesian manifold of known data & work well within it, but fail outside it. True discovery requires creating new maps, which LLMs can't do

Meet Kimi K2.6 agent - Video hero section, WebGL shaders, real backends. From one prompt. 🔹 Video hero sections - cinematic aesthetic, auto-composited 🔹 WebGL shader animations - native GLSL / WGSL, liquid metal, caustics, raymarching 🔹 Motion design - GSAP Framer Motion 🔹 Backend database: Kimi wires up auth database backend in one pass. 🔹 Website stack - React 19 TypeScript Vite Tailwind shadcn/ui 🔹 3D w/ physically-based lighting - Three.js React Three Fiber

Demis Hassabis’s “Einstein test” for defining AGI: Train a model on all human knowledge but cut it off at 1911, then see if it can independently discover general relativity (as Einstein did by 1915); if yes, it’s AGI.