Hi! this is Rohan Anil, I develop training algorithms among other things - my work trains various frontier models. At CoreAuto, I was fascinated by the problem of training on compute islands, eventually in space training. I hope you all enjoy this write up!🪐🛰️

I've always believed that intelligence should come from simple, general building blocks that learn their own abstractions directly from raw data; in other words, the bitter lesson. I think we're finally seeing a real step towards AI that can learn from all data. Language modeling on raw bytes is a microcosm of this ideal, and to me it's a central problem in AI: if we can't build LLMs that learn end-to-end from bytes, how will we handle the rest of the world's data that we understand even less? Last year we introduced hierarchical networks (H-Nets) as a first step. Our new version finds up to 3-5x efficiency gains in controlled scaling law experiments, and at larger scale matches an open-source model trained on 11T tokens with ~10x less data and compute. The posts below walk through what we found and why we think it works.

Introducing Kimi K3.1 Pro: our most capable 3T-class open-weight model. 🔹 1M-token context, 2.8T total / 104B active params (MoE), Multimodal. 🔹 Full-project coding: games in Unreal, Unity & Godot. 🔹 Excels at parametric 3D modeling in Blender & CAD. 🔹 More reliable at finance work: Accurate modeling with hallucination down by 65%. 🔹 Matching the strongest proprietary models across coding, agentic and knowledge benchmarks. Kimi K3.1 Pro and K3.1 Pro Highspeed (2–3× faster) are now live on Kimi Work, Kimi Code, and the Kimi API. Open weights by Nov 4, 2026. 🔗 API: 🔗 Tech blog:

DeepSeek just open sourced its high-performance, large-scale inference framework DeepInfer in Apache-2.0 license. Supported features: - Optimized kernels, model and sampler runtime - Parallelism: Attention DP / balanced CP, MoE EP + EPLB - DSpark and confidence-scheduled verification - SWA bounded replay - Asynchronous scheduling (zero-overhead batch scheduler) - Batch-invariant inference - Deterministic sampling - Structured output

interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic, because neural networks appeared to fold into arbitrary tangled knots. years later, I started openai's first large-scale automated interpretability project, finding about 2% of neurons legible with the primitive tools we had with modern agentic tools and a much closer look, the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works

life update: I've recently left @Kimi_Moonshot! The past year at Kimi has been the best of my life so far. I'm deeply grateful to have worked with and learned from such amazing and supportive colleagues. Some of my favorite memories include (in chronological order): optimizing KDA's kernels and trying every design possible with @yzhang_cs; coming up with and refining Attention Residuals with @yzhang_cs @Jianlin_S; working on optimizers, which by far taught me the most about training dynamics (and where I pulled the most all-nighters); creating the cases for K3's release with @yzhang_cs @Xinyu2ML while watching the World Cup; and more, and more. Quoting my earlier tweet about Kimi K3’s release: it is truly incredible to have built such an open (and strong) model with a group of genuine people. My days at Kimi are memories I'll forever cherish... thanks for everything! (especially @Yulun_Du for believing in and hiring a 16yo back then :P). More things to explore :) 🫡

Can LLMs generate novel research ideas? We propose Priced Guidance to quantify this via compression! We lower bound P(LLM forecasts a future paper’s idea X). Instead of sampling ideas until X is matched, the LLM plays “20 Questions” to recover X, charging each hint by surprisal.

Say hello to Echo, the best writing model at style imitation. Echo beats frontier models at writing tasks ranging from fiction to technical explanations, despite costing less than $5K to train.

Can you pace the frontier? @danrobinson and I made a multiplayer game inspired by a game theory paper from @andrewjkoh where you play an AI lab. Racing ahead increases your profits… but also the risk Try to beat the computer — or your mutuals online :)

We (@PantheonInc) have been working on a robotics data quality pipeline that uncovered a series of major problems in public robotics datasets, especially for world modeling. To improve the quality of data available to open-source robotics, we're publishing annotations for four of the most popular datasets. Some examples of issues, and our report 🧵

Today we announced the Claude-led discovery of a molecular machine that we suspect could represent a new gene editing mechanism. Its precise function, biotechnological utility (if any), or level of significance is not yet clear, but at minimum it is work I would have been proud to do as a PhD student. The work was done mostly, though not entirely, by Claude: our life sciences team suggested a broad area of research, Claude read through the literature and a bunch of genome data and discovered something interesting, then Claude proposed experiments to verify the discovery and our team carried them out. It’s easy to dismiss this as a one-off or curiosity, but we’ve repeatedly seen a pattern where AI performance in new intellectual domains goes from weak to superhuman in a matter of a few years. In 2023 models struggled to do math at the level of an average high-school student. In 2024 they started to do well on math competitions for the best high-schoolers in the country, in 2025 they started to solve minor open problems, in early 2026 more significant open problems, and in late 2026 they are beginning to solve the top few open problems in all of mathematics. We believe AI for biology is on a similar exponential trend. The main difference between biology and mathematics, of course, is that math can be done purely theoretically, while biology requires experimentation. Some have used this to draw the conclusion that AI’s utility in biology will be limited. We think this is wrong. As we’ve demonstrated today, humans can collaborate with AI to perform the experiments, validate key results in a few weeks and, if necessary, work with the AI to iterate on what they find. Eventually it may even be possible for Claude itself to safely perform the experiments by autonomously controlling lab equipment, with appropriate safeguards in place, but we aren’t doing that today (our lab is also a BSL1/BSL2 facility that doesn't handle materials dangerous to humans). More broadly, biomedical advancement has many stages — from fundamental biology discoveries, to translational research, to drug discovery, clinical trials, and finally the actual delivery of medicines and health care to patients. We are also interested in these later stages, but even simply accelerating the first stage of fundamental biological discoveries has the potential to speed up and broaden the entire pipeline. Improving our understanding of biology and sharpening biologists’ tools can drive forward all of the later stages, for example by identifying new drug targets, finding new therapeutic modalities, allowing for more precise measurement, and speeding up the experimental loop which itself further accelerates our understanding of biology. This will not in itself speed up clinical trial times, but if it succeeds it could greatly increase the number of promising candidates that go into the pipeline — an increase in throughput even though latency remains. In Machines of Loving Grace, I wrote about AI’s potential to “cure most diseases in 5-10 years” — a goal that sounds impossible, but one I believe is just barely possible if AI is applied to every stage of the pipeline. The first step is showing that AI can first help with, and then drive, biological discoveries. Claude’s discovery is the latest in a line of related prior work that goes back decades, beginning with systems like CRISPR, and continuing with discoveries like the bridge recombinase and VIPR in the past few years. Recently, there has been heightened interest in systems based on reverse transcriptase (RT) enzymes, the enzyme underlying the system Claude identified. And most recently, a Stanford team working independently described a novel RT system with an associated non-coding array that is in some ways similar to the one Claude found, though they are distinct systems that evolved independently from each other. I believe that we’re at the very beginning of finding such systems and developing them into powerful tools for biotechnology. I’m proud of the resources Anthropic has invested in accelerating the public benefits of AI through the life sciences, and we’re aiming both to grow our life sciences team and to work with other scientists to extend this approach to a broad range of problems. If you have a proposal for a research collaboration or are interested in joining our life sciences team, please reach out.

50+ pages of prose and formulas, zero experiment, but code in the github. Is this a slop-grenade, an unvalidated theoretical proposal, or a rushed flag-plant to stand out in 50k+ iclr submissions?

Sharing my first of hopefully many research blog posts! This one is the kind of educational blog post I wish I'd had when I started with RL for LLMs. I tried to make it as open as possible. Every rollout is browsable, the code is open source, and I walk through my entire thought process, from learning rate sweeps to reward shaping.

i do think a lot of people on the pro-open-source side are having a bit of a knee-jerk reaction to the pacing statements today, as we're used to viewing the closed labs as power-seeking. but i think their hands are somewhat forced here, and this is just another chapter on the fairly inevitable path towards decently-fast decently-safe decently-commoditized intelligence abundance. ask any F500 exec or swing voter. the world doesn't really want super-fast-takeoff superintelligence owned only by two companies, and thus we won't get it. it'll happen at the pace that the world can accommodate it, which means reaching some sort of confidence consensus that the models are aligned enough that we won't be dealing with scary new incidents all the time. this will trickle out broadly, in the form of best practices and distillation. the smarter a model is, the more it has a "personality", and the less effective strict rules are. there will be awkward compromises and moral tensions. but we ultimately just want the models to be reasonable, and to do the sorts of things reasonable humans would do if our brains were faster and less error-prone and had more working memory. i think we'll get there. the labs will build mac and windows, the rest of us are building linux. everyone's gonna do great. weird stuff will keep happening, but we'll still wake up and go to work, until the work does itself in a manner the world finds acceptable.

This is misleading from the lens of practical model architecture research. Some clear limitations that weaken the conclusion: - While SWA can inherit weights learned by full attention, you don't assume the same for linear attention and train them on top of the QKV proj matrices learned by full attn with zero ablations on whether this is close to optimal (in fact this training method is far from optimal). - Apart from how you train it, the linear attn variants tested here are also years behind the current sota. No frontier model uses pure linear attn. It has repeatedly been shown that the optimal config is a hybrid of linear and full, which beats both pure linear and pure full - e.g. Kimi Linear's architecture with 3 kimi delta attention to 1 full attn is, to date, still by far the best-performing arch among the ones I've tried. And yes, SWA hybrids were included in the comparison. - "Training linear attention models is expensive and most software/hardware is made for Softmax attention" This is untrue. There are many fast open-source kernels written for linear attn models whose speeds are even faster than training full attn models under fair context length. Speaking of which, I don't see a training sequence length specified in the paper. - I understand the limited compute. You don't have to do from-scratch or large scale training, but even under this switching setting a principled work still need to properly compare the variants. Overall, I think it is important to have a full understanding of the current state of method X before claiming its false inferiority. If one lacks the compute/infrastructure/data/etc., open research is still very appreciated, but overly strong and unprincipled claims could be misleading.