Dòng tin

simonw released 1.0a32 at simonw/datasette

datasette 1.0a32

R to @simonw: Excellent news, this was a mistake, it's coming back! https://x.com/PhilippSpiess/status/2061189796689494339

RT by @simonw: Sorry about this! The button was unintentionally removed during some recent menu restructurings. It'll be back with the next hotfix release. You can bind a custom keyboard shortcut to the "Copy as Markdown" command and use that as a workaround!

RT by @_akhaliq: The new Suyra OCR 2 model is top trending on http://paperswithcode.co Also can you spot the new "Conferences" tab? 👀

RT by @dair_ai: 😂PewDiePie building his own agent orchestrator and releasing it was not on my 2026 bingo card. Own the agent. Own the harness. It's not that hard, folks.

R to @simonw: I'd been letting Codex Desktop update itself, now I feel like I need to hold out on upgrades in case they remove key features that I'm depending on!

RT by @dair_ai: // The Efficiency Frontier // Cool paper on context management. As agents reuse the same documents and histories across many turns, the cheapest context strategy is not fixed. This work describes a principled rule for picking one per deployment instead of defaulting to whatever topped a benchmark in isolation. Retrieval and compression methods are almost always benchmarked on accuracy and cost separately, so you never learn when one actually beats another under real load. The Efficiency Frontier models context strategy selection as a single cost-performance problem, with a log-utility term for diminishing returns from extra context and a reuse parameter N that amortizes preprocessing across repeated queries. Sweep N and the optimal strategy changes, exposing crossover regions where retrieval, compression, or full context each wins. On 5,000 HotpotQA instances, deployment-aware selection cuts effective token usage about 25 percent at the same performance, and amortized memory compression runs over 50 percent cheaper than full-context prompting in higher-performance settings. Paper: https://arxiv.org/abs/2605.23071 Learn to build effective AI agents in our academy: https://academy.dair.ai/

Pinned: The Top AI Papers of the Week (May 24 - May 31) - SkillOpt - AutoScientists - The Efficiency Frontier - Language Models Need Sleep - Adapting the Interface, Not the Model - Forecasting Scientific Progress with AI - Compiling Agentic Workflows into Weights Read on for more:

x.com/i/article/206078258805…

RT by @_akhaliq: Top AI Papers of The Week (May 25-31): - Gamma-World: Generative Multi-Agent World Modeling Beyond Two Players - SkillOpt: Executive Strategy for Self-Evolving Agent Skills - Qwen-VLA: Unifying Vision-Language-Action Modeling across Tasks, Environments, and Robot Embodiments by Alibaba - LocateAnything: Fast and High-Quality Vision-Language Grounding with Parallel Box Decoding - AgentDoG 1.5: A Lightweight and Scalable Alignment Framework for AI Agent Safety and Security - DVAO: Dynamic Variance-adaptive Advantage Optimization for Multi-reward Reinforcement Learning - Lens: Rethinking Training Efficiency for Foundational Text-to-Image Models - WBench: A Comprehensive Multi-turn Benchmark for Interactive Video World Model Evaluation - Rethinking Cross-Layer Information Routing in Diffusion Transformers Find them below:

RT by @_akhaliq: Want to try Step 3.7 Flash without touching a line of code? It now has a hosted demo — open it right in your browser, no install needed. Built on Gradio by @_akhaliq 🙏, now live in our Hugging Face org. Give it a try 👇

RT by @dair_ai: The efficiency frontier! Where do you think GPT-5.6 will land?

RT by @jeremyphoward: Codex just found a “workaround” of not having sudo on my pc…

RT by @ylecun: I'm MAGA's twisted sense of reality and morality, a scientist who developed treatments and vaccines that have saved hundreds of thousands of lives (without getting rich from it) should be sent to Gitmo. "In addition to continuing work on vaccines already in clinical trials for hookworm as of 2010 and schistosomiasis, Hotez led a team of researchers developing vaccines against other diseases including leishmaniasis, Chagas disease, SARS, and MERS, As of 2020, he was also working in development of a Coronavirus vaccine. With Maria Elena Bottazzi, he led the team that designed COVID-19 vaccine named Corbevax." https://en.wikipedia.org/wiki/Peter_Hotez

The end will begin when humanity turns away from humanity

RT by @ylecun: Exxon Says We’re Two Weeks From Petrol Armageddon Right. So it turns out closing the world’s most important oil chokepoint has consequences. Who knew. Exxon’s Senior VP Neil Chapman stood up at a Bernstein conference this week and said what everyone in the industry is thinking but politely avoiding at dinner parties: global oil inventories are approaching “truly unprecedented” lows. His words. And when a Senior VP at Exxon starts using phrases like “truly unprecedented,” you should probably pay attention. The IEA has already called Iran’s Strait of Hormuz blockade the largest oil supply disruption in recorded history. Over a billion barrels lost since late February. A billion. That’s not a rounding error. That’s a civilisational inconvenience. Chapman’s estimate: Brent crude hits $150 to $160 per barrel once stockpiles reach critical levels. Two to three weeks away. Wood Mackenzie, never a firm known for cheerful forecasts, went further and suggested $200 by year-end if the strait stays shut. With a global recession as a bonus prize. So there we have it. The world spent decades building an energy system with a single point of failure, handed Iran the keys to it, and is now surprised that the keys are being used. Genius, really. https://open.substack.com/pub/gandalv/p/your-grocery-bill-is-going-to-rise?r=3v7cjb&utm_medium=ios

RT by @_akhaliq: We just adopted a super cool new space template for LocateAnything, made by @_akhaliq the great. Thank you AK! Try it out: https://huggingface.co/spaces/nvidia/LocateAnything Credit to AK's space example: https://huggingface.co/spaces/akhaliq/LocateAnything

RT by @drfeifei: GPIC should be the new standard benchmark for generative modeling. Training 1 epoch on GPIC is the same cost as 100 epochs on ImageNet, but is a much better proxy for real-world problems. If you work in generative modeling, try GPIC for your next project!

Did you think about the Roman empire today?

Einstein on (not) using NL for invention: "The words or the language, as they are written or spoken, do not seem to play any role in my mechanism of thought"

I’m very excited by this new benchmark dataset for visual generation that is suitable for the modern era of large scale generative models!🤩

RT by @_akhaliq: This week, I got our GitHub Actions to use @HuggingFace Jobs instead of the default GitHub CI runners, making workflows run on much more reliable CPUs or even on serverless GPU (that cost less than a penny per run)! Here's what you need to do this for your own repos ⤵︎

I have difficulty arguing that most of the writing changes suggested in gmail now aren’t improvements, but it does tend to wipe out my particular authorial voice.

RT by @ylecun: .@ylecun’s definition of what is a world model.

RT by @demishassabis: Gemini Omni 🐦 prompt in 🧵

R to @jeremyphoward: I've largely switched over to using GPT-5.5 in recent weeks, which I like nearly as much as Opus 4.6 and 4.7, and is *very* reasonably priced (since I can use my subscription with the @OpenAI API.)

Has @AnthropicAI completely given up on making API usage reasonably-priced? Following the token-usage changes recently they announced various updates to *subscription* usage to make it more reasonable. But they've done NOTHING for API users. The cost is insane at this point.

Worked on some code this morning using Opus 4.8 and so far I'm really liking it. Much more cooperative than 4.7 and less "over agentic". Stops and asks for my input when needed in places 4.7 (and GPT 5.5) would just foolishly blast ahead. (Still WAY too expensive.)

R to @jeremyphoward: (This is using our in-house harness, Solveit.)

RT by @ylecun: Took me a while to figure out what all the ESMFold2 rage was about. At first, the benchmarking data didn't look super remarkable to me but it turns there are many impressive aspects: - Fully open source, open weights + massive ESM Atlas (1.1B structures vs 0.2B for AF3). - SOTA performance despite no MSA use. MSA search and triangular attention were simply taken out of the base model. - Direct consequence, super low latency inference: 1024-residue protein structure prediction in 9 secs, still outperforming prior models on antibody-antigen tasks. - Best in class PPI and antibody-antigen results. 65% pass rate on antibody-antigen benchmarks after inference-time scaling, significant improvement over AF3. - Tons of experimental data, in particular with lab-validated miniprotein binders plus single-chain antibodies across 5 targets in cancer and immunology. Binding affinities consistent with therapeutic activity. - Inference-time scaling benefits PPI: Multiple seeds + selection by confidence show real gains on challenging antibody-antigen predictions, leading to comments/hypotheses that it has learned an energy-function-like behavior via the folding module. - Base model works without MSAs, but providing them further boosts prediction quality on difficult protein-protein interaction cases. One caveat: No true scoring for protein-protein interactions, making it harder to assess which specific residues or domains are reliably involved in binding.

RT by @ylecun: Woah. A profound shift in American science is coming. Every federal research grant could soon require sign-off from a political appointee. Scientific progress depends on funding decisions grounded in merit and public health need, NOT ideological approval.

RT by @demishassabis: After AlphaGo, the skill of human Go players noticeably improved. I suspect we will see a similar pattern in math.

It’s a real honor to receive an honorary doctorate of science from @BrownUniversity . 😍

The MiniMax M2 series was one of the most widely used open-weight LLM series earlier this year. Now, we got a technical report with some interesting tidbits. I summarized some of them below: 1. Full attention as an anti-trend?: They tried hybrid sliding-window attention variants (like so many others, like Xiaomi MiMo, Laguna, Gemma 4, Arcee, Olmo 3, etc.). But even though there were efficiency gains, they said that the production-quality tradeoffs were not worth it for M2. 2. Linear and sparse attention deployment issues: They found that linear and sparse attention are attractive on paper because they reduce the cost of long-context attention, but they are harder to make work well in a production agent system. In particular, they found that these efficient attention variants may be more fragile when KV-like state or intermediate memory is stored in lower precision. Also, they have worse prefix caching support, which matters a lot when using coding agents (which reuse a lot of the context). 3. Fine-grained Mixture-of-Experts (MoEs) are useful: Finally a recent MoE ablation study! It's only on the 2B-active parameter scale, but hey, better than nothing. Concretely, they compare a baseline with 32 experts and top-2 routing against a fine-grained setup with 128 experts and top-8 routing. The fine-grained setup improves MATH from 19.6 to 24.1 and HumanEval from 29.7 to 32.5. That's clearly a win for more fine-grained experts (confirming what the DeepSeek MoE paper reported ~2 years ago). 4. Sophisticated agent pipeline It's probably no surprise, but this papers confirms that training for agent-like behavior on software engineering task is now a big component of the training pipeline. They mine GitHub pull requests, builds runnable Docker environments, extracts task-specific test rewards, etc. 5. Interleaved thinking for context management Interestingly, they found that removing reasoning blocks from previous turns results in worse performance, especially in multi-step agent tasks. (Another point why long-context support is so important these days). 6. Speed rewards It's common to have token usage penalties, but what's interesting is that the MiniMax team adds a task-completion-time reward that depends on wall-clock time. This is to minimize unnecessary (slow) tool calls. Also, I'm thinking that this would encourage agent parallelization (if supported by the harness) 7. Self-evolution Looks like self-evolution is also already a big design component of open-weight LLMs. E.g., the paper says that M2.7 already handles 30 to 50 percent of the daily RL iteration workload, modifies its own scaffold, and completed a 100-round autonomous scaffold optimization cycle with a 30 percent gain on internal evaluations.

RT by @demishassabis: When I was asked by the American Academy of Arts and Sciences to write an essay on my thoughts on how AI will accelerate Science, I felt honored but also felt that it would require a lot of thoughtfulness and diligence to distill my thoughts on paper. The essay has now been published and I cannot be more thankful to the >@americanacad and @GoogleDeepMind teams for their feedback and encouragement during the process. Key reflections from my essay: 🔭 AI is our newest revolutionary lens: Just as the telescope and microscope expanded our physical perception, AI is extending our cognitive reach, allowing us to decipher the immense complexity of the data-universe. 🧬 The rise of "machine intuition": AI is not just a computational engine. By detecting hidden structures across disciplines—from protein folding to extremal combinatorics—it acts as an ultimate bridge, accelerating the interdisciplinary breakthroughs that modern science depends on. 🏗️ From puzzle-solvers to architects of questions: As we transition toward open-ended, agentic AI systems that actively generate novel hypotheses, the burden of reasoning is shifting. We are evolving from being the solvers of intricate puzzles into the architects of profound scientific questions. ✨ Expanding human potential: AI won't replace scientists; it expands what we can imagine and achieve. Just as the telescope didn't make astronomers obsolete, AI is giving us the stars. Read the full essay here: https://www.amacad.org/publication/daedalus/unlocking-scientific-intuition-reasoning-at-digital-speed

RT by @demishassabis: Gave google omni a sketched camera path and asked it to generate drone POV footage.

I have been very impressed by @SemiAnalysis_ . I think of myself as a wide ranging systems engineer, looking for value at every level from the chip specs to the user interface, but SA exposes me to additional levels of "the system", both above (datacenters) and below (semiconductor fabrication). It probably puts me in "just knows enough to be dangerous" territory. Neat things I learned today: Some of the 800VDC datacenter design choices leverage parts commoditized by electric vehicles. There is now a SiC MOSFET that can operate on 10kV electricity, opening up the possibility of working directly with medium (ha!) voltage AC power transmission lines without stepping down.

It is easy enough to make your own, but I think standard relu should have been defined as passing the value at zero, so gradients flow backward through it, allowing some things to be zero weight initialized when symmetry breaking isn’t an issue.

RT by @demishassabis: omni continues to blow my mind (left original / right generated) waymo looks sick in matte black

How good is AI memory?

RT by @drfeifei: OpenArt now lets you turn a single image into a persistent 3D world creators can direct with precise control. Wide shots, top-down views, over-the-shoulder framing. All from the same environment, acting like a permanent virtual set. Powered by the World Labs API. Learn more ↓

Developer productivity is hard to manage.

Creativity feeds on constraints

Memorial Day. Today we honor those who gave everything to defend our democracy. Democracy isn't guaranteed; it's a precious inheritance that requires our constant care. 🇺🇸

Added a DeepSeek Sparse Attention (DSA) from-scratch implementation to my LLMs-from-scratch repo thanks to an awesome new reader contrib. With motivation, overview, and GPT-style model reference implementation as standalone example code: https://github.com/rasbt/LLMs-from-scratch/tree/main/ch04/09_dsa

AI Dev 26 x SF | Ara Khan: Evals Are Broken Use Them Anyway

The new White House policy requiring green card applicants to apply from outside the US is a capricious attack on legal immigration. It will hurt families, leave us with fewer doctors, teachers and scientists, and hurt American competitiveness in AI.

Semantic Search Starts With Embeddings

R to @rasbt: PS: it can be found in recent Qwen models since Qwen3-Next

Gated DeltaNet has been one of my favorite "hybrid attention" newcomers in the good old transformer stack. Excited to see Gated DeltaNet-2. Adding it to my reading stack. In the meantime, I have a primer on Gated DeltaNet here: https://magazine.sebastianraschka.com/i/177848019/26-gated-deltanet

It's been *almost* a bit quiet around LLM architecture releases in the past two weeks 😅 Interesting tidbit is the parallel block design. Via the Cmd-A the tech report "equivalent performance but significant improvement in throughput compared to the vanilla transformer block."

RT by @drfeifei: The winners of the first ever World Jam have arrived! 🏆 Step into the Museum, an interactive archive showcasing the winning creations from the World Jam. Where every portal opens into a new spatial experience. 👉 http://jam.worldlabs.ai

We would love to see more collaboration and research in the field of human-AI interactivity. Check it out!

It is a shame that the simple act of transferring a large block of data as fast as possible over the internet is not handled effectively by the primitive operating system calls. You either multiplex over parallel persistent TCP connections to combat head-of-line blocking and slow starts, or reinvent reliable delivery and flow control over UDP. QUIC has a lot going for it, but it is a large library (six figure LoC!) and conflates security and performance in a way I don’t love. There is also fundamental information about competition with other processes and link layer congestion that should be useful, but is unavailable to user libraries. You should be able to just write(really_big_buffer) and it is all taken care of for you.

Personal update: I've joined Anthropic. I think the next few years at the frontier of LLMs will be especially formative. I am very excited to join the team here and get back to R&D. I remain deeply passionate about education and plan to resume my work on it in time.

RT by @ID_AA_Carmack: Doom Just Received One Of The Highest Cultural Honors In The US http://dlvr.it/TSbrMH

I only recently read more about the concept of system accidents by Charles Perrow, very insightful and relatable.

RT by @lilianweng: 2. (Real time fact checking) - The Interaction Models hear you speak and fact-checks you in real time — like having a teammate who's always paying attention.

Pinned: In the past few months, we had a lot of fun (and stress 😅) to produce 12 versions (+ many subversions) and 137 pages in our training run log book. Turns out human-human collaboration is important to improving human-AI collaboration. 😊

This works really well btw, at the end of your query ask your LLM to "structure your response as HTML", then view the generated file in your browser. I've also had some success asking the LLM to present its output as slideshows, etc. More generally, imo audio is the human-preferred input to AIs but vision (images/animations/video) is the preferred output from them. Around a ~third of our brains are a massively parallel processor dedicated to vision, it is the 10-lane superhighway of information into brain. As AI improves, I think we'll see a progression that takes advantage: 1) raw text (hard/effortful to read) 2) markdown (bold, italic, headings, tables, a bit easier on the eyes) <-- current default 3) HTML (still procedural with underlying code, but a lot more flexibility on the graphics, layout, even interactivity) <-- early but forming new good default ...4,5,6,... n) interactive neural videos/simulations Imo the extrapolation (though the technology doesn't exist just yet) ends in some kind of interactive videos generated directly by a diffusion neural net. Many open questions as to how exact/procedural "Software 1.0" artifacts (e.g. interactive simulations) may be woven together with neural artifacts (diffusion grids), but generally something in the direction of the recently viral https://x.com/zan2434/status/2046982383430496444 There are also improvements necessary and pending at the input. Audio nor text nor video alone are not enough, e.g. I feel a need to point/gesture to things on the screen, similar to all the things you would do with a person physically next to you and your computer screen. TLDR The input/output mind meld between humans and AIs is ongoing and there is a lot of work to do and significant progress to be made, way before jumping all the way into neuralink-esque BCIs and all that. For what's worth exploring at the current stage, hot tip try ask for HTML.

I think modern LLMs are p-zombies without moral patienthood—on par with insects, at best, in my moral calculus. But I also I think we should establish norms for treating models well *before* models with patienthood exist—i.e. now. We should want to have this right from day one.

LLMs soon: “So, what are some good qualia for someone just getting into not being a p-zombie?”

R to @goodside: Also I should clarify: I don’t mean God literally, I mean natural selection. And by “bad idea” I mean “bad for inclusive genetic fitness.”

R to @goodside: (Several replies have noted this is the premise of Peter Watts’s 2006 novel Blindsight, which I haven’t read but will now.)

I believe in the Festivus School of prompt engineering, which says all prompts used in production naturally iterate toward an airing of grievances—a list of all the ways the model has disappointed you in the past year.

This is the the quote I've been citing a lot recently.

Fireside chat at Sequoia Ascent 2026 from a ~week ago. Some highlights: The first theme I tried to push on is that LLMs are about a lot more than just speeding up what existed before (e.g. coding). Three examples of new horizons: 1. menugen: an app that can be fully engulfed by LLMs, with no classical code needed: input an image, output an image and an LLM can natively do the thing. 2. install .md skills instead of install .sh scripts. Why create a complex Software 1.0 bash script for e.g. installing a piece of software if you can write the installation out in words and say "just show this to your LLM". The LLM is an advanced interpreter of English and can intelligently target installation to your setup, debug everything inline, etc. 3. LLM knowledge bases as an example of something that was *impossible* with classical code because it's computation over unstructured data (knowledge) from arbitrary sources and in arbitrary formats, including simply text articles etc. I pushed on these because in every new paradigm change, the obvious things are always in the realm of speeding up or somehow improving what existed, but here we have examples of functionality that either suddenly perhaps shouldn't even exist (1,2), or was fundamentally not possible before (3). The second (ongoing) theme is trying to explain the pattern of jaggedness in LLMs. How it can be true that a single artifact will simultaneously 1) coherently refactor a 100,000-line code base *and* 2) tell you to walk to the car wash to wash your car. I previously wrote about the source of this as having to do with verifiability of a domain, here I expand on this as having to also do with economics because revenue/TAM dictates what the frontier labs choose to package into training data distributions during RL. You're either in the data distribution (on the rails of the RL circuits) and flying or you're off-roading in the jungle with a machete, in relative terms. Still not 100% satisfied with this, but it's an ongoing struggle to build an accurate model of LLM capabilities if you wish to practically take advantage of their power while avoiding their pitfalls, which brings me to... Last theme is the agent-native economy. The decomposition of products and services into sensors, actuators and logic (split up across all of 1.0/2.0/3.0 computing paradigms), how we can make information maximally legible to LLMs, some words on the quickly emerging agentic engineering and its skill set, related hiring practices, etc., possibly even hints/dreams of fully neural computing handling the vast majority of computation with some help from (classical) CPU coprocessors.

RT by @karpathy: New work with @AlecRad and @DavidDuvenaud: Have you ever dreamed of talking to someone from the past? Introducing talkie, a 13B model trained only on pre-1931 text. Vintage models should help us to understand how LMs generalize (e.g., can we teach talkie to code?). Thread:

Building technologies for better human-AI collaboration on next gen hardware at scale. Exciting.

May 2021 Gwern.net Newsletter

April 2021 newsletter