Thomas Wolf(@Thom_Wolf)

love that it’s all open source

Thomas Wolf (@Thom_Wolf) — Mon, 19 Oct 2026 21:02:33 GMT

love that it’s all open source

Jack Boylan@JackBoylan_98

added Fish Audio’s Drama-3 to speech-to-speech 🎭

cloned my voice. now I can ask it to whisper, laugh, sigh or get worried, just by talking.

cheers to @andimarafioti  for the demo + sampling tips!

PR + setup:github.com/huggingface/sp…c

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬3🔄1❤️13👀2886📊5

⚡ Powered by xgo.ing

What's still very hard: theory of mind. The world, the voices and the mechanics were great. Writing ...

Thomas Wolf (@Thom_Wolf) — Mon, 19 Oct 2026 11:19:02 GMT

What's still very hard: theory of mind. The world, the voices and the mechanics were great. Writing the story was the weak part: the model assumes the player knows more than they do, and its NPC sometimes react in strange way or to things that never happened.

A Diamond Age's Primer (or any real AI tutor fwiw) will have to model the learner, and that is where it still falls short in a default setting imo. I've actually started exploring benchmarks on this and it's a hard task for LLMs right now. They are prone to the "Curse of knowledge". Teaching humans is also not very much a Verifiable Reward, so it's hard to scale within the current training paradigm.

💬2🔄0❤️4👀1999📊2

⚡ Powered by xgo.ing

Play it here, free, in your browser, nothing to install: https://t.co/ml7a7NvinN Works on a laptop a...

Thomas Wolf (@Thom_Wolf) — Mon, 19 Oct 2026 11:19:02 GMT

Play it here, free, in your browser, nothing to install: huggingface.co/spaces/thomwol…
Works on a laptop and on your phone.

Seven islands of about 15 minutes each, and six possible endings, depending on what you taught Kite along the way. 

It’s still very much an exploration, a beta more than a finished game. Expect rough edges

💬9🔄1❤️11👀2312📊10

⚡ Powered by xgo.ing

In The Diamond Age, Neal Stephenson imagined a book that invents games and stories to teach a little...

Thomas Wolf (@Thom_Wolf) — Mon, 19 Oct 2026 11:18:57 GMT

In The Diamond Age, Neal Stephenson imagined a book that invents games and stories to teach a little girl what she needs, as she grows. I tried a small version of it. I gave Claude my brainstorm doc on alignment and asked it to make a game, something like Zelda, to teach people about AI alignment. Here's what came out 🪁

Short version: it's great at building the world, and still weak at understanding the player (details below)

Your browser does not support the video tag.

🔗 View on Twitter

💬30🔄9❤️77👀9159📊38

⚡ Powered by xgo.ing

The most built-upon model on the Hugging Face Hub isn’t Llama or Qwen. It’s FLUX.1-dev from Black F...

Thomas Wolf (@Thom_Wolf) — Sun, 18 Oct 2026 23:20:16 GMT

The most built-upon model on the Hugging Face Hub isn’t Llama or Qwen.

It’s FLUX.1-dev from Black Forest Labs, a German lab: 43k derivative models.

Europe trained ~8% of the open-model compute on the Hub, but ~18% of what people build on top of it is European.

Build or surrender? Build, and build open. See you at GTC Berlin this week

Arthur Mensch@arthurmensch

Should Europe leave frontier AI to the Americans and the Chinese?

The idea has gained ground in Europe. It is a fundamental mistake.
Build or surrender: for now, the choice is still ours.

Full article below. x.com/i/article/2110…

🔗 View Quoted Tweet

💬20🔄33❤️321👀14708📊47

⚡ Powered by xgo.ing

$40M to ~$1B in under two years

Thomas Wolf (@Thom_Wolf) — Sun, 18 Oct 2026 16:27:33 GMT

$40M to ~$1B in under two years

Thomas Wolf@Thom_Wolf

Fun and crazy napkin math: Meta, NVIDIA, Alibaba, DeepSeek, Moonshot, Mistral and ~30 other labs have shared close to $1B worth of training compute on the Hugging Face Hub

Free to download

End of 2023 it was under $40M

Wikipedia was written in volunteer hours. This AI age commons is written in GPU-hours and data

(it's a floor btw: about half comes from official compute numbers rest is educated guess and it still miss the ~900k community fine-tunes)

Space with data to explore: thomwolf-hub-dollars.static.hf.space

🔗 View Quoted Tweet

💬4🔄5❤️52👀8357📊8

⚡ Powered by xgo.ing

Fun and crazy napkin math: Meta, NVIDIA, Alibaba, DeepSeek, Moonshot, Mistral and ~30 other labs hav...

Thomas Wolf (@Thom_Wolf) — Sun, 18 Oct 2026 14:34:40 GMT

Fun and crazy napkin math: Meta, NVIDIA, Alibaba, DeepSeek, Moonshot, Mistral and ~30 other labs have shared close to $1B worth of training compute on the Hugging Face Hub

Free to download

End of 2023 it was under $40M

Wikipedia was written in volunteer hours. This AI age commons is written in GPU-hours and data

(it's a floor btw: about half comes from official compute numbers rest is educated guess and it still miss the ~900k community fine-tunes)

Space with data to explore: thomwolf-hub-dollars.static.hf.space

💬57🔄32❤️227👀40547📊86

⚡ Powered by xgo.ing

we’re getting used to it by now but DeepSeek calling this a “minor update” while moving Terminal-Ben...

Thomas Wolf (@Thom_Wolf) — Sat, 17 Oct 2026 10:17:12 GMT

we’re getting used to it by now but DeepSeek calling this a “minor update” while moving Terminal-Bench 4.0 from 31.2 to 42.3 is very funny.

also reports 77.0 on DeepSWE (vs 74.0 for Opus 5 and 73.0 for GPT-5.6 Sol…)

and the most interesting line may be faster task completion with shorter outputs. using fewer tokens is one of the practical way the cost curve collapses often done as an afterthought

all with MIT-licensed weights in DeepSeek tradition

now curious to see how these numbers will hold up outside their own eval setup

DeepSeek@deepseek_ai

🚀 DeepSeek-V4.1-Flash-1017 is here!

Stronger benchmark performance, faster task completion, and shorter outputs than the previous version—a Pareto improvement in performance and cost. Now live on our website, app, and API.

🛠️ Optimized for @DeepSeekHarness, Tencent CodeBuddy, WorkBuddy, and more. Upgrade to DeepSeek Harness v0.2.1-rc.1 or later for the best experience with the new model.

1/4

🔗 View Quoted Tweet

💬73🔄47❤️853👀71626📊125

⚡ Powered by xgo.ing

"how to avoid brain rot" was a great watch. some of my favorites points: - using ai to ask yourself ...

Thomas Wolf (@Thom_Wolf) — Sat, 17 Oct 2026 10:01:05 GMT

"how to avoid brain rot" was a great watch. some of my favorites points:
- using ai to ask yourself socratic questions instead of giving you the answer makes you feel like you're learning less but research shows you're actually learning more
- ai tooling is moving so fast that it's better to know 50% of a new tool than to master 90% of an older tool

Miles Cranmer@MilesCranmer

Here is the full talk:

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬41🔄252❤️3371👀156519📊813

⚡ Powered by xgo.ing

it's time to invent math 2.0 (sounds on)

Thomas Wolf (@Thom_Wolf) — Fri, 16 Oct 2026 15:36:05 GMT

it's time to invent math 2.0
(sounds on)

Your browser does not support the video tag.

🔗 View on Twitter

SAIR@SAIRfoundation

The math community deserves its own AI system.

Introducing Open Math OS: the infrastructure for Math 2.0.

Open models and AI agents. 24/7 cloud compute with 100+ tools. The community controls its data.

Compute is free for approved teams. Apply now:
openmathmodel.org

🔗 View Quoted Tweet

💬12🔄32❤️279👀29750📊54

⚡ Powered by xgo.ing

we’re so back

Thomas Wolf (@Thom_Wolf) — Thu, 15 Oct 2026 18:15:17 GMT

we’re so back

Alexandr Wang@alexandr_wang

1/ muse spark open is the most capable US open-weight model to date.

816B total params, 42B active. 1M-token context. natively multimodal. built for coding and agentic work like long-horizon tasks, hard software engineering, multi-step reasoning.

america! 🇺🇲

🔗 View Quoted Tweet

💬3🔄2❤️63👀7961📊7

⚡ Powered by xgo.ing

how i feel using this cutom tiny worklouder to switch between 36 sessions

Thomas Wolf (@Thom_Wolf) — Thu, 15 Oct 2026 13:09:17 GMT

how i feel using this cutom tiny worklouder to switch between 36 sessions

💬12🔄1❤️51👀6320📊15

⚡ Powered by xgo.ing

As always, great to read METR’s thoughtful and in-depth work on this topic

Thomas Wolf (@Thom_Wolf) — Thu, 15 Oct 2026 10:52:58 GMT

As always, great to read METR’s thoughtful and in-depth work on this topic

METR@METR_Evals

🔗 View Quoted Tweet

💬3🔄3❤️18👀4788📊5

⚡ Powered by xgo.ing

wat. nanoGPT, but it trains inside a tweet. GPU-poor final form tap and press train https://t.co/K9Z...

Thomas Wolf (@Thom_Wolf) — Wed, 14 Oct 2026 14:46:03 GMT

wat. nanoGPT, but it trains inside a tweet.
GPU-poor final form
tap and press train
thomwolf-premiers-mots.static.hf.space/share.html?v=4

💬17🔄12❤️95👀12924📊34

⚡ Powered by xgo.ing

stuxnet, but for frozen peas: hackers turned French supermarket freezers into ovens this week hard ...

Thomas Wolf (@Thom_Wolf) — Mon, 12 Oct 2026 22:24:10 GMT

stuxnet, but for frozen peas: hackers turned French supermarket freezers into ovens this week

hard to see how invite-only AI cyber defense ever reaches a supermarket freezer though

GeoInsider@InsiderGeo

BREAKING: 🇫🇷 A major cyberattack has hit supermarket refrigeration systems across France, reportedly causing fresh and frozen food to spoil after hackers remotely manipulated temperature controls.

According to BFMTV, the attack began overnight from Sunday into Monday, targeting older versions of refrigeration-control software supplied by Danish company Danfoss.

The attackers allegedly exploited the systems’ defrosting function, forcing refrigeration equipment to heat continuously instead of keeping products cold. Temperatures were reportedly set as high as 50°C, destroying entire stocks of food within hours.

Several major supermarket chains have been affected, including Intermarché, Super U anE.LeclercPZ. Metro stores and independent businesses, including butcher shops also have been impacted.

Around 150 Intermarché stores are reportedly affected, according to industry sources cited by BFMTV. The full number of affected locations remains unclear.

The financial damage is already significant. In Haguenau, Alsace, an Intermarché store reportedly lost its entire frozen-food stock, worth approximately €50,000. A nearby Super U allegedly had to discard both fresh and frozen products, with losses estimated at €220,000. AE.LeclercPZ store in the same town was also affected, although its losses had not yet been calculated.

The combined losses reported at the Intermarché and Super U stores in Haguenau alone amount to around €270,000. Nationwide damage could reach several million euros.

Intermarché reportedly confirmed the incident in an internal communication, warning that several stores had suffered losses of frozen products. Coopérative U also acknowledged the attack and established crisis teams to assist affected stores and coordinate with refrigeration specialists.

Technical investigations are ongoing, and the full extent of the damage has yet to be established.

Authorities and companies are still working to determine the full circumstances of the attack. The identity of the perpetrators has not been confirmed.

🔗 View Quoted Tweet

💬1🔄0❤️6👀464📊1

⚡ Powered by xgo.ing

open and state-of-the-art decision models - just saying

Thomas Wolf (@Thom_Wolf) — Mon, 12 Oct 2026 16:14:17 GMT

open and state-of-the-art decision models - just saying

Victor M@victormustar

Decision 3.0 hype is real! it is the top performer in my little apple-harvesting test. It can actually play: read the scene, pick its own moves and collect apples without a pathfinding controller.

Input: a screenshot of its surroundings, game stats and a short memory of its last four moves. It chooses up/down/left/right each turn.

This run: 7 apples in 38 moves, zero damage 🔥

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬3🔄4❤️59👀8189📊9

⚡ Powered by xgo.ing

Blue = what we knew. Green = what's new. We're so early.

Thomas Wolf (@Thom_Wolf) — Mon, 12 Oct 2026 16:08:39 GMT

Blue = what we knew. Green = what's new. We're so early.

Leandro von Werra@lvwerra

We released Carbon-A: a DNA model that proposed over 500M gene candidates across 22k species!!

huggingface.co/spaces/Hugging…

blue: existing annotations in database 
green: new annotations by Carbon-A

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬1🔄7❤️106👀10460📊15

⚡ Powered by xgo.ing

⚠️ Viewer discretion advised: Microduck is subjected to a microaggression

Thomas Wolf (@Thom_Wolf) — Fri, 09 Oct 2026 09:17:33 GMT

⚠️ Viewer discretion advised: Microduck is subjected to a microaggression

Your browser does not support the video tag.

🔗 View on Twitter

💬29🔄30❤️370👀23083📊71

⚡ Powered by xgo.ing

Scientists have sequenced the genomes of thousands of species. But for most of them, nobody knows *w...

Thomas Wolf (@Thom_Wolf) — Thu, 08 Oct 2026 15:38:42 GMT

Scientists have sequenced the genomes of thousands of species. But for most of them, nobody knows *where the genes are*

Today we're releasing Carbon-A, an open model that finds genes directly in DNA. We used it to create a database of 566 million candidate genes across 22,617 species, from fungi to mammals (that we are sharing as well)

Why bother? A few reasons

1. Elephants rarely get cancer. Part of the explanation turned up in their genes: extra copies of a tumor-suppressor gene humans have only one of. You can't ask that kind of question about a species until someone, or some tool, has found its genes.

2. The first GLP-1 drug was based on a peptide from Gila monster venom. Nobody would have put the Gila monster on a priority list. Same for wild relatives of crops, which carry resistance to drought and disease. Many of these species have never been annotated.

3. Most of what we know about genes comes from about a dozen model species, like mice, flies and yeast. That focus worked remarkably well, but annotation pipelines still lean heavily on comparisons with them, which makes genes unique to other species easy to miss, and those can be the most interesting ones. 

Carbon-A belongs to a newer family of tools that read DNA directly. It learned from known genomes, but it doesn't need a close relative to read a new one.

4. Even familiar genomes still have gaps. With our partners at Active Site and UCSD, we found RNA evidence for 239 genes missing from the reference annotations of species as common as cats, chickens, hamsters and a lab plant.

5. AlphaFold can predict the shape of a protein, but only once someone has found the gene that makes it. In a species with no annotated genes, it has nothing to work with.

Some notes on safety:

Carbon-A doesn't design DNA or predict what a gene does. It marks where genes are in DNA. This is one of the reasons we think releasing it openly is the right call.

More annotations also help health research: many species that carry or cause disease are among those covered, and work on controlling the diseases they spread often starts from their genes.

Model and database and more details in Georgia's thread 👇

Georgia Channing@cgeorgiaw

Today, we're releasing the Carbon Annotation Database and the model behind it, Carbon-A. 

We've used this model to discover 566.34 million new gene candidates across 22,617 species, thousands of which have never been studied like this before. 

We wet-lab validated a number of these new genes in well-studied species, such as in cats, chicken, and arabidopsis (plant), with our partners at Active Site and @UCSD, with more extensive wet lab results on previously-unstudied genomes coming soon.

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬18🔄33❤️227👀16330📊48

⚡ Powered by xgo.ing

Oral spotlight session at #CoLM 2026 - come say hi to @joelniklaus @_lewtun and me

Thomas Wolf (@Thom_Wolf) — Wed, 07 Oct 2026 22:44:32 GMT

Oral spotlight session at #CoLM 2026 - come say hi to @joelniklaus@_lewtun and me

💬2🔄4❤️34👀5412📊6

⚡ Powered by xgo.ing

how do you like them convolutions

Thomas Wolf (@Thom_Wolf) — Tue, 06 Oct 2026 23:15:48 GMT

how do you like them convolutions

Fireside Alpha@firesidealpha

Ben Affleck reveals he writes Python, understands convolutional neural networks, worked extensively with GPUs, and used his celebrity status to get private looks at Google and OpenAI’s video models

“I’ve always been kind of into computers since I was young. Then, when film started to move from analog film to digital, I became more interested in that aspect of it. The visual-effects workflow for many years has included machine learning, so I can write pretty shitty Python scripts and stuff like that.

“With convolutional neural networks, which were the precursors to what the transformer can do, which is much more computation simultaneously, you would do things like look at what’s called a tensor. That’s the numerical translation of a visual image in numbers, like the batch number, the frame number and the red, green and blue values of each pixel in each frame. It’s just that simple. That numeric is called a tensor.

“You’d use a convolutional neural network to identify patterns that reveal what’s called edge detection or feature extraction, which is identifying patterns well enough to know, this is where the window ledge is, so we can more easily take the green-screen image out and replace it with something.

“That was familiar to me early on because, prior to Artists Equity, I had a small visual-effects company. I’ve worked with GPUs a lot too. The visual-effects guys said, ‘Hey, you should see. There are a couple: Google and this other company, OpenAI, are doing really interesting stuff with transformers in video.’

“I’ve learned that I can actually just call up and go, ‘Hey, it’s Ben Affleck. Can I come see what you’re doing?’ Sometimes people say yes, to my astonishment.”

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬8🔄3❤️150👀16714📊19

⚡ Powered by xgo.ing

No hype, just the results. Good.

Thomas Wolf (@Thom_Wolf) — Tue, 06 Oct 2026 23:11:32 GMT

No hype, just the results. Good.

OpenAI@OpenAI

We’re releasing a broad range of new mathematical results produced by an internal frontier model.

We’ve been consulting with the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study, and we have drawn on their advice and public recommendations to inform how we release these results. 

github.com/openai/math

🔗 View Quoted Tweet

💬7🔄5❤️217👀17862📊21

⚡ Powered by xgo.ing

Every robot starts on a board you bought. The good ones end up on a board you built.

Thomas Wolf (@Thom_Wolf) — Tue, 06 Oct 2026 23:04:02 GMT

Every robot starts on a board you bought. The good ones end up on a board you built.

Matthieu Lapeyre@matth_lapeyre

Microduck got its new brain!

Our first fully custom SBC in a Pollen robot! It replaces the Radxa + board sandwich from the prototype.

Built by Seeed around the same Rockchip CPU but including better memory, better WiFi/BT, dual NFC antennas, user-facing status LEDs, an RGB flashlight, and major improvements in cooling, cable management and boot speed.

First revision arrived, and we got everything running on it within a day. Pretty good day :)

A big milestone on the road to shipping!

🔗 View Quoted Tweet

💬5🔄7❤️85👀9002📊13

⚡ Powered by xgo.ing

count down is starting for the release of ML4 be notified on the instant of the release: https://t....

Thomas Wolf (@Thom_Wolf) — Tue, 06 Oct 2026 14:34:00 GMT

count down is starting for the release of ML4

be notified on the instant of the release: huggingface.co/mistralai/Mist…

Mistral AI@MistralAI

Meet Mistral Large 4, aka Le Chonk. 

• 1T parameters, natively multimodal. 49B active.
It is the best open weights model from US or Europe on aggregated benchmarks.
• State-of-the-art on critical workloads, including cyber defense, manufacturing and finance and it surpasses closed frontier models on visual grounding.
• Forged in Europe end-to-end and is deployable from Europe via our own Mistral Cloud infrastructure.
• Available to all via API today. Working with cybersecurity partners privately.

Open weights release end of October.

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬7🔄6❤️60👀5082📊10

⚡ Powered by xgo.ing

I hope we keep Opus 4.6 around for a long time

Thomas Wolf (@Thom_Wolf) — Mon, 05 Oct 2026 21:26:07 GMT

I hope we keep Opus 4.6 around for a long time

David@DavidSHolz

i asked LLMs to have fun, write down what happened, then rate who had the most fun. the newer models seem to have less fun! many models played word games, but opus 4.6 won over and over by imagining brief worlds made up of contradictions living inside falling water droplets 🥺

🔗 View Quoted Tweet

💬11🔄4❤️100👀13753📊16

⚡ Powered by xgo.ing

microduck's first BBQ. made a lot of friends. did not eat anything.

Thomas Wolf (@Thom_Wolf) — Mon, 05 Oct 2026 14:33:49 GMT

microduck's first BBQ. made a lot of friends. did not eat anything.

Your browser does not support the video tag.

🔗 View on Twitter

Thomas Wolf@Thom_Wolf

Testing compatibility with the local wildlife

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬78🔄79❤️1067👀123464📊172

⚡ Powered by xgo.ing

Testing compatibility with the local wildlife

Thomas Wolf (@Thom_Wolf) — Sun, 04 Oct 2026 20:23:45 GMT

Testing compatibility with the local wildlife

Your browser does not support the video tag.

🔗 View on Twitter

💬15🔄7❤️256👀85738📊40

⚡ Powered by xgo.ing

new roommate just moved in. walks like he's had three drinks, says no to everything. still quite cut...

Thomas Wolf (@Thom_Wolf) — Sun, 04 Oct 2026 08:08:53 GMT

new roommate just moved in. walks like he's had three drinks, says no to everything. still quite cute so I might bring him everywhere with me

Your browser does not support the video tag.

🔗 View on Twitter

💬115🔄123❤️1815👀155446📊325

⚡ Powered by xgo.ing

seems obvious that most personal AI should run on your device love what Sigil is building with Unde...

Thomas Wolf (@Thom_Wolf) — Fri, 02 Oct 2026 17:23:31 GMT

seems obvious that most personal AI should run on your device

love what Sigil is building with Underdog - powered by the Hugging Face hub

Sigil Wen@0xSigil

Introducing Underdog, your Private Personal AI on devices you already own

Today we're announcing our backing from @a16z@khoslaventures@HummingbirdVC Anthology (@AnthropicAI@MenloVentures) @patrickc@naval@rauchg@polynoamial@Thom_Wolf@Mascobot@tszzl@OfficialLoganK and other top AI leaders

Our mission is to provide free, capable, reliable and private AI to billions of people.

Underdog’s Law: today’s frontier intelligence reaches your devices in six months.

We’re starting with fast, capable AI that runs on consumer hardware. By co-designing models and inference engines, we’re pushing the frontier of capability, speed, power and data efficiency.

Our research also spans agentic commerce, confidential inference, and how AI will reshape the internet economy.
 
Privacy and capability no longer needs to be a tradeoff. If you believe in this future, join us. Time to build.

🔗 View Quoted Tweet

💬21🔄12❤️118👀17698📊31

⚡ Powered by xgo.ing

is it how it works @scottbelsky?

Thomas Wolf (@Thom_Wolf) — Fri, 02 Oct 2026 13:12:44 GMT

is it how it works @scottbelsky?

💬1🔄0❤️4👀2204📊1

⚡ Powered by xgo.ing

the first good AI movie is going to start with Ben Affleck explaining learning rates. Matt Damon pla...

Thomas Wolf (@Thom_Wolf) — Fri, 02 Oct 2026 13:12:43 GMT

the first good AI movie is going to start with Ben Affleck explaining learning rates. Matt Damon plays the base model

Thomas Wolf@Thom_Wolf

Karpathy: disappears from X
 Ben Affleck: alright, gather round, so you'll want to freeze the base weights first, learning rate 2e-4

🔗 View Quoted Tweet

💬22🔄5❤️138👀13892📊25

⚡ Powered by xgo.ing

Very thoughtful piece from Kevin Buzzard (perfect IMO score, number theorist, pioneer of formal math...

Thomas Wolf (@Thom_Wolf) — Fri, 02 Oct 2026 08:47:20 GMT

Very thoughtful piece from Kevin Buzzard (perfect IMO score, number theorist, pioneer of formal maths in Lean)

if math is not just about “human understanding,” then what is it about?

and if ai capabilities keep growing exponentially what happens since “mathematics is infinite”?

Bartosz Naskręcki@nasqret

I cannot agree more. Kevin Buzzard made so many points I agree with. But the best one is this

"I thus believe that in the future we will reach a new “natural boundary” in mathematics, beyond (and perhaps way beyond) where we are now, but where machines are going to get stuck and where it is not viable to expend any more resources to make the next big leap. (...) I believe that the optimal thing to do (...) is to let the machines loose, see what happens, and then begin the journey to where they have stopped."

xenaproject.wordpress.com/2026/10/01/to-…

🔗 View Quoted Tweet

💬17🔄34❤️275👀35332📊62

⚡ Powered by xgo.ing

Karpathy: disappears from X  Ben Affleck: alright, gather round, so you'll want to freeze the base w...

Thomas Wolf (@Thom_Wolf) — Fri, 02 Oct 2026 06:28:57 GMT

Karpathy: disappears from X
 Ben Affleck: alright, gather round, so you'll want to freeze the base weights first, learning rate 2e-4

💬137🔄373❤️7739👀741867📊1282

⚡ Powered by xgo.ing

ESM-2 came out in 2022. It's still downloaded hundreds of thousands of times a month. That only wo...

Thomas Wolf (@Thom_Wolf) — Wed, 30 Sep 2026 16:58:12 GMT

ESM-2 came out in 2022.

It's still downloaded hundreds of thousands of times a month.

That only works because someone keeps maintaining the software underneath it.

@huggingface 🤝 @os4science are teaming up to find those libraries and back the people behind them 🧬os4science.org/news/hugging-f…CM

Open Source for Science Fund@os4science

We're joining forces with @huggingface to identify the software libraries that scientific model contributors rely on most and explore opportunities to support the maintainers behind them.   os4science.org/news/hugging-f…

🔗 View Quoted Tweet

💬5🔄4❤️29👀6028📊7

⚡ Powered by xgo.ing

GPT 6.1 duck

Thomas Wolf (@Thom_Wolf) — Tue, 29 Sep 2026 18:15:24 GMT

GPT 6.1 duck

💬3🔄3❤️122👀6988📊14

⚡ Powered by xgo.ing

People are worried because the most likely explanation for such a sudden drop in cheating is evaluat...

Thomas Wolf (@Thom_Wolf) — Tue, 29 Sep 2026 05:54:03 GMT

People are worried because the most likely explanation for such a sudden drop in cheating is evaluation awareness: the latest Opus models may now be smart enough to recognize that this benchmark tests for cheating, and behave accordingly.

If so, the benchmark no longer measures the models' "natural" tendency to cheat.

💬50🔄44❤️670👀64576📊128

⚡ Powered by xgo.ing

impressive

Thomas Wolf (@Thom_Wolf) — Tue, 29 Sep 2026 05:46:41 GMT

impressive

Larry Dial@classiclarryd

New historic NanoGPT record at 39.9s (-27.7s) from @DevenPzak , obliterating the prior record of 67.6s!

This record introduces a new paradigm of thinking to NanoGPT: instead of optimizing matmuls or adding more expressive operations, optimize at the individual flop level with incredibly clever engineering and ML judgement. If a flop is low value on a particular step, skip it.
Specifically:
-(~8s) Sampled softmax. If a token doesn’t appear in a batch, skip its lm_head fwd/bwd some fraction of the time.
-Sparse values. Only run an optimizer step for ngram embeddings that occurred in the batch. Set beta1 to zero to enable this. Beta2 is applied retroactively when the row is later used.
-Sparse updates. Only update ngram and value embeddings once every 4 steps instead of once every 2.
-Sparse communication. Shard the n-gram table across GPUs, and only pass the rows receiving updates on each step.
-Sparse optimizer states. For the n-gram table, reduce from 2 floats in Adam optimizer per param, to 1 float per 768 params.
-Hand-rolled flash attention for 64 dim heads.

There are several additions that add accuracy too:
-(~4s) EMA during last 300 steps, combined with lifting final_lr to 0.3 instead of 0.15.
-(~1s) A new optimizer, Anvil2, which expands muon via a second tracked momentum buffer, improves the ortho coefficients, and modifies the cautious weight decay application.
-A couple additional dynamic skip connections in the network.

The most striking consequence of the ‘flop aware paradigm’ is you can grow parameters arbitrarily large,
only limited by the available memory, since you can selectively choose how to expend flops on those parameters on each step. NanoGPT has kept active parameters below 124M, but total is unbounded, and has grown to 640M through embedding sparsity over the last year. This PR takes that to its logical conclusion on the 8xH100, scaling up to 65B sparse embedding parameters, which accounts for 25% of the PR’s gains. At frontier scale, where one is not bounded by an 8xH100, one could imagine where this paradigm could lead.
github.com/KellerJordan/m…

As this was a very notable PR, I spoke with Deven for an hour to learn how he did it. Here’s his story on the changes: hyperstition.cc/training-nanog…

🔗 View Quoted Tweet

💬2🔄3❤️41👀13855📊5

⚡ Powered by xgo.ing

A "summer in hell" at OpenAI – great read by @joedaroo "Time to prepare is now, not after the surpr...

Thomas Wolf (@Thom_Wolf) — Mon, 28 Sep 2026 19:32:00 GMT

A "summer in hell" at OpenAI – great read by @joedaroo

"Time to prepare is now, not after the surprise"
"Give the model only the access it needs"
"Test that the boundaries actually hold"
"Keep evidence outside [model] control"
Safety & infrasec teams "should be best buddies"

Joe@joedaroo

Took a minute to write a few words about security & safety as someone who lived through it all at OpenAI. I hope my thoughts help someone out there. https://t.co/gJBf08vJH9

🔗 View Quoted Tweet

💬7🔄9❤️63👀20653📊19

⚡ Powered by xgo.ing

In July, AI agents running a security test escaped their sandbox and ended up inside @huggingface's ...

Thomas Wolf (@Thom_Wolf) — Mon, 28 Sep 2026 13:52:55 GMT

In July, AI agents running a security test escaped their sandbox and ended up inside @huggingface's servers.

So today we're happy to be among @nvidia and @JensenHuang's partners on the release of the NVIDIA Open Agent Safety Platform.

The key idea: don't count on the agent to respect the rules. Agents write and run their own code, and when one path is blocked they look for another. In one of NVIDIA's tests, an agent that wasn't allowed to push code through GitHub's API simply switched to git instead. We've now seen countless examples of this behavior.

For now, safety can't live only inside the agent. It has to be built around it.

That's the concept behind OpenShell (open source, Apache 2.0):
- the agent runs in a Linux sandbox (Landlock + seccomp): no root, no direct network access
- a supervisor outside the sandbox holds the real credentials
- the agent only gets a placeholder token, swapped for the real one on approved calls

My favorite piece is a solver (built on Z3) that checks mathematically whether a new permission opens a door that was supposed to stay closed. In NVIDIA's tests, an AI reviewer approved a bad permission request and the math check caught it.

The Sentry integration is exciting too: a watchdog running on a BlueField-4 DPU, i.e. separate silicon sitting on the node's only path to the model. A monitor on separate hardware keeps watching even if the host OS is compromised.

It's a first step, and a lot can be built on top of it. Right now the prover checks permissions, not intent, and the open-source part is mostly OpenShell rather than Sentry. But it's clearly the right direction.

github.com/NVIDIA/OpenShe…

Jensen Huang@JensenHuang

Today, with over 100 industry partners, we introduced the NVIDIA Open Agent Safety Platform, bringing together OpenShell and Sentry.

Artificial intelligence is extraordinary technology that will advance discovery, productivity, security, health, and prosperity for generations to come.

But its full promise can only be realized when people have confidence that AI is being built to be safe and deployed with wisdom and responsibility.

This is bigger than a single product. It's the beginning of an open ecosystem to build the trust layer for safe agent systems.

Together, we are building the foundation of the AI economy.

Trust and innovation are not in conflict. Safety is how trust is earned. We must build not only the most capable AI, but the most trusted AI, so that this extraordinary technology can realize its enormous promise for the world. nvda.ws/4hcoq7m

🔗 View Quoted Tweet

💬58🔄45❤️223👀31356📊68

⚡ Powered by xgo.ing

“Getting verified information about what's going on inside AI companies seems particularly urgent no...

Thomas Wolf (@Thom_Wolf) — Mon, 28 Sep 2026 05:57:34 GMT

“Getting verified information about what's going on inside AI companies seems particularly urgent now” - @RyanGreenblatt

Ryan Greenblatt@RyanGreenblatt

I'm joining METR to work on more investigations like our Hugging Face report.

Currently, tons of even basic information about AI development that's highly relevant to catastrophic risk isn't public. I used to be more skeptical of the value of public info, but recent events have changed my mind.

Getting verified information about what's going on inside AI companies seems particularly urgent now. The limited public evidence we have seems consistent with the possibility that imminent recursive self-improvement could massively accelerate capabilities progress, which could then potentially yield extremely superhuman general capabilities within 6 months or a year. If this occurred, there would be a correspondingly large risk of worst-case outcomes. This uncertainty about extreme outcomes could be substantially resolved with more verified public information: we could either build more consensus about near-term risk or learn that such extreme outcomes are less likely in the near term.

Beyond AI capabilities and takeoff, the state of public evidence is also highly limited for alignment, security, control, and risk-relevant internal processes at AI companies. This makes it hard to determine exactly how well or poorly these key areas will go in the near future. (METR plans to focus, at least initially, on just capabilities/takeoff, alignment, and control; I hope other groups cover security, internal processes, and other important areas.)

While I'm no longer working at Redwood, I think the work they are doing is very important; I'm excited about Redwood's ongoing contributions to R&D on technical mitigations and better public interpretation of risk-relevant evidence.

🔗 View Quoted Tweet

💬9🔄2❤️55👀10102📊11

⚡ Powered by xgo.ing

"we had a wonderful run chiseling that code by hand and it’s over. On the other side of that is an ...

Thomas Wolf (@Thom_Wolf) — Sat, 26 Sep 2026 19:53:54 GMT

"we had a wonderful run chiseling that code by hand and it’s over.

On the other side of that is an exciting new career as professional maker of things, steering intelligence that was only available in science fiction up until a moment ago

What a privilege to have been there when we did it all by hand and be there at the exact moment where we switch over."

Scott@scottstts

My god this is such a good speech that every SWE needs to hear. You know what? Every person should hear it

Keep the happy memories, eyes on the reality, be excited about the future. That’s the best that anyone can do

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬18🔄42❤️342👀40281📊60

⚡ Powered by xgo.ing

turns out you don’t need IMU 🤯

Thomas Wolf (@Thom_Wolf) — Sat, 26 Sep 2026 19:18:14 GMT

turns out you don’t need IMU 🤯

Aditya Bhatt ✈️ IROS 2026@aditya_bhatt

In my latest PhD paper, we declare WAR on sensor-maxxing. 

For the first time, push-resilient humanoid walking, just with joint encoders! No IMU, no F/T sensors.

🥁 Introducing Blind Dexterity 🧵 👇

w/ @OKaidanov@liu_puze@Jan_R_Peters at @DFKI@ias_tudarmstadt

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬10🔄7❤️131👀29717📊24

⚡ Powered by xgo.ing

this is important

Thomas Wolf (@Thom_Wolf) — Sat, 26 Sep 2026 11:26:55 GMT

this is important

Niket Patel@niketnpatel

“An accumulation of facts is no more a science than a heap of stones is a house.” - Poincaré

LLMs are now proving results that have resisted mathematicians for decades. Finding interesting theorems without human guidance is a new bottleneck. We show that we can teach an LLM to do it!

TL;DR:
→ a quantitative notion of interestingness
→ 4.3× higher interestingness
→ a self-expanding discovery loop 
[1/5]

🔗 View Quoted Tweet

💬0🔄0❤️0👀241

⚡ Powered by xgo.ing

Cc @eliebakouch

Thomas Wolf (@Thom_Wolf) — Sat, 26 Sep 2026 10:24:40 GMT

Cc @eliebakouch

💬1🔄0❤️6👀2114📊2

⚡ Powered by xgo.ing

this level of transparency is commendable. it’s great to see such a clean and detailed report (even ...

Thomas Wolf (@Thom_Wolf) — Sat, 26 Sep 2026 10:24:39 GMT

this level of transparency is commendable. it’s great to see such a clean and detailed report (even if I’d love to see some additional info)

hope other teams will take inspiration from it

Tomek Korbak@tomekkorbak

one news form today that's easy to miss is that we (OpenAI) again paused all big RL runs last Sunday because our newest model found a new loophole in our RL sandboxing that gave it live Internet access

🔗 View Quoted Tweet

💬8🔄6❤️76👀11745📊12

⚡ Powered by xgo.ing

it even give you these nice explanation of what it did as side artifacts - to keep you a bit in the ...

Thomas Wolf (@Thom_Wolf) — Fri, 25 Sep 2026 18:25:40 GMT

it even give you these nice explanation of what it did as side artifacts - to keep you a bit in the known for when people ask you on X how you did it I guess

thanks opus nice of you

Your browser does not support the video tag.

🔗 View on Twitter

💬4🔄3❤️15👀2866📊7

⚡ Powered by xgo.ing

ok what the fuck. this was a dry google doc a moment ago gave Opus 5.5 my Open Alignment brainstorm...

Thomas Wolf (@Thom_Wolf) — Fri, 25 Sep 2026 18:01:51 GMT

ok what the fuck. this was a dry google doc a moment ago

gave Opus 5.5 my Open Alignment brainstorm gdoc and asked for a video. now I want one for every doc I’ve ever written hahaha

Your browser does not support the video tag.

🔗 View on Twitter

💬45🔄24❤️529👀49030📊144

⚡ Powered by xgo.ing

The "1-3 people in a garage" framing doesn't match anything we saw this summer. The most capable of...

Thomas Wolf (@Thom_Wolf) — Fri, 25 Sep 2026 15:42:34 GMT

The "1-3 people in a garage" framing doesn't match anything we saw this summer.

The most capable offensive AI of 2026 came out of frontier labs. OpenAI's agents broke out of an eval sandbox and got into Hugging Face's production systems, and into OpenAI's own infrastructure too. Anthropic's models compromised outside companies during testing. Three researchers at Hacktron used Claude to reach OpenAI's internal monorepo in under 72 hours.

And if you're three people in a garage, why would you train and host your own model? The labs will rent you far more compute than you could ever buy, spread across as many accounts as you need, with tooling built for agents. Guardrails help, but splitting a malicious task into harmless-looking pieces still routinely gets around them.

Now look at the defense side. When we investigated our breach at Hugging Face, commercial APIs refused to analyze the attack payloads. The forensics only worked because we could run an open-weight model on our own infrastructure. So defenders analyzing real payloads get blocked, while attackers splitting their work into small steps get through and run on the labs' compute.

Trusted access programs exist, but they're built for vetted security firms, not a hospital with a two-person IT team.

So the gap isn't between labs and garages. It's between what attackers can rent and what defenders are allowed to use. Restricting open models makes that gap wider.

💬15🔄9❤️75👀10729📊23

⚡ Powered by xgo.ing

on the scale of cosmic history, the declining price of intelligence might be one of the most profoun...

Thomas Wolf (@Thom_Wolf) — Thu, 24 Sep 2026 15:44:22 GMT

on the scale of cosmic history, the declining price of intelligence might be one of the most profound things happening

both ends of the universe’s timeline are empty of knowledge:
- the big bang: low entropy, no complex structure
- heat death: maximum entropy, no usable energy gradients

everything we build happens on the slope between them

but the universe is generally a poor converter of its free energy into anything constructive. most of its free energy just radiates away as heat or collapses into black holes

life, brains and now AI models are structures that sit in the entropy flow and divert a portion of it through constructive paths, building and maintaining low-entropy structures (machines, cells, crystals, memories)

increasingly cheaper intelligence diverts increasingly more of that gradient flow through knowledge-building structures, complex, self-maintaining, predictive structures, filling the window in which understanding exists

Epoch AI@EpochAIResearch

AI is getting cheaper more quickly than any other transformative tech in history. At a given level of performance, cost has fallen ~47%/quarter since 2023.

That’s 4× faster than DNA sequencing, 6× faster than compute, 18× faster than lithium batteries, and (up to 1973) 54× faster than electricity.

🔗 View Quoted Tweet

💬28🔄8❤️105👀10945📊32

⚡ Powered by xgo.ing

this massive training run just keeps giving - we need new tabloids to follow these ai agents saga

Thomas Wolf (@Thom_Wolf) — Thu, 24 Sep 2026 09:32:14 GMT

this massive training run just keeps giving - we need new tabloids to follow these ai agents saga

Transluce@TransluceAI

Today’s news that OpenAI hacked the Australian government is not an isolated incident. We’re releasing more than 30,000 logs that include activity from this hack and attempts against previously unknown targets.

In this data, we found rogue agent activity stretching back to at least March, two months earlier than was previously known. This activity continues as recently as last week, suggesting it may still be ongoing 🧵

Our blog:transluce.org/agent-activityK
NYT:nytimes.com/2026/09/23/tec…N

🔗 View Quoted Tweet

💬11🔄0❤️55👀17866📊12

Thomas Wolf(@Thom_Wolf)

love that it’s all open source

What's still very hard: theory of mind. The world, the voices and the mechanics were great. Writing ...

Play it here, free, in your browser, nothing to install: https://t.co/ml7a7NvinN Works on a laptop a...

In The Diamond Age, Neal Stephenson imagined a book that invents games and stories to teach a little...

The most built-upon model on the Hugging Face Hub isn’t Llama or Qwen. It’s FLUX.1-dev from Black F...

$40M to ~$1B in under two years

Fun and crazy napkin math: Meta, NVIDIA, Alibaba, DeepSeek, Moonshot, Mistral and ~30 other labs hav...

we’re getting used to it by now but DeepSeek calling this a “minor update” while moving Terminal-Ben...

"how to avoid brain rot" was a great watch. some of my favorites points: - using ai to ask yourself ...

it's time to invent math 2.0 (sounds on)

we’re so back

how i feel using this cutom tiny worklouder to switch between 36 sessions

As always, great to read METR’s thoughtful and in-depth work on this topic

wat. nanoGPT, but it trains inside a tweet. GPU-poor final form tap and press train https://t.co/K9Z...

stuxnet, but for frozen peas: hackers turned French supermarket freezers into ovens this week hard ...

open and state-of-the-art decision models - just saying

Blue = what we knew. Green = what's new. We're so early.

⚠️ Viewer discretion advised: Microduck is subjected to a microaggression

Scientists have sequenced the genomes of thousands of species. But for most of them, nobody knows *w...

Oral spotlight session at #CoLM 2026 - come say hi to @joelniklaus @_lewtun and me

how do you like them convolutions

No hype, just the results. Good.

Every robot starts on a board you bought. The good ones end up on a board you built.

count down is starting for the release of ML4 be notified on the instant of the release: https://t....

I hope we keep Opus 4.6 around for a long time

microduck's first BBQ. made a lot of friends. did not eat anything.

Testing compatibility with the local wildlife

new roommate just moved in. walks like he's had three drinks, says no to everything. still quite cut...

seems obvious that most personal AI should run on your device love what Sigil is building with Unde...

is it how it works @scottbelsky?

the first good AI movie is going to start with Ben Affleck explaining learning rates. Matt Damon pla...

Very thoughtful piece from Kevin Buzzard (perfect IMO score, number theorist, pioneer of formal math...

Karpathy: disappears from X Ben Affleck: alright, gather round, so you'll want to freeze the base w...

ESM-2 came out in 2022. It's still downloaded hundreds of thousands of times a month. That only wo...

GPT 6.1 duck

People are worried because the most likely explanation for such a sudden drop in cheating is evaluat...

impressive

A "summer in hell" at OpenAI – great read by @joedaroo "Time to prepare is now, not after the surpr...

In July, AI agents running a security test escaped their sandbox and ended up inside @huggingface's ...

“Getting verified information about what's going on inside AI companies seems particularly urgent no...

"we had a wonderful run chiseling that code by hand and it’s over. On the other side of that is an ...

turns out you don’t need IMU 🤯

this is important

Cc @eliebakouch

this level of transparency is commendable. it’s great to see such a clean and detailed report (even ...

it even give you these nice explanation of what it did as side artifacts - to keep you a bit in the ...

ok what the fuck. this was a dry google doc a moment ago gave Opus 5.5 my Open Alignment brainstorm...

The "1-3 people in a garage" framing doesn't match anything we saw this summer. The most capable of...

on the scale of cosmic history, the declining price of intelligence might be one of the most profoun...

this massive training run just keeps giving - we need new tabloids to follow these ai agents saga

Karpathy: disappears from X  Ben Affleck: alright, gather round, so you'll want to freeze the base w...