Impressive jump on agentic coding according to its benchmarks! Now on par with Claude Opus 4.1 (from 5 months ago!), K2 Thinking, and GPT-5.2 Codex, at a lower cost. A bit overshadowed by DeepSeek, whose DSA mechanisms achieve great cost cuts.

There are few benchmarks yet for @OpenAI’s fresh GPT 5.2 Codex model. Initial benchmarks from the announcement imply a drop below Gemini 3 Flash in agentic coding. In fact, the performance seems close to DeepSeek V3.2 at a 50x price jump.

Gemini 3 Flash from @GoogleDeepMind is a big step up in general knowledge, across a wide price range (modulated by the thinking time. At the highest, it surpasses the recently-released GPT-5.2 in both cost and quality.

GPT-5.2 from @OpenAI has a marginal but welcome improvement in reasoning, securing its place at the very top.

With Devstral 2, Mistral pushes the envelope on two fronts: ① what an agentic coding model can do, with no reasoning! There’s a quality gap to reasoning models, expectedly. The positive: it is cheaper; potentially even cheaper in practice than indicated in this chart.

The path of the Mistral 7B is nice to see! The OG one topped open models of that size. For the first time, a local model felt usable on consumer hardware. Not only is the latest Ministral 8B on the Pareto frontier for knowledge vs. cost (and for search, math, agentic uses)…

DeepSeek released V3.2 (and V3.2 Speciale, a math-oriented model). New model, new benchmarks! The biggest jump for DeepSeek V3.2 is on agentic coding, where it seems poised to erase a lot of models on the Pareto frontier, including Sonnet 4.5, Minimax M2, and K2 Thinking.

So, how is Gemini 3 on this new leaderboard? Its intrinsic knowledge is unmatched, surpassing 2.5 and GPT-5.1. bsky.app/profile/espa...

Unveiling a new LLM leaderboard: metabench.organisons.com Why? Company C1 releases model M1 and discloses benchmarks B1. Company C2 releases M2, showing off benchmarks B2 which are distinct. Comparing those models is hard since they don't share benchmarks!

Am I using the Gemini APIs wrong? I keep getting 429's. The key was fresh from aistudio.google.com. gemini-embedding-exp-03-07 is the only embedding model in the market that I can’t benchmark because of it. The quota in the Console says I'm at 0.33% usage…

Our latest open-source speech-to-text model just claimed 1st place among streaming models and 5th place overall on the OpenASR leaderboard 🥇🎙️ While all other models need the whole audio, ours delivers top-tier accuracy on streaming content. Open, fast, and ready for production!

Isn’t there a better way to handle screens than asking a *language model* to guess the number of pixels to the left and top of a UI widget?

Talk to unmute.sh 🔊, the most modular voice AI around. Empower any text LLM with voice, instantly, by wrapping it with our new speech-to-text and text-to-speech. Any personality, any voice. Interruptible, smart turn-taking. We’ll open-source everything within the next few weeks.

Search > Recommendation. I find more interesting, high-signal things from querying what I like, than linearly going through a feed that learnt from my navigation. Generally, giving users the ability to send reliable signals beats extracting signals from their background noise.

It is critical for scientific integrity that we trust our measure of progress. The @lmarena.bsky.social has become the go-to evaluation for AI progress. Our release today demonstrates the difficulty in maintaining fair evaluations on the Arena, despite best intentions.

I wonder what the story was for Phi-4 Mini. Its tokenizer for conversation is completely different from Phi-4.

New paper: Simulating Time With Square-Root Space people.csail.mit.edu/rrw/time-vs-... It's still hard for me to believe it myself, but I seem to have shown that TIME[t] is contained in SPACE[sqrt{t log t}]. To appear in STOC. Comments are very welcome!

Censorship is when the government silences speech. With Mr Musk being in government, doesn’t that make every X suspension or shadow ban, censorship?

Preventing political opponents from joining elections, by removing their diploma and putting them in prison with unjustified charges, is not democratic. Is there a shred of reason behind Ekrem Immamoglu's jailing? apnews.com/article/turk...

We've kept pushing our Open-R1 project, an open initiative to replicate and extend the techniques behind DeepSeek-R1 And even we were mind-blown by the results we got with this latest model we're releasing: ⚡️OlympicCoder [1/3]

Is there an economic reason for which the tariffs established during Mr Trump’s first term didn’t cause a recession, but those established now did?

We can get GNSS spacial positioning all the way to the moon, given the right receiver! Greatly simplifies space travel. I still believe we should set up a separate GNSS on every planet. ntrs.nasa.gov/api/citation...

Italy will reintroduce nuclear energy through SMRs and fusion research. Decarbonization fights against an existential risk. I approve! www.mase.gov.it/comunicati/n...

LLMs get better at tool use and search. Model memorization is thus less useful than reasoning. Yet a lot of benchmarks still focus on the former.

It is a bit sad that codec programs gave up on using GPGPU / CUDA, which is much more widespread than hardware acceleration.

Mistral Chat Pro being so fast to generate messages is really nice. I would love to see how it feels if they release a reasoning model.

Surprisingly, bigger Llama 3 models are worse at learning from relevant context, and giving a good answer, than smaller ones. Unsurprisingly, base models evaluate the probability of a good answer better than instruct models, which will give a low probability to speech that doesn't match their style

We just released Hibiki 🟢, a real time speech-to-speech translation 🇫🇷 -> 🇬🇧. It preserves the voice of the user, and the smaller variant can run on iPhone as showed by Neil in this video. Find the code on github github.com/kyutai-labs/... and the weights on HF and give it a spin!

Bittersweet to see the latest Codestral so close to the open-weights version, yet to see both are so close to Claude.

US presidential actions related to energy: • Eliminate EV mandate • Terminate the Green New Deal • Stop funding EV charging stations • Eliminate taxes on fuel and gas-powered vehicles Doesn’t that negatively impact Tesla?

Why does the default PyTorch DataLoader collation prevent having multiple dataset rows with subfields being lists of varying length?

The year is 2027 and after exhausting its appeals Google loses and has to sell Chrome to a DOJ-approved owner. As defined in the suit, "Chrome" includes all backend systems essential to the browser. This means Sync. Elon Musk buys Chrome. He now has the full browser history of 3 billion people.

Mistral AI has added features to make it the best chat interface (file interface like Claude Artefacts, Web search like Gemini, image generation like ChatGPT), and it is now getting hammered by users worldwide, 500 errors galore. mistral.ai/news/mistral...

The issue with this kind of login form: when I come back, I have no idea which one I picked, and I won’t try them all until I find something I made before.

How dare mosquitoes bite the hand that feeds them

I want to make Firefox’ UI very minimal, only summoned through Ctrl+L. Here is my solution so far. (Requires adding the CSS file in a chrome/ folder in the root dir in about:profiles, then about:config > toolkit.legacyUserProfileCustomizations.stylesheets > true.)

Paying online subscriptions should always use virtual cards. When your physical card gets stolen, all your online subscriptions fail at once.

What are your tips when you lose a Yubikey, for removing it from all websites where it was 2FA, and adding a new Yubikey?

“Large” language models is a moving target. The first GPT had a hundred million parameters, and called itself large already. GPT-4 has almost two trillion.

Wordle optimal play, finally on print! I explore a bit in depth how I achieve it iteratively, relying on reinforcement learning principles, using a new tree search algorithm which is generic enough to be applicable to other games. espadrine.github.io/blog/posts/w...

Every technology lives and dies on whether it is more useful than what it chose as its competition.

The Mistral Chat, chat.mistral.ai, is quite good! Its contents will be used for training though. The terms mention a paid version will allow opting out. Paid customers may wrongly assume they’re opted-out by default, though. I’d rather use thumbs-up/down to opt in.

If you squint, decoder-only transformers are RNN: its state is the KV cache, and each token generation only uses the state and the token input, to yield the predicted output.

I’ve been into generic optimizers lately. Solving Wordle perfectly is a pretty good testbed in my opinion for some of the required algorithms! New article, the first of two, towards achieving perfect play with it: espadrine.github.io/blog/posts/f...

The Firefox Primary Password has a weird security model: it prevents other users of a shared device to read your passwords, but not to authenticates into your banking site, since cookies are in the clear. support.mozilla.org/en-US/kb/use...

Subtle login UI design: if the password is wrong, don’t disable the password input field while your retry timeout runs; disable the button instead and indicate the timeout on the side. At least the issue is logged with KDE: bugs.kde.org/show_bug.cgi...

Is Kullback-Leibler divergence a better categorical loss than cross-entropy sometimes? I am reading arxiv.org/pdf/2007.12509 which used the former.

Textual inversion is finicky but magical when it works. I tokenized my drawing style off eight examples in about 15 minutes of training time.

Reddit’s images URL somehow serve either a JPEG (if they detect that it is loaded from within a Reddit page), or HTML (if they detect that it is loaded from the browser). They failed to realize that there is a good reason to want the JPEG directly: their app UI is so crowded that the image is tiny.

Amazon video selects among users a group of reviewers for test screenings, if I understand correctly the email I got.