New amazing interpretability/SAE work on Baguettotron! Almost surprised how much the high entropy section are actually connected to analytical features: lyramakesmusic.github.io/bread-slicer/

Also to further clarify our current position: open source > sovereign ai.

First blogpost of my cofounder post-Delhi on the deafening silence of public/political sphere as we come closer to another industrial revolution. anastasiastasenko.substack.com/p/the-countr...

Hmm. Main French scientific institution now forbidding to use any "ai generative tool" save from their own variant of Mistral, Emmy. Includes even open source

Oh amassing large enough datasets with provenance for language model training is totally doable. Just when you do that you feel lonely (and unpaid) as people don’t really care.

Announcing the release of new Common Corpus, and the first really global update. For the Indian AI Impact Summit we added 270B tokens of permissibly license collections, more than half of content from non-Western countries. pleias.fr/blog/blogcom...

Pleias on an open source AI panel with Jimmy Wales (Wikimedia) and OpenUK tomorow.

+1. Existing agentic models could be considerably better at art/illustration design (not just creating an image in one batch, but iteratively enhancing it, adding details, etc.) if there was the incentives for it. But programming/law/management are just much more lucrative.

If you're ever in Delhi next week: Pleias has been invited to the AI Impact Summit and my co-founder will give several conferences on SLMs and building AI infrastructure as a commons.

As a result, did the long delayed/much needed update of personal web page: vintagedata.org

I’m not sure if it’s limited to the EU but sex censorship of the web is getting absolutely ridiculous. 1950s level.

i guess i’m really an ai researcher now

Meanwhile, further independent evaluation confirms SYNTH performs better than Nanochat/Fineweb across nearly all benchmarks. (Humaneval to be expected: we did not out any code inside yet)

It took me weeks, but finally it's there: an overlong blogpost on synthetic pretraining. vintagedata.org/blog/posts/s...

maybe it's awfully european of me, but i'm not convinced dumping a mass of non-contextualized private documents is ever a good thing

seems we’re dangerously close to a point where someone will tweet you can slurp three gas town juices in a moltbot - and kill a few ai bubbles.

Common Corpus paper is going to @iclr-conf.bsky.social !

If I see another headline about Yann le Cun "contrarian" bet, I’m killing a non-verbal model.

Now starting. From what I see generally compared to Joyce and Kafka but for now the obvious comparison point is Robert Musil (did Musil read it?)

If you ever wondered if SYNTH could be usable in mid-training for larger models: Step-DeepResearch (from StepFun) is now out as private beta. stepfun.ai/deep-researc...

It’s a bit soon for follow Friday, but still recommending to follow this cat.

Sans surprise. C'est probablement le secteur le plus exposé à l'IA agentique : plein d'entreprises sont en train de découvrir qu'elles peuvent ré-internaliser la génération de code et que la dette technique n'est pas pire que du SaaS mal contrôlé.

Baguettotron is now taught in the classroom.

maybe model collapse is back after all: claude starting to cite grokipedia (obviously flat wrong…)

Meanwhile… I’m afraid Greenland will be the least of our concerns.

sometimes you look at the data and regret it.

Really like this retrospective of how the original Qwen series came to be: "could you also make a 3-4b for us".

Following on our partnership with Wikimedia Enterprise, very happy to see Pleias featured in Wikipedia 25th anniversary post. wikimediafoundation.org/news/2026/01...

and so, back to train something new

The real shift I'm seeing with Claude Code & co: "programming" is now mostly about knowledge infrastructure management.

I feel now structuralism is now getting comforted a bit of everywhere: obviously linguistic/culture but also even archeology/comparative mythology. Except almost no one is left to claim it.

not your usual pull request rejection.

I'm afraid many left/liberals remind me of that classic Lubitsch dialogue (from Cluny Brown)

Just finished it. It’s super weird it’s probably the only novel about digital humanities, as a social environment (working with a computer in a lit/ssh lab: should I learn program?) and a recursive theme (text is constantly self-conscious of lexical patterns). And it predates the word for decades.

So we already have the first major paper of 2026, DeepSeek mHC: Manifold-Constrained Hyper-Connections. arxiv.org/pdf/2512.24880 Short thread: this is actually an engineering paper, taking as a starting points ideas already exposed in an original Hyper-Connections (HC) paper from ByteDance.

So happy new year and happy public domain day! Lots of entries in 2026: Thomas Mann, Paul Claudel Albert Einstein and slowly getting in the heart of classic Hollywood talkies with movies from 1930 (Blue Angel, Animal Crackers, original design of Pluto…).

So 2025 wrapped. *Trained SOTA small models. *Published open datasets reused by Nvidia, Anthropic, Stepfun, Jina… and maybe more importantly, many independent researchers. *Proved full synthetic training is viable in small size range (I have little doubts it will scale). *Shipped a few blogposts

Glad someone is finally calling it out. Always been disappointed in many AI ethic works, not because I disagree with the premises but because it’s just bad social science.

Fortunately pipelines are ready for it, but MoE fine tuning looks much more data hungry.

lol

DOI another day.

So this Christmas we have opera at home.

Just launched the Wikidata dumps parsing. No more refactors, time for big plans.

Maybe it’s me but being a mess of different historical periods feels pretty on brand for a Homeric adaptation.

Funnily enough, my partner just made the same point to a semi-well-known French researcher complaining about widespread training on copyrighted data. And this was the answer.

AI discourse on the good site. AI discourse on the bad site.

French investigative media Mediapart rediscovering water apparently: been known for +1 year, when Meta entered discovery and since then Meta won the case.

I recently started playing Chopin Sonata 3 again and there was a puzzling mystery: my favorite recording (Arrau) diverges significantly from the sheet music. Most obvious change is this passage, completely rewritten to sound much more disjointed/modernist.

My take on this: science is already "automated" to a large extent and this goes back to early digitization if not prior. Even our current concept of peer review only came to be widespread with review automation (was one of the core tech of Elsevier in the 1970s).

Pour ce prix-là on entraîne plusieurs LLMs à l’état de l’art. GPT-OSS c’était trois millions, DeepSeekv3 10, Sonnet-Opus quelques dizaines.