Alexander Doria

And new Pleias paper: for multilingual generation, who should judge the judge? LLMs are commonly used for assessing language quality but with lower resource language that becomes much less obvious: do you need a frontier model? is it even enough?

We then ask then potential judges to spot the original and translationese versions. Performance generally maps with model size (Fable) or documented multilingual performance (Gemini) with some surprises: Jev turned out unexpectedly strong.

And new Pleias paper: for multilingual generation, who should judge the judge? LLMs are commonly used for assessing language quality but with lower resource language that becomes much less obvious: do you need a frontier model? is it even enough?

Went through it but zero mention of an obvious contender: data shape. Latest models like Qwen-3 or Kimi are trained on vast amounts of "clean" synthetic data/environments and this impacts superposition/features.

Did not include it in the SYNTH paper but we did some jspace analysis of Baguettotron and the commit zone topology is radically different from Gemma-270m (+very high level of connection between features). Looks to be more an effect of synth pretraining than deep arch.

And new Pleias paper: in collaboration with AI Bridges and @wikimediade.bsky.social we developed a new harness for complex semantic search over the largest knowledge database, @wikidatacommunity.bsky.social arxiv.org/pdf/2610.06650

While our initial focus was on the harness, we see a wider potential for making smaller effective models for search (is it cognitive core?): recursively exploring graphs and texts is largely a skill in itself that can be exercised and learned with much less parameters

And new Pleias paper: in collaboration with AI Bridges and @wikimediade.bsky.social we developed a new harness for complex semantic search over the largest knowledge database, @wikidatacommunity.bsky.social arxiv.org/pdf/2610.06650

Even though it’s the original web of agents, semantic web remains largely untapped. Our research aimed to leverage an apparent convergence between search depth in knowledge graph and recursive depth in Recursive Language Models

After a long wait, releasing the SYNTH paper! It’s not pretraining, mid-training or post-training, it’s just training: a fully synthetic single-stage pipeline to train workable reasoning models with unprecedented data efficiency. arxiv.org/abs/2609.378...

SYNTH paper is also a contribution to pretraining science. Thanks to controlled environments, we discovered that epistemic calibration (does the model knows it doesn’t know?) is an emerging capability from 300M parameters onwards.

After a long wait, releasing the SYNTH paper! It’s not pretraining, mid-training or post-training, it’s just training: a fully synthetic single-stage pipeline to train workable reasoning models with unprecedented data efficiency. arxiv.org/abs/2609.378...

We show the recipe continues to scale, by releasing two additional baguette models (including the 600m that is currently powering Paris subway).

Something I hardly see addressed in the goncourt/ai discourse: it's still pretty hard to generate a workable novel (let alone good), especially considering he seems to have used a previous generation model (gpt-4o style + natural time lag before publishing).

current read, immediately joining my list of retroactive llm literature.

thanks to synthetic environments we can run matrix in reverse. www.anthropic.com/research/ali...

won’t age well

So confirms Astra is current SOTA on private multimodal tasks (segmentation/hard manuscript), and that’s not close.

Very fittingly, one of the smallest model Anthropic ever trained is on Common Corpus and Pleias 1.2B tokenizer: 2.9M model artificially expanded to 331M to study weights interference for the new Transformer Circuits. transformer-circuits.pub/2026/interfe...

Probably the best model report this year: multiple infrastructure pieces cohering to fix computation and network to hold 3T parameters.

And new technical blogpost by Pleias application team on deploying small reasoning models for edge devices : featuring cache context management on Rasperry, designing system orchestration under constraints (reranker, chunking) and model specialization. pleias.ai/blog/local-a...

We describe several practical use cases on the field, including a legal assistant for conflict-related sexual violence (CRSV) survivor networks, powered by a retrained version of Baguettotron (600M variant) on a synthetic environment in humanitarian law.

Alors deux minutes d’explications : 1. Les modèles à poids ouverts sont protégés par la norme safetensors : par définition ils n’embarquent pas de code. Jamais vu un cas de backdoor.