CSO/Co-founder @maisaAI_. Head Contrib/ Ambassador🤗 @huggingface. Research 🌸@bigsciencew/@BigCodeProject | ex @narrativaAI
Murcia - SF
Joined January 2011
- Tweets45.8K
- Following2.6K
- Followers20.2K
- Likes50.3K
Manu Romero retweeted
Santander is scaling AI process automation with Maisa.
We’re working alongside Santander’s teams to deploy AI for more complex processes across its main markets in Europe and the Americas, combining reasoning, execution and traceability with specialized models that can run on the bank’s own infrastructure.
Our collaboration with Santander reflects our commitment to bringing AI into production at scale in one of the most demanding industries, with governance, reliability, and cost efficiency as core priorities.
Read the full story on the @bancosantander website:
santander.com/en/stories/san…
Do you remember encoder-decoder archs such us BART, T5, etc
You already gave the LLM the answer. So why is it writing it out from scratch?
Introducing a new AI model architecture: Constrained Decoding Models (CDMs)
LLMs were built to write. Most production workloads just need an answer: fast, consistent, and affordable.
CDMs are purpose-built for 1 task each. They read your input once and skip the token-by-token generative loop that makes LLMs slow and expensive.
Not just a decision model, but a family of 3 CDMs live today:
• Classify: define your labels and get back a calibrated probability + reason for each. Used for routing, tagging, and intent detection.
• Extract: describe the entities in natural language and get back well-formed JSON with confidence scores + citations. Used for web scraping, document processing, and PII redaction.
• Summarize: pass in long text, images, or PDFs and get back a consistent summary under 1,000 tokens. Used for context compaction, metadata generation, and transcript digests.
Built for production:
• 1M tokens of multimodal (text + image) input, 31x larger than Jev
• 77.2% on DecisionBench vs Jev's 72%
• 200x faster, 467x cheaper than frontier LLM APIs
• $0.05/M input tokens. Output is free (yes, even for summarize)
Live on your favorite platforms today:
@OpenRouter @vercel_dev @LiteLLM @LangChain @TrustedRouter @itsArthurAI @useRouterPlus
Get 50M input tokens for free (for the first 50 hours) at link below đź‘€
🧠New model: Typed Decision, built on EmbeddingGemma 2
Give it a state + any schema of questions → typed, calibrated answers (choice · score · noul). No text generation, no output parsing.
76.4% on typed-decisions, ~6 pts behind a 27B LLM with ~100Ă— fewer params.
🤗 huggingface.co/mrm8488/embed…
Manu Romero retweeted
The intuition: models trained on different data seem to converge to a similar geometry (the Platonic Representation Hypothesis).
Distances between concepts like food, people and animals look alike in both spaces. So is this shared geometry enough to align them?
Keep plain SFT, but first transform each expert trace into a nearby trace that the model itself would plausibly write, while making sure it still reaches the expert's answer.
The best thing about Jev is that it has made classification models cool again.
It felt like if you weren’t building LLMs, you had nothing to say in the AI world.
Manu Romero retweeted
🤔 What if LLMs learn to reason not by copying perfect solutions, but by learning what NOT to do? ❌
đź’ˇIntroducing Negative Self-Distillation (NSD) -- a label-free method that trains LLMs to explicitly avoid flaws!
🧵[0/n]
See you there!
We're hosting a Career Day in Madrid to hire AI Automation Consultants for team Maisa.
Get an opportunity to meet every decision-maker and go through every interview round in 1 day!
15th October in Madrid 🇪🇸
Get all the details and apply here: maisa.ai/aac-hiring-day
Manu Romero retweeted
"Memory Decoder at Scale: A Pretrained, Parametric Long-Term Memory"
Most LMs still store knowledge and reasoning in the same weights, so making the model remember more usually means scaling the whole backbone.
This paper makes long-term memory a separate pretrained module. It trains a parametric memory to imitate kNN retriever distributions offline, then plugs it into a frozen decoder through next-token probability interpolation.
They found that memory scaling beats backbone scaling, with Pythia-410M plus a 6.9B memory beats Pythia-12B across 17 tasks with 39% fewer total parameters.
The cool part is that memories can be swapped by domain. Biology, law, and finance memories improve Qwen3 models by over 9 points on average while the base model stays frozen.
Read more here:
alphaxiv.org/abs/2607.27919
Manu Romero retweeted
Most motion papers tailor one controller to one specific task. This year at SIGGRAPH, our research team asks: can motor control itself be pretrained and reused?
Generative Pretrained Controllers, or GPC, turn motor skills into a vocabulary of discrete tokens and train a transformer-based generative controller through next-token prediction. Just like GPT, the same pretrained controller can then be fine-tuned to solve new tasks.
Trained on 600+ hours of motion, GPC runs in real-time inside a physics simulation, producing natural and physically grounded behaviors for interactive control.
Many things to come!!!
We’re sharing the next major milestone in our non-invasive brain-to-text decoder research: Brain2Qwerty v2.
Building on v1, which was published today in @Nature, Brain2Qwerty v2 is the highest-performing end-to-end pipeline capable of real-time sentence decoding from raw brain signals. It advances beyond character-level performance to decoding words and semantics, enabling accuracy for overall communication.
We believe this research has the potential to make a real difference for the millions of people who suffer from brain lesions or disorders that prevent them from communicating.
🧵👇