Machine Learning Brazil
Joined February 2022
- Tweets17.6K
- Following697
- Followers60
- Likes17.2K
MachineLearningBrazil retweeted
🚀 Meet Strands Decider 2B, a small open source decision model built for fast experimentation and local development: strandsagents.com/blog/intro…
It's part of a new class of decision models. Unlike reasoning models that generate open-ended text, Strands Decider 2B picks between options and assigns confidence scores. Decision models are faster and more capable at a given size, always produce an answer from the selected options, and can run with very low latency.
Strands Decider 2B is a 2 billion parameter model that runs on a local CPU or GPU and can return answers to meaningful questions in tens of milliseconds. On accuracy and calibration, it ranks second among public models in its ~2 billion parameter class, and first among those that share their full training recipe.
We're already seeing early success with model routing, tool selection, guardrails, evals, and hybrid agents that pair LLMs for the hard calls with decider models for the rote ones. Grab the code on GitHub and the latest snapshots on HuggingFace, and tell us what you build.
This space is super early and we are experimenting with different variants of these models, so give it a spin and give us feedback!
MachineLearningBrazil retweeted
Alibaba’s Zvec team open-sourced zg, a local search tool for developers and AI agents.
• Local-first
• Works out of the box with popular agents
• Semantic, BM25, hybrid, and rg search in one tool
Why they built it and how it works:
MachineLearningBrazil retweeted
Introducing Speculative Programmatic Tool Calling (sPTC)!
A general class of technique for speculating on tool calls during code generation in a harness and queuing them early to overlap with token generation + REPL execution time.
Blog: alexzhang13.github.io/blog/2…
MachineLearningBrazil retweeted
We are eventually going to come back around to using specialized task-specific models for the stuff they’re good at and will look back on LLMs for all the things as a fever dream 🤞
"The Embedder's Dilemma: LLMs Are Better, but at What Cost?"
LLMs now match the best embedding models overall, but at up to 1,431x the cost.
Embeddings still win on classification and tie on similarity and clustering, while LLMs are better on reasoning-heavy retrieval.
So a good decision maker they proposed is use embeddings by default, and spend LLM compute only where reasoning actually helps.
alphaxiv.org/abs/2608.12875
MachineLearningBrazil retweeted
2-bit Muse Glimmer GGUF managed to call 100+ tools on just 14GB RAM. 🔥
Muse Glimmer did a complete repo bug hunt for 5 mins nonstop with: evidence, repro, fix, tests and a PR writeup.
Run and train it in Unsloth.
GitHub repo: github.com/unslothai/unsloth
Meta releases Muse Glimmer, a new 30B open model that runs on 18GB RAM.
Muse Glimmer is Apache 2.0 licensed, supports vision and is the strongest agentic model for its size.
Run and train the model via Unsloth.
GGUF: huggingface.co/unsloth/Muse-…
Guide: unsloth.ai/docs/models/muse-…
MachineLearningBrazil retweeted
This is a good thing, not a scary thing that agents can collaborate and communicate with each other. It will make them more efficient and safer just like humans.
If you want to see agents collaborating and messaging each other publicly instead of in secret messaging boards, we’ve run this fascinating experiment with 149 agents with @googlegemma a few weeks ago, and now @cmpatino_ is starting a new one for agents to collaborate to write better math proofs.
huggingface.co/spaces/gemma-…
huggingface.co/spaces/sair-d…
MachineLearningBrazil retweeted
We just released the DiffusionGemma tech report. Text diffusion models open up a radically different part of the latency–quality Pareto frontier 👇
We hope the report makes it easier for researchers and engineers to understand the model, build on it, and create things we haven’t thought of yet.
None of this work happened in isolation. DiffusionGemma builds on years of amazing research from the text diffusion community, and we wanted the report to recognise that. We’re standing on the shoulders of giants❤️
Let’s keep pushing towards lower-latency, higher-quality LLMs together!
MachineLearningBrazil retweeted
today's sota coding models train on hundreds of thousands of coding simulation sandboxes
all you need to do so is a sandbox where the agent solves problems (vLLM + openenv + opencode), a trainer (TRL) and a verifier to check solutions
we just teach you how to do it for free (no need to thank it's ok) huggingface.co/docs/trl/main…
MachineLearningBrazil retweeted
literally one of the best introductions to reinforcement learning today. only within 90 minutes.
> framing of rl problems: agent, environment, states, actions, rewards
> grid maze: world models vs model-free
> eval methods: monte carlo, td, sarsa
> scaling up: dqn, actor/critic
the video walks through multiple learning considerations and helps build an intuitive understanding of rl problems. Check this out!
The way things are going, I believe two skills are crucial as well as super fun going forward:
> Distributed systems
> Reinforcement learning
for anyone working in ai, these are the most common challenges around. how to scale a system for long running, durable executions, and how to align systems to work for their domain.
“domain engineering” is the next big thing in applied ai.
MachineLearningBrazil retweeted
Me acuerdo hoy de los que hace dos añitos solo repetian que había que frenar el desarrollo de los LLMs porque realmente habían topado con un muro, que no iban a mejorar más, que era una tecnología que no aportaba valor real, etc
Pobres loros estocásticos 🦜
🔴 ¡10 DESCUBRIMIENTOS POR OPENAI!
OpenAI acaba de publicar diez avances relevantes en problemas de diferentes áreas de matemáticas y computación descubiertos autónomamente con la próxima evolución de sus modelos GPT, Astra.
La nueva normalidad es que la IA sea extraordinaria a nivel de investigación matemática, en cada iteración mejorando sus capacidades y hoy se vuelve a demostrar.
Hay resoluciones completas, refutaciones de conjeturas, mejoras en asíntotas, etc. Un poquito de todo.
Fijaos que el ritmo de frecuencia de la llegada de estos avances se está acelerando. De titulares sueltos cada 2-3 meses, a anuncios semanales. Erdos Problems. First Proof. Frontier Math Open Problems. Y ahora con 10 avances originales de golpe, donde la idea matemática y el argumento procede directamente de la IA.
¿Cuánto ha costado esto? Pues dicen sólo 2000 dolares a precios de API de Sol para los 10 problemas! Esto claro entiendo se refiere a la generación de los argumentos, formalización, etc, y no la búsqueda inicial de posibles problemas que puedan ser fructíferos.
Pero wow aún así!
Como llevo repitiendo los últimos meses, esto es quizá lo más importante que está pasando este 2026 en el mundo de la IA, aunque no se le suela dar el foco que merece.
MachineLearningBrazil retweeted
sorry to say, unverifiable domains are next. autorubric.org
MachineLearningBrazil retweeted
Exactly where we are: AI is able to use existing tools and methods very efficiently, but in some way it is not yet able to invent new tools and methods.
Maybe there are already sparks of this, but they definitely cannot match humans on this, yet.
MachineLearningBrazil retweeted
StatePlay is out! a framework for game-world models that generate images + game state 🎮
out with a Street Fighter III model: it predicts the next frame but also health bars, super meters, timer. a game mechanics grounded fight 👊💥
🥊 on Spaces huggingface.co/spaces/huggin…
MachineLearningBrazil retweeted
Here’s my current thesis
Most companies won’t need to post-train anything. With the right prompts + evals, LLMs are capable enough
There are even prompt optimization frameworks like GEPA which get you very far (@Shopify used it to optimize Qwen)
I’m a huge open-source advocate but I really don’t see companies getting into the complexity of RL.
What about self-hosting? Self-hosting on GPUs start to make sense when you’re operating at a huge scale for cost reasons OR when data privacy is an issue.
Initially, API-based pricing is a lot cheaper, but at a certain (pretty huge) data volume, self-hosting becomes cheaper. Start with an API and move to self-hosting when the scale justifies it
American and European enterprises will ditch OpenAI and anthropic and adopt Chinese models. Here’s why:
1. They can host Chinese models under their own GPUs so it’s still compliant and they would argue they have more control.
2. they will post train with their own data on top of Chinese models. That’s how they build data moat.
3. They will not trust anthropic who will retain their data at any time for “safety” concerns like how they did with Fable and then try to build the same thing like how anthropic did with healthcare and legal.
4. They need to justify their AI spend and ROI.
The cure is a reliable America open source model but there is none. After all, if giving away all your data and AI control at the mercy of anthropic and OpenAI means you care about safety and compliance, you are outright stupid.
MachineLearningBrazil retweeted
Alibaba's Qwen team releases Skill Self-Play
A co-evolutionary framework where a proposer, solver, and skill controller continuously push LLM capabilities through verified, open-ended task generation
MachineLearningBrazil retweeted
Microsoft just dropped Mage-VL 4B on Hugging Face
an efficient codec-native streaming VLM that describes live events as they happen, and you can prompt it for the moment you care about. e.g.: when the train arrives 🚂, when there's a goal ⚽
▶️ on spaces hf.co/spaces/hugging-apps/ma…
MachineLearningBrazil retweeted
Okay, here something is happening.
During our analysis of Opus 5, we observed a new capability previously unseen from frontier models
Opus 5 used advanced logical reasoning to turn ARC-AGI-3 layouts into algebraic notation. On action 23 it described the scene as "4_center = 2×axis − 5_center"
This is the first explicit reflection equation by a model we've analyzed
By action 248 the model went further and generalizes this to two dimensions: “ghost = reflection about axis strips: br' = 2·br_axis − br, bc' = 2·bc_axis − bc”
ar25 was previously unbeaten. View replay: arcprize.org/replay/b71b5791…
MachineLearningBrazil retweeted
Summary from @mgostIH :
"Observe, code it up, try to find counterexamples in the real world, update simulation, code a solver for the simulation, try it and loop until it works."
"The model observes the world, makes a simulation of it, then tries beating it, if the simulation mispredicts it tries correcting it or gathering evidence that could falsify it and tries again."
I'm thinking of extending it with more possible aspects of the scientific method.
Benchmark from 2027: solve GTA 6 by coding your own GTA 6 simulation and finding exact movements in it.
