We’re happy to announce 2 releases today: - 🧠Brain2qwerty v1 is published at @NatureNeuro - 🚀 Brain2Qwerty v2 is now publicly released Explore how we decode sentences from non-invasive brain recordings: facebookresearch.github.io/b… Links: 📄v1 Nature Neuro: nature.com/articles/s41593-0… 📑v2 Meta preprint: facebookresearch.github.io/b… 💻Code: github.com/facebookresearch/… 📊Data: huggingface.co/datasets/bcbl… 📝Blog: ai.meta.com/blog/brain2qwert… 🧵Thread: x.com/JarodLevy/status/20715…

I’m in Seoul 🇰🇷 this week for #ICML2026 If you’re interested in our Brain2Qwerty work or brain decoding more broadly, I’d love to chat! Feel free to reach out. I’ll also be at the @AIatMeta booth presenting a demo of TribeV2 from our team. 🧠 See you there!

This is exciting and all, but the most interesting part for me is this: Auto Research, powered by @cursor_ai agent. The AI agent independently wrote code, ran experiments, analyzed the results, and improved the model's Word Error Rate by up to 19.8%, vastly outperforming traditional hyperparameter search algorithms (Optuna). And that's because agents weren't just tweaking hyperparameters. They autonomously discovered and coded ML techniques to make the brain-decoder better. Multiple agents independently invented strategies like "modality dropout" (forcing the AI to rely more on brain signals rather than its own language predictions) and good old beam search decoding. Vibe-science era, what can i say.

🚀 Brain2Qwerty v1 is in Nature Neuroscience, and today we're releasing v2 😎! 🧠⌨️ Same goal, but v2 slashes the best subject Word Error Rate from 52% to 23%, avg 39%—significantly narrowing the performance gap with invasive BCIs. 📄 Meta AI Blog: ai.meta.com/blog/brain2qwert… 👇

Claims highest resolution of human brain, 3d snippet shows a macaque brain 🧐

🧠➜🖼️ Boost brain-to-image decoding (up to 68% 🚀!) by adding synthetic brain data (#TribeV2) in the training set. Our latest paper is out: - 📝 Paper: arxiv.org/abs/2606.06345 - 🧵 Details: 👇

We have a new postdoc position to work on Neuro AI and human intracranial recording with Pierre Bourdillon and I: Apply here: docs.google.com/forms/d/e/1F… Lab info: bourdillon-titan-lab.fr/ feel free to RT ;)

We did it! Thrilled to announce that with my team at FAIR Meta we released 25 auto-formalized mathematics textbooks covering analysis, algebra, geometry, topology, combinatorics, probability, statistics, PDEs, number theory, and theoretical computer science - the largest such effort to date.

Backprop misaligns with the brain: checkout our latest study👇:

Announcing 💃 DANCE 🕺to Detect and Classify Events in EEG, by the one and only @JarodLevy

Announcing: a new interactive tool for a quick and simple start of encoding or encoding: 🧠 fMRI, EEG, MEG, iEEG, spikes… preprocessing 💬 text 🔊 audio ▶️ video 🏞️ image… embeddings 📦 pip install neuralset 🔍facebookresearch.github.io/n… #NeuroAI #OpenSource

Meta AI just released NeuralBench — a unified, open-source framework to benchmark NeuroAI models. The Problem: EEG foundation models are being evaluated on inconsistent pipelines, narrow task sets, and incompatible metrics — making cross-model comparisons nearly meaningless. The Solution: NeuralBench standardizes everything under one interface. NeuralBench-EEG v1.0 at a glance: 📊 36 downstream tasks across 8 categories (cognitive decoding, BCI, clinical, evoked responses, and more) 🗂️ 94 datasets · 9,478 subjects · 13,603 hours of EEG data 🤖 14 architectures: 8 task-specific models 6 foundation models (REVE, LaBraM, LUNA, BENDR, BIOT, CBraMod) Key takeaways: → Foundation models (up to 157M params) only marginally outperform task-specific models trained from scratch — CTNet at just 150K parameters ranks competitively with LUNA at 40.4M → Cognitive decoding tasks (speech, video, sentence, word decoding from EEG) remain far from solved — performance stays close to dummy level even for the best models → REVE, pretrained exclusively on EEG, outperforms all models on MEG typing decoding — a strong early signal for cross-modality transfer Structure: Built on NeuralFetch NeuralSet NeuralTrain (PyTorch-Lightning, MNE-Python, HuggingFace). MIT-licensed. CLI-first. Installable via pip. Full analysis: marktechpost.com/2026/05/07/… Code: github.com/facebookresearch/… Paper: ai.meta.com/research/publica… @AIatMeta @Meta_Engineers @hubertjbanville @stephanedascoli , @DahanSimon , J. Rapin, M. Careil, @BenchetritYoha1 , @JarodLevy , @saarang_p , @Antoine_RtkAI , @LucyZ47712090 , E. Cascardi, @klbegany , @teonbrooks , @JeanRemiKing .

As neuro-ai shifts toward brain foundation models, the need for a good, standardized benchmark has never been greater NeuralBench is finally here! Grateful to have helped on this massive effort by @hubertjbanville and the team! Can’t wait to see where the community takes it!

... or directly explore foundation models' ability to - decode motor signals, - detect parkinson's disease, - predict subjects' age - and so many more! just from brain activity 🧠: facebookresearch.github.io/n…

💫Very happy to release NeuralBench, to benchmark Neuro AI models and datasets in the open! 🧵Thread, 💻Code, 📝White Paper below:

✨🧠 Tribe v2, our latest model of human brain responses to sound, sight and language can now be (partly) explored on your phone📱: ▶️demo: aidemos.atmeta.com/tribev2/ 📄paper: ai.meta.com/research/publica… 💻code: github.com/facebookresearch/…

Meta FAIR Releases NeuralSet: A Python Package for Neuro-AI That Supports fMRI, M/EEG, Spikes, and HuggingFace Embeddings Every other tool supports some. NeuralSet supports all. Key Points: → One unified PyTorch DataLoader for fMRI, MEG, EEG, iEEG, fNIRS, EMG, and spike recordings → Native HuggingFace integration: DINOv2, CLIP, Wav2Vec, Whisper, GPT-2, LLaMA, VideoMAE — out of the box → Stimulus embeddings are always temporally aligned with neural recordings — no manual alignment code → Pydantic validation catches config errors at initialization, not hours into a cluster run → Same script runs on your laptop and a SLURM cluster — one config flag change → Hash-based caching means running a large language model over an entire corpus happens once, then never again The core design principle is structure–data decoupling. The entire experiment is represented as lightweight event metadata — a pandas DataFrame. No raw signals are loaded until a PyTorch DataLoader actually needs them. You can filter, explore, and recombine terabyte-scale datasets without touching a single file. 📦 pip install neuralset ↗ Full analysis: marktechpost.com/2026/04/29/… ↗ Docs: facebookresearch.github.io/n… ↗ Paper: kingjr.github.io/files/neura… @AIatMeta @Meta_Engineers @Meta @JeanRemiKing @JRaugel @JarodLevy @EvansonLinnea @LucyZ47712090 @juliengadonneix @asantosrevilla @SHouhamdi98568 @BenchetritYoha1 @stephanedascoli @DahanSimon @hubertjbanville @teonbrooks @klbegany @shubhkhanna__ @PierreOrhan @alexisthual @honualx ...

I will be in ICLR this week! 🇧🇷 Come by pavillon 3 (poster 1626) this afternoon to chat about TRIBE! We’ll also be presenting our TRIBE demo at the Meta booth throughout the week.

Come have a chat with @stephanedascoli and I about TribeV2, at ICLR, poster P3 #1636 💫

Excited to share NeuralSet from the team at Meta FAIR! 🧠NeuralSet is a high-performance package for brain & AI research that makes it dramatically easier to work with brain data at scale. 💻 Code: facebookresearch.github.io/n… 📄 Paper: kingjr.github.io/files/neura…