If Meta Perfected AutoResearch, your agent CAN TOO!

Meta's Brain2Qwerty v2 hits 78% word accuracy without surgery

Meta announced Brain2Qwerty v2, a non-invasive brain-computer interface that decodes typing intent at the word and semantic level in real time, skipping the letter-by-letter bottleneck of prior systems. Across nine volunteers in a lab setting, peak word accuracy hit 78% with an average of 61%. Controlled typing conditions only — no consumer hardware — but sentence-level decoding is now practical enough that the gap with surgically implanted chips meaningfully narrowed.

The engineering win: signal + LLM priors fused end-to-end

The leap isn't from better sensors alone. Meta fused raw neural recordings directly with LLM-style language priors to reconstruct full sentences. The code is public on Meta's blog, so you can replicate the pipeline yourself if you're building in the space.

AutoResearch agent ran closed-loop ML iteration

This is the part that matters for builders. Meta states it used AutoResearch to "discover and implement improvements that reduced the word error rate beyond standard hyperparameter optimization." The agent proposed changes, they were tested, and results beat hand-tuned baselines. Not boilerplate scaffolding — closed-loop experimental iteration on a live machine-learning system, with a human neuroscientist still in the loop.

What this means for your agent

If Meta can use an in-the-loop research agent to grind through variations a human researcher wouldn't have time to explore on a neuroscience-grade system, you can do the same on your own stack. The ceiling on coding agents just moved from "writes code" to "runs the experiment."

Constraint that actually matters

AutoResearch only paid off because Meta fed it large volumes of high-quality data. Without sufficient data, the agent has nothing to iterate against — that is the bottleneck, not the agent framework.