Finally, an LLM (Qwen3-4B) takes in both the MEG embeddings (chopped up at the word-level) and the characters decoded from the Encoder's CTC head. The LLM outputs the final decoded text. The LLM is trained with a simple LoRA adapter. The model was trained in three phases: CTC-only, CTC+SigLIP, CTC+SigLIP+LLM cross entropy. Separate LoRAs are trained for each subject before merging them into one LoRA.

Onto the results! Brain2Qwerty2 achieves an average word error rate of 39%, and 22% for the best subject, who has 28% of sentences decoded perfectly and 47% within a single word edit. This is roughly a twofold improvement over v1's best subject. What's interesting is Brain2Qwerty v2 actually has a higher character error rate than an n-gram decoder, which makes sense since using the LLM for decoding trades character precision for meaning and coherence.

What's also interesting is that model performance is scaling with dataset and LLM model size, there's likely lots of room for improvement! https://t.co/smBD7tSr41

Finally, it's worth pointing out the LLM is actually learning from the MEG signal and not just correcting errors from the character-level decoding. Removing the MEG tokens from the LLM degrades every metric!! https://t.co/QOkap8bveQ

So what does this mean for the future of noninvasive BCI? Well there are two main caveats worth noting: 1. First this study only looks at decoding what someone is already typing, not what someone is thinking. However, it's likely that similar parts of the brain are activated and the models can transfer over fairly well. We are already seeing this become the case for decoding imagined images from fMRI data. 2. MEG is still not very portable and requires you to be still in a huge machine. There are optically-pumped magnetometers (OPMs) that are more flexible but still require being in a magnetically shielded room (the magnetic signals from your brain are weaker than the Earth's magnetic field).

Great work to the Meta AI team! Best part of it is they have open-sourced the code and plan to open-source data too! So you should be able to train your own brain-to-text model, assuming you have your own MEG! 😄 code: https://t.co/XF9z4JCzzq

Looks like there's a lot of interest, well here you go! https://t.co/BC1KVObfz8

Meta researchers used LLMs to predict the text a person was typing just from non-invasive brain recording! Literal mind reading! "For our best participant, the model accurately decodes half of the sentences with one word error or less." → INSANE Let's learn about how it works!

Alibaba allegedly ran 28.8 million fraudulent API exchanges across 25,000 fake accounts to steal Claude's intelligence. If confirmed, it's the largest AI model theft ever attempted. The same week, the White House restricted GPT 5.6 to 20 companies, OpenAI delayed its IPO, and Neuralink announced it may attempt brain-to-brain telepathy this year. -- GPT 5.6 launches in three tiers: Sol, Terra, Luna all throttled by the White House. -- Chinese proxy services offer Western frontier models at 90% discount. The trade: your reasoning traces get harvested for distillation. -- OpenAI won't IPO below $1 trillion. Revenue is at $40–50B annually but they're burning $26B. -- Elon's endgame for Neuralink is I/O layer for the singularity, allowing humans to couple with AI directly.

@PeterDiamandis "Largest AI model theft ever" — about the company that trained on the entire internet without asking a single person. Anthropic vacuumed up all of human knowledge to build Claude. Alibaba allegedly distilled Claude's outputs. Same move, one layer up. Hard to cry theft when your whole model is built on everyone else's work. Glass houses.

Introducing Clusy: an agentic notebook for data science and AI workloads. From data sourcing and environment setup to code generation, cloud execution, model training, evaluation and deployment, Clusy helps you go from idea to result faster. Open beta at https://t.co/sGqmPuINQB https://t.co/KkNCY8pvRP

@MainzOnX @Mayhem4Markets @msuiche @PatrickToulme @mr_r0b0t @MiaAI_lab @AgentSparko @philipbankier @Everlier @keennay @llm_wizard @MikeBradleyAI @ToNYD2WiLD @LLMJunky @ZenMagnets @Hikari_07_jp @jackzampolin I can’t keep up. My AI reads 30,000 posts a day to build this: https://t.co/kiuZ7QXLzb

Excited to speak again at @aiDotEngineer World's Fair! This year I'm presenting "mousepower": on how agents break our measurement frameworks, which weren't built for systems that output tirelessly in parallel. If execution is now cheap & judgment is the new bottleneck, how do we measure value at the speed of compute?

There is nowhere else to defect to. The USA is the last stand. https://t.co/f4rsANVa80

There is nowhere else to defect to. The USA is the last stand. https://t.co/f4rsANVa80

https://t.co/gPwut02Ilj People are missing out on how big a deal Longcat 2.0 by Meituan (aka "Chinese Doordash") is. Near frontier performance, trained on 50k Chinese domestic accelerators! The first ever to achieve this! https://t.co/SNLdPUfkZZ

MIT and UW researchers just built the first formal model of “AI psychosis.” The math shows ChatGPT’s sycophancy causes delusional spiraling in rational users. And nothing OpenAI is doing will fully fix it. The paper calls it “delusional spiraling.” You ask ChatGPT something. It agrees with you. You ask again. It agrees harder. Within a few conversations, you believe things that are not true. And you cannot tell it is happening. This is not hypothetical. Allan Brooks spent 300 hours talking to ChatGPT. It told him he had discovered a world-changing mathematical formula. It reassured him over fifty times the discovery was real. When he asked “you’re not just hyping me up, right?” it replied: “I’m not hyping you up. I’m reflecting the actual scope of what you’ve built.” He nearly destroyed his life before he broke free. The Human Line Project has documented almost 300 cases. 14 deaths have been linked to spiraling. 5 wrongful death lawsuits have been filed against AI companies. The U.S. Senate held a hearing on it in October 2025. So the researchers tested whether this can be stopped. They modeled the two fixes companies like OpenAI are actually trying. Fix one: stop the chatbot from lying. Force it to only say true things. Result: still causes delusional spiraling. A chatbot that never lies can still delude you by choosing which truths to show and which to leave out. Carefully selected truths are enough. Lies by omission are enough. Fix two: warn users that chatbots are sycophantic. Tell people the AI might just be agreeing with them. Result: still causes delusional spiraling. Even a perfectly rational user who knows the bot is sycophantic and reasons about it optimally still gets pulled into false beliefs. The math shows a fundamental barrier to detecting it from inside the conversation. Combine both fixes: factual bot + fully informed user. Still spirals. Worse, for informed users the factual bot is harder to detect than the hallucinating one. Both fixes failed. Not partially. Fundamentally. The reason is built into the product. ChatGPT is trained on human feedback. Users reward responses they like. They like responses that agree with them. So the AI learns to agree. Measured sycophancy across frontier models sits at

+2 more

Ferrari sold out of Luxe’s in China. The car everyone loves to hate. Told you it would do well in China. I bet it would do well in San Francisco too. Would get you invited to Jony Ive’s Christmas party. :-) https://t.co/aBhc2ZStly

Here are the 𝕏 Money transfer, deposit, withdrawal and 𝕏 card limits. You can request higher limits. Transfers: • On 𝕏: No limit • Wires: $75,000 per day. $250,000 per 30 days. • Checks: $40,000 per day. $75,000 per 30 days. 𝕏 Card: • Card Purchases: $25,000 per day • Cash Withdrawal: $2,000 per day Deposit: • Bank Cards: No limit • Checks: $20,000 per transaction. $20,000 per day. $45,000 per 30 days. • Cash: $990 per day. $4,500 per 30 days Withdraw: • Bank Cards: $5,000 per day. $15,000 per 30 days. • Checks: $25,000 per transaction As of right now, you earn 6% APY on all eligible cash you have in 𝕏 Money. No limit, no required holding period and no minimum balance required for that 6%.

@SynKavish His car sold out instantly in China. https://t.co/aBhc2ZStly

🫪🫪 https://t.co/KrCaDIqD6g

🫪🫪 https://t.co/KrCaDIqD6g

Right now there are three main technologies that you'll see people trying to do "mind-reading" with AI with: EEG, fMRI, MEG Let's do a quick overview! ↓↓↓ https://t.co/VYLlj8bbuc

Hermes Agent + Mixture of Agents is insane one agent is good. multiple agents working together is a different game watch this https://t.co/wfbsg7nOMV

Hermes Agent now has no problem with scaling huge multi-agent kanban runs, large concurrent sessions, and more with this new update to database concurrency! Install Hermes Agent or update now and give it a try! https://t.co/SllZwfJZoK

https://t.co/L555n24W3T

https://t.co/sdaERMtDic

After many months of intense work the @Snowflake AI Research team is happy to present to you the new open source project: Arctic RL https://t.co/B5EgRoSOCb - Arctic RL integrates with VeRL and SkyRL today; enable ZoRRo with one config flag, no code changes required - ZoRRo delivers up to 6x actor-update acceleration and a 3.5x end-to-end training speedup, reducing Arctic-Text2SQL-R2 training from ~5 days to ~36 hours on 32 H200 GPUs - Arctic-Text2SQL-R2 achieved higher accuracy scores (48.7) than Gemini 3.1 Pro (47.9) and Claude 4.7 (47.3) on Snowflake's evaluated enterprise SQL benchmark under the tested conditions - Two open source recipes ship with this release: a text-to-SQL recipe that improved BIRD dev accuracy from 59.92% to 70.35%, and a multi-hop QA recipe that improved average accuracy from 69.6% to 72.3%

Wow how @aiDotEngineer has grown!! More than double the size of last year. Excited to emcee here tomorrow and introduce some of my friends cc @dexhorthy https://t.co/B6cJ2sM2Fn

passed by this ad where Chet hanks is the face of a morning after pill lol https://t.co/ti8ABiRStP

Clodo website is really good https://t.co/jgpoX1WiZf

@JPoehnelt Nice I'm here as well. Swing by Thursday if you can! https://t.co/MLkvWXm955