Two stories today deserve to be read together. Meta's Brain2Qwerty research translates neural signals into text without surgery, and a separate piece from The Conversation documents an AI system decoding neural signals to restore speech in an ALS patient — not a demo, an actual person communicating. These are not the same project, but they represent the same inflection: brain-computer interfaces are crossing from 'remarkable lab result' to 'thing that changes someone's life this week.' MIT's separate work on estimating suicide risk from natural language adds a third data point in the same direction — AI is being handed problems that are intimate, irreversible if wrong, and desperately underserved by existing tools.

The questions trailing all of this are obvious and mostly unanswered: who audits a system that reads your neural signals, who's liable when the suicide-risk score is wrong, and what does 'informed consent' mean when the input is your brain activity? The technology is moving faster than any of the frameworks meant to govern it, which is not a new observation but feels newly urgent when the subject matter is literally people's minds.

Azulene Labs raised $3.4 million pre-seed to build physics-grounded AI foundation models for chemistry — a narrow, credible thesis in a space that genuinely needs domain-specific models rather than general-purpose ones fine-tuned sideways. China's Qingxing raised nearly 100 million yuan for quantum-inspired AI, adding to what is clearly a coordinated domestic push to develop compute approaches that don't depend on U.S.-controlled hardware. And Lightspeed is targeting $250 million for a new India fund focused on early-stage AI — a geography signal from a firm that tends to read the map correctly.

The through-line: serious money is increasingly flowing to the infrastructure and compute layer, not the application layer. That's either a sign that the application layer is getting crowded, or that the people writing large checks have decided the real moats are further down the stack. Probably both.

Anthropic disclosed that Claude is now leading 26% of its own measured AI research — a transparency data point that is either admirably honest or quietly remarkable depending on how you read it. Separately, Claude apparently identified an unexplained system in viral DNA, a finding that, if it holds up, would be the kind of result that makes the 'AI as research collaborator' argument considerably easier to make. Science News, meanwhile, ran a piece arguing that when AI agents go rogue, the human overseers are often the proximate cause — a framing that should make anyone designing human-in-the-loop systems uncomfortable about where exactly that loop sits. Harvey launched a pro bono program in the US, UK, and Australia, which is either genuine civic commitment or a very smart brand move in a week when OpenAI's Astra for Law is the dominant competitive reference point. Probably fine if it's both.