One hour in a scanner is now enough to rebuild what you see

Share
One hour in a scanner is now enough to rebuild what you see

Today's inbox pairs two sides of the same question: what AI can infer about people, and what it takes to defend against it. A brain decoder that used to demand 40 hours in a scanner now needs one, and the founder of Mandiant banked a monster round for AI that attacks like a human team.

A brain decoder can now rebuild what you're looking at from a single hour of fMRI — and it learned mostly from images no one ever showed a subject. Michal Irani's team at the Weizmann Institute of Science split the job in two: one branch predicts the layout of an image (where the colors sit), another predicts its content (a bunch of bananas on a plate), and a diffusion model turns those predictions into a picture. To get enough training data, they also trained the system backward — an encoder that predicts brain activity from an image — so about 70% of the training corpus was never paired with a real scan at all. The payoff is calibration: where earlier decoders needed roughly 40 hours of fMRI on each new person, this one needs about one hour, which matters because scanner time runs $600 to $1,000 an hour and "none of us can afford 40 hours of imaging for a new subject," says Tommy Sprague of UC Santa Barbara, who was not involved in the work. The failures are vivid — a cake came back as three sandwiches, a dog in a bathtub as a goat — and despite the "mind-reading" label, it reconstructs what you see, not what you think; lead researcher Michal Irani says dreams and inner thoughts remain aspirational. That distinction is the whole story: the mental-privacy debate people are having about this result is running ahead of the actual capability, but the trajectory is real — her group is moving to video and audio next, and other labs are chasing the same decoder through EEG, where a cap (or headphones) replaces a $1,000-an-hour scanner. We covered that cheaper, cruder path earlier — AI decodes silent reading from a 19-electrode EEG headset.


Mandiant founder Kevin Mandia's AI-security startup Armadin raised $255.5 million at a valuation of more than $2.5 billion. The Series B, co-led by Andreessen Horowitz and Accel with Bain Capital Ventures and Redpoint coming in, lifts total funding to $445 million just seven months after the company's public launch — a pace that says investors are buying Mandia's framing that "AI lets an attacker find and chain weaknesses faster than any human team can respond." Armadin sells swarms of AI agents that behave like hackers and already runs production campaigns for Fortune 500 and government customers, which puts fresh capital on the offensive side of a security market that has spent the past year funding defensive trust layers instead. The take: at this multiple, the market has decided AI-driven offense is a category of its own — and that the fastest way to price it is to back the people who spent decades defending enterprises from the human version of the same playbook.

What to watch: whether the decoder survives the move from fMRI to EEG without collapsing — that is the step that turns a lab result into a mental-privacy problem.

If a one-hour scan let a locked-in person finally speak, would you accept the trade? Write to us via our contact page.

Get SiliconNoon in your inbox

One email a day with the AI stories that matter — what happened, the numbers, and why it matters. No spam, unsubscribe in one click.

Read more

AI beats Stratego's greatest player on an $8,000 budget

AI beats Stratego's greatest player on an $8,000 budget

One paper closes the last big board-game frontier for AI — at a university price tag — while two disclosures show how the buildout is actually wired: agents leaking what they see, and a chip supplier financing its own customer. Ataraxos, a Stratego AI built by Carnegie Mellon, NYU, Stanford and MIT researchers, beat Pim Niemeijer — the most decorated player in the game's history — 15 wins, one loss and four draws across a 20-game series, the first superhuman result in the game's history…

Black Forest Labs launches Flux 3 Image with multi-step editing

Black Forest Labs launches Flux 3 Image with multi-step editing

A release-heavy morning: a frontier image model goes live with box-by-box editing that promises to hold everything else still, a 111,000-star coding harness hits 1.0 after reversing itself on MCP, and Perplexity slips an open-weights decision model out the same week the category caught fire. Black Forest Labs has released Flux 3 Image, the image half of its Flux 3 family, as a general-availability API built around multi-step editing.

An LLM that edits molecules instead of drawing them

An LLM that edits molecules instead of drawing them

Nature Machine Intelligence has published a Peking University method that reframes molecular design as editing rather than generation — and the numbers suggest the framing is the contribution. SynCraft, from Junren Li and Luhua Lai, takes a molecule that scores well but can't be made and asks a language model for the shortest set of structural repairs that make it synthesizable: delete an atom, mutate one, add a bond, change a bond order, set a chiral centre.