AI News Today
← All episodes
Episode 1 · March 27, 2026 · 12:05

Meta's AI Can Now Predict Your Brain

Meta's New AI Can Read Your Brain: Tribe V2 ExplainedMeta has unveiled Tribe V2, a revolutionary AI model that predicts how the human brain responds to sights, sounds, and text with 70x more resolution than its predecessor. This "digital twin" of neural activity bypasses expensive fMRI scans, opening new doors for neuroscience, healthcare, and the future of artificial intelligence.00:00 - Intro: Meta's Brain-Reading AI01:04 - Introducing Tribe V201:34 - Massive Scaling: V1 vs. V203:20 - How the AI Architecture Works04:27 - Cleaner Signal than fMRI?06:55 - Future Applications in Science09:02 - Mind Reading vs. Prediction10:31 - The Power of Scaling Laws

Full transcript

Meta's AI can now read your brain. Meta just built an AI that predicts exactly how your brain reacts to anything you see, hear, or read, not in theory, in practice. Right now it's called TRIBE V2 Trimodal Brain Encoder, and it just changed what's possible in brain science, medicine, and AI itself. Here's why this matters, and why it's moving faster than most people realise.

Brain science had had a bottleneck for decades. To study what your brain does, you need an fMRI machine. Now those machines cost between $1.5 and $3 million to buy. They charge up to $620 an hour to use, and the average study scans around 30 people.

30 people. That's a data problem, and a scale problem. You can't understand the human brain from 30 people any more than you can understand human behaviour from 30 Twitter users. So for decades, brain science moved slowly, expensive, small, hard to repeat.

One experiment could take a year to run, analyse, and publish. TRIBE V2 changes that math completely. Feed it a video, an audio clip, a piece of text, and it tells you how a human brain would respond across 70,000 individual data points in the brain, without putting anyone in a scanner. But Twitter's team calls it a digital twin of neural activity.

I'd call it a flight simulator for the brain. Same idea, test everything in software before you ever touch a real thing. Let me give you the numbers because they're absolutely wild. The original TRIBE model, version 1, was trained on data from just 4 people.

Four volunteers who each spent over 80 hours lying in a scanner, watching movies like The Bourne Supremacy and a BBC nature documentary called Life. That model had 1 billion parameters and could predict brain activity across roughly 1,000 areas of the brain. TRIBE V2 trained on 700 people, over 500 hours of fMRI recordings, more than 1,100 hours of total training material, and instead of 1,000 brain areas, it now covers 70,000. That's a 70x jump in resolution.

Then there's the zero shot thing. This is the part that really gets me. TRIBE V2 can predict how a brain it has never seen before will respond. New person, new language, new type of content it wasn't specifically trained for.

No retraining required. Meta measured the improvement 2-3 times better than previous methods for both movies and audiobooks. And then their team found a scaling law. The same kind that drives chat GPT in every major language model.

The more fMRI data you need and feed TRIBE V2, performance keeps going up. Log linearly. No ceiling in sight. That's not incremental progress.

That's a model that's going to keep getting better for years. If you want to dig deeper into how to use AI breakthroughs like this one to understand them, stay ahead of them, and actually apply them to your business, the AI Profit Boardroom is where that happens. Four weekly coaching calls, daily step-by-step tutorials, 30-day roadmaps, and a community of 2,600 business owners already doing this. Link in the comments and description.

Or go to the AIProfitBoardroom.com to get access. Now let me explain how this thing actually works because the architecture is very clever. Think of TRIBE V2 as a three-stage machine. Stage 1.

Three separate AI models read one sense. Meta's VJEPA 2 watches the video. WAVE to VEC to BERT listens to the audio. And LLAMA 3.2, the 3 billion parameter version, reads the text transcript.

Each one produces a dense numerical description of what's happening moment to moment. Step 2. Those three streams get merged. An 8-layer transformer, same basic architecture as GBT, figures out how they all connect.

Because your brain doesn't process sight, sound, and language in isolation. It weaves them together. And TRIBE V2 learns to do the same. Stage 3.

The merged output gets mapped onto brain activity. Not just regions, individual voxels, 70,000 of them, across the whole brain. The result, a predicted fMRI scan for any content, for any person. Here's the thing that surprised me most about this research.

TRIBE V2's predictions are sometimes more accurate than an actual brain scan. And I know that sounds backwards, but here's why it's true. Real fMRI scans are noisy. Your heart beating creates artifacts, for example.

Moving your head slightly changes the signal. Your attention drifts. The scanner itself introduces distortions. TRIBE V2 is trained on 700 people.

It learned the underlying pattern. What the brain actually does. And averaged out all the noise. So when it predicts a brain response, it's giving you the clean signal, not the messy individual scan.

The lead researcher, Jean Remy King, a CNRS neuroscientist on detachment at META in Paris, put it directly. Non-invasive recordings are notoriously noisy and can greatly vary across recording sessions and individuals. TRIBE cuts through that. To be clear, it still only explains about 54% of explainable brain variance.

Nearly half of what the brain does remains unpredicted. This isn't a complete brain map, but 54% at 70,000 voxel resolution across 700 people with zero-shot generalization is a genuine leap forward. How did META know it worked? Well, they entered in the Algonauts 2025 competition.

They took it as the Olympics of brain modeling, run by researchers at MIT, Universal Berlin, and Goethe University in Frankfurt. 263 teams entered from labs around the world. And the challenge was predict how the brain responds to nearly 80 hours of movies, including a trick test, a silent black and white Charlie Chaplin film. No audio.

Just to see what happens and see what models do when a modality disappears. TRIBE V1 won first place out of 263 teams. And the version they just released, TRIBE V2, scales that winning architecture with 70 times more brain resolution and 175 times more subjects. META released everything publicly.

The model weights, the full code base on GitHub, the research paper accepted at ICLR 2026, and an interactive demo anyone can use right now. This isn't a press release. This is actual real science done in the open. So where does this actually go?

What changes? Three directions, and they're all moving at the same time. First, neuroscience itself. Right now, a researcher who wants to understand how the brain processes sarcasm, fear, or music has to build a study from scratch.

Ethics, approval, recruiting volunteers, scanning time, right? And you've got months of work for one experiment. With TRIBE V2, you run the same experiment computationally in hours. You test 500 audio clips instead of five.

You can screen your hypothesis before you spend a dollar on lab time. A content agency studying emotional response patterns in ad creatives could use this kind of model to pre-test brain engagement before production. Second healthcare. META's team explicitly mentions neurological conditions.

Compare TRIBE V2's predictions for a healthy brain against real scans from patients with aphasia or PTSD or sensory processing disorders, and the gaps tell you where neural signaling has broken down. That is diagnostically useful in a way that wasn't affordable before. And third, drug development. Drugs that target the brain at brutal rates in clinical trials, partly because we can't predict what they'll do to complex neural systems.

So they slow down in progress. A model that predicts brain response at 70,000 voxel resolution could help screen compounds earlier, before human trials, before the expensive failures. And then there's the feedback loop that almost nobody's talking about. TRIBE V2 was built using AI models, transformers, Lama V Jepa.

And now it helps us understand biological brains. The insights from that go back into designing better AI. Your brain runs on 20 watts of power whilst processing sight, sound, and language simultaneously in real time. If AI researchers can figure out how it does that, the efficiency gains for artificial systems could be enormous.

The brain is teaching AI how to be better. And AI is teaching us about the brain. Both loops are running at the same time. Now let me be honest about what this isn't.

Because there's real hype around this and you deserve the full picture. TRIBE V2 does not read your mind. It predicts brain reactions to external stimuli. It cannot decode your private thoughts, your intentions, your emotions, your internal monologue.

The gap between your brain lights up when you hear the sound and I know what you're thinking is enormous and we're nowhere near crossing it. The model explains 54% of explainable variants. That's impressive. But it also means that 46% of what your brain does remains unpredicted.

So there's plenty of unknown territory left. Euro-marketing researcher Thomas Zerger Ramsey made a sharp point. TRIBE was trained on people watching movies. That's a controlled environment.

The real world is very much messier. And he warned against reverse inference. Just because a brain region activates doesn't mean you know the psychological state behind it. Privacy is a live issue too.

Multiple US states are passing neural data privacy laws right now. The UN's special reporter and Anna-Brian Negras has raised concerns directly about brain prediction technologies and their potential for manipulation. These aren't hypothetical worries. Meta released TRIBE v2 under a non-commercial license.

But the model and code are publicly available, which means the technology is out there. Here's what I keep coming back to. The scaling law finding is the most important thing in this paper. And it's getting the least attention.

Every major AR model we have, GBT, Claude, Gemini, followed a scaling law. More data. More compute. Better performance.

No ceiling until you hit one. And so far, most of these models haven't hit any. TRIBE v2 follows the same curve. More fMRI data means better brain predictions.

And we've barely scrapped the surface of how much fMRI data exists or could be collected. The models team at Meta is sitting on a trajectory that looks a lot like the early days of large language models. In 2019, GBT-2 was a party trick. In 2022, GBT-4 was a job threat.

In 2026, TRIBE v2 explains 54% of brain variance. But what does TRIBE v5 explain? That's a question that should be on everyone's mind. The AI tools moving fastest right now aren't just the ones making content or writing code.

They're the ones rebuilding fundamental scientific infrastructure. TRIBE v2 is one of those tools. And if you want to stay ahead of what AI is actually doing, the moves under the surface, not just the headlines, then the AI Profit Boardroom is the place to be. Four weekly calls, daily tutorials with real step-by-step guides on implementing AI, 30-day roadmaps, a community of 2,600 business owners, prompts for everything, and a local map so you can connect with people in your city.

The link's in the comments' description, or you can go to the AIprofitboardroom.com. This is Meta's AI reading your brain, and it's only getting sharper.

More episodes

Browse all episodes →