TRIBE v2: Meta Built Something Genuinely Strange, and I Can't Decide How to Feel About It
Meta released a brain prediction model trained on 720 people's fMRI scans. It can predict how any human brain responds to video, audio, or text. It is open source. I keep coming back to this one.
I read about Meta’s TRIBE v2 a few weeks ago and then kept coming back to it. That does not happen often with AI announcements. Most of them follow a predictable pattern: bigger model, better benchmark, new capability, repeat. This one felt different.
Not because it is the most powerful thing Meta has ever released. But because of what it is actually doing, which is strange in a way I had to sit with for a while before I could articulate it.
What TRIBE v2 Actually Is
TRIBE stands for Trimodal Brain Encoder. Version 2 was released in March 2026.
Here is the plain version: it is an AI model trained to predict how a human brain will respond to almost any piece of content, whether that is an image, a video, an audio clip, or text.
It does this by generating a detailed map of which brain areas would activate when a person processes that content. Not a vague heatmap. A 70,000-voxel prediction of neural activity across the entire cortex.
To train it, Meta collected over 1,000 hours of fMRI recordings from 720 healthy volunteers who were exposed to a wide range of media: images, podcasts, videos, and text. The model learned to connect features extracted from that content to the brain activity patterns it recorded.
The technical backbone uses LLaMA 3.2 for text encoding and V-JEPA2 for video. A transformer network processes the combined features and a specialized mapping layer translates them onto the voxel space.
The result is a model that can take new content it has never seen before and, for a subject it has never scanned, still generate a plausible prediction of how that brain will respond. Zero-shot generalization. It is 70 times higher resolution than comparable models.
What Meta Says It Is For
Meta is clear in the announcement. The intended use cases are neuroscience and clinical research: testing hypotheses without human subjects, accelerating work on neurological disorders, using brain-inspired patterns to improve AI systems.
This is real and legitimate. Running fMRI studies is expensive, slow, and logistically complex. If a model can run experiments computationally that would otherwise require recruiting participants and booking scanner time, that is genuinely useful for researchers.
I believe that framing. I also think it is only part of what this opens up.
The Part That Keeps Coming Back to Me
The thing I keep thinking about is this: if you can predict how a brain responds to content, you can, at least in theory, pre-test content against that prediction before publishing it.
Upload a video. Get a brain response map in seconds. See which regions light up around attention, reward, emotional processing, and cognitive load. Iterate.
Nobody is doing this at scale yet. A month after release, the adoption in content creation and marketing is still minimal. A few experimental tools, some Reddit experiments, some Medium posts. No mainstream integration.
But the direction is clear.
The weird thing is that it is not entirely new as a concept. Neuromarketing has existed for years. Eye tracking, galvanic skin response, attention heatmaps. What TRIBE v2 does is make the brain itself the measurement device and then replaces the measurement with a prediction.
You are not scanning anyone anymore. You are running the brain as software.
The Open Source Part
Meta released TRIBE v2 with full model weights, the training and evaluation code, and an interactive demo. It is on Hugging Face at facebook/tribev2 and GitHub at facebookresearch/tribev2. CC BY-NC license, so free for research and non-commercial use.
This is the detail I keep coming back to more than anything else.
When something this close to human cognition gets open sourced, the applications stop being controlled by the people who built it. The research community gets it. Builders get it. People with motivations Meta never anticipated get it.
That is not a criticism of the decision. Open sourcing this kind of model is probably the right call for research velocity and for democratic access to tools that would otherwise only exist inside well-funded labs or marketing agencies.
But it means the neuromarketing question, the content optimization question, the “what happens when creators can pre-validate content against brain responses” question, those are not hypothetical anymore. They are just a matter of time and tooling.
Where I Actually Land
Honestly, I do not have a clean take on this.
Part of me finds it genuinely exciting. A model that understands something about how brains process information could make AI systems that work better with human cognition instead of against it. That matters.
Part of me finds the content optimization angle a little unsettling. Not because prediction is wrong, but because there is a difference between content that resonates because it is true or useful, and content engineered to activate reward circuits. We already live in an environment shaped heavily by engagement optimization. Adding neural prediction to that pipeline makes it more powerful, not necessarily more honest.
The thing I keep reminding myself is that tools do not determine outcomes. Who uses them, how, and toward what ends does.
I am not building with TRIBE v2 right now. But I am watching it. The interactive demo is worth spending time with if you want to understand what this actually feels like in practice.
More than anything, I think it is worth understanding what this model is before the discourse around it gets loud and either dismissive or alarmist. It is an early signal of what happens when AI models start getting trained not just on text and images, but on the biological response to text and images.
That is a different category of capability. And I think it is worth taking seriously.