Meta's TRIBEv2 predicts how video, audio, and text drive human brain activity, then uses it to reproduce established neuroscience experimentswithouttouching a human.
ByEjaaz Ahamadeen· 5 min read · AI Circle
4 people laid in an fMRI scanner and watched every episode of Friends, seasons one through six, plus four feature films. While they watched, a machine recorded their whole-brain activity, in real-time. That footage of the brain (roughly 65 hours of it) became the training ground for a competition to answer one question: given a piece of content, can you predict what the brain does in response?
The team that won was Meta's and their model called TRIBE beat 262 other entries in the Algonauts 2025 challenge by a clear margin.
But what peaked our curiosity for this deep dive is what this model can do. TRIBE is worth understanding because it's the foundation-model playbook for neuroscience, and it works.
Fig. 1 · The arc, 2023–2026
From reading one image to modelling the whole brain
Meta FAIR, Brain & AI (team led by Jean-Rémi King). Sources: arXiv 2310.19812 · 2412.17829 · 2507.22229 · 2605.04326.
Decoding vs encoding
What's on your mind?
For the last few years the brain-and-AI discussion has been about decoding: Read the signal in someone's brain and reconstruct the thought (mind-reading!).
A meta 2023 paper reconstructed the images a person was looking at from magnetoencephalography (MEG), a technique that samples the brain thousands of times a second. Shown a photo of a giraffe, it would generate something giraffe-shaped and giraffe-colored. Meta was blunt about it's limits in its own write-up: the images "remain less precise than the decoding obtained with fMRI," and the model "often generates inaccurate low-level features by misplacing or mis-orienting some objects."
TRIBE runs the process the other way. Instead of reading a thought out of your brain, it predicts the brain's response to an input. You give it a clip of video, its audio, and a transcript of the dialogue and it predicts, across the whole cortex in real-time, how a given person's brain will light up.
This is called an encoding model and its the biggest breakthrough in neuroscience thats been accelerated by AI. Encoding gives you a brain you can run experiments on without touching another human being.
Fig. 2 · Decoding vs encoding
The interesting move is the reverse one
Decoding: Benchetrit, Banville, King, ICLR 2024 (arXiv 2310.19812). Encoding: TRIBE, d'Ascoli et al. (arXiv 2507.22229). Decoding looks like telepathy; encoding is the half you can run experiments on.
Under the hood
What TRIBE actually is
The newest version TRIBEv2, announced by Meta in 2026, is a trimodal foundation model: video, audio, and language in, brain activity out.
It takes frozen embeddings from off-the-shelf pretrained models for each modality, feeds them through a transformer (like the one your chatbot runs on), and adds a per-subject layer to account for the fact that no two brains are wired identically. That's roughly how it works.
The impressive part is the training data. It was built on more than 1000 hours of fMRI across 720 subjects, stitched together from multiple datasets. On new subjects, after fine-tuning, it predicts brain responses 2-4 times more accurately than a standard encoder trained from scratch - the method neuroscience has leaned on for years. Huge progress.
1st/ 262
Where TRIBE finished in the 2025 Algonauts brain-modeling challenge.
1,000+ hrs
Of fMRI recordings behind TRIBEv2, across multiple datasets.
720
Subjects whose brains the foundation model was trained on.
2–4×
More accurate than a standard encoder trained from scratch, on new subjects.
The result is the model reproduces neuroscience researchwithout being guided.The fusiform face area responds to faces, the parahippocampal place area responds to scenes, a dedicated region for reading, the expected left-side dominance for language - evidence goes on and on.
Decades of hard-won experimental findings can now be recovered from a model that only ever learned to predict responses to episode of Friends.
Fig. 3 · Where the model wins
Combining senses only helps where the brain combines them
Schematic. Reflects the qualitative finding in TRIBE (d'Ascoli et al., arXiv 2507.22229) that multimodal models outperform single-sense models in higher-level associative cortex. Bar lengths are illustrative, not measured per-region values.
Implications
So why does this matter and who should use this?
For people building at the frontier - TRIBE is evidence that the brain is now tractable to the same tools that built language and vision models: big pretrained encoders, a transformer to fuse them and then enough data to generalize to things it never saw.
The question of "can we build a working simulator of a brain" is moving from no to partially, faster than most people actually realize. This has major implications on the models and products you're building.
You can ask it which edit of a film holds attention, which explanation lands in the language network, how a deaf and a hearing brain diverge on the same scene, all without touching a human brain. For clinical work, brain-computer interfaces, and basic science, that is a genuinely meaningful.
Fig. 4 · In-silico neuroscience
Run the experiment on the model, not the person
TRIBEv2 / "A foundation model of vision, audition, and language for in-silico neuroscience," Meta AI, 2026 (arXiv 2605.04326).
But of course there are questionable uses too.
Predicting how content stimulates our brains is a short step away from "optimize this ad for his brain".
But the models aren't capable of this. While TRIBE can predict cortical responses to movie clips it can't tell you whether someone will buy, click, or remember an ad.
The model's researcher's are clear about this:
Where the model stops
It is bound by the capabilities of fMRI, which "cannot capture the millisecond dynamics of neuronal firing." It basically sees the brain in slow motion.
Its inputs are only visual, auditory, and semantic which "omits primary sensory modalities such as olfaction, balance, somatosensation." i.e. most of what a body feels, it never actually experiences.
It "currently treats the brain as a passive observer."
TRIBE is an encoding model only. It predicts brain activity from content and cannot run in reverse to reconstruct content from your brain.
Outlook
What to watch
The obvious next move for Meta is to close the loop, and to push encoding toward the active brain rather than the passive one, to accelerate this they've open sourced the code, which means the method isn't staying inside 1 lab.
So thats it. The brain is now a modeling problem like any other foundational model, and the people who understand foundation models are now the people with an edge in understanding how our brain might work.
Join AI Circle
Enjoyed this piece? Join us.
AI Circle is a vetted network of practitioners, researchers, and founders from across the top AI labs. Our membership gives you access to roundtables, in-person events in your city and a private community where the frontier getsdiscussed first.
