Stablecoin issuer Tether is promoting three studies involving researchers from Tether Evo, its brain-computer interface division, as evidence that AI could learn across different brains and reduce the need to build a new system for every patient.
Three Evo researchers authored the speech paper, while Tether-affiliated researcher Matteo Ferrante co-authored the visual and music studies with university researchers. And in all three cases, the teams developed new decoding methods using brain recordings collected for earlier research.
But Tether presented the papers together in a paid TechCrunch article that said "a single model" could interpret brain activity across people through speech, vision and music.
The papers instead describe three separate systems. And only two tested whether information learned from several subjects could be reused.
The visual result came from 100 choices
The biggest gap appears in the visual study. Tether said the model identified an image "out of thousands" with 70% accuracy. The macaques had viewed more than 20,000 images, but most of those were used for training.
- The reported 69.3% result came from choosing among 100 held-out images, each shown 30 times and averaged to strengthen the signal.
- And although the study used recordings from two macaques, it didn't test whether a decoder trained on one could work on the other.
- Its authors instead listed cross-subject generalization as future work.
The speech model still needed calibration
The speech study offered stronger support because researchers trained a shared decoder across participants.
Still, each new person had to provide brain recordings for an individual alignment step. The results varied.
- For one participant, the shared model produced a 53.3% phoneme error rate, compared with 30.6% for a model trained from scratch.
- The authors acknowledged that the benefits of shared pretraining were "not yet universal."
Tether also said adaptation could take "minutes to hours." But the paper discussed that timeline as a possible outcome for larger future models rather than a result tested with new patients.
The music study did show cross-person decoding across five participants. Still, its main findings have been public since June 2024.