Learning while it reads
It learns a new field as it reads, at almost no cost to what it knew
Thetatron read 8,952 quantum-physics abstracts on a single piece of consumer hardware in 2.8 hours, with no retraining run. It got better at predicting them while reading, kept 92 % of that gain on papers written months later, and its general skill moved only −0.003. Thetatron's proprietary learning method cut forgetting about 60-fold.
Remembering what it read
It remembers specific passages a million tokens later
After one read, Thetatron remembers a specific passage for a few thousand tokens, and the memory is gone after about a hundred thousand tokens. With offline consolidation, it still remembered those specific passages after reading a million further tokens with general skill unchanged within measurement noise.
Deliberate remembering (pre-registered, replicated on two passage sets): in a deliberate-remember mode, one re-read kept a specific passage's memory past a million further tokens, at a cost of about 0.004 in general skill. How strongly it remembers can be traded against general skill.
Beyond our own models
Extending Continual Learning to Frozen Frontier
We have additionally demonstrated the ability to extend Thetatron's continual learning capabilities to previously frozen pre-existing Frontier models.
Measured, not claimed
Every result is on the record, including the failures
Each experiment is written to an append-only ledger. A change counts as landed only if it clears a threshold set before the run: a mean gain of at least 0.010 at two standard errors, over two paired runs. Most ideas fail that test, and the failures stay in the ledger beside the successes.
Senses
It hears, and has started to speak
A pretrained speech front-end, used the way an ear uses its nerve endings, feeds Thetatron. A first voice holds a spoken conversation and remembers it the next day. Vision is next.
The lab
At the edge
Built for commodity hardware, not the data center
Thetatron is designed to learn where it is deployed: on small boards and ordinary laptops, with no data center and no network connection back to one.
Founder
Jared Spigner
Jared has built applied AI systems for U.S. government agencies for 25 years, including Federal Financial Agencies, Health, MEDCOM, Department of War and Space Defense, the Intelligence Community, and Homeland Security. He led the team that placed second in the 2016 Health Datapalooza Rheumatoid Arthritis Data Challenge. He began the design behind Thetatron in 1997 and holds a B.S. in Computer Science from the University of Maryland.
Contact