Memory outside the weights
What it reads is stored byte for byte, so it can quote the exact file and position it came from instead of reconstructing it from weights.
BerylAI reads, remembers what it read word for word, checks its own answers, and tells you when it doesn't know. It runs without a data center and without a trillion parameters.
Building an AI usually takes billions of dollars, teams of engineers, and whole data centers just to supply the compute and training data.
We're taking on that challenge with nothing but a laptop. Instead of more parameters, more data and more compute, we change the design: keep knowledge outside the weights, check answers instead of trusting them, and keep learning after launch.
What it reads is stored byte for byte, so it can quote the exact file and position it came from instead of reconstructing it from weights.
If nothing it has read answers a question, it says so. Maths answers come with a certificate checked in exact arithmetic and a Lean 4 proof check.
New material goes into memory by day and is replayed into the model at night, so it keeps learning without forgetting what came before.
A small reasoning model trained only on strings up to 40 symbols long reverses, doubles and sorts strings up to 1,000 symbols, exactly.
Developed and trained on an RTX 3080 laptop with 16 GB. Answers are served on the CPU.
Every experiment goes into a numbered lab record, wins and losses both. The comparison below shows real replies with nothing edited out.
Pick a prompt or press Send. Both replies are real and unedited.