Engram: brief training fixes long-run models' increasingly inscrutable notes
scott_linderman · x · 2026-11-05
jxmnop shares a glimpse of work at Engram: they're excited about models that understand our data better than we do, but found that long-running models take increasingly inscrutable notes — and small amounts of training can fix this. More to come.
More from Research
- Scientists fuse graphene and glass into glaphene, a 2D semiconductor with new electronic behavior — acmoytoy · 2026-11-05
- 'Joules-to-GDP' framework: the highest-value AI tasks aren't the most energy-efficient — arena · 2026-11-05
- UTokyo team 3D-prints lifelike musculoskeletal robot legs via parametric design — repligate · 2026-11-05
- FinForesight: A Backtesting Benchmark for Deep Research Financial Forecasting — Fengyuan Liu · 2026-11-05
- X-Hinges at UIST 2026: 3D-printable robot hinges with embedded sensors grip your finger — repligate · 2026-11-05
- Follow-up: model gap in MNIST weight programming, more data doesn't help — RobertTLange · 2026-11-05