Spotlight architecture adds growing memory to LLMs with constant-cost token generation
teortaxesTex · x · 2026-10-03
Christos Tzamos's team unveiled Spotlight, a new architecture claiming LLMs can gain knowledge and capabilities without retraining: models get growing memory while token generation stays constant-work regardless of stored information, marrying attention's capacity with the linear cost of fixed-state recurrent models.
Commenter teortaxesTex notes the constant factor looks large — each access touches 9 cells, each storing a full dk×dv matrix per head — though that may be inherent to the continual-learning design.
More from Models
- Steve Yegge: Two Weeks With Opus 5.5 — Precision Rivals Fable, Recall Trails on Open-Ended Tasks — Steve_Yegge · 2026-10-03
- OpenAI's Codex global reset appears to skip Business accounts, support suggests buying credits — AdventurousFeeling19 · 2026-10-03
- Mystery 'iguana_necktie' Field Spotted in Anthropic Usage API — bytebot · 2026-10-03
- Insider teases 'new SSI model,' calling Ilya 'a truly remarkable human being' — iruletheworldmo · 2026-10-03
- Exploit Bench results are highly harness-sensitive: GLM 5.3 Flash beats 4.1, best in Claude Code — teortaxesTex · 2026-10-03
- Reddit user says Opus 5.5 burns weekly cap at 300M tokens, down from 1-2B before — Chemical-Ad-7982 · 2026-10-03