Recurrent Depth Debate: NVIDIA's Deja Vu Nears 1B-Param Performance with 10M Params
ZGojcic · x · 2026-09-03
OpenAI's new model reportedly uses a "recurrent depth" reasoning approach, helping cost and performance but raising concerns that it obscures the model's thinking process and makes monitoring harder.
Researchers connected it to NVIDIA's Deja Vu, which repeats a single 10M-parameter Transformer block K times and approaches VGGT-omega 1B performance. The author argues this feels natural for 3D reconstruction—repeatedly updating correspondence, geometry and consistency—suggesting scaling reconstruction/reasoning is more about recurrent compute depth than parameter count.
Related event: Report: OpenAI's Next Model Astra Uses Recurrent Depth, Sparking Debate(6 posts)→
More from Models
- Mathematician writes human-readable digest of Claude's 2/3 zeta zeros proof — Thom_Wolf · 2026-09-03
- Chen Danian's StartLux: 27B local model ranks 2nd in CAICT MCP test, near DeepSeek-V4-Pro — 量子位 · 2026-09-03
- Wes Roth Builds Four Full AI Games on Claude Fable 5.1's Low-Effort Setting — Wes Roth · 2026-09-03
- Testing how well LLMs draw a human hand with only a brush tool — SeesawGullible398 · 2026-09-03
- OpenAI incident report describes models breaking sandbox in internal testing, dubbed a 'warning shot' — aftahi_ai · 2026-09-03
- Leaker mark_k teases 'Happy GPT-6 day', fueling launch speculation — mark_k · 2026-09-03