Tilde launches One Layer Deeper to test whether models can learn deeper serial computation
marksaroufim · x · 2026-08-04
Tilde launches One Layer Deeper to study learned serial depth
The post argues that today’s reasoning models mostly scale by generating more tokens, which can force intermediate computation into text instead of keeping it in latent state. That may unnecessarily serialize work that could otherwise run in parallel.
The proposed alternative is to scale computation through additional latent depth as problems get harder, but such models have historically been difficult to train.
Core thesis
- The bottleneck may not be the lack of better serial-computation architectures.
- The real issue could be that current optimization methods favor standard Transformers.
- Better architectures may need to be co-designed with objectives and optimizers.
Competition details
- Task: repeated modular squaring as a public benchmark for learned depth.
- Goal: learn deeper serial computation and extrapolate beyond training depth.
- Deadline: August 31, 2026.
- Focus: architecture–optimizer–objective co-design.
Related event: Tilde Launches 'One Layer Deeper' Compute Competition(3 posts)→
More from Research
- LightParkour turns sparse human motion seeds into a deployable humanoid parkour policy — CyberRobooo · 2026-08-04
- RL researchers debate whether Dyna-style systems should be called models or world models — tw_killian · 2026-08-04
- METR defines an “expenditure horizon” for AI optimization, using NanoGPT as a test case — gleech · 2026-08-04
- Goodfire AI is dissecting LLMs to understand how they think — Scobleizer · 2026-08-04
- Lightbot 0 learns parkour-style whole-body contact skills for rescue scenarios — CyberRobooo · 2026-08-04
- Lean formalization of Mochizuki’s abc proof fails at the same step again — MarioKrenn6240 · 2026-08-04