Noam Brown — 'your fancy AI scaffolds will be washed away by scale' (Latent Space)

TL;DR

A Latent Space podcast clip surfaces a quote attributed to Noam Brown (OpenAI): “your fancy AI scaffolds will be washed away by scale.” The accompanying framing is that routers, harnesses, and complex agentic systems are getting replaced by base models that work better out of the box, with reasoning models cited as the existing precedent. The tweet itself contains no benchmark numbers, no paper link, and no extended argument — it is a position claim being amplified, not new evidence.

Key claims

• Complex agentic scaffolds (routers, harnesses, multi-stage systems) are positioned as transient — competitive against current-generation base models but expected to be obsoleted by the next scale step [tweet body].
• Reasoning models are cited as the existing existence proof: capabilities that were previously stitched together with prompt-time scaffolds (CoT prompting, plan-and-execute, self-critique) collapsed into the base model once trained with reasoning RL [tweet body].

Method

Not applicable — single tweet, no methodology. The artifact is a podcast clip embed (no transcript in the tweet itself) plus a 4-line opinion summary by @latentspacepod. The underlying Noam Brown source (talk, podcast episode, conference clip) is not linked in the tweet.

Results

No results. Position statement only.

Why it’s interesting

This is the bitter-lesson framing applied specifically to agentic infrastructure, and lands on a live tension inside the wiki. Agentic Software Engineering is currently organized around the opposite assumption — that cross-scaffold tool-call format diversity, execution-grounded verification, and closed-loop synthesis are load-bearing post-training recipes (see Qwen3-Coder-Next Technical Report §4.2.2 on training across Cline / Qoder / OpenCode / Claude Code / Qwen Code formats specifically because cross-scaffold transfer is weak). If the Noam Brown framing is right, that recipe is a transitional fix for a scale-shaped hole. It also resonates with The flavor of the bitter lesson for computer vision (normative anti-intermediate position for vision) and On the Slow Death of Scaling (the counter-position that pure scaling is decelerating).

See also

• Agentic Software Engineering — the concept page whose central recipe (scaffold-format diversity, multi-stage post-training, closed-loop synthesis) this tweet’s position directly challenges
• The flavor of the bitter lesson for computer vision — same bitter-lesson move applied to CV hand-crafted intermediates instead of agent scaffolds
• On the Slow Death of Scaling — counter-position arguing pure-scaling returns are slowing, making scaffolds more rather than less load-bearing
• MiroThinker: Pushing the Performance Boundaries of Open-Source Research Agents via Model, Context, and Interactive Scaling — argues interaction depth is an independent scaling axis (scaffolding-shaped), a concrete claim against the “scale alone” framing
• Agentic AI's OODA Loop Problem — argues the OODA-loop / harness structure is itself the security and reliability boundary, not just a capability crutch