Researcher's viral thread: frontier models still fragile as autocomplete for real work

On Sept 20, researcher lateinteraction posted a long thread taking stock of the current moment: daily, under the 'Ultra' tier, he wrestles 'barely passable' work out of frontier models, supplying more detailed context and fine-grained feedback than he has ever given any human colleague — even though writing feedback for models is his day job. He describes 2026's frontier models as having 'jagged capabilities': even after 8 straight hours of Astra/Fable-style agentic products, if your bar is high-dimensional (beyond one or two easily verified dimensions), the model remains as fragile as autocomplete, with no cross-generation improvement.

Confirmed

He compared the capability mismatch to hailing a taxi and being handed a trebuchet pre-aimed at a few destinations — it never lands where you want to go
The thread resonated widely and sparked a reply chain
Ex-OpenAI researcher doodlestein responded that the root cause lies in process, advising iteration on a planning document before letting the model write code

Why it matters

This is a concentrated practitioner critique of whether frontier models can handle real work, targeting the reliability gap of agentic products on high-dimensional, hard-to-verify tasks
Multiple peers joined with mitigation strategies, showing that 'how to use still-fragile frontier models well' has become a shared practical concern

2026-09-20 ~ 2026-09-20 · 5 related posts

Primary sources

AI Researcher: Even 8 Hours of Agent Work Feels Brittle and 'Autocomplete-Like' — lateinteraction · 2026-09-20 05:38 (UTC+08:00)
Ex-OpenAI researcher: iterate on the plan document with multiple frontier models before writing code — doodlestein · 2026-09-20 06:31 (UTC+08:00)

Voyage AI founder Tengyu Ma: frontier models at Ultra settings are still fundamentally dumb — lateinteraction · 2026-09-20
lateinteraction: 8 hours of frontier-model 'Ultra' work still collapses like autocomplete — lateinteraction · 2026-09-20
[source] AI Researcher: Even 8 Hours of Agent Work Feels Brittle and 'Autocomplete-Like' — lateinteraction · 2026-09-20
ColBERT Author Slams Frontier Models: 'A Pre-Aimed Catapult That Never Lands Where You Want' — lateinteraction · 2026-09-20
[source] Ex-OpenAI researcher: iterate on the plan document with multiple frontier models before writing code — doodlestein · 2026-09-20