Researcher's rant on brittle frontier models sparks resonance

On September 20, a researcher posting as lateinteraction published a long tweet as a "checkpoint" on the current moment: at the "Ultra" tier, he squeezes "barely acceptable" work output from frontier models every day, and it's painful—he has to provide more detailed context and finer-grained feedback than he has ever given any human colleague in his life, and his day job is literally writing feedback for models. He described 2026's frontier models as having a "jagged capabilities" problem: even after 8 straight hours using Astra/Fable-style agentic products, as long as the work demands quality across high dimensions (more than one or two easily verifiable dimensions), the models remain as fragile as autocomplete, with no improvement across generations.

Confirmed

• lateinteraction used a metaphor for the mismatch in model capabilities: it's like hailing a cab and someone hands you a catapult pre-aimed at a handful of destinations—it never lands where you actually want to go
• The long tweet resonated widely and sparked a threaded discussion
• Former OpenAI researcher doodlestein responded that the root cause is process-related, and offered practical advice: iterate with the model on a planning document over multiple rounds first, then let it write code

Why it matters

• This is a front-line AI researcher's concentrated rant on "whether frontier models can handle real work," directly pointing at the reliability shortfalls of current agentic products on high-dimensional, hard-to-verify-in-real-time tasks
• Multiple peers joined the discussion with coping strategies, showing that "how to make good use of still-fragile frontier models" has become a shared practical concern among practitioners

Note: Multiple posts give conflicting descriptions of lateinteraction's real identity (one says Voyage AI founder Tengyu Ma, others say ColBERT author Omar Khattab or Charles Packer); the sources contradict each other and the identity cannot be confirmed.

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

• [source] 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
• AI Researcher: Even 8 Hours of Agent Work Feels Brittle and 'Autocomplete-Like' — lateinteraction · 2026-09-20
• [source] 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