OpenAI researcher lists LLM bottlenecks: context poisoning, missing video encoders, costly perception

willdepue · x · 2026-09-22

OpenAI's willdepue tells Nat Friedman the current LLM bottlenecks: context poisoning (hesitance to cram everything into one long chat), lack of video encoders making perception slow and expensive, and perception/OCR that is still mediocre and costly.

Related event: OpenAI Researcher Lists LLM Shortcomings: Real-Time Voice, Context, Full Multimodality(3 posts)→

Original post →

More from Models

Models channel →