Burkov: OpenAI's Luna seems deliberately weak as a multi-step agent
burkov · x · 2026-10-12
ML author Andriy Burkov claims OpenAI's Luna appears to have been deliberately trained not to act well as an agent. It excels at one-shot requests—he moved his product ChapterPal entirely from Gemini Flash to Luna—but cannot handle problems requiring multi-step, iterative tool-call sessions that drive a harness.
He frames this as a training choice rather than a capability gap. The claim is personal observation and unconfirmed.
More from Models
- Anthropic rumored to be testing a new model, suspected early Opus 5.5 checkpoint — iamfakhrealam · 2026-10-12
- Open-source leaderboard whatmodel.sh shows the best local models your machine can run — jakecoolguy · 2026-10-12
- 'Go use Opus 4.8 for one day': X post captures breakneck AI pace — adonis_singh · 2026-10-12
- Running Qwen3.8-Flash-Next on six 3090s: 7 users at 262k context with a patched vLLM — flynth92 · 2026-10-12
- ChatGPT's traffic share fell from 71.8% to 56.4% in a year, Similarweb data shows — gaganghotra_ · 2026-10-12
- Rumors claim Google, OpenAI and Anthropic are all near recursive self-improvement — FinanceYF5 · 2026-10-12