NYC AI dinner: inference providers' perf optimizations mostly done by agents, code review declining
IgorCarron · x · 2026-10-10
IgorCarron relays takeaways from a NYC AI dinner:
- Performance engineers at major inference providers (Together, Fireworks) now mostly manage agents and RSI loops; most perf optimizations are found by agents.
- RL-ing too much on code and math hurts models, especially their language ability.
- Attendees were impressed by @bot and Cursor's acquisition by xAI; consensus is xAI is a legitimate frontier lab.
- Most participants have agents writing, reviewing, and submitting PRs 24/7.
- Code review is no longer super productive.
- Fable 5.5 will be stunning, but OpenAI is expected to beat it in December (truncated).
More from AGI Musings
- Museum visit laments: talks about industrial history now fixate only on job losses and inequality — Afinetheorem · 2026-10-10
- AI isn't killing creative jobs — it's changing what clients pay for — kevinsurace · 2026-10-10
- AI recipes aren't AI creations: the original author was simply erased — gerardsans · 2026-10-10
- KKT Points in Imperfect-Recall Games Correspond to CDT+GT Optimal Policies — jessi_cata · 2026-10-10
- AI consciousness meme: classicalists throwing darts at the blank edge of the 'space of minds' chart — eigenhector · 2026-10-10
- Three Body as an ASI allegory: was Liu Cixin on LessWrong in 2008? — abhiadesai · 2026-10-10