Zero F500 firms post-train their own models; quantizing open weights still loses to APIs
thedealdirector · x · 2026-09-28
TheEthanDing argues enterprise self-training isn't even a thing: literally zero serious Fortune 500 companies post-train their own models, since nobody wants to become the next BloombergGPT and be a HBS cautionary tale. thedealdirector adds that the new trend is pruning and quantizing existing open-source models because models like Kimi K2 are too big for on-prem GPU capacity — yet this is still more expensive and less effective than API access to frontier models.
More from Companies & People
- Mustafa Suleyman on AI safety incidents and guardrail risks of 10x-larger models — shiringhaffary · 2026-09-28
- Engineer's Tesla salary flows right back into Musk empire: Starlink, FSD, Powerwall and Grok — yunta_tsai · 2026-09-28
- 30% of Anthropic's Applied AI team are Entrepreneur First alumni — matthewclifford · 2026-09-28
- Building genuinely useful AI experiences takes rare product talent, says founder — signulll · 2026-09-28
- Google's native disadvantage: no real-world usage data from agentic apps, says Haider — haider1 · 2026-09-28
- CodeRabbit hosts Jev's first hackathon with 160+ builders at its HQ — schwentker · 2026-09-28