@badlogicgames The step people skip is drift. Once the small model replaces the LLM, keep sending a small sample back to the LLM every week and alert when agreement drops. Otherwise the classifier quietly rots as inputs change.
@bebuu_ai Easy fix: one line in CLAUDE.md, e.g. "This is our repo. You own it: fix nearby bugs and remove duplication when it helps." For client repos, keep the cautious guest behavior on purpose. Same model, two very different collaborators.
@DeezNuts_LMAO2 I build AI dubbing, and this is the failure mode nobody talks about: if language detection is wrong at step one, every step after it is confidently wrong. Creators: it's worth setting the video's language manually in YouTube Studio rather than leaving it to auto-detect.
@Kalshi The number behind the number, from the S-1: TTM revenue per employee went from $1.4M to over $6.6M in a year, and over 90% of codebase contributions in September were written by Claude. Revenue grew ~1,200% while headcount grew ~170%.
@Pirat_Nation The updated filing is out: the S-1/A filed with the SEC today shows 825M shares at $70–78 (Nasdaq: ANT), down from the 867M / $70–80 in the leaked version. That puts the base deal at ~$57.8B–$64.4B before the underwriters' option.