'Training your own models is an incredibly stupid idea' — why RAG beats fine-tuning pitches
rseroter · x · 2026-10-21
- Blogger Devansh argues Silicon Valley startups are misleading enterprises into wasting millions on the "train on your proprietary data" pitch — citing Databricks' $1.3B MosaicML acquisition, Harvey's $550M raise, and Gartner's $4.9B specialized-model market forecast for 2026.
- Once you accept that philosophy, every problem points to more training: more data, changing data, lingering mistakes — a bottomless commitment that hands your operating margins to cloud providers.
- A growing body of research suggests injecting proprietary knowledge into weights is the most inefficient way to make it useful; "your company's knowledge doesn't need to live inside a model's weights — your AI needs to know how to use it."
More from AGI Musings
- AI called a store for me — but no transcript means zero transparency, argues Allie Miller — alliekmiller · 2026-10-21
- Dev hasn't touched his MacBook screen in weeks: agents do everything, iPad is the interface — MarcJSchmidt · 2026-10-21
- UAE becomes first country to make AI a mandatory K-12 subject, diverging from Western bans — alexvoica · 2026-10-21
- Bittensor miners speed up REAX decision model nearly 3× within one evening of competition — airesearch12 · 2026-10-21
- Chollet: Engineers Hit Record Numbers as Agents Write ~100% of Code, Bottlenecks Remain — manuelkoelman · 2026-10-21
- Fields medallist says OpenAI's 700+ AI math manuscripts left him in tears — miniapeur · 2026-10-21