Altman: OpenAI's Best Model 'Can't Responsibly Be Open-Sourced' Yet
External_Mood4719 · reddit · 2026-10-24
Speaking at Stanford GSB's 'View from the Top,' Sam Altman laid out his stance on open weights and safety.
Key points
- Open-weight models are important for society up to a fairly high capability level; banning open source entirely would be 'an extremely high cost to innovation and liberty' — we shouldn't do it.
- But above some threshold models can do real harm: 'I don't think we could responsibly open source our current best internal model anytime soon,' pending stronger global cyberdefense and biosecurity.
- He admits he underestimated the alignment problem in 2020; current models are better aligned than expected, but 'the rapid progress of the past six months has tested the strength of my convictions.'
- In narrow domains like mathematical reasoning, 'we already have superintelligence now — it happened faster than I expected.'
- He called for serious industry discussion on pacing AI progress, arguing current governance frameworks can't sustain the present speed.
Context: Altman reflected that OpenAI was founded as a pure research lab assuming commercialization would sort itself out; around 2020 the company pivoted, and ChatGPT's explosive growth turned it into a massive company virtually overnight.
More from Companies & People
- Palantir CEO Alex Karp: Europe barely matters in the AI race between China and the US — 2C_ornot2C · 2026-10-25
- "Stop pretending Jensen believes in anything but stock price": open-source drama erupts — zetalyrae · 2026-10-25
- Peter Thiel backs founders who speak in definite future tense, skips the probability talkers — 2C_ornot2C · 2026-10-25
- Do Chinese AI Labs Offer OSS Maintainer Programs Like OpenAI and Anthropic? — DueHearing1315 · 2026-10-25
- Data broker procurement postings are a public roadmap for frontier labs — sytelus · 2026-10-25
- Audi bets $4B on first US plant in Georgia days after unveiling robotics initiative with NVIDIA — NinaDSchick · 2026-10-25