alt Hacker NewsIt’s easy to dunk on them because of incentives, but consider for a moment a world in which they have access to internal, non-lobotomized models, have run evals on them and have been genuinely concerned by what they saw.
Replies
CamperBob2 • today at 2:58 AM
but consider for a moment a world in which they have access to internal, non-lobotomized models, have run evals on them and have been genuinely concerned by what they saw.
That world arrived with GPT-2: https://techcrunch.com/2019/02/17/openai-text-generator-dang... , https://www.weforum.org/stories/2019/02/amazing-new-ai-churn...
These people have been distinctly unreliable at predicting the robot apocalypse, yet they demand a degree of control that shouldn't be entrusted to a genuine psychic oracle.
If that's the case, why not publish that research?
Until they can show receipts, we're forced into a binary situation of "Do you trust the CEO of a lab with a trillion dollar valuation quickly approaching their IPO?"
Maybe he's right, but from an outside perspective it just looks like an attempt at regulatory capture to pull the ladder up behind them.