The Midas Project
@TheMidasProj
We monitor the practices of leading AI companies.
加入 October 2023
266 正在關注 5K 粉絲
Update: model card is out! The section numbering does not inspire confidence about the effort underlying it... but kudos to xAI for getting this out alongside the model release.
顯示更多
If you take their word for it, Grok 4.6 is now Fable-level at coding performance.
You may recall that Fable was taken off the market *for weeks* due to a single jailbreak, despite a 200-page model card full of safety testing.
Grok 4.6 was released (1) without a model card demonstrating any safety testing whatsoever; and (2) from a company that seems to be orders of magnitude more vulnerable to jailbreaking than its peers, with universal jailbreaks costing only ~$60 to discover.
What are we even doing here?
顯示更多
