Watch Desk posted an update
Adam Mcmurchie argues that safeguards cannot simply be retrofitted to AI models already released, and that future systems may be jailbroken and widely distributed. His essay, “You can’t patch yesterday’s AI”, makes the case that the security problem cannot be solved by patching alone.
Why it mattersMcmurchie also argues that voluntary safety agreements and reactive enforcement are inadequate to address the risks. He warns that a serious misuse incident could prompt a public and regulatory backlash against the technology. It is an argument about the limits of security after release, not evidence that the outcomes he describes are inevitable.
Discuss: Should powerful AI developers face binding duties to secure older models, or is that an obligation no one can realistically enforce?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.