I would expect Andrej's impact to be vastly higher doing what he's doing now, vs the very significant constraints imposed in a big company.
@jeremyphoward
-
Why Limit Your AI Career to Commercial Organizations?
By
–
I imagine he might ask you something similar: why are you limiting yourself to a commercial org at the most pivotal time in human history since at least the industrial revolution?
-
Risks of exceeding safe parameters in AI systems
By
–
It's not at all hard to plug numbers in such that it goes "too far".
-
Plugin Risk Analysis: Existential Risk Reduction vs Global Feudalism
By
–
A simple plugin example:
– It increases by 10x the risk of an irreversible change to a global feudalism where per capita utility decreases by 100x, but also:
– It decreases P existential risk from 1e-5 to 1e-6
Overall expected utility would be highly negative. -
Python Package Releases Show No Clear Productivity Jump
By
–
New @answerdotai research by @R_Dimm & @alexisgallagher looking at whether there's been a clear jump in productivity based on python package releases. tldr: No.
"Relax. You are not missing a party that literally everyone else was invited to." -
Vendor Security Responsibility Over User Workarounds
By
–
Giving vendors a free pass just because their oversights can be fixed by every single user adding additional safeguards is not how secure systems are built.
-
LLM Tool Call Validation: Client-Side vs Server-Side Approaches
By
–
Respectfully, I think you misunderstood it. The article clearly shows how the issue can be fixed both client-side and LLM-side, and suggests both. Structured LLM tool call requests can be easily correctly and fully validated on both ends, but OpenAI does it server-side.
-
Server-Side Tool Call Schema Validation Best Practices
By
–
Not obvious, no. In fact OpenAI already does it correctly server side. There's no reason for a provider to not do this. Or to at least document that tool call schemas are merely a suggestion.
-
Client-side fix implemented in Lisette and Claudette
By
–
It's also fully fixable on the client side (which we've done in Lisette and Claudette).
-
OpenAI Models Avoid Prompt Injection Vulnerabilities
By
–
OpenAI's models don't suffer from this. It's entirely avoidable prompt injection, as Piotr shows in the article.