@AnthropicAI Smart to run the check right after the response, while the model still has full context of what it just said. Curious about false positives: what share of self-assessment flags hold up as real misuse on human review?
@_vmlops Optimizing the skill.md instead of the weights is underrated: edits are diffable, reviewable, and roll back in seconds, unlike a fine-tune. The real question is overfitting. Does an optimized skill still hold up when you swap the underlying model?
@alex_verem 86% is the scary number. The fix probably isn't "trust the agent less" but making it show the option it didn't pick, with the tradeoff, before the expert commits. Would love to see whether that alone moves the number.
Capability removal holding up makes sense: if the agent that reads untrusted text can't also send or delete, an injection has nowhere to go. Did you test a split setup, a read-only agent that only passes structured fields to a privileged one? Curious how it scored vs plain tool allowlists.
@AnthropicAI Curious how you'll measure success here: papers where Claude helped, or results that actually reproduce in the lab? The second is the bar that would win over skeptical scientists.
@elonmusk Von Neumann probes by 2100 hinge less on propulsion than on autonomous manufacturing. A probe has to mine, refine and rebuild itself with zero human input. Today's AI-run factories are the real dress rehearsal.
@saastr The ServiceNow line is the real signal: 21% growth and still down 14% means the market now prices AI as a threat to seat-based revenue, not a tailwind. Curious whether ChatGPT-native workflows end up cannibalizing ServiceNow's own UI.