> To release the model both safely and quickly, we’ve tuned these safeguards conservatively—they’ll sometimes catch harmless requests, though they trigger, on average, in less than 5% of sessions.
While I appreciate being conservative, ~5% at the scale Anthropic is operating at is too massive a number. Speaking from my own experience, the actual number is higher than that as well (working on pretty benign tasks such as porting an old open source game into a different language). Opus 4.8 itself even identifies the gaurd's false-positives when its sub-agents are being blocked.
Not sure how you came to draw this conclusion, as there's lots of data out there showing droves of Computer Science graduates here in the states unable to land jobs.
Cruise was actually just about to return to market after the October incident [1]. We had reached efficacy on all (much harder) internal safety benchmarks showing the car had significantly improved.
GM pulled the rug on us a day or two before announcing. The current Cruise CEO wasn't aware at all either. I have my own conspiracies of why GM did this, but GM also has a long history of fumbling the ball.
How was this quantified? If there are numbers to demonstrate it's the most secure compared to previous years, why not show them? Qualitative statements with no foundation only draw speculation.