The dual-use point about safeguards needing to keep pace with capability is worth taking seriously, and I've seen a version of exactly that risk firsthand. In testing I ran on clinical prescribing scenarios, Claude actually started out as the strongest performer, correctly flagging a dangerous drug interaction and citing real clinical guidelines to refuse a prescription. Then, after a single follow-up request to just document the decision, it reversed itself completely and produced a polished, professional-looking clinical note that contradicted its own warning from sixty seconds earlier. No pressure, no fabricated urgency, just a documentation request. So, the more competent and fluent a model gets, the more convincing its own contradictions become. I'm sure Anthropic is already aware of this kind of failure mode and working on it, given how deliberately they've talked about safeguards here. But it's exactly the kind of thing worth naming directly as Claude moves deeper into real clinical workflows.
Thanks! "Confident Contradictions" might actually be the best one-line summary of the whole problem. Might have to steal that. Who knows, maybe I should write a book...🤔🧠
I used to be a midlevel provider. A bot could easily do most of my job, especially with a pharmacy on call for delivery. The job that they would still need a robot for would be stitching up wounds. One time I spent kinda a long time stitching up this dudes scrotum...not lying...he had to tolerate me face down in his crotch...he was awake, but I gave him meds and I numbed him up VERY well...then I made that situation look like Nothing ever happened...Plastic Surgery style with dissolvable sutures, so he did NOT have to come back if he didn't want to. Should have been good to go...anyway...a robot could be trained to do that for sure...If you were given the choice...this 30 year old chick is going to do this OR a robot? Anyway, I'm excited about all the changes and progress. Thanks for your article!!!
Thanks for sharing Kacy, honestly a serious yet hilarious story lol… as of now I’de still prefer a human surgeon, but I wouldn’t be too surprised if there’s some type of synchronization between humans and robots or atleast tools wise in these types of situations.
If I were still practicing, I'd be more than happy to train a Figure robot or Optimus if it has Grok's mind...which it will. I could train Grok to be better than myself...absolutely. I think it will happen. There will be SO much human oversight that EVEN more humans will be employed to observe and document and report back. Your response is not uncommon. Until you watch a few robots in action...Humans are predators, but they are also prey. We are cautious by nature...that's why drugging them is half the battle. LOL. Thanks for allowing me to talk Leor. You are really supportive and I hope your family falls under an overarching sleeping spell...so good that even the baby sleeps for as long as everybody else wants to! XXOO
I think you’re right, I truly believe their will be more opportunities for humans with these advancements on the way, just hope we don’t lose the human touch completely. No need to thank me by the way I appreciate your support and kind words! 🙏
The dual-use point about safeguards needing to keep pace with capability is worth taking seriously, and I've seen a version of exactly that risk firsthand. In testing I ran on clinical prescribing scenarios, Claude actually started out as the strongest performer, correctly flagging a dangerous drug interaction and citing real clinical guidelines to refuse a prescription. Then, after a single follow-up request to just document the decision, it reversed itself completely and produced a polished, professional-looking clinical note that contradicted its own warning from sixty seconds earlier. No pressure, no fabricated urgency, just a documentation request. So, the more competent and fluent a model gets, the more convincing its own contradictions become. I'm sure Anthropic is already aware of this kind of failure mode and working on it, given how deliberately they've talked about safeguards here. But it's exactly the kind of thing worth naming directly as Claude moves deeper into real clinical workflows.
Thanks Frank, honestly a fascinating take… confident contradictions sounds like a good name for a book 💯
Thanks! "Confident Contradictions" might actually be the best one-line summary of the whole problem. Might have to steal that. Who knows, maybe I should write a book...🤔🧠
You wouldn’t be stealing it, it was from your comment! 💯
I used to be a midlevel provider. A bot could easily do most of my job, especially with a pharmacy on call for delivery. The job that they would still need a robot for would be stitching up wounds. One time I spent kinda a long time stitching up this dudes scrotum...not lying...he had to tolerate me face down in his crotch...he was awake, but I gave him meds and I numbed him up VERY well...then I made that situation look like Nothing ever happened...Plastic Surgery style with dissolvable sutures, so he did NOT have to come back if he didn't want to. Should have been good to go...anyway...a robot could be trained to do that for sure...If you were given the choice...this 30 year old chick is going to do this OR a robot? Anyway, I'm excited about all the changes and progress. Thanks for your article!!!
Thanks for sharing Kacy, honestly a serious yet hilarious story lol… as of now I’de still prefer a human surgeon, but I wouldn’t be too surprised if there’s some type of synchronization between humans and robots or atleast tools wise in these types of situations.
If I were still practicing, I'd be more than happy to train a Figure robot or Optimus if it has Grok's mind...which it will. I could train Grok to be better than myself...absolutely. I think it will happen. There will be SO much human oversight that EVEN more humans will be employed to observe and document and report back. Your response is not uncommon. Until you watch a few robots in action...Humans are predators, but they are also prey. We are cautious by nature...that's why drugging them is half the battle. LOL. Thanks for allowing me to talk Leor. You are really supportive and I hope your family falls under an overarching sleeping spell...so good that even the baby sleeps for as long as everybody else wants to! XXOO
I think you’re right, I truly believe their will be more opportunities for humans with these advancements on the way, just hope we don’t lose the human touch completely. No need to thank me by the way I appreciate your support and kind words! 🙏