Emergencies aren’t just worse versions of bad days-qualitatively different cognitive states, team dynamics, failure modes.
You can’t prepare for them by just doing more of what you already do.
You have to train (and train your systems) specifically for crisis.
Systems that swarm need to be attuned to these psychological aspects of swarming before the next crisis starts, not just during it.
Would love to hear what you think.
#SangfroidLabs#EmergencyMind#SwarmTeams
https://t.co/pBtpY1UfpC
Let's talk psychology of swarm teams.
Operating as part of a swarm team is different than working on a normal team. And swarming is a distinct skill that requires thought and practice.
https://t.co/pBtpY1UfpC
[3] Meaning: When the team dissolves, swarmers are left to process what happened alone.
No formal elders, no debrief structure, no shared community.
This can be lonely and challenging, and requires a proactive response by the groups that swarmers return to.
LLMs are affecting what people do in emergencies (and not for the best).
More safety research needs to be done by medical professionals. I want to see @ddworkis working with frontier labs (@OpenAI / @AnthropicAI) to improve the way LLMs are trained in this space.
@HeartNews Our recent work looking at model-to-model differences in how AI systems responded to a potential acute stroke, including (incorrectly) different behaviors for weakness vs numbness.
https://t.co/ozimcJ40Ul
@EricTopol@DrHelenOuyang Insightful but challenging since which model you ask and exactly how you describe your symptoms can lead to widely different types of answers even for critical concerns.
https://t.co/ozimcJ40Ul
We really don't know how patients actually describe what they're feeling when having a stroke.
We know "weak" and "numb" are scary, but we really don't understand at the verbatim level of granularity we need.
@AbridgeHQ | @AmbienceAI | @corti_ai - big opportunity to help here!
If you woke up with a numb or weak arm and asked an AI what to do, would it tell you to call 911?
It should. But the answer is: depends which one you asked. And on exactly how you asked it.
https://t.co/ozimcJ40Ul
A thread on #AIsafety, #healthcareAI, and #stroke ⬇️
All three models got distracted by a potential alternative explanation. They also showed strong word-to-word differences, recommending 911 far more often for "weak" than for "numb."
Which model they use and what words they happen to choose to describe what they're feeling could send them in completely different directions with irreversible consequences.
Preprint below - curious what you think.
https://t.co/ozimcJ40Ul
If you run a resuscitation program, train residents or nurses, or think about team readiness in high-stakes environments — try it and tell me what you find. #FOAMed
Live app: https://t.co/0ajtTX4Bar
Code: https://t.co/OkrAujxIf2
Introducing HALOSim! >> Most ER and ICU providers go months between live cardiac arrests.
We published the data: 98% of nurses exceeded a 90-day gap between real exposures. That's a structural exposure problem that most systems have no way to see.
https://t.co/YyqlVTwRfs
You put in your unit's event rate, your team's shift schedule, and your training program (or use simulated data).
HALOSim models critical gaps between exposure, on-shift readiness, and how different training programs might make a difference: