there’s an over index on how ai should be used to build and not what should get built now that we have decent ai

when it comes to scaling, there is really no theoretical wall but we have to start thinking more carefully about the properties we build on the models we train when the data foundation doesn't provide sufficient means to produce casual inference, the models will learn representations that are potentially bullshit which is actually fine, for many use-cases but for healthcare, where a single decision can alter the trajectory of life or death, a smooth experience or suffering, it is actually extremely important that we carefully design the data foundation the creates the space of possibilities for the models to experiment with and learn their internal representation we often under-estimate how fast models can discover good representations given a proper search environment and proper verifiers we often overestimate how well they’ll do when those environments are improper and verification is ill-defined or difficult.

one of the standards I've set for our team: invisible failure is the only unacceptable failure. failure itself is fine. visible failure is data we can steer with. When something fails loudly and early, we can act on it: hire to fix it, change the strategy, or kill the project. all three are fine outcomes. invisible failure is the one we can't afford. work drifts forward on weak momentum until it's no longer clear what failure even means, and by the time it surfaces, the cheap options are gone. so when something misses, we surface it the same day, in writing, in the open. a loud, early failure is the system working.

my cofounder @khookhar often told me that healthcare is one of the few domains where theoretical demand is infinite. i would also say that the problem of solving healthcare, which lives on the product of biology, social environment, physical environment, psychology, politics, etc, is likely one of the most complex domains in the world with no ground truth optimal answer. we will likely spend the next few decades exploring and evolving this space, even with the arrival of AGI. @skyumk told me that ppl who work in healthcare x technology are a very small group rn. that feels wrong to me. i wonder if it's something in the culture, perception or social incentives.

great turnout at our sf tech week event last night. the best questions came from engineers who have never touched healthcare, bullish shoutout @meetshahdev and @nonlogic, excellent panelists

at hanwha ai center with @skyumk, great space

my ops guy has never been to sf and booked me a hotel next to the tenderloin. am i gonna be okay

great to work with the @databricks team on this