I just want to orchestrate them lol

But lots of people use it! I am so not a fan of the “cool kids” thing like “go all in [modern data tool]” every 3 years.

I don’t want to work in SSIS to be clear

LLMs are valuable, and they can cause a seismic financial crisis if we are over levered and value very boring things like the depreciation timeline of NVIDIA GPUs (or even the volatility!) incorrectly.

Inspired by the great 2 part series on Search Engine, Colossus

overcast.fm/+ABBVQRXFIz8

If you have any experience in finance or financial history, you’ll know that financial crises are not tied to the underlying value of the assets, not really.

Subprime mortgages were and are valuable, just a little less valuable (and a lot of volatile!) than financial institutions predicted.

Lots of other great things, give it a listen!

overcast.fm/+ABQHd82OHLs

AI also allows you to move further up the stack. James gives the example of submitting a PR to a project instead of merely filing an issue.

I just did that last week with Dagster! Customer had an issue, I submitted a high quality PR and got it merged instead of getting into someone’s backlog.

2/3

With LLMs, maybe we can just bang out all the items in the backlog, and maybe with agents “you” aren’t even doing it.

But how much does that matter to a business or endeavor if most of it isn’t all that important anyway?

Used to be a big Cowen fan, _especially_ of his interviews, which are often surprising and delightful.

Haven’t been listening to them as much lately and tbh I think it’s the way he’s become AI pilled. I use LLMs a ton at work and am generally bullish but somehow feels different

This is the perennial question. Can teach tech but can you teach enthusiasm / curiosity?

Years ago, one of my coworkers was talking about learning this or that new technology, and she said "I thought I was done learning this kind of thing in college" 💀

Probably mostly the latter — especially if you are trying to create lists of patients that have a certain diagnosis, the precision usually isn’t worth the extra explanations / effort, and modeling the ontology onto tabular data is also challenging (which is how everyone thinks)

Ok one last thing! We had access to SNOWMED CT (a true, academic ontology) but didn’t use it much for clinical or claims analytics because it was much more complex and hard to talk about. Talking about ICD codes for diagnoses or CPT codes for procedures was _much_ easier, so that’s what we did

Entered diagnoses using Intelligent Medical Objects diagnosis names (which is close to an ontology). IMO _can_ map to both ICD-9 and ICD-10, so we too the IMO diagnoses for a given ICD-9 set, got the IMO diagnoses, then checked the ICD-10s that matched that and did some analysis for weird cases

I helped with the migration from ICD-9 to ICD-10.

People use ICD as a way to identify patients with specific sets of clinical or billing diagnoses to create patient cohorts for analysis.

When switching from ICD-9 to ICD-10, there was no 1:1 crosswalk for the diagnosis codes 😱

But clinicians ..