Leonard Cohen's Hallelujah.

However, I don't know why. I really don't.

imagine the sankey diagram

*sigh* I think I was wrong about this one. I still think you should write something yourself if you want intellectual engagement, but actually sitting down to read this and the punchline is clever and effective.

word macros usually don't come with delusions of grandeur though

Yes, couldn't agree more.

This is a good case of unknown unknowns - genuinely hard to grasp/maybe impossible if you haven't been taught.

Gotcha. I was just elaborating on the way one might approach this *without* these technologies, sorry if that muddied the discussion or was annoying.

I have no dog in this fight but this is the kind of thing that when I tackle it at work I do have a bunch of lawyers or analysts that I more or less completely rely to navigate whilst I develop the subject matter expertise myself. It's very hard.

is positive or negative or whatever.

Versus an LLM approach, which might out of the box try to do all of that at once.

or if you were to do sentiment (I think an LLM might actually be the best thing here) you might want to do aspect based sentiment, so you'd have a little thing that detects the entities ("who") and the topic (the "what") and then a third that that just figures out how whether that relationship

this is a more generic answer about these tasks, but for things like financial docs or news it can quite useful to have things that are really good at extracting specific kinds of tables or figures (and only those) or just handling one specific type of problem that might arise in a task.

for, even now. classical ml and deep learning have a lot of use for this kind of thing I think - anything deterministic you can tune and random/grid search or train.

yes, I would expect so!

I also think a big generalized out-of-box LLM is a perfectly reasonable baseline that will suck a lot, but it is a fantastic baseline to start with.

but yeah, I think about these problem sets a lot. you can use llm solutions but they're not always the right tool to reach

source: I've been doing this kind of thing for a living since 2014.

you could probably do *something* like a parameter sweep with the right dataset by plugging in different models/approaches that all measure using the same metric. certainly more plausible now with CC.

imho its almost certainly a pipeline. I think you'd want to break this out into a bunch of different tasks, and then the work would be figuring out what's deterministic and what isn't, what each of those methods looks like, etc..

Yes. At minimum I think you'd want something fine tuned for purpose. My career's focus has been on extracting data from unstructured text, so my guess is a bunch of classical techniques would get you quite far and then a fine tuned LLM could get you credible results "generally"

not trivial tho.

(Or they have things I didn't want to do myself but approved, like mermaid diagrams or ascii diagrams).

I *am* OK with reading things like this if they state, at the top "PRODUCED IN CONVERSATION BETWEEN *author* and CLAUDE (insert model here), but even then those documents are more usually for the machine than for other engineers/intellectual peers.

Makes sense to me. I'm still working on expressing my objections to things like this, but it basically boils down to:

- must expend effort in order to expect it
- your ideas are your value; being unwilling to express them makes me question that
- if I wanted Claude's ideas I'd ask Claude.

(Disclaimer: I'm in AI and I use Claude to write 100% of my code).

One thing that's banned on my team is using the AI to write things that require real intellectual engagement - briefs, proposals, design documents, etc. I think if you're gonna ask for intellectual effort you have to give it, too.

you want to have the machine write reference docs, fine, but a proposal, a design doc, something where you're expecting real intellectual engagement... gotta do it yourself imho.

I am a someone who works in AI - I really think this is the wrong usage for it. If should I expend the intellectual effort of reading your work if you didn't expend the intellectual effort of writing it?

feels like a violation of a social contract to me. we banned this sort of thing on my team.

rather embarrassed to say it crosses my lips all the time, but I am an MLE instead of SWE. Now I'm wondering if it is offputting to some of my audience - will have to consider.

I can't actually *recommend* this to anyone, it is that harrowing, but it is a very very very effective anti-war film.

Come and See

yes, yes they are. glad to hear you're doing well!

take care faine

The difference between even *top engineers* in my org and *average engineers at my org* is actually getting sharper. Top engineers are literally just putting massive projects on their backs and soaring to deadlines. Now it feels like the biggest blockers are the folks who can't keep up.

Doll, I think you've done it again.

I feel like the comic went tragically off the rails at some point, but I can't put my finger on the exact moment. Could be Loup, but it could also be as early as Tony coming back.