Is the drum in alexdeng.github.io/public/files... or is there another resource you'd recommend (until your paper comes out)?

As a _very_ minor point at typical sample sizes, the ANCOVA will certainly penalize you for the two degrees of freedom that you're otherwise stepping out of the model (slight change in error terms). But, I'm looking forward to seeing the paper he is putting together to explain the rest!

@phdemetri.bsky.social would you mind checking my restatement? cc: @chelseaparlett.bsky.social

I have used d2 and mermaid for diagrams with it mostly. True re: GIMP. A web based tool it might have a shot at steering, but it wouldn't be cheap.

Does it do any better with the graphs, graphics, and illustrations if you bounce it through writing code to produce them? Maybe giving it hints about technology that tends to produce more appealing outputs? Or it is just at a loss?

🤦‍♂️Give me the power to give myself the grace I aspire to give to others.

It's a little less insane than it looks on the surface if you consider that someone needs to be able to wake a screensaver from a blueooth mouse or keyboard. But... yikes... it wasn't one of the first 10 error messages I tried to run to ground to solve the problem.

I kind of love using a model from a different family to vet what the first model said. It has seemed to work better than reprompting the initial model with a fresh context window.

Now right an LLM as a judge to evaluate the prompt and evaluate the evaluation...

Relearned lesson for me: Hosted LLMs can sort through a mountain of stuff quickly - so quickly you don't always recognize it is happening. Don't let it happen to you. If you've put in more than 5 minutes trying to shape an Assistant to finish an 'easy' task.. it's not going to get easier.

So, like an animal I looked myself and found out that the relevant data was server side rendered by Svelte. The actual information to render is in a script tag to be rendered locally by Svelte? IDK, but that gave me the hint I needed to hand off and extract the table. tinyurl.com/mwm6wkze

In trying, lots of full page reads and cycling through minor changes in Javascript, like a developer who got stuck and now is just trying things without any grounding hoping that something works.

Challenge 2: Now on the site, Claude has to read the full page content. It tried and thrashed to figure out relevant Javascript, but got nowhere. Full page reads could easily be 100kB each. Claude tried in vain to extract the data we needed useing Javascript and got nowhere.

Solution 1: Used DuckDuckGo's equivalent, duckduckgo.com?q=\orion+spa... where the \ prefix does 'I'm Feeling Lucky' with no extra redirect page (nb of course the URL need to URL Encoded). I expect this may save a number of you some time and cost.

Solution: That was a good first step, but navigating and loading a Claude could navigate that using my Chrome (MacOS feature). But, still, processing 1.5 kb to get a url? That's crazy.

Challenge 1: Claude's default search methods ate through tokens like crazy :money_with_wings:. The search terms and target site are clear, we don't need to search. Google's I'm Feeling Lucky should be great for this. Except they broke it by showing a redirect page.

I wanted to know the range of VRAM used by certain models under different quants (story for another day). There were many better ways to do this I'm sure. But, thinking it would be easy & fast, I asked Claude. Mistake.

If only the messaging were more consistent. Shutting down a lawless executive vs protecting healthcare vs Republicans have the majority in both houses of Congress and can't get their stuff together are really different. But the talking points seem to be all over the map.

#2 seems like a lock. Combining #1 and #3, does that mean 35 days to context change? Any speculation as to what that might be?