mdritch

I recently did some analysis on how different style prompts like "be succinct" and "avoid mannered prose" affect LLM outputs and I also saw a ~circular relationship with a 2d dimensional reduction (MDS) of the similarity matrix. I wonder if it's some artifact of PCA/MDS + Gemma's output dist

Glad you got something out of it! Yes, it's hard to tell when something should be in claude.md or added to each subsequent prompt in a session. I need to do a similar analysis here for multi-turn settings and especially coding settings.

Sure! Roughly we are using hierarchical models for reads on underlying count or prevalence data. We use those higher-order means or other fit params to kick off remediation tasks at different levels of that hierarchy depending on those higher level params. We assume some correlation between sibling nodes in that hierarchy.

Question: Can one or another of those thresholds in sibling or parent nodes ever be met if some number of the samples are below some floor reading? Or, how many zeros does it take to silence a threshold check on the node itself, a sibling, or a parent?

To make this tractable I have tried gridding fit parameters, freezing randomness, and using simplified algs like original Metropolis-Hastings

I love z3. I used it for the first time for Jane Street's puzzle last year involving a hashing alg disguised as a neural network. I use a lot of MCMC at work and I have made a few small investigations into MCMC model formal verification via z3, but nothing real yet. This has inspired me to pick that back up.

It reminds me of a bit I did a few years ago trying to find the "longest straightest" line in Atlanta [PDF]: https://matthewritch.com/pages/LongestStraightestLine.pdf

hilarious thank you

I tried looking and couldn't find a proper price per token for the chat model. It claims to be free in some places. I did find these prices for the other services: Text to Speech (Bulbul v3): ₹30 per 10K characters Text to Speech (Bulbul v2): ₹15 per 10K characters Sarvam Vision: Free per page Speech to Text: ₹30 per hour Speech to Text with Diarization: ₹45 per hour Speech to Text & Translate: ₹30 per hour Speech to Text, Translate & Diarization: ₹45 per hour Sarvam Translate V1: ₹20 per 10K characters Translate Mayura V1: ₹20 per 10K characters Transliterate: ₹20 per 10K characters Language Identification: ₹3.5 per 10K characters

I do this with claude frequently. The best tactic I have found is to include a finished file as a "style guide" alongside your content prompt. To find an appropriate style guide for your subject matter, find paper on the arxiv that covers similar material and click the "TeX Source" link.

I do this with latex beamer for presentations. The problem is exporting what I make there to .pptx for my coworkers to edit. I think people use the templates to more easily convert to powerpoint later

yeah working in "random" (in a domain that sounds interesting and hiring undergrads) labs is a great tactic. do that asap, it helped me

I was using it around that time yesterday and experienced no issues. I am on the east coast though

you could set up a "team admin" user on google drive and then invite your individual team member to that user's files? No matter what service you use, you will always need to set up an admin account

I like the idea. About question 2, I think you need some way to publicly benchmark your stripped-down models' performance. Your models probably won't be able to perform on the standard benchmarks, but the big models will probably be able to work on your custom eval sets, such as those Hindi math problems.

I would publish: 1) your domain specific eval set 2) your model's results on that eval set 3) biglab's model's results on that eval set

That would give users a way to determine if your model is actually capable in that reduced domain

ah too bad.

Not sure, but it looks like the creator is still active at https://x.com/lynnetye you could ask her there