Dan Turkel

Templates support interpolation in fragments by daturkel · Pull Request #1305 · simonw/llm With this change, you can have a template with fragments: - $input or fragments: - github:$input This works beautifully for a url summarization template, which can be as simple as this: system: You...

Here's a stab at it: github.com/simonw/llm/p...

Alright I've tried combining my github.com/daturkel/llm... with a template gist.github.com/daturkel/4fc... but I cannot for the life of me get claude to consistently stop saying "Based on the webpage content," or "I'll fetch that webpage and summarize it for you" :(

ha it was the process of writing the bash script that made me want to see if I could simplify. the script workaround works fine for now but I might hack around to see if it's possible to make this work in a template

@simonwillison.net Is there a way in llm to have a template pass $input to a fragment loader? E.g. if I want a url summarizer, I want llm to load the input url as a fragment and pass that as context. Just setting $input as a fragment doesn't work.

The market is saturated, but for products which differentiate themselves with features or performance, people seem to be willing to convert at the drop of a hat. Lots of cursor fans switched to claude code seemingly overnight. This feels like a marketing failure (then again, i haven't used q either)

I got web search with exa working pretty easily too: github.com/daturkel/llm...

llm-tools-rag-demo.md GitHub Gist: instantly share code, notes, and snippets.

@simonwillison.net I'm working on rag tool for llms own embeddings db. It works, though I've had trouble getting good results out of it (it really doesn't work well with symbex --nl output; it works ok with docs). Curious what you think

github.com/daturkel/llm...

gist.github.com/daturkel/b27...

I may give it a shot if you're open to it!

GitHub - daturkel/llm-fragments-site-text: Run URLs through Trafilatura to get plaintext Run URLs through Trafilatura to get plaintext. Contribute to daturkel/llm-fragments-site-text development by creating an account on GitHub.

@simonwillison.net I made an llm fragment plugin to use a library (Trafilatura) to locally convert a site to markdown rather than using the Jina API. The experience was super easy and fun.

Is it possible to pass a fragment into a chat though?

github.com/daturkel/llm...

Transforming Location Retrieval at Airbnb: A Journey from Heuristics to Reinforcement Learning How Airbnb leverages machine learning and reinforcement learning techniques to solve a unique information retrieval task in order to…

Loved this post from Dillon Davis at Airbnb on how they evolved their location retrieval system from heuristics to a NN-powered bandit.

Somehow I'd never heard of Monte Carlo dropout, but it looks like a great way to get uncertainty estimates from NNs (eg for bandits).

medium.com/airbnb-engin...

I thought Succession was based on the Murdochs but it turns out that the Murdochs are based on Succession.

YouTube video by Apple Apple Intelligence | Create memory movies | iPhone 16

even apple is selling genai as the way to half-ass a gift for your spouse. there are so many appealing non-dystopian use-cases...why pick this? youtu.be/A0BXZhdDqZM?...

Your marketing worked on me—I was going to grab maybe 6 or 7 and then I decided screw it I'll get the bundle.

Someone can take my photo in public and sell it, and we generally are on board with this. But we probably wouldn't want a billion dollar corporation to monetize photos taken of all people in public at all times. Same principle, different vibes.

Unless your site is very complicated I think you should have more or less the same experience. It's python, not a punch-card powered mainframe!

Benedict Evans has a good essay about how principles don't always hold up at massive scale, which has been useful for me in thinking through some of these issues. www.ben-evans.com/benedictevan...

It does have hot reloading, and it's fairly fast. docs.getpelican.com/en/latest/pu...

You probably found that Pelican and Nikola are the two dominant python choices. They're not as popular as Hugo and some others, but I've been using Pelican for a few years now with no problem. I find it a bit more hackable when I'm fluent in the language, which I'm not for Go etc

That makes sense, I actually recall someone even saying that the research seemed to be outpacing the accessible implementations for sparse neural embeddings. Maybe once it's in sentence transformers, splade will be everywhere 😅

This is not how the industry will win the hearts and minds of its critics...

Bluesky might not have some of the specific subsets of trolls that are now rampant on Twitter, but I don't think they'll solve the fact that people can be much meaner behind a screen and keyboard than they would be "irl" 🤕

If all of this means better standards going forward for opt-outs, that could be a win for everyone.

I think the fairly indiscriminate use of internet data to train commercial generative models has brought awareness to and shifted opinion against similar (if more innocuous) practices that have existed for a while, and the industry may have to reckon with the resulting bad vibes.

Whether I agree with it or not (and I'm still thinking it through), I think the reactions are to the principle of the collection and publication of these datasets in general, not exclusively to the HF dataset. I doubt the posters you're referring to would be thrilled about this one.

Hey Tom, thanks so much for your work on sentence-transformers, it's a terrific library.

When I was at SIGIR this year, SPLADE and sparse embeddings came up a *lot* in talks, but no one was able to point to much adoption of these methods in industry. Have you seen it? If not, why do you think?

A screenshot of a python snippet demonstrating a configuration pattern.

Yeah, it's not perfect and it's a little verbose, but it gracefully handles a few different scenarios (see attached).

(also, love the blog post! this is a hard and common problem and I don't think anyone's solved it super well, including our design described above)

we add methods to our pydantic model which return a dict of all the kwargs needed in a certain situation, so you end up with something like

some_fn_or_obj(**cfg.get_kwargs_for_some_fn_or_obj).

I prefer not passing the entire config object in order to reduce the size of the interface 2/2

We manage configs with hydra.cc structured configs (which are pydantic models) but we don't use its ability to dynamically instantiate objects. Instead, since multiple objects and functions may need overlapping fields from the config... 1/2