 I still don't like Accept headers I just found out if you hit https://pypi.org/simple/pydantic/ with "Accept: application/vnd.pypi.simple.v1+json" you get back super-useful JSON about that package... but I can't link to a demo because I can't include the Accept header in a link! Here's a Gist instead: https://gist.github.com/simonw/8cf8a850739e2865cf3b9a74e6461b28 Posted some notes on the new PyPI digital attestations feature released today, providing digital signatures that help demonstrate that the package you are downloading from PyPI was built from a specific version of the underlying code on GitHub https://simonwillison.net/2024/Nov/14/pypi-digital-attestations/  @simon I understand what this does, but I don’t understand the value of it. It provides validation that the build happened on MS’s server and that they used used a specific checkout. But if builds are not reproducible (eg: use unchecksumed external resources), this guarantees nothing. If builds are properly reproducible, what value does the attestation add? Shout out to @simon ’s shot-scraper for grabbing web browser screenshots from the command line. I redid dozens of screenshots for the upcoming book update, and shot-scraper made it way easier than my previous approach with Firefox’s screenshot tool. On the one hand, I'm very sympathetic to the argument that "AI" is an over-hyped buzzword that is rapidly losing all meaning, if it ever had any (beyond being a combination of science-fiction and an academic discipline from the 1950s) On the other hand though, I'm building a feature where LLMs help a user build a SQL query using an English-language question and I need to decide what label to put on that button, it's hard to come up with anything that's as clear as "Use AI to write this query"
Show previous comments
@simon looks like in Chrome, Google is going for "Ask AI" and "AI Assistance" https://developer.chrome.com/docs/devtools/ai-assistance/network @simon “Generate” (or “autogenerate”)? Thanks to the combo of Ollama and the llm-ollama plugin you can now run Meta's Llama 3.2 Vision image model (7.9GB) on a Mac and use it to run prompts against images https://simonwillison.net/2024/Nov/13/ollama-llama-vision/  @simon the 90B (55GB) might confuse people. You do need ~88GB of RAM, not counting your context window, just to run the 90B model size. So 128 GB of RAM, or else you are going to get 1 token per 30 to 45 seconds or more of output while everything swaps around. That small model is going to run very, very well on any M-series Mac with enough RAM. Wrote up some notes on the new Qwen2.5-Coder-32B model, which is the first model I've run on my own Mac (64GB M2) that appears to be highly competent at writing code
Show previous comments
@simon besides offline use and additionally privacy, did you detect any other advantage running locally?  @simon qwen is amazing. It’s the best performing local model in the Home Assistant AI benchmarks. https://github.com/allenporter/home-assistant-datasets/tree/main/reports The more experience I gain as a software developer the less tolerance I have for the idea that something doesn't need documenting if you can go and read the source code instead (That's despite getting much, much better at reading source code to answer my own questions as I gain experience)
Show previous comments
@simon The thing that makes me sad about this discussion is that it has gone on since the late 1950s; there have not been any new arguments in 30 years and the only good explanation was Knuth’s Literate Programming book but nobody ever really understood it. If you do not have at least as many comments in your source code as the actual compiliable source code, you are making a mistake. I should clarify: when I talk about documentation here I'm not talking about code comment style docs - I'm talking about "this is how to use this library / API" docs If your code is clearly written and nothing else ever needs to call it then I don't particularly mind if there's no additional documentation - but if I'm expected to use call your library from my own code I'm very much not keen on being told I have to read all of that code myself just to use it! Like a lot of people I'm really concerned about what the incoming regime is going to do, so here's one small way I'm trying to help: https://jacobian.org/2024/nov/11/digital-security-checkup/ Jamelle Bouie's TikTok account is one of my favorite sources of political commentary right now - he's a columnist for the New York Times who has basically perfected the very different art of TikTok Here's his latest, about tariffs and domestic supply chains https://www.tiktok.com/@jamellebouie/video/7435779269323803947   my other idea that I have _zero_ chance of actually making happen is that Crooked Media or someone like them needs to poach Matt Levine from Bloomberg and stand up a full-fledged competitor to CNBC where they cover economic and market issues from a progressive standpoint. If everyone gets all their financial and economic news from conservatives of course the population is going to keep trusting conservatives on the economy  @glyph I'm down. Building bridges across my experiences, communities, and identities is absolutely the plan. Got distracted digging around in the belly of the MDN browser compatibility tables, and found out their API is served with access-control-allow-origin: *... so now I've built my own little browser support timeline viewer tool! https://tools.simonwillison.net/mdn-timelines#ViewTransition More details here: https://simonwillison.net/2024/Nov/11/mdn-browser-support-timelines/
Show previous comments
@simon you might also be interested in https://bcd-watch.igalia.com to get updates when new APIs get browser support (based on the same MDN data you’re using!)  @simon You might be interested in what we're building over at https://webstatus.dev/ (which does have an API, though it needs docs). You can already do some pretty interesting queries. I want to enable comments on my blog again, but (I'm current possibly overthinking things in that) I'm worrying if I need a privacy policy, or how I should think about things like GDPR, and should users be able to delete their comments? Never thought about this stuff for a second back in the 2000s!
Show previous comments
@simon If you go the federated route, I like how these daily prompts work. See ,https://kmcd.dev/posts/daily-prompts/ for details and then click on the /prompts section to see them in action. (there is pretty low engagement) That said, I saw your post about using GitHub Auth and that's what I default to these days. The stakes are higher for not being a jerk plus you have GH's moderations rules/team should you need to have to report someone. Saw dolphins in Half Moon Bay on Thursday! (I'm pretty sure this is a dolphin and not a porpoise, I think porpoises are smaller) Kicking off our first ever Discord chat + video + live demos Datasette Public Office Hours in 10 minutes time, details here: Here are detailed notes from our public office hours, showing how myself and @alexgarciaxyz imported San Mateo County election results into Datasette, cleaned them up and then used them to build geospatial visualizations in an @observablehq notebook https://simonwillison.net/2024/Nov/9/visualizing-local-election-results/ Wrote up a few notes on trying out ChainForge, a Yahoo-Pipes-style "visual programming" tool for evaluating prompts against different LLMs https://simonwillison.net/2024/Nov/8/chainforge/ @simon Same style, different purpose: https://github.com/comfyanonymous/ComfyUI I am finding myself turning to gpt-4o-mini a whole lot more since they added prompt caching last month - where you get an automated 50% discount if you send the same tokens twice or more It is fantastic for use-cases like answering questions about a medium sized codebase Hearing about the death of June Spencer at the age of 105 - who was in the Archers from 1950 to 2022 - made me curious as to the world record for longest time playing the same character Depends on how you count: June started in 1950 but took some breaks for family, while Patricia Greene's Jill Archer started in 1957 and has a 67 year uninterrupted run https://en.m.wikipedia.org/wiki/List_of_longest-serving_soap_opera_actors  Currently enjoying browsing stock photos of pelicans on their nests https://www.alamy.com/stock-photo/brown-pelican-nest.html?sortBy=relevant A thing I have learned about voting in US elections in California is that you should do vote by mail, but be sure to return the ballot before 1st November If you instead of drop off your pre-filled ballot on 5th of November it doesn't get counted for several more days (still waiting here) This is particularly frustrating when one of your hyper-local elections only attracts a few thousand votes total and two of the candidates are within 100 votes of each other! I interviewed Rajiv Sinclair about his team's new project, VERDAD - an outstanding piece of data journalism that tracks 48 US talk radio stations (many in Spanish), archives their audio, transcribes it and uses Gemini 1.5 to help identify potential snippets of misinformation - then presents the results in a UI for human review I'm hoping to turn this into a series of YouTube interviews with people building cool data projects where we nerd out about what they've built and how they built it, so I'm optimistically thinking of this as episode one! https://www.youtube.com/watch?v=t_S-loWDGE0 @simon That is a superb use of LLMs. I've seen a lot of text-classification tasks (that previously required expensive model training) can now be done rather cheaply using LLMs + engineered prompts. Cost and development velocity has improved quite a bit with this new LLM-as-rater approach compared to previous approaches of custom-model-training. The next bottleneck is human evals, but I guess we can't completely remove them until LLMs stop making mistakes. |
Gregory's Server
@simon http content-negotiation sucks. It always gets in the way.
Many thanks to the github people who showed how to do good url design.
@simon FYI: ?format= works.
https://pypi.org/simple/pydantic/?format=application/vnd.pypi.simple.v1+json
@simon yeah the RDF / Linked Data community somehow got itself addicted to content negotiation as a favoured way to publish data views. Horrible!