Also got a reminder of how freakishly capable the tiny (1.9GB) Llama 3.2:3B model is at summarization and question answering against documents you feed into it via its impressive 128,000 token context

@simon 1/2: I find AI deeply unscientific the way you interact with it. Knowledge in modernity took a turn epistemologically speaking from god to an agreement like relationship between theory and practice. Knowledge is not what agrees with someone or is true but what represents an agreement between theory and practice. All scientific work revolves around that. These sophisticated text generators are little more than lemma generators.

@simon 2/2: You can look up a mix of words that may be true believing to have gained insight without engaging into building your own model of something or performing your own experiment. Now if it were to force you to do just that... Its a shortcut to mediocrity. I usually use it for tedious tasks that I am unable to automate or ask it to scrutinize my premises, models or experiments but, to no surprise, I rarely find out anything that's not contained in a well curated article or dictionary.

@simon I don’t think AI’s useless. I think it’s being used for all sorts of things it’s not appropriate for because of hype, i think almost everyone who stands to make any money from that happening is annoying, and most importantly i think that it’s more damaging than any use is worth. Like, say, fossil fuels (which is not a bad analogy). This year has only multiplied the above sentiments more. But I suppose this means the poll doesn’t apply to me!

Here's what it said when I asked it to count those pelicans

@simon

It's a fairly good system, I asked it some confusing questions about unlikely images. Like what is a unicorn made out of (chocolate, correct).

Fairly impressive for 64GB. lmk how your local trials go.

@FreakyFwoof @simon When reasoning goes to a path of hallucination, it says all kind of things beyond visual. lol Here's an example: "I can also imagine the sounds that might be present in this scene: the gentle lapping of water against the shore, the creaking of the boat as it moves with the current, birds chirping in the background, and maybe the distant rustling of leaves in the wind."

@simon Amazing! That's the hands-on practical work we need right now!

It's fascinating that a small team can come along, use a few tricks of the trade, train it on eight GPUs, and massively push state of the art. That's similar to what happened with resnet in the vision domain where folks went back a few times over the yeas and showed how it can be improved considerably based on what we learned.

Looking forward to having this power a ColBERT-like retrieval setup!

@simon you could also load it into https://editor.swagger.io

@simon I wish it was standardized as a standalone LLM API. That way it can be expanded with use cases that OpenAI doesn’t cover, and standalone SDKs can be made.

Reusing an API will eventually go wrong as owner of API will only make changes that fit their use cases.

Just found out the hosted Swagger explorer has a ?url= parameter that can load an external schema, so here's that OpenAI OpenAPI description rendered using that https://petstore.swagger.io/?url=https://raw.githubusercontent.com/openai/openai-openapi/refs/heads/master/openapi.yaml

@simon Hah, not a linkblog, but just this minute I pushed a new Notes section on my website (and your blog is among my inspirations/aspirations): https://danburzo.ro/notes/a-note-on-feeds-a-feed-on-notes/

@simon I used to do this in a section on my blog called “RE” https://kevingimbel.de/blog/category:RE

But I’ve wanted to connect it to my LinkAce instance which I use for storing links & sharing them on Mastodon and Bluesky. I agree it’s a good thing to have these sorts of posts!

@simon Thank you for the inspiration. I opened my linkblog with Snac, a minimal ActivityPub server https://href.literarymachin.es/raffaele
Notes on how I run it: https://literarymachin.es/linkblog-fediverse/

@simon I watched it twice; once solo and once with my family. It was still riveting. Great seeing them go through the high-adrenaline season finale!

@simon I rewatched #Severance immediately after first viewing. Damn, it is good TV. Later, having heard it was renewed, I rewatched it again in the expectation it would arrive about a year after the first, only to check Wikipedia and exclaim "_2025?!_ WTF?"

I'll try to catch you up. Currently frantically rewatching S1 of Silo, because I hadn't noticed S2 had started.

Quality US TV arrives all at once like buses. Must be something to do with some inscrutable custom like sweeps week.

The Web was a mistake.

@simon Another thing that is annoying about it is that on mobile if you have extensions to darken the page, they don’t work when in the canvas version.
Also selecting text and images is just not really possible.
It’s not the way I’d like the web to go, BUT it’s a very interesting project nonetheless!

@simon I mean yes, but blind people hate this trick

@simon I need to look into this for maybe my office hours or some other live blogging things that could be fun.

OK, all done (I just went through and added alt text to the images with the help of Claude)

@simon if it’s any consolation, I thought your one-shot Python scripts post was simply top notch. Amazingly helpful, so thank you! 🎁

(I ended up making my own little word counter thing (via Claude) to help me fit my proposal in the 300 word recommended limit https://tools.simonwillison.net/word-counter )

40 minutes left! Still time to get something in if you hurry!

The most impressive example I've seen so far involves vision input:

llm -m gemini-2.0-flash-thinking-exp-1219 \
-a http''s://storage.googleapis.com/generativeai-downloads/images/geometry.png \
"What's the area of the overlapping region?"

Here are my custom instructions which I'm using as part of a Claude Project, but I expect they'll work the same way with other LLMs too

@simon

I'm a big fan of this style; below a certain level of complexity, single-file standalone scripts with the inline requirements - plus uv - are a game changer in being able to easily share and deploy things for non-python developer users.

@simon A couple weeks ago I switched GH Copilot to Claude, but not tried it yet (ChatGPT kept making large messes of my code and I’d manually update and use Claude to debug if needed).

I was really close to dropping paying for Copilot as I pay for Claude.

Free with GPT doesn't get me far.

Have you tried the switching of Copilot to Claude yet?

@simon Your prompt may have become part of the things they do QA on before releasing now. 😅

@simon thanks!

I pasted a Markdown page, and hints seems to be shifted somehow? They don't show the "passive voice" (see screenshot). Anyway this is already useful with just the highlighted words so not critical ^^ Thanks for sharing!

@simon have you seen https://vale.sh/ - a linter for prose. I use it to edit my blogs