Ben Schmidt

@Armavica@fosstodon.org No it's definitely right! And I'll take an awkwardly-explained log chart over an easy-to-understand but useless line chart any day of the week.

The log that dare not speak its name.

@agoldst@mastodon.social I actually did have a physical Dave and Buster's token sitting on the table behind me as I wrote this -- not sure where it came from/why we didn't get a card. Yeah a word is kind of like a letter if you stretch it: but a piece of type is weird. Each individual type sort representing the letter 'Q' in a type set is actually just one token of the abstract type. (I.e., you have many 'Q's). Although… Peirce also knew about *typewriters*, which only have one instance of each letter.

@agoldst@mastodon.social True though that a numismatic token is not a Peircian token AFAICT. This is just something that popped into my head looking through some coin metadata at Dumbarton Oak a few years ago… Suddenly relevant because I realized that probably people seeing our company bills 'by tokens' think we're like a video arcade making up its own fake money, not realizing we're Extra Serious Corpus Linguists.

@agoldst@mastodon.social Mostly because 'type' in numismatics has almost exactly the same meaning as 'type' for Peirce, while a piece of type really doesn't overlap much with a word? And 'token' has a clear numismatic meaning, while I'm not aware of a typographic one.

Anyone know what the metaphor underlying Peirce's type-token distinction is? I think I always sort of assumed it was about typography, but it's actually numismatics, right? Is that widely understood?

@simon@fedi.simonwillison.net Perfect, thank you!

@simon@fedi.simonwillison.net I think I remember seeing you in 2022 or early 2023 write about how LLMs are really good at helping API design, since they have seen all the APIs in the world and know what developers will be expecting. Do you know where that is? Writing an internal doc about API design, wanted to link to it.

@fotis_jannidis@fedihum.org I mean, my company makes it, so yes… But also I'd say that the difference is:

1. It's better especially for non-technical users -- we have a really GUI so it's actually reasonable for non-programmers
2. We've got some neat local indexing stuff built in for RAG on private documents
3. We have more supported models (all the community finetunes on huggingface+llama.cpp)

OTOH ollama tends to have the latest and greatest models faster than we do.

Today Nomic released version 3.0 of GPT4All, the easiest and best way to run LLMs on your own computer without the cloud. Fully FOSS. If you haven't ever run a private local LLM -- or even if it's been 6 months -- try some of the llama3 or mistral derivatives, and chat locally with your own documents. It's remarkable how good open-source local quantized models have gotten, even as commercial models have been stuck at GPT-4 level for a year.

https://www.nomic.ai/gpt4all

Come work at Nomic! I'm hiring a front end/Web engineer to build the next generation of data interfaces for curating, exploring, and model-building from text and image data. Apply here or pass it on: https://jobs.ashbyhq.com/nomic.ai/42e7a74f-9a26-42fb-9b48-5af6b0698045

I appeared on a podcast with Paul Butler talking about browser technologies: https://podcast.browsertech.com/episodes/ben-schmidt-of-nomic-talks-webgl-and-webgpu-for-ai-embedding-visualization

@eldang@mapstodon.space @JensWiesehahn@fosstodon.org
Fascinating. Yeah I guess some kind of in-process A/B or regression must be it. Both my personal and my work accounts, and my wife's account on her work computer, are back on Web Mercator.

Does anyone know why and when Google Maps switched back from using a globe at the farthest-out zoom levels, and back to Web Mercator? (I would guess the answer is A/B testing revealed people got confused and they turned t off quietly, but maybe there's a story somewhere?)