tkellogg

looks like they have it now

https://optuna.readthedocs.io/en/stable/reference/generated/...

I recall when my team chose MongoDB ~2011, Postgres & friends didn't have JSON columns, so there was a lot of extra data modelling that probably was unnecessary.

The biggest use case for MongoDB was for huMongous data. Obvs MongoDB was a good fit, because of the name.

Yeah, I'd argue that transformers created such capital saturation that there's a ton of opportunity for alternative approaches to emerge.

OTOH having the chart start at zero would REALLY emphasize how saturated this field is, and how little this announcement matters.

wow, so easy!

It depends what problem you're solving. If it's a high frequency request, like a chat response, it's far too inefficient. Most web APIs would consider it bad practice to read 2MB of data on every request, even worse when you consider all the LLM computation. Instead, use RAG and pull targeted info out of some sort of low-latency database.

However, caching might be a sweet spot for these multi-modal and large context LLMs. Take a bunch of documents and perform reasoning tasks to distill the knowledge down into something like a knowledge graph, to be used in RAG.

costs rise on a per-token basis. So you CAN use 10M tokens, but it's probably not usually a good idea. A database lookup is still better than a few billion math operations.

researchers don't read tutorials, they cross check each other's work. You need details to do that.

are you pushing PHI/PII through github actions?

Why did WPF go out of style? (or did it? idk) It seems like a good pattern, but it's always good to look at historical cases for the gotchas

nah, it's just proof that I didn't LLM my blogs into existence ;)

FYI for windows, I made this a long time ago:

https://github.com/tkellogg/Jump-Location

Which was fun and all, but eventually replaced by a pure PowerShell implementation that's become far more active:

https://github.com/vors/ZLocation

I made fossil for this purpose. Sometimes the good stuff is regular social posts, sometimes it’s links. The only way to separate the wheat from the chaff is AI (imo)

https://timkellogg.me/blog/2024/01/03/birb

OP here. Sorry you had that experience. I recently started doing it. Prior to the AI images, I got a lot of complaints about a “wall of text”, so I was looking for a way to break it up, but also not invested enough to spend much time on it. AI images gives me a sort of 2-for-1, in that I get decent visuals, as well as accurate alt text. idk, maybe i should back off a bit. Suggestions welcome.

OP here. Yeah, it’s also extremely low-risk. The sensitive part of the algorithm is the embeddings. This part is just making the group human readable.

The reason I didn’t do mistral the first time is because I’m lazy. I gave myself 3-4 hours to get a first pass done, and getting a local model running seemed unnecessarily difficult. It would be a great addition though.

anyone out there generating fediverse feeds for their static site?

not sure, but you could also look at the inverse. e.g. a 90% to 95% improvement could also be interpreted as 10% failure to 5% failure, i.e. half the amount of failures, a very big improvement. It depends on a lot of things, but it's possible that this could feel like a very big improvement.

Here's the lobste.rs post about this post https://lobste.rs/s/5vs9tv/lobsters_hacker_news

Hard to read that and not hear him alluding to the fact that most of the United States is monolingual, and thus linguistically poor. Rural US also frequently exhibits the identity poverty that he talks about. He never explicitly mentions it. I think there's something to this, needs more thought.

did HN become generally more negative over the same time?

the environment

imo it's the form factor that works. ChatGPT is good for doing standalone things like generating a script. CoPilot is good for making incremental changes to an existing code base. Honestly, its REALLY good at it. One of my favorite parts is that it mimics the style of the surrounding code.

ah! that's true, I do wrongly conflate those a lot.

An alarming trend I see with LLM skeptics is this general direction, where whole categories of real humans are about to be classified as sub-human because an AI has now surpassed their abilities. Every time an LLM skeptic uses the "it's not actually intelligent" argument, it comes dangerously close to dehumanizing actual humans.

yes! absolutely.

> I think interpretable is a overloaded term.

Author here, this is basically the tl;dr of the paper I kept referencing throughout the post. My take, I hope I was clear, is that understanding the inner workings isn't very helpful, except for ML engineers trying to debug a model.

I think I'd break the terms down something like

- debuggable: The traditional definition of interpretability

- trustable: What I talk about here

One such project is RWKV[1]. On the open source leaderboard it lived in the middle of the board for a while, so it really is a legit approach, it's just not hot.

[1]: https://huggingface.co/blog/rwkv

Bigger ad-based social media NEEDS to be big. Their investors need to see growth, so they're always investing in new features to expand their user base or grow into new businesses. If it starts shrinking, it collapses.

The fediverse can simply be. Growth isn't required for survival. Operating a mastodon instance can be quite cheap because all you have to pay for is hardware costs & power, etc. No salaries, no R&D investments. Only build features that users want, there's no hidden incentives.

They're difficult to compare

Like laser eyes?

aaah, yeah, i changed the date on it before posting... sorry about that https://news.ycombinator.com/item?id=30990438

Alright, I'm half-way there. It now write JSON logs to `~/.config/dura/logs/*.json`