Weekly Web – 02/16/2025

NOTE: It’s been hectic time the last couple of weeks, so I’m not able to keep up with too much of the news at the moment. Today will be another light week.

Introducing Perplexity Deep Research

I know this was available with ChatGPT Pro and with Gemini, but I’ve been using Perplexity for quite a while, so I’m looking forward to trying this out inside of Perplexity. It feels like this should be be a really good use case for LLMs.

Run LLMs on macOS using llm-mlx and Apple’s MLX framework
By Simon Willison

I still run most of my LLM (or any AI model) tests on an M1 MacBook Pro. I need to look into using MLX more. It sounds like inference is getting really good using that framework. I know CUDA on Nvidia cards will still outperform it in most, if not all, cases, but I’d like to see how good MLX works on my machine. I also keep meaning to try out Simon’s LLM cli. I usually use Open WebUI and Ollama.

WASM will replace containers
By Creston

I thought this was an interesting idea. I don’t know that I agree with this. I can see it being true in some contexts, but it’s not a universal replacement. There are certainly valid use cases, but I don’t think Docker and the like are in any danger anytime soon. There is some good discussion on Hacker News about this article here.

Weekly Web – 02/02/2025

NOTE: This was another light week where I didn’t get to read up on as many posts as I wanted so not too many links today.

Qwen 2.5 VL!
By Simon Willison

I haven’t played with this model yet, but looking forward to another multimodal model I can run locally on my home server.

Aider

This is something I’m adding to my toolkit to try out in the near future. I’ve seen people mention how great it is online, but I haven’t actually tried it for myself since I have access to other cloud hosted AI tooling. I’m interested to see how this works with my locally hosted models though and to see how this works for me given I’m usually using an IDE for dev work.

Cursor Project Rules (via X)
By Michael Feldstein

Interesting reading through some of these project rules to help Cursor be more helpful without having to put a lot of content directly in the prompt. This seems more flexible than shoving everything into the .cursorrules file while still letting Cursor see whatever context it might need.

VGHF Digital Archive

This isn’t AI or even work related, but if you’re around my same age and were into video games growing up, you might find this interesting. It’s scans of all kinds of popular gaming magazines from the 90s/2000s. I used to *love* reading so many of these. It’s awesome to see them available on the web.

Weekly Web – 01/26/2025

Deekseek R1 on Ollama

This isn’t what I’d normally link to, but the new Deepseek R1 model is available on Ollama. This is such an interesting model because you can see how it reasoned through your prompt to derive its final answer. I haven’t played with it enough to see how useful the small models are for real world use cases, but it’s really interesting to see it question itself and work through the problem.

What I’ve learned about writing AI apps so far
By Laurie Voss

I thought these were interesting observations. I don’t know if I’ve had the same experience using LLMs in every case, but it’s still interesting to see what challenges others are hitting in this space. I think this post undersells LLMs, but I still thought it was an interesting perspective.

AI Mistakes Are Very Different from Human Mistakes
By Bruce Schneier

This was pretty insightful. I hadn’t really thought too much about the mistakes LLMs make. I see them, but I hadn’t really thought about how they do compare to human mistakes other than they do seem to be more random. This isn’t so much a post of actionable things you can do to prevent/catch mistakes, but it’s more about being aware of the types of mistakes you can expect.

Anthropic’s new Citations API
By Simon Willison

I haven’t tried this API yet, but I do plan to work on a RAG type system that has a need like this sometime in 2025. This is pretty interesting, particularly when I think about using something like this for internal support documentation and things like that where you are going to want to pull out specific, accurate information even if you do use an LLM to generate it’s take based on that text as well.

Ignore the Grifters – AI Isn’t Going to Kill the Software Industry
By Dustin Ewers

This is a great read if you’re in software development (as a developer or a manager). I think this is largely the way it’ll play out as well. AI absolutely helps in a lot of cases, but it’s not going to outright replace software developers. There will be impacts, but I think people are too optimistic about what it’ll do at least in the short term. This post feels pretty realistic.

A selfish personal argument for releasing code as Open Source
By Simon Willison

I feel this. I don’t know that I’ve written a ton of code and I wish a lot of it was open sourced for many reasons. Not because it’s amazing code, but I’ve always disliked the idea of never being able to even see my code again to remember how I solved a problem. I don’t think I’ll be able to do this anytime soon, but I still like the idea.