January 2025

There were 4 posts published in January 2025 (this is page 1 of 2).

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.

Weekly Web – 01/19/2025

Quick Note – I didn’t read through as many items this week, so this is an extra short list. I’ll get back to my regular reading schedule in a week or two.

o1 isn’t a chat model (and that’s the point)
By Ben Hylak

I thought this was an interesting observation about how to get the most out of o1/o1 pro. I hadn’t really been making too much of a distinction between o1 and 4o as far as how I prompted. I’ve tried this approach a few times now and I do get more useful answers. It’s not bad with a “regular” prompt, but it does seem to be more helpful when you don’t guide it to the answer.

I Ditched the Algorithm for RSS—and You Should Too – Joey’s Hoard of Stuff
By Joey Einerhand

I’ve been using RSS for a very long time. I usually use NetNewsWire, but my usage is really just blog feeds. I like the idea about even pulling RSS for Reddit and several other sites. I have an automated MailBrew-like tool that pulls Reddit & Hacker News and emails me each day, but I’d prefer that to be in RSS…I hadn’t really thought about whether that was possible or not. This gives me several ideas about new RSS sources I should be using.

Weekly Web – 01/12/2025

NVIDIA Puts Grace Blackwell on Every Desk and at Every AI Developer’s Fingertips

I don’t know if I’ll personally try to buy one of these or not, but I did sign up for the email list to see when it’s available. This sounds like a great idea for anyone wanting to do experiments locally with larger AI models. I’m running some larger models on a 96GB of RAM server, but it’s all on the CPU so it’s super slow. This Digits device seems relatively cost effective (for what it is). It sounds like a 5090 might be more powerful, but way less RAM, so this is a good balance of GPU power with unified RAM. Yes, you can get an Apple machine that can match the unified RAM, but at a higher cost AND I’ll be very, very shocked if this machine’s performance doesn’t blow Apple Silicon out of the water for these AI model specific tasks (it may be worse in every other way, but not that one).

34 Life Lessons from 34 Years The Curiosity Chronicle
By Sahil Bloom

Lots of good advice in this one. I always enjoy seeing what general life lessons learned posts. I can certainly relate to quite a few of them, but I always find a few things that make a lot of sense, but I’ve never really thought about before.

Agents
By Simon Willison

This is Simon’s commentary on another article. You should read the original article too. I’m just getting started understanding agents. I certainly can’t process everything in the longer article, but I’m going to come back to this once I understand some basics better.

I Quit! The Tsunami of Burnout Few See
By Charles Smith

I think this is an interesting take on why we see so much burnout in the tech industry (I’m sure it’s everywhere, but I’m most familiar with tech). I think there are quite a few good points in here about what’s happening and how it’s not just “overwork” that causes all of this – it’s larger than that.

Mastering Long Codebases with Cursor, Gemini, and Claude A Practical Guide
By wheattoast11

I haven’t tried this yet, but I do a good bit of work with Cursor, so I’m really interested in giving this a try for some larger codebases I work with. It sounds like a great idea and I can see how it would help. I’m hopefully going to get a chance to try this out in the next couple of weeks.

Okay, Gemini 2.0 might actually be better than ChatGPT Pro. (Twitter Thread)
By Leo Grundstrom

This is a Twitter thread, but I thought it was interesting enough to include here. It’s a handful of use cases where Gemini 2.0 beats ChatGPT Pro even though Gemini is free (at least free enough for smaller experiments). I don’t work with Gemini much, but I have wired it into my home Open WebUI server so I’m going to give some of these a spin.