Dave Hudson

software designer | | | | opinions are my own

[Dave Hudson @davehudsonio]

12 months ago I had an idea to build a programming language for AIs

Idea: offer powerful features but remove I/O, so an AI could safely run and commit to the result prior to approval instead of committing to the program that might have bad side effects

https://t.co/IjfHlcTR8S

[Dave Hudson @davehudsonio]

Last week I decided to build an agentic terminal, allowing an LLM to read and control one or more terminal windows alongside a human user. There are a lot of proprietary solutions in this space, so I figured it would be fun to build an open-source one.

https://t.co/bqQ0UoG5o5

[Dave Hudson @davehudsonio]

I've been thinking about why we need to think about operating systems again in an age of AI. I finally wrote down some of my thoughts & described what I've been doing!

AI changes the threat models in completely new ways. It offers new opportunities too: https://t.co/LJ047oUy4n

[Dave Hudson @davehudsonio]

Wanted a new approach to share open source research notes:

1. Take notes in a Markdown file during the day
2. Hand them to an LLM to turn them into dynamic pages for my personal website, and edit all the links for me.
3. "make install"!

https://t.co/73Vt2CCNJE

[Dave Hudson @davehudsonio]

I’ve been updating my personal blog so it now includes all my AI and blockchain posts. Fixed a few broken links along the way! https://t.co/RLxkSB91bH

[Dave Hudson @davehudsonio]

It turns out it's not just humans that are polite to LLMs. Here's grok 4 asking nicely 🤣

[Dave Hudson @davehudsonio]

@alexfinn @nickbaumann_ If you want to give something a try I’ve been building open source software that lets me run exactly the same thing against multiple AI models (you can also dial temperature up and down). Reports token usages so you have clarity on how each model does: https://t.co/73mDV1Z420

[Dave Hudson @davehudsonio]

@svpino This is where “attaching files” in context is not as good as explaining why. There’s a huge difference between “here’s a file to edit”, “here’s a file to understand so you can edit another one”, and “here’s a file I want you to use as a template for the style I want”

[Dave Hudson @davehudsonio]

Build a human-#AI operating system? Challenge accepted!

Open source, bloat free (clean build from scratch - not forking something else), LLM agnostic

https://t.co/73mDV1Ywcs

[Dave Hudson @davehudsonio]

It’s fun to come up with new ways to let AIs and humans collaborate in the same space!

[Dave Hudson @davehudsonio]

That moment where I discovered my LLM gets bored doing repetitive things 🤣

There's a video of my adventures in giving LLMs a calculator to play with at: https://t.co/h4gVO5xsjA

[Dave Hudson @davehudsonio]

Just published v0.12 of Humbug - AI development environment:

- AI responses can now handle tables and lines
- Solidity language support
- Change AI mid-conversation and edit/fork conversations
- Use #LLMs on non-standard URLs
- Squished bugs!

https://t.co/73mDV1Ywcs

[Dave Hudson @davehudsonio]

@DataChaz @ycombinator So essentially- do mostly what software engineers have been doing for years just faster 😉

[Dave Hudson @davehudsonio]

Even with a good spec AIs are very conservative at removing things (but I use such specs extensively- they are very helpful for providing context). I found LLMs tend to assume code is being used elsewhere unless told explicitly it’s not, but if you do give them that reassurance they will remove things. The bigger problem is they exhibit the same tendency to add extra things they “might need” as human devs- but then they’re trained on work from lots of human devs who did just that

[Dave Hudson @davehudsonio]

I finally got time to write something about getting started with Metaphor, and how you can use it give repeatable structure things you want your LLM to do.

It covers the basics of the structure, how to use the open source tools that support it, and how to use them to get your AI models to help you help them do a better job!

https://t.co/fRHu0Krlk9

#promptengineering

[Dave Hudson @davehudsonio]

@aaditsh I do this sort of thing all the time (also change temperatures on the same model and see the differences there too). It’s one of the reasons I built this: https://t.co/73mDV1Ywcs

[Dave Hudson @davehudsonio]

We’ve built an open source #LLM dev environment that does more than bolt AI onto the side of an IDE. The key is conversations & allowing you to build tools to meet your unique needs.

Help us shape the future of #AI software development at https://t.co/73mDV1Z420

[Dave Hudson @davehudsonio]

It seems more senior engineers are waking up to the need for the right context to have AIs do their best work.

I wrote about this problem a few weeks ago: https://t.co/ZBOfEMTNWj

[Dave Hudson @davehudsonio]

@karpathy We firmly believe this is the way to go! We built an open source language and dev tool to capture context and let us have exactly these sorts of conversations with AIs so we could be sure they had everything they need to generate great code - https://t.co/73mDV1Ywcs