HN user

jpe90

502 karma

https://jeskin.net

Posts8
Comments143
View on HN

I recently wrote a blog post exploring the idea of interfacing with local LLMs for ambiguous tasks like this. Doesn't that make more sense than coding the neural network yourself? Using something like llama.cpp and evaluating whether a small model solves your problem out of the box, and fine-tuning if not, then programmatically interfacing with llama.cpp via a wrapper of your choice seems more pragmatic to me.

I started blogging about stuff I've learned about that I am likely to forget down the road. It feels so good when you throw up a post, never share it, and still manage to get a thank you email for writing it years later. If you have a folder full of notes, I would recommend turning them into a blog on a rainy day.

I recently learned Fennel and cranked out a Love2D game as well. Not only that, but I discovered it’s trivial to run the game on some spare nintendo consoles I have lying around, because Lua is insanely portable.

In contrast, I spent a ton of time trying to build other lisps that are designed to be embeddable on those consoles and got nowhere.

Fennel rocks. Such a pragmatic and well designed mini-language.

I’ve been exploring this project recently, I’m having a great time with it so far. The many examples and great documentation are much appreciated. I’m very tempted to try it for some upcoming projects building workflow systems at work.

Would it be feasible to fine-tune a large, capable model (like the recent LIMA) on the source code (and maybe a few high quality libraries) of a niche language, such that it's much better at helping you write and understand it?

Imagine how many doors it would open if you could fine-tune models capable of writing language bindings for you and keeping them up to date.

The OpenAI chat completion endpoint encourages the second-person prompting you describe, so that could be why you see it a lot.

You're talking about system prompts specifically right? And I'm assuming the "encouragement" you're referring to is coming from the conventions used in their examples rather than an explicit instruction to use second person?

Or does second person improve responses to user messages as well?

The excellent backwards compatibility rocks too. This weekend I wanted a library to diagram the class hierarchy of a Java project I put together.

I found a 10 year old library that worked with no issues, I was done in minutes even though I rarely use Java or Clojure so my familiarity is low. There are very few languages where I would have a similar experience.

https://github.com/stuartsierra/class-diagram

This sounds super cool, I’m very curious how you implemented it.

I’m only vaguely familiar with the API.. if I had to guess I would say you send

- a system instruction that its job is to filter unwanted content

- examples of unwanted content

- an instruction like “filter the following html:”

For every web request you want to filter, you would re-send all of those messages followed by the page HTML as the final message. Is that close?

Yeah I think people are reading this the wrong way. Software was originally considered unpatentable under 35 U.S.C. § 101, but language was developed to allow software to be patented by narrowing the scope of the claimed inventions to a physical processor which are programmed to contain specific algorithms.

I don't think this is a showstopper in terms of copyright registration of A.I. works, but rather a first step in a back-and-forth process of figuring out the legal structure of how to makes rules and enforcement work.

Use GNU Emacs 3 years ago

I came back to Emacs after trying alternatives for a while. I ripped out most of my config and found it much more pleasant to run close to stock.

I'm having better luck using my web browser for email, sublime merge for Git, debuggers in external programs, etc. If I need to do some edits with multiple selections I open up Kakoune, and if I have a burning need for LSP I switch to VS code.

that was an interesting read, i really need to learn more about compilers!

by the way, the expanding hamburger in the navbar is a little buggy on mobile safari. it only expands at the top of the page, so if you press it while you’re midway scrolled it just pushes the content down a bit.

I actually just finally tried out plan 9 in a VM and liked it so much that I’ve been using acme for almost all my development on mac via plan9port for a few weeks now.

I never bothered checking it out before because I had always seen it described as unusable so I figured it would be a waste of time.

Before I read SICP I had used OOP, functional programming, and lazy evaluation to varying degrees.

In school and independent study I learned how to use those concepts to write programs, but I didn’t really have a clear conceptual understanding of what they really are or how they work under the hood.

I read SICP to learn more about lisp and was surprised to learn a lot about all of those concepts as well. It’s all explained very clearly and there are code examples you can toy with to wrap your head them.

I found it such a gratifying book to read through. And I really regret spending time learning Haskell before reading it, so much needless head scratching.

Copilot is a horrible crutch. I will never use it, because I know how to read the documentation/source and write my own code.

I think it depends on how you use it. I never let Copilot make design decisions, because it kind of sucks at that, and it that carries a lot of risk of wasting more time than it saves in the long run.

But there are some tasks that I think it's well suited for. In a recent commit[1] to one of my side projects, I swapped out one graphics library for another, which involved replacing a bunch of library calls and would have been tedious and boring to write by hand. I turned Copilot on for this, and after I did a couple substitutions by hand it was able to do most of the work and was easily a net positive.

For my own usage, I keep copilot off in emacs, and opt to flip it on with a keybinding if and when I know exactly what the code I want to be written should look like, but actually writing it out would be a pain.

[1] https://github.com/jpe90/model-view-projection/commit/1ed4fb...