Joel Martin

If you are into programming languages, or your work/play is polyglot, you might enjoy the mal/make-a-lisp project: github.com/kanaka/mal It is one application (interpreter for the mal Lisp language) implemented in 88 different languages by 36 unique authors.

This is another nice size comparison map. While New Guinea and the UK appear to be of the same size on a Mercator map, NG is three-times the size of the UK. Source: buff.ly/3kx978C

OpenAI certainly made mistakes, and this removes none of their responsibility. But some people think that if we just used LLMs in some safe way, everything would be fine. There is no such option. This technology is extremely capable, and not entirely predictable.

After Mythos/Glasswing and now the Erdős unit distance conjecture, I decided to reread ai-2027.com. Honestly a bit unnerving that reality seems to be well ahead in many ways. 😬

I can't find an option in the Gmail iPhone app to view raw email headers, which is deeply frustrating when I am trying to evaluate a suspicious email to figure out if it's a phishing attack or not

Babashka 1.12.215: Revenge of the TUIs blog.michielborkent.nl/babashka-1.1... One of the most exciting babashka release thus far! (If you encounter the blog post link on HN, please upvote it (please don't share a direct HN link as this is penalized by HN's algorithm). Thank you!) #clojure #babashka

What portal from hell opens and issues forth hoards of ladybugs in the middle of New England January? I need to know

UNIX V4 (written in C in the 1970s) found on 50+ year old 9-track tape is restored, compiled, then boots. www.theregister.com/2025/12/23/u...

How long until people start running an inetd (look it up, kids) for our growing piles of local mcp servers?

I wrote a post about GPT-5.2-Codex-Max porting Photoshop 1.0 to the web with only a few human turns of the crank here: photoslop-1.com/port and the result is here: photoslop-1.com I was inspired by @simonwillison.net 's recent post about sending the agent-LLM off to port an HTML parser

Zawinski's law (www.laws-of-software.com/laws/zawinski/) should read "Every program attempts to expand until it can *chat with an AI.* Those programs which cannot so expand are replaced by ones which can." I'm not sayin' I like it, but it is what it is

Folks: Use AI in hard mode. Use it to be stronger, not as a substitute for your strength. And for goodness sake, don't let it be your voice. It's a great audience and editor to give you feedback and to train and refine your voice. Your voice is you — don't lose it!

Prototypes are the most powerfully tool I know of for increasing the productivity of conversations about what software should do and how it should work They're tools for thinking

I added a basic tool use chapter to the rlhfbook if you've long heard about it and never spent the time to learn the fundamentals of what it actually is and how the data is formatted. (I'm becoming more tool-pilled)

"Design Patterns for Securing LLM Agents against Prompt Injections" is an excellent new paper that provides six design patterns to help protect LLM tool-using systems (call them "agents" if you like) against prompt injection attacks Here are my notes on the paper simonwillison.net/2025/Jun/13/...

By surveying workers and AI experts, this paper gets at a key issue: there is both overlap and large mismatches between what workers want AI to do & what AI is likely to do. AI is going to change work. It is critical that we take an active role in shaping how it plays out. arxiv.org/pdf/2506.06576

Lots of neat stuff in this paper showing 30% of US python commits use AI As of the end of 2024: “the annual value of AI-assisted coding in the United States at $9.6−14.4 billion, rising to 64−96 billion if we assume higher estimates of productivity effects reported by randomized control trials”

39 years ago today.... nedbatchelder.com/blog/202506/...

Completely agree - code is such an interesting application because you get a basic form of "fact checking" for free, unlike with prose simonwillison.net/2025/Mar/2/h...

I think the one counterexample is that it's good at writing a lot of code, which is creating more text. My rough rule of thumb is that LLMs are good when it's easy to verify the output as correct, which is quite a lot of things

mal/make-a-lisp just hopped up to 89 different languages with the addition of a Hare implementation (harelang.org). Contributed by github.com/einsiedlersp...

New release of LLM, accompanied by a new plugin - you can now use llm-video-frames to turn a video file into a sequence of JPEGs and feed those into a long-context vision model like GPT-4.1-mini simonwillison.net/2025/May/5/l...

In which our hero fine-tunes when GitHub actions are run, and battles mysterious YAML and GitHub oddities along the way: nedbatchelder.com/blog/202505/...

@emollick.bsky.social You might be interested in this: bsky.app/profile/kana... I made it to be lower barrier than other agent tutorials I've seen: no API sign-up/credits needed (via free GitHub Copilot) and hopefully it is approachable for less experienced python devs (incremental, diagrams, etc)

The postmortem OpenAI just shared on their ChatGPT sycophancy behavioral bug - a change they had to roll back - is fascinating! My notes here: simonwillison.net/2025/May/2/w...

A 6 step tutorial on creating an LLM coding agent: kanaka.github.io/blog/litellm.... * Uses LiteLLM * About 140 lines of python when finished * Can use the free plan of GitHub Copilot (no API sign-up/credit is needed to work through the steps) * Step 5 of the agent is used to create Step 6

I think the key bit here is that LLMs don't make me X% faster at everything, it makes me able to solve a specific subset of tasks instantly, *including* many tasks that then boost my regular programming. So even though programming is high-leverage, creative, nonfungible, LLMs boost my productivity

Have you sailed "all the seven seas"? Well, what are the seven seas?

🔮 The jagged reality of AI – o3’s real world test https://www.exponentialview.co/p/the-jagged-reality-of-ai-o3 (comparing the best models on real-world tasks) #AI

@simonwillison.net I forked your llm-fragments-github plugin: github.com/kanaka/llm-f... Supports local files/dirs and git working copies. Also adds file path prefix (useful context when working with lots of files). Question: Can fragment plugins add/use options?

The Github Copilot premium requests multipliers (docs.github.com/en/copilot/m...) is probably the most honest information about the real cost of various models. That 50x multiplier on GPT-4.5 is certainly eye watering.

Doing Real Work With LLMs: How to Manage Context https://www.jonstokes.com/p/doing-real-work-with-llms-how-to (this is so good) #AI #prompting #ContextWindows #search

Quantitative LLM benchmarks should probably be generated on the fly rather than a static collection. Think property-based testing (PBT) rather than a pile of individual tests. Reduces memorization cheating and reduces the need for hidden test sets.

Today's rust that surprisingly works github.com/rust-cross/c...

I use a lot of the software I write. I am the OG dumber model.

😆. Particularly for FOSS projects, being able to do synthetic user usability testing across a range of user capabilities is pretty interesting. For many of these applications we're looking to maximize how smart the model is but this is a case where we might not want that

This, interestingly may be a use case for using "dumber" models in the future, even when more capable models exist, serving as a proxy for a lower range of user capability during this sort of synthetic user testing

Props to Google for including O4-mini in their Flash 2.5 release. A model released *yesterday*, while some companies only compare to their own models. Looking good gemini.

My notes on the GPT-4.1 series so far - 1m tokens (across all three models) and GPT-4.1-nano being their cheapest model yet are particularly notable https://simonwillison.net/2025/Apr/14/gpt-4-1/

I'm having a much less productive experience with a greenfield domain bsky.app/profile/baue... I'm wondering if removing the mal implementations from the explicit context is enough to negate the highly likely case that 3.7 Sonnet has mal in its training set