Max Woolf's Avatar

Max Woolf

@minimaxir

Senior Data Scientist at BuzzFeed in San Francisco // AI content generation ethics and R&D // plotter of pretty charts https://minimaxir.com

2,435
Followers
97
Following
1,251
Posts
03.07.2023
Joined
Posts Following

Latest posts by Max Woolf @minimaxir

Preview
a pixel art drawing of a man in a suit and tie ALT: a pixel art drawing of a man in a suit and tie
10.03.2026 21:13 πŸ‘ 2 πŸ” 0 πŸ’¬ 0 πŸ“Œ 0

ouch

10.03.2026 18:13 πŸ‘ 1 πŸ” 0 πŸ’¬ 0 πŸ“Œ 0
Preview
Meta to Acquire Moltbook, Viral Social Network for AI Agents The world’s largest social media company is buying what may be the world’s strangest social network.

LOL www.bloomberg.com/news/article...

10.03.2026 15:06 πŸ‘ 31 πŸ” 1 πŸ’¬ 2 πŸ“Œ 2

That's my secret: I'm never ok.

09.03.2026 18:29 πŸ‘ 0 πŸ” 0 πŸ’¬ 0 πŸ“Œ 0
Hi all!

Sorry for the belated updates: for once in the past few years, I have been too productive and forgot the update entirely.

I was indeed more productive with my agentic experiements, first releasing miditui and then ballin, which were fun terminal experiments but nothing to change the world.

I was planning on releasing more, but I made a discovery: agents can benchmaxx hyper optimized code in Rust that works correctly, which enables a number of applications I thought would not be possible. So I spent a lot more time researching it and found some interesting things.

I did learn enough such that my opinions on agents shifted, so I wrote my longest ever blog post at 33 minutes: An AI agent coding skeptic tries AI agent coding, in excessive detail. It didn't go megaviral but it received good, human feedback. :)

As noted in that post, I am also now investigating a "port" of scikit-learn to Rust with many features I wish it had: no ETA on when it's done, but it'll be done when it's done.

I got sidetracked and I might have another cool project to release this week as a result of my benchmaxxing research, and maybe a quick blog post this month too.

Again, this Patreon has been very helpful as my AI research costs have increased significantly over the past couple months πŸ˜…

As I noted in the blog post, I'm not burnt out on AiI yet.

Max W.

Hi all! Sorry for the belated updates: for once in the past few years, I have been too productive and forgot the update entirely. I was indeed more productive with my agentic experiements, first releasing miditui and then ballin, which were fun terminal experiments but nothing to change the world. I was planning on releasing more, but I made a discovery: agents can benchmaxx hyper optimized code in Rust that works correctly, which enables a number of applications I thought would not be possible. So I spent a lot more time researching it and found some interesting things. I did learn enough such that my opinions on agents shifted, so I wrote my longest ever blog post at 33 minutes: An AI agent coding skeptic tries AI agent coding, in excessive detail. It didn't go megaviral but it received good, human feedback. :) As noted in that post, I am also now investigating a "port" of scikit-learn to Rust with many features I wish it had: no ETA on when it's done, but it'll be done when it's done. I got sidetracked and I might have another cool project to release this week as a result of my benchmaxxing research, and maybe a quick blog post this month too. Again, this Patreon has been very helpful as my AI research costs have increased significantly over the past couple months πŸ˜… As I noted in the blog post, I'm not burnt out on AiI yet. Max W.

Very Belated January and February Update on my Patreon www.patreon.com/posts/152627...

09.03.2026 17:54 πŸ‘ 1 πŸ” 0 πŸ’¬ 1 πŸ“Œ 0

just high

09.03.2026 16:21 πŸ‘ 0 πŸ” 0 πŸ’¬ 0 πŸ“Œ 0

I legitimately don't understand how people have enough quota to abuse subagents without subscribing to a Max plan.

08.03.2026 20:09 πŸ‘ 1 πŸ” 0 πŸ’¬ 0 πŸ“Œ 0

GPT 5.4 consumes waaaaaay more quota than expected and I can't explain why.

08.03.2026 19:53 πŸ‘ 3 πŸ” 0 πŸ’¬ 3 πŸ“Œ 0
Post image

me irl

06.03.2026 17:22 πŸ‘ 26 πŸ” 1 πŸ’¬ 1 πŸ“Œ 0

I told GPT 5.4 to download/run against their test suite and it passed the ripgrep test cases it found, but for some reason it does not want to run a comprehensive suite. Investigating.

06.03.2026 16:47 πŸ‘ 1 πŸ” 0 πŸ’¬ 1 πŸ“Œ 0

Apparently this implementation beats ripgrep 2-4x in every benchmark

06.03.2026 08:38 πŸ‘ 2 πŸ” 0 πŸ’¬ 0 πŸ“Œ 0
Post image

I saw a joke tweet saying "grep is the most important agent tool" so I decided to benchmaxx a faster grep tool in Rust.

It worked lololol

06.03.2026 03:11 πŸ‘ 25 πŸ” 1 πŸ’¬ 4 πŸ“Œ 0

fuck it, feed the paper to Codex/Opus and let them figure out how to implement it

05.03.2026 19:17 πŸ‘ 1 πŸ” 0 πŸ’¬ 0 πŸ“Œ 0
Preview
Introducing GPT-5.4 Introducing GPT-5.4, OpenAI’s most most capable and efficient frontier model for professional work, with state-of-the-art coding, computer use, tool search, and 1M-token context.

god damn it let me relax for one day please openai.com/index/introd...

05.03.2026 18:10 πŸ‘ 21 πŸ” 2 πŸ’¬ 2 πŸ“Œ 1

Everytime GPL license drama pops up I am reminded why I exclusively open-source using MIT.

05.03.2026 15:53 πŸ‘ 8 πŸ” 0 πŸ’¬ 0 πŸ“Œ 0

I really wish Claude Code had native queueing.

05.03.2026 00:59 πŸ‘ 1 πŸ” 0 πŸ’¬ 0 πŸ“Œ 0
Post image

now this is software development

05.03.2026 00:54 πŸ‘ 1 πŸ” 0 πŸ’¬ 1 πŸ“Œ 0

"I will be more virgin."

04.03.2026 23:45 πŸ‘ 9 πŸ” 0 πŸ’¬ 0 πŸ“Œ 0

Yeah this is exactly the thing. Increasingly you are telling people to not believe their lying eyes when you tell them LLMs are useless. Their actual lives will directly contradict this. They will likely discount most other unrelated things you say as a result. This can’t become a left/lib dogma.

04.03.2026 19:11 πŸ‘ 537 πŸ” 73 πŸ’¬ 40 πŸ“Œ 11
Explain the benefits of Rust's memory safety as if I was a nontechnical teenager.

---

Okay, imagine your computer's memory is like a giant wall of lockers in a school hallway.
Every program running on your computer needs to use some of those lockers β€” storing numbers, text, images, whatever it's working with. The problem is, most programming languages basically hand you the keys and say "good luck." You might accidentally open someone else's locker, forget to lock yours when you're done, or even try to use a locker that's already been emptied out. These mistakes cause crashes, security holes, and weird bugs that are really hard to track down.
Rust is different. It acts like a really strict but fair hall monitor. Before your program even runs, Rust checks your code and says things like:
"Hey, you're done with that locker β€” I'm locking it up for you." You don't have to remember to clean up after yourself. Rust handles it automatically, so you never leave a locker hanging open and wasting space (this is called a "memory leak" in programming).
"Hold on, two people can't write in the same locker at the same time." If one part of your program is changing some data, Rust makes sure no other part can mess with it at the same time. This prevents a whole category of nasty, unpredictable bugs.
"That locker is empty β€” you can't read from it." Some languages let you peek into a locker that's already been cleared out, and you'd get garbage data or a crash. Rust simply won't let you do that.

Explain the benefits of Rust's memory safety as if I was a nontechnical teenager. --- Okay, imagine your computer's memory is like a giant wall of lockers in a school hallway. Every program running on your computer needs to use some of those lockers β€” storing numbers, text, images, whatever it's working with. The problem is, most programming languages basically hand you the keys and say "good luck." You might accidentally open someone else's locker, forget to lock yours when you're done, or even try to use a locker that's already been emptied out. These mistakes cause crashes, security holes, and weird bugs that are really hard to track down. Rust is different. It acts like a really strict but fair hall monitor. Before your program even runs, Rust checks your code and says things like: "Hey, you're done with that locker β€” I'm locking it up for you." You don't have to remember to clean up after yourself. Rust handles it automatically, so you never leave a locker hanging open and wasting space (this is called a "memory leak" in programming). "Hold on, two people can't write in the same locker at the same time." If one part of your program is changing some data, Rust makes sure no other part can mess with it at the same time. This prevents a whole category of nasty, unpredictable bugs. "That locker is empty β€” you can't read from it." Some languages let you peek into a locker that's already been cleared out, and you'd get garbage data or a crash. Rust simply won't let you do that.

Opus 4.6 is not good at metaphors.

04.03.2026 05:33 πŸ‘ 5 πŸ” 1 πŸ’¬ 0 πŸ“Œ 0
Post image

just vibecoded a sick-ass confusion matrix

03.03.2026 15:21 πŸ‘ 13 πŸ” 1 πŸ’¬ 0 πŸ“Œ 0
Post image Post image

update: annoyed it enough

02.03.2026 04:15 πŸ‘ 6 πŸ” 0 πŸ’¬ 0 πŸ“Œ 0
Post image

Here's Opus's optimization results from that prompt, which naturally raised an eyebrow. I tested out the app afterwards and it turned out Opus disabled most of the physics.

02.03.2026 01:31 πŸ‘ 14 πŸ” 0 πŸ’¬ 1 πŸ“Œ 1

Asked Claude Opus 4.6 to make my terminal ball physics app faster and it decided "the rapier 2D physics crate is too slow, I'm implementing my own physics solver from scratch dammit"

It then gave up when it realized how much work that it would be.

01.03.2026 20:53 πŸ‘ 71 πŸ” 3 πŸ’¬ 8 πŸ“Œ 2

I suspect I will be banned from /r/rust.

28.02.2026 20:08 πŸ‘ 1 πŸ” 0 πŸ’¬ 0 πŸ“Œ 0

The biggest mystery of Hacker News I've never solved is how submissions that fell off the front page still get upvotes.

28.02.2026 16:45 πŸ‘ 4 πŸ” 0 πŸ’¬ 2 πŸ“Œ 0
Katy Perry signs up for a Claude pro plan

Katy Perry signs up for a Claude pro plan

what

28.02.2026 06:46 πŸ‘ 398 πŸ” 43 πŸ’¬ 27 πŸ“Œ 16

Sparse arrays are scary

27.02.2026 21:22 πŸ‘ 0 πŸ” 0 πŸ’¬ 1 πŸ“Œ 0

there's a nonzero chance I butchered the Rust technical explanations

27.02.2026 19:41 πŸ‘ 0 πŸ” 0 πŸ’¬ 1 πŸ“Œ 0
Preview
An AI agent coding skeptic tries AI agent coding, in excessive detail No vagueposting here, just look at the Estimated Read Time.

New blog post up: the culmination of my past few months working with agents Opus 4.5 and beyond, and the *many* things I learned. Also, the discovery of an optimization trick with promise.

As a bonus: this post will make Rust engineers very mad. minimaxir.com/2026/02/ai-a...

27.02.2026 18:10 πŸ‘ 58 πŸ” 9 πŸ’¬ 7 πŸ“Œ 2