I'm starting to believe that ExitStack() is a code smell.

I made an agent that unsubscribes you from spam. It controls a web browser to navigate an unsubscribe page. The agent runs JavaScript to perform actions on the page, and gets to see a screenshot after each action it takes. Blog post: blog.aqnichol.com/2025/09/20/u...

Don't use floats for money.

Damn birds are striking again?

There was a ~month where, instead of opening this app, I was learning graph theory. My mind felt so much more at peace. What should I learn next?

I don't understand laws around music. 👎 copy a few words from someone else's song. 👍 re-record your entire album word-for-word so the old version that you sold to a company for millions of dollars becomes worthless to them.

Why don't we have a programming language for deep learning where dtypes and shapes are part of the type system?

The "baseball Karen" has a narrow opportunity to start a podcast and scam millions of people with a memecoin.

completely messed up that I can try to type `10.9.0.4` and instead type `10.9.04` and it still works.

Pure evil: "you just checked out, you have exactly two minutes to get 20% off something else"

Had a dream the other day that I met someone who went insane after being unable to pin down a memory corruption issue in a Flash game 10 years ago. I tried to help them by digging up the old code, and the first step was even figuring out how to run Flash games anymore.

The other day I noticed that a service I wrote (and self-host) has a memory leak and has been getting OOM killed once every ~15 days for the last six months. I only noticed because my server was slower than usual, and I found that almost all memory and swap was full.

Apparently I am anonymous.

Tbh, I don't even know how to get to a chat UI where I can interact with Gemini. Never looked into it and was never presented with it. Google completely failed with this.

Neural networks with multiple features dimensions aren't safe. What happens when one feature disagrees with the other 511 of them? Which feature wins? That's why we only use one-dimensional neural networks here at ...

In elementary school, we were given a list of electives to sign up for. I was the only person to choose "Solitaire". The teacher asked if I still wanted to do it, and I remember saying something like "isn't it *supposed* to be a solitary game?"

I feel sad that AI has ruined the emdash for my wife.

Just saw an ad for Claude for the first time.

Sudoku is just a graph coloring problem. The sudoku board is a graph where rows, columns, and 3x3 blocks are fully connected.

I implemented graph coloring using dynamic programming on a tree decomposition. The US map is quite "tree-like", with a treewidth of 6. This makes it really easy to compute colorings and other (typically NP-hard) problems on it.

Someone tell me if you can use Genie 3 to generate an image from on top of an airplane while it sits on the tarmac. E.g. generate an airplane on a tarmac and climb on top of the plane or onto the wing.

It's kinda wild that, for free without a GPU of your own, you can use a pretty solid voice cloning model. huggingface.co/spaces/mrfak...

{{chain}}

If you're worried that AGI will turn the world into a paperclip machine, have you met Jewish mothers? Mention your love of a certain food once, and the observable universe will soon be made entirely of this food.

So what ever happened with that neuralink lossless compression challenge?

Finally glad to see programming Twitter hating on Rust.

Implemented a little LP-based solver for the traveling salesman problem. It can solve a 52 city problem optimally in a few seconds. I also implemented the LP solver itself.

Fun little exercise: take a graph that happens to be planar, and find valid 2D coordinates for all of the points to turn it into a straight-line drawing.

A new argument for air conditioning in Europe just dropped.

I wonder if it'll try, given that it clearly found out the algorithm is unknown to be in P or NP-complete.

If you want to implement WYSIWYG in your website, just put 💩 emojis all over your code because it's likely a shitty website. Most websites are.

So there are no remaining reasons why MLX couldn't have just been a metal backend for PyTorch or JAX?

The only show we actually need.

I'm kind of curious: who actually uses graph theory in their day jobs? As a follow-up, who is using the networkx package in Python?

Who cares? Please PLEASE stop making me take 5 laptops and tablets out of my bag and put each one in a separate bin.

TIL about the disjoint-set data structure. I've implemented a less efficient (maybe original?) nlog(n) structure for this at least a few times without thinking about it. en.m.wikipedia.org/wiki/Disjoin...

When I was a teenager, I had an OCD compulsion to solve the subset sum problem for various things in my head (e.g. evenly divide words in a sentence into two halves). At the time I didn't know what this problem was called or that it was NP-hard. github.com/unixpickle/B...

Amazing question and answer. This is why we still need online forums, regardless of how helpful chatbots are. Nothing beats an answer directly from the author of seminal papers.

Why is every paper about tree decomposition behind a paywall? And not only that, but every paper basically says "and for this key component, see [3]" where [3] is also behind a paywall.

I only trust ugly, low quality footage. When it's high quality and aesthetically pleasing, it's probably AI-generated 😅

Maybe it's just me, but I kind of hope that superintelligence isn't controlled by a group of bribable people 😨

If humans weren't dependent on fossil fuels, would we ever have discovered the Chicxulub crater?

Kinda offended that meta didn't try to recruit me 😂

Every time I close the door of a Tesla, it feels light, cheap and flimsy, like it'll break at any second. Apparently this Uber driver noticed the same thing.

What the heck was my Uber texting about?

Palindrome search with LMs. blog.aqnichol.com/2025/06/24/f...

I could have sworn there was a fork of Go (or a compiler that targets it) with different syntax that included nicer error handling. Did I just hallucinate this project?

I implemented the Christofides algorithm for approximately solving TSPs. Writing a maximum weight matching algorithm based on an old paper was one of the harder things I've ever done.

A Markov model and a transformer can both achieve the same loss on the training set. But only the transformer is smart enough to be useful for other tasks. This invalidates the claim that "all transformers are doing is memorizing their training data".