williamstein

> In cryptography, most of the problems believed to be exponentially hard, such as factoring, don't have proofs of an exponential lower bound.

There are subexponential algorithms for integer factorization, e.g., the general number field sieve and Lenstra’s ECM.

LEAN.

If only one company had AI and nobody else did, then that one company would be expected to start seeing an explosion in growth/revenue. As it is, everybody has access to AI, so extensive use of AI should just result in keeping up with the competition. The ROI to spending on AI is "not dying".

Turns into an Ad for exe.dev

> This means that the model can support defensive cybersecurity work while still blocking vulnerability discovery in compiled binaries, which is more commonly used offensively.

Annoyingly, this is a concrete argument that open source software may be easier to attack.

Agreed, it's definitely Kevin. His writing style is unmistakable.

https://www.swiftlatex.com/editor.html for the wysiwyg editor says "We are working hard to fix the editor." It has said this for many years. I think I tried it once when it was live and it was pretty cool. My guess is people observed it could corrupt documents, so it was taken down.

I strongly disagree with "Claim 1: Every goal of code review can be served by agents at lower cost and higher throughput." They define the goals and include this one:

> "Knowledge transfer. Agents can actively generate on-demand explanations, architectural summaries, and updated documentation at merge time, which is a more reliable and scalable mechanism for propagating knowledge than the incidental commentary of a busy colleague.

Knowledge transfer is much more than what they claim. An critical goal of code review is making sure humans do deeply understand overall what software actually does. Code is far more precise and detailed than any "on-demand explanations, architectural summaries, and updated documentation."

This interview https://youtu.be/oWOz2htozfI?si=qdQ0uZRoZOYeThOn from 2 days ago with a top researcher from OpenAI directly addresses the bitter lesson argument and the importance of scaling for the history of their models.

> Some of the numbers that you saw from the number of students who receive failing grades were because we caught them (cheating) and prosecuted them and are sending their cases to the center for student conduct,” Garcia said. According to Garcia, nearly 30 students in CS 10 were caught cheating on take-home exams in spring 2026.

It seems like a big problem that xtermjs has is iOS copy/paste, so I fixed that today [1].

[1] https://github.com/xtermjs/xterm.js/issues/3727

This seems like a useful discussion of the relationship between wterm and xtermjs: https://github.com/agent-of-empires/agent-of-empires/issues/...

Wow, finally an alternative to xterm.js?

I maintain a website of scans of things Grothendieck wrote, pictures, etc. https://wstein.org/sga/

He was NOT reading the code: "For 7 months I'd been prompting and shipping without ever sitting down and actually reading the code Claude wrote."

Mojo is still NOT open source (the standard library is but not the compiler). Open source is table stakes for a modern programming language.

What does " Not mentioning the resets every 1 million customers" mean?

This guy is: https://youtu.be/sxX8BMscce0?si=1MuE3_cCH_uDabrT

SEA with node.js "works" for nearly arbitrarily general node code -- pretty much anything you can run with node. However you may have to put in substantial extra effort, e.g., using [1], and possibly more work (e.g., copying assets out or using a virtual file system).

[1] https://www.npmjs.com/package/@vercel/ncc

The current OpenClaw GitHub repo [1] contains 2.1 million lines of code, according to cloc, with 1.6M being typescript. It also has almost 26K commits.

[1] https://github.com/openclaw/openclaw

FWIW, the Joint Mathematics Meeting is bigger, based on number of registered attendees [1].

[1] https://jointmathematicsmeetings.org/meetings/national/jmm20...

This exact thing solves a huge problem with SEA binaries as he points out in his post. You can include complicated assets easily and skip an ugly unpack step entirely. This is very useful.

I tried to make a version of this using Zig once, but ran out of steam: https://cowasm.org/

Is this open source or source available?

He is trying to use a different phrase “write-only code” to define exactly the same thing Karpathy defined last year as “vibe coding”.

For what it is worth, in my experience one of the most important skills one should strive to get much better at to be good at using coding agents is reading and understanding code.

> Checkpoints run as a Git-aware CLI. On every commit generated by an agent, it writes a structured checkpoint object and associates it with the commit SHA. The code stays exactly the same, we just add context as first-class metadata. When you push your commit, Checkpoints also pushes this metadata to a separate branch (entire/checkpoints/v1), giving you a complete, append-only audit log inside your repository. As a result, every change can now be traced back not only to a diff, but to the reasoning that produced it.

The context for every single turn could in theory be nearly 1MB. Since this context is being stored in the repo and constantly changing, after a thousand turns, won't it make just doing a "git checkout" start to be really heavy?

For example, codex-cli stores every single context for a given session in a jsonl file (in .codex). I've easily got that file to hit 4 GB in size, just working for a few days; amusingly, codex-cli would then take many GB of RAM at startup. I ended up writing a script that trims the jsonl history automatically periodically. The latest codex-cli has an optional sqlite store for context state.

My guess is that by "context", Checkpoints doesn't actually mean the contents of the context window, but just distilled reasoning traces, which are more manageable... but still can be pretty large.

It's a reference to https://news.ycombinator.com/item?id=46723990

I strongly agree. The memory and cpu usage of codex-cli is also extremely good. That codex-cli is open source is also valuable because you can easily get definitive answers to any questions about its behavior.

I also was annoyed by Theo saying that.

There's a plugin that evidently supports ChatGPT Pro with Opencode: https://github.com/sst/opencode/issues/1686#issuecomment-349...