skybrian

I have sympathy for both sides. I don't blame Chinese firms for attempting distillation attacks and I don't blame Anthropic for trying to stop them.

They're just competing companies with conflicting interests.

If algorithmic theorists don't care about practical performance then why should we care about what they're doing? It's just a strange game they're playing, of no interest to anyone else.

It seems like rather than coming up with more "galactic" algorithms, a more useful thing to do might be to come up with a replacement for big-O that's not so exploitable, and to convince theoreticians to play that game instead.

True, but they'd need to be specifically checking a thermometer or it might warm up a fair bit before being noticed.

When a freezer stops working and nobody notices for a while, the food will spoil then, too.

If it were monitoring-only, rather than connected to the power, it would probably be fine.

Shouldn’t we wait until the S-1 is officially released before jumping to conclusions? Currently the financial data we have is based on leaks.

Oh, absolutely. Just saying there’s something for mathematicians to do before publishing results.

To avoid lockout, you want to have more than one key to open a door however you do it. At least, for accounts that are important to you. It might be that one is a passkey and one is something else, though.

There is a reason that banks send increasingly strident text messages that they will never ask for this code.

It's not that hard. Chrome and something else. (Such as iOS.) A different kind of backup authentication method would work too.

Ideally, passkeys should be cattle, not pets. You should have more than one per account, saved to more than one password manager. Then it's fine if they're not copyable because you can get another.

Is it a correct answer, though? Even with a Lean proof, someone has to be verify that it proved the right theorem.

> see first-hand that large, complex projects that used to take them weeks/months can now be completed by agents with a prompt

Really? I start with a conversation for maybe 5 turns or so, where I ask it what would need to change, what the API might be, any database schema changes, URL schemes, and so on, and finally ask it to break it down into commits. Then I let it go, implementing 3-10 commits at a time via subagents. It usually gets the UI somewhat wrong, so there are followups to fix it. This is with Sol and Luna subagents.

Is that what other people see?

It might superficially look like answers, but apparently someone still needs to interpret the results? Maybe it would be better to treat it like data.

I like this metaphor better: the mathematics community is like a group of ancient astronomers that are unexpectedly getting the output from a modern telescope. "Why even look at the sky anymore," some say. "And we don't even own the telescope!"

Meanwhile, the Rubin observatory is creating an archive that grows by 20 terabytes each night, and nobody really worries about this. Dealing with that amount of data is definitely a problem, but also a nice problem to have, provided that there's a scientific community that can deal with it. Mathematics has unexpectedly become Big Science and there needs to be a better relationship between the funders of the scientific instruments and the scientists themselves.

It's a bit confusing that the output from the scientific instrument looks like badly-written academic papers. Maybe it should be treated more like raw data? The format is not ideal, but historians (for example) often have to deal with far worse.

If you don’t use AI to scan for security bugs yourself, someone else will likely be using AI to find them first. So in that way, at least, LLMs could be considered inevitable.

You still need a runtime for your CLIs though. How does local development work for celld and workerd? Are you moving back to Node for that?

JSR packages can run on Node.js, Bun, Cloudflare Workers, etc. The maintainer uploading the package decides which runtimes are supported.

What does "never use an eyebrow" mean here? I don't see any eyebrows.

I have 11 VM’s running on exe.dev for $20/month. There’s one I use every day for my links website [1] and the rest are mostly-idle development machines. One of them has the git worktree and preview website for a static website that’s hosted on Netlify. [2]

Checking out a git repo in a Linux VM and hosting prod elsewhere (automatically deployed via a push to GitHub) seems like a nice pattern. Since exe.dev doesn’t charge anything to leave the dev machines running, I leave them running.

This is all new. I started using coding agents for development at the end of December and exe.dev looked like a nice way to try it out.

[1] https://skybrian-links.exe.xyz/ [2] https://pages.skybrian.com/

Sure, why not stay on 8-bit? You’re just asserting HDR is important, not giving a reason.

I think you must either have a very narrow idea of how people earn money, or are ridiculously extrapolating what AI might do someday.

Yes, but that’s still a niche use case. Web pages look fine using the colors we already have. Somewhat more vibrant colors are a luxury.

This seems to come down to object model and tooling differences between sandboxed JavaScript and sandboxed Unix. You can make Unix commands do whatever you like, including special commands to send things back to the coding agent. But the output of a Unix command is arbitrary text (by default) and the AI will have to pipe things into some other command when there’s a lot of output. In the coding harness I use, I see OpenAI’s LLMs writing a lot of pipelines, often processing the output with python. Also the whole thing needs to run in a VM.

A JavaScript sandbox works somewhat differently since the output is an object. Tools are functions that naturally return objects. Composing functions and async calls work differently. If a function returns a very large object, the coding harness could be smarter about presenting the result to the LLM. Maybe JacaScript works better than bash for calling mcp APIs?

But you can have both! The LLM can get direct access to a JavaScript sandbox, which in turn provides an API to access a Linux sandbox. This seems to be what pi is doing with code mode?

You’re throwing a lot of bits away, though. If you used four bits per number then you could pack 16 random digits into one 64-bit draw. But a rejection-sampling branch will be taken much more often.

Random numbers aren't precious, but they do take time to generate. Rejection sampling adds a branch and it can matter how often it's taken. In my property-testing framework, generating a large, random array of numbers more efficiently improved performance.

For me a big open question is what sort of progress will we see in robotics. I won't even attempt to speculate, but it does seem hard in different ways than proving math theorems.

A custom Mac application seems limiting compared to a web app. I use a web browser to connect from my laptop, tablet, and phone to use Linux VMs on exe.dev.

But it might be nice to have a nice way to run VMs on my Mac mini, too.

This seems to be running on an FPGA board that costs ~$300? Anyone know more about the hardware?

I believe they are saying that forbidding AI-generated bugs means you don’t find out about security bugs, but they’re still out there.

Whatever the original intention, this is clearly a bug and should be fixed. But should ChatGPT sign its cartoons with its own name or leave them unsigned?