yuzuquat

I'm curious how this will ultimately pan out. Presently, we scrutinize frontier labs because they're the ones able to easily deploy large agent swarms. It's not hard to imagine that as costs come down and models become better, smaller and smaller labs/organizations will increasingly have this same capability.

As some back of the envelope math, Openai claimed to have used ~300b output tokens to solve navier-stokes. 4x$200 plans allows me to burn ~2-3b input/output tokens a day. What's to stop a few highly motivated individuals with a reasonable background in mathematics from spinning up their own massive agent swarms in 6-12 months?

I think this fragile truce between ai-powered mathematics and overarching mathematical community is just that - fragile. In the same way the stable diffusion broke the dam with the artist community, we're seeing the same here.

reminds me back in the day when we used to do version control with google drive. you generally don't need to do this anymore i don't think. it should be fairly trivial to open the codex cli in the project directory, ask it to do some analysis and planning -> save to md. then opencode with deepseek to read that md and start implementing. consider that some cross-model subagent workflows already exist: i use claude/fable to plan and there are openai plugins to directly spin up codex subagents. this latter approach gives claude/fable more directly control to babysit the codex agents whereas the former is more of a clearcut handoff

would this not be trivially solved by say - removing the websearch skill from the main orchestrating agent and have it always delegate to some subagent? a subagent sans knowledge of any pii would categorically be unable to exfiltrate any information. granted, populating the subagent with useful context stripped of any pii might require a bit of work and not be perfect, i feel like it would take us 90% of the way there. am i missing something?

I'm curious how multiple accounts come into play here. Does anyone know?

my understanding is that careful means cleaning up the dead-man’s switch before revoking

I thought the premise was that he’s an exceptional interviewer (and maybe good engineer), but did no work so as to eventually get pip-ed and laid off. If he was exceptional at getting work done, then you’re right and a lot of the critique is misplaced. Would love to be corrected if that was indeed the case

looking at the docs, this uses swc under the hood

If you're willing to entertain MPAs, there's small little library called swup that creates seamless transitions between routes (eg. /home, /faq) despite them being 2 different domains. I'm trying it out for www.lona.so and it looks pretty interesting so far

Effectively this it what it does - intercept links to other pages, loads them in the background, then swaps out the contents of the current page with the new page. Your MPA effectively becomes an SPA with just some minor rejiggering

I'm building a completely native HTML spa app over at lona.so (just html + js web components and i bundle it myself). Still very much so early days but check it out!

My feeling is that the browser already does this in that it considers all DOM apis within a single 16ms (requestAnimationFrame?) as a single transaction.

The trouble for browsers, is if certain DOM apis have a dependency on the layout of another element. My naive and unvalidated understanding:

    // Good: These DOM calls in a single frame will trigger layout-paint-composite (1 loop)
    - e.style.backgroundColor = "red";
    - e.style.width = "20px";
    - e.style.transform = "translateX(10px);

    // Bad: These DOM calls in a single frame will trigger layout-?-layout-paint-composite (2 loops)
    - ...
    - e.style.height = otherElement.offsetWidth + 200 + "px"
    - ...

The reason being that without knowing the width of "otherElement", there's no way for the js runtime to execute the "e.style.height" line and execution needs to be paused while layout occurs.

If you're looking for a transactional syntax (similar to what you've proposed) that also addresses this though, fastdom looks like a good option:

    fastdom.mutate(() => { element.style.width = "20px" });

I'm not a browser expert though so if I"m misunderstanding something, would love to know.

I’m currently early on working on a webapp/spa with no node and just some tooling I wrote myself (www.lona.so). I plan to document my process and post it at some point. My two points of advice if you’re looking to go the native route is to 1) look into web components and 2) find something to syntax highlight html strings (es6-string-html on vscode).

If you’re totally new and getting started, just throw in some code into an html file and double click it!

This is pretty anecdotal (6 years back) but I've had one professor tell me they were paid $800 for a NYT illustration and $400 for a smaller publication. I might be off with the specific numbers but these are the general ballpark numbers I remember. Nothing about MTG, etc that I can give concrete numbers for.