AST graph in Three.js that lets you explore dependencies in real time through a living 3D space. Experimental code play traversing structures, relations, and flows as if the codebase.
Costs still matter. If the expenses are mainly coming from API usage, then if GPT-5.6 is at least close in quality and available within the subscription plan, it'll likely become more popular to use.
How that affects revenue directly is obvious, but looking at it indirectly and from a long-term perspective, if the same situation happens again and GPT catches up in performance, they'll either have to include it in the subscription plan or usage will drop dramatically.
I jak idzie ? Testowałes Fable ? Od dzisiaj chyba w claude code też zagnieżdżanie subagentow ciekawe jak to się sprawdzi. Szkoda że Fable do 22czerwca w planie bo użycie go przez API to pewnie na większym projekcie z 20tys dolarów lekko pójdzie.
I tak to będzie wyglądać obstawiam w przyszłości wszystko pójdzie przez API a koszty są ogromne.
@TheRealAdamG That's a shame that they do that. I especially don't understand what difference it makes to them. It just means the limit gets used up twice as fast compared to Opus.
Lately I've been using Claude Code Remote Control to orchestrate Codex CLI. I also do it the other way around, but the simplicity of connecting remote control through Claude Code is something you should bring to Codex CLI.
There is basically zero setup. I just run "/remote-control", wait a few seconds, and I'm coding from my phone. Claude brainstorms with Codex, and Codex brainstorms with Claude. It creates a very fluid workflow between the two.
@andrzejdragan Powinno być ambitniejsze zadanie jak w ogóle zrobić żeby się podobało. Poprzeczka się podnosi razem z narzędziami. Inaczej zostaje żłobienie w kamieniu. Można oznaczać wtedy bez komputera i sztucznych narzędzi.
I'm not sure what the difference is. Whether you load the context through memory or directly through the prompt, it shouldn't really affect token efficiency. The model still has to process that information either way. The only practical difference might be that the user spends less time typing and burns fewer calories entering the same context over and over again.
@argofowl@cursor_ai For $20/month, Claude actually gives you a pretty solid Opus usage limit. You can get around eleven 5-hour sessions out of it, and those sessions are quite productive. It's worth prompting efficiently and keeping the conversation active so you don't lose the cache. 😀
Even without moving anything, if the liquid glass effect actually samples the background underneath, then an image vs a video is a big difference.
With a static image, the background doesn’t change, so the browser may reuse or cache part of the result, which means less work.
With a video, every frame is a new background, so it has to sample the area under the button again, apply blur/displacement/refraction again, write intermediate buffers again, and composite the final result.
So yes: it has to distort/blur the current background buffer in real time, because the pixels under the glass are changing every frame.
In short: liquid glass over a static image can potentially be cached, while liquid glass over video has to react to new pixels every frame, so it’s more expensive.
I’m not sure if I understood you correctly or explained myself clearly. Maybe you’re already doing this, but if the issue is recalculating the movement itself, I’d move the element with "transform: translate3d(var(--x), var(--y), 0)" and update "--x" / "--y" on pointer drag, instead of recalculating the SVG/filter values.
Something like:
.element {
--x: 0px;
--y: 0px;
transform: translate3d(var(--x), var(--y), 0);
}
const el = document.querySelector(".element");
el.addEventListener("pointermove", (e) => {
if (e.buttons !== 1) return;
el .style.setProperty("--x", `${e.clientX}px`);
el .style.setProperty("--y", `${e.clientY}px`);
});
So the filter can stay relatively static, and only the composited element moves.