samrolken

Cute how Apple is dinged for having a “proprietary” App Store but not Google.

How can I retrieve an API key from ChatGPT to use my subscription in other tools then? This seems like it could be useful.

Claude Code does a lot regarding optimizing context usage, tool output, sub-agent interactions, context compaction, and stuff like that. I don't imagine OpenCode has the same financial incentive to decrease the token cost Anthropic takes on under the subscriptions.

Claude Code does a lot of work in optimizing context usage, how much output is included by tools and how that's done, and when to compact. This very well may make the cost of providing the subscription lower to Anthropic when Claude Code is used. It's well within the realm of possibility if not likelihood that other tools don't have the same incentive to optimize the buffet usage.

Not sure where that goes in the analogies here but maybe something about smaller plates.

Wow, thanks everyone. First HN post ever and it’s this intentionally terrible experiment that I thought was the dumbest weekend project I ever did, and it hit the front page. Perfect.

I’ve been reading through all the comments and the range of responses is really great and I'm so thankful for everyone to take the time to comment... from from “this is completely impractical” to “but what if we cached the generated code?” to “why would anyone want non-deterministic behavior?” All valid! Though I think some folks are critiquing this as if I was trying to build something production-ready, when really I was trying to build something that would break in instructive ways.

Like, the whole point was to eliminate ALL the normal architectural layers... routes, controllers, business logic, everything, and see what happens. What happens is: it’s slow, expensive, and inconsistent. But it also works, which is the weird part. The LLM designed reasonable database schemas on first request, generated working forms from nothing but URL paths, returned proper JSON from API endpoints. It just took forever to do it. I kept the implementation pure on purpose because I wanted to see the raw capabilities and limitations without any optimizations hiding the problems.

And honestly? I came away thinking this is closer to viable than it should be. Not viable TODAY. Today it’s ridiculous. But the trajectory is interesting. I think we’re going to look back at this moment and realize we were closer to a real shift than we thought. Or maybe not! Maybe code wins forever. Either way, it was a fun weekend. If anyone wants to discuss this or work on projects that respond faster than 30 seconds per request, I’m available for full stack staff engineer or tech co-founder work: sam@samrolken.org or x.com/samrolken

This was an unserious experiment meant to illustrate the gap and bottlenecks that are still there. I agree that there's a lot that could be done to optimize this kind of approach. But even if you did, I'm not sure the results would be viable and I'm pretty sure classic coding (with LLM assistance and all) would still outperform such a product.

Most of today’s top models do a decent job with assembly language!

As an unserious experiment, I deliberately left this undefined for max hallucinations chaos. But in practice you could easily add the schemata for stuff in the application-defining prompt. Not that I’m saying that makes this approach any more practical…

No, I wouldn’t say that my hypothesis is that non-deterministic behavior is good. It’s an undesirable side effect and illustrates the gap we have between now and the coming post-code world.

Exactly. It even includes built-in prompt injection as a "feedback form".

I thought about doing that, or having the LLM create and save HTML components, but for this particular experiment I wanted to keep it as pure and unfiltered as possible.

As soon as they put a persistent Salesforce brand banner across the top which did nothing but waste space and put that ugly logo in our face every day, my team started our transition off Heroku pretty much right away.

I might consider something like this, but I can connect four monitors to my MacBook Pro. How easy is it to find a Linux-friendly laptop which can also do this?

I’d guess they are using Haiku for this.

They are saving this for some future release I would guess. A “personalization”-focused update wave/marketing blitz/privacy Overton window shift.

Not just that, but even:

However, tools developed for Hollow will need to be open source to be approved

Weirdly contradictory.

I have a lot of luck using 4o to build and iterate on context and then carry that into o1. I’ll ask 4o to break down concepts, make outlines, identify missing information and think of more angles and options. Then at the end, switch on o1 which can use all that context.

I think it’s possible that it can be a useful step. Fleshing out vocabulary can help with all the real-world use, because you’ll have more context and can fill in meanings, and get more out of real world use. But probably only as a booster to real world use. By itself, it does very little.

Why must LLMs or “AI” beat or match the smartest and most capable humans to be considered to solve a real problem? There’s been a lot of technology invented and in widespread use that solves real problems without having human-like intelligence.

I have a bash script which is very similar to this, except instead of dumping it all into one file, it opens all the matched files as tabs in Zed. Since Zed's AI features let you dump all, or a subset, of open tabs into context, this works great. It gives me a chance to curate the context a little more. And what I'm working on is probably already in an open tab anyway.

Context for LLMs, and use cases uniquely enabled by LLMs, mostly I think.

Threshold cryptography might be what you're looking for.

I’m not a fan per se of TE, but I did get the OP-Z as a continuation of playing with the POs.

I still like it but I’m already trying to find something to eventually replace it with. But is there really something out there with similar size, features, and price as the OP-Z? I would like to find something.

I've been maintaining my own script that does something like this. A few things I've found to be useful: It can also pull in context from git history (prioritize based on which files have been most recently worked on, include recent git commit messages which can help the LLM know more about what's going on), and optionally it can go do multiple stages... for long files, first summarizing them, then including the summary in the final prompt.

After a while of following him, some bot on his account DMs me about some crypto or whatever stuff.

Comments are most helpful when they explain something that is not obvious by just looking at the code. Comments that just explain what the code plainly does don't add much value. This kind of technology can still be useful in other forms as other comments here note, but this kind of auto-commenting, I really don't see it catching on.

If you have a git client for iOS like Working Copy, then it's pretty straightforward to set it up.

You can sync just fine without a subscription. As you already mentioned iCloud, just put your vault there and get syncing for free.

I have a few X220s and I really love them.

I would never say they could be my daily driver or come close to the capabilities of today’s machines.

But for me they really have a place. For hobby and random uses they are great. For a machine to hook up to various amateur radio equipment or some arduino gizmo or to try out that random Linux distro or hobby/toy OS… In this kind of role this machine absolutely rocks.

OpenAI's products and DALL-E 3 can't really transform existing images. GPT-V lets it see images, and DALL-E 3 can make images. But they aren't directly connected.

We'll analyse your photo, describe it in words, and generate a new image from scratch.

I've tried the same technique to get around "copyright policy" refusals, but never get anything close to what I want.