sausagefeet

We do this with OCaml, and it works pretty great. I have a particular style that I like and it takes some working of skill files to ensure it adheres to it, but in terms of just writing OCaml, despite being a relatively unpopular language, I have not had issues in any model that I've used.

Personally, I recommend installing the rpiv suite, I like it a lot because it adds a lot of structure, better than plan mode on CC, IMO.

An Anthropic sub is $200/mo, max. I only use open weight models, but the reality is that $200/mo Anthropic sub is way better value per dollar than an open weight model. I use open weight models for other reasons, but I have to limit myself to flash, and I am very controlling of which providers I use to keep the price down, and I still come out a little worse than Anthropic for my monthly bill.

I don't know how you use your agents and what kind of work you are doing, but my experience is that in terms of pure output per dollar, you cannot beat an Anthropic subscription with an open weight model.

This law seems really hard to write, doesn't it? And what's Bending Spoons done that should actually be wrong? I could imagine such a law helping them. As a small company, companies Bending Spoons is making opportunity for me to build competing products at a price people actually like (in my case, it's actually IBM, but similar disregard for customers). So please, Bending Spoons, buy more companies so we can get some churn on these companies. As long as we can avoid regulatory capture.

I've only used CC and Pi and the point of pi is to use extensions. Luckily you can write them yourself pretty easily.

I cannot get behind CC is the best harness. Once u switched to pi I knew I was never going back and occasionally I need to use CC for my subscription and find it unpleasant. CC is just this incoherent hodge podge of functionality that seems to have no vision. Subjective, of course.

How much electricity do you think a few DGX Sparks consume?

My experience has been that building the thing is not modern "plan mode".

I do not use an explicit "plan mode" that tries to limit what the LLM can do. But I do tell it to explore and research and produce an artifact of what needs to happen to finish a large ticket and then use that as input to the actual implementation. I do have it even write small experiments to verify some-parts of the total design to connect it to reality.

The reason is that there can often be lots of decisions that need a large amount of context about the code base in order to reasonably determine what the right design and implementation is and this doesn't happen (in my experience) with the little Interview dialogue that pops up when it isn't sure. Or, even worse, Opus will just make decisions for you that are incoherent with the goals of the codebase.

So, I think, "plan mode" as an explicit mode is dead, but the act of planning is still a good thing to do.

Is that entirely true? If you have a replication protocol like Viewstamped Replication you have the option to make a trade off for how many machines you the will fail concurrently vs forcing an fsync.

I think I'm a bit of an outlier here on that k never really understood why durability requires fsync. It doesn't help you if the storage dies. If durable storage is important to you, then isn't the only solution to synchronously replicate to a quorum of machines at which point an fsync isn't that relevant? What am I missing?

I wish I could find the quote, I think it was on C2 Wiki, but alas, it was something like:

If you're an artist, programming is your clay, if you're an engineer, programming is your building material.

Point being, which I always connected with, was that programming is a medium for expressing a lot of things. Just like you can use paper and pen to write poetry or a dry technical manual. And everyone gets to decide how they see it and use it.

I have been settling on using Mimo 2.6 flash for discovery/research and glm 5.3 flash for implementation. So far I've been pretty happy with the results, but mimo can be quite slow. If speed is an issue, I just switch to glm 5.3 flash.

Actually, I do have an important fix for a deal we are trying to close that does need to go out. And I am paying GitHub to host this. Once or twice, I can see, but GitHub's SLA is getting worse than just hosting it myself, and that's the entire reason I pay GitHub.

Again, I just don't think that is possible. If you have, even, 100 customers, you're going to have 100 unique code bases? How can a human tell if a prompt is reasonable for a particular codebase? And who is going to train those support people? Who is going to ensure they are adding tests that actually test real and correct things?

Maybe I'm a luddite, but this seems far fetched, to me. I just don't see a path for LLMs to make all kinds of changes to a codebase from non-technical people, correctly. Yes, LLMs can do amazing things, but they still don't have a mental model of what a thing is doing and make all sorts of weird decisions that are not what you want.

I believe LLMs are going to make the bar for a SaaS you would pay for, as a company, higher.

At least they are honest about it:

> The planning is happening openly, including a voluntary separation window. That creates real uncertainty for our team over the next few weeks, but we believe the outcome will be better for it.

No good way to execute lay-offs, my preference would be to do it like a band-aid. What use is it to do it in open unless they plan on having gladiatorial matches to keep your job. Otherwise it's just like a painful game of Duck Duck Goose.

Curl is a lot more complicated than, I believe, you think. Most people know of it simply as a CLI to hit an HTTP(S) endpoint and write it out. But:

1. It supports basically any file transfer protocol.

2. It is a library that is designed for long running processes.

3. Because it's designed for long running processes, it makes use of every trick it can to pipeline and re-use connections and resources.

4. It has an asynchronous API so it can be integrated into any existing event loop.

Is a web browser or database more complicated? Most certainly, they solve really massive problems. But curl is certainly more complicated than probably most application code that uses it.