loveparade

Fair point, I was more trying to make a statement about the amount of training data available, not the "difficulty" of the task. I just used Tailwind as an example because it is so ubiquitous with so much training data for LLMs to learn from, while any niche application doesn't have that.

Used by Anthropic to sell tokens

These studies are meaningless because speedup is heavily dependent on the kind of work you're doing. No doubt that you can do mechanical refactors 100x faster with AI, and also no doubt that using AI will be slower for tasks where it's less about writing code and more about context/world knowledge or building understanding. Averaging across these tasks doesn't make sense because everyone's work consists of a different distribution of tasks.

A frontend dev doing tailwind integration for his day job is gonna see very different speedups than someone working in a niche scientific codebase. Taking the average makes about as much sense as taking the average of the speedup from calculators for a mathematician, a farmer, and an elementary school student.

Senior people get away with it because they understand the impact of their changes since they are intimately familiar with all the systems. So they know that their crappy shortcut doesn't break anything.

A junior employee doesn't have that knowledge, so they are expected the follow rules to ensure that they don't inadvertently break something.

ai generated slop

If you used the same architecture as GPT2 today you're in for a bad time training a new frontier model. It's only because we have dozens of breakthroughs

What exactly are these dozens of breakthroughs? Most frontier models architectures today still look very much like GPT2 at their core. There were various of improvements like instructgpt, finetuning techniques, efficiency improvements with kv caches, faster attention, lora, better tokenizers, etc. Most of these are for making things run faster. The biggest differentiator has probably been data curation and post-training data and the ability to fit more into the model. But I think we had few breakthroughs that would fall into the category of different technologies.

The more AI-generated AI bad stories we get the more likely LLMs will produce more!

I haven't had a drink in 6 months or so. Not because I wanted to stop, I just have not had the desire to drink recently.

Now I would love to tell you about all the amazing magical healthy benefits that have come with that, but unfortunately there are none. I feel no difference at all.

Doesn't matter who reads it. The point is that you will probably never learn to do "high level system design" well if you do not have enough experience writing and refactoring code yourself. It's like you wanting to become the chef of a kitchen and giving instructions without having ever prepped food.

There is indeed something useful about trying to write elegant code. Not because others read it. But because that's how you learn about the engineering tradeoffs and abstraction that exist everywhere.

I give it one to two more years before open source models have fully caught up. Products are commodities and models are commodities too. GPUs cores are still hard to get for inference at scale right now. They need a platform with lock in but unsure what that would look like and why it wouldn't be based on open source models.

Good lucking getting GPUs.

Not really, when OpenAI was formed in 2015 there were no LLMs, at least none that worked well. It was a regular AI research lab mostly doing Reinforcement Learning on game environments like Atari similar to DeepMind. Once they struck gold with LLMs (2019 or so?) and saw there is money to made everything changed, as expected when a bunch of SV types get involved.

I watched the video but I don't quite get it. I feel like I'm missing something? A nicer git workflow is not what I need because I can ask an LLM to fix my git state and branches. This feels a bit backwards. LLMs are already great at working with raw git as their primitive.

I'm curious what their long term vision they pitched investors is.

Yeah, it looks like a model issue to me. If the harness had a (semi-)deterministic bug and the model was robust to such mix-ups we'd see this behavior much more frequently. It looks like the model just starts getting confused depending on what's in the context, speakers are just tokens after all and handled in the same probabilistic way as all other tokens.

For me the reason to add dependencies to my projects is exactly because they are maintained upstream and I don't need to worry about maintaining them myself. If I need to fork and maintain it myself I'd rather write my own version of it that perfectly fits my use case, or use another dependency that is maintained.

Understandable, I'd probably do the same in his position. Still sucks, we've seen this pattern a thousands times before and what happens next is pretty obvious.

I was prototyping something with pi under the hood for a personal project, going to switch off it now.

Are you sure they are not just refusing to solve your UI bug due to safety concerns? They may be worried you'll take over the world once your UX becomes too good.

It really seems it's mostly AI comments on this. Maybe this topic is attractive to all the bots.

You could manage your subscriptions in an RSS reader, that's what I used to do. Each channel has multiple RSS feeds associated with it for different types of videos (live, vod, etc).

The whole Youtube experience has gotten so bad over the years. I love the youtube content, but I wish I didn't have to deal with the UI/UX and recommendations that the YT app forces on me.

Annoying Shorts. I'm trying to keep my watch history clean to "steer" recommendations, but YT keeps adding things to it that I didn't actually watch just because I happened to hover my mouse over a video, etc.

Interesting, I don't like codex exactly because of its built-in sandboxing. If I need a sandbox I rather do a simple bwrap myself around the agent process, I prefer that over the agent cli doing a bunch of sandboxing magic that gets in my way.

What do you recommend? I've tried both pi and opencode and both are better than claude imo, but I wonder if there are others.

I see these analogies a lot, but I don't like them. Assembly has a clear contract. You don't need to know how it works because it works the same way each time. You don't get different outputs when you compile the same C code twice.

LLMs are nothing like that. They are probabilistic systems at their very core. Sometimes you get garbage. Sometimes you win. Change a single character and you may get a completely different response. You can't easily build abstractions when the underlying system has so much randomness because you need to verify the output. And you can't verify the output if you have no idea what you are doing or what the output should look like.

I think it just really depends. There is no fixed rule to how PhD programs are supposed to work. Sometimes your advisor will suggest projects he finds interesting and wants to see done, he just doesn't have time to do it himself. That's pretty common. Sometimes advisors don't have that and/or want students to come up with their own projects proposals, etc.

That has been my experience as well. Most of the value of writing docs or a wiki is not in the final artifacts, it's that the process of writing docs updates your own mental models and knowledge so that you can make better decisions down the road.

Even if you can get an LLM to output good artifacts that don't eventually evolve into slop, which is questionable, it's really not that useful, especially not for a personal wiki.

I assume this means we can no longer use Claude code sessions in editors like zed because it also wraps claude cli via ACP?

Yeah and it doesn't help that the claude CLI itself IMO isn't that great. It feels a bit like a sloppy vibe coded app. So they are forcing me to use an inferior product.

That's why I am using Codex. I slightly prefer Claude in terms of code quality, but it's close, but not being able to use my subscription with other CLIs and apps ruins Claude for me.

Codex, it's much more generous. And doesn't lock you into using their CLI.

Still I'm a bit surprised you burn through tokens that quickly. I rarely ever reach my limit.

I'd disagree. You're not learning anything if you close your eyes and tell your LLM "set up home manager" - but you'll learn a lot if you read the code it produces, ask clarifying questions, and actually try to understand what is happening. It's just a tool that helps you avoid doing tons of research manually by searching google, and helps you avoid dealing with ugly nix syntax.