loveparade

For me the most useful thing about plan mode in Claude or Codex is not the plan, I usually barely read it, it's that it forces models to gather as much context as possible before implementing. Without plan mode models tend to implement wrong solutions because they are reading not enough context before writing code.

Another useful property is the readonly nature. I can easily let multiple agents plan in parallel without having to worry about annoying worktrees or conflicts and then I can come back to the plans later.

Of course this can be done with just another prompt, but that's exactly what plan mode is. It's nothing more than a predefined prompt in the harnesses with maybe some extra guardrails (that don't always work)

This is totally on you.

I will forgive it using load bearing as long as it finds the right seams.

I think the scary part is that someone gives gmail access to Claude.

I also thought that's crazy. Why even bother for these kind of bounties.

That's what I thought too. I used speckit quite a bit before and I've had mixed results at best. You are just moving ambiguity and code review from one place to another without really gaining anything.

If you could easily benchmark the quality of code then models would be trained on these benchmarks/metrics.

I have been quite disappointed with Astra. I switched over a week ago and I didn't notice a massive difference compared to Sol at first, but I figured I'd use it anyway because it surely can't be worse. Then I saw the bill, it's burning my subscription 10x faster than Sol for essentially no benefit. Not only is it more expensive per token, it also seems less token efficient. And not obviously any better. I'm back to a combination of Sol + Claude.

I also use Astra at work where I don't need to worry about token cost on highest effort and same story there, I don't see any difference in everyday work other than it being more expensive. Of course my experience is highly subjective, but with how meaningless/overfit the benchmarks are, subjective experiences are imo what matters.

Horrible advice. This may have been good advice 10 years ago, but not today. There are no positions for people who "kind of understand how toy LLMs work" because so many engineers do these days. Most of the real LLM optimization work is at the edge of research and highly proprietary and not something you could ever do without infra that costs millions.

But of course, 10 years ago this wasn't obvious.

I built my own pipeline for this because in my experience the YouTube auto transcripts are quite bad.

I download the raw audio track with yt-dlp, give it to qwen asr for raw transcription, then DeepSeek flash for transcription cleanup, formatting, corrections with web search enabled, etc. I use openrouter to call the models. It's a few cents per video (mostly the asr model at ~12 cents per hour) and quality is great. You could probably use a local asr model if you want to save money, I just don't bother because it's so cheap already.

That only makes sense for videos that don't have official (not auto-generated) transcripts of course. If there are official transcripts that are good you can just get them directly with yt-dlp.

You can do remote control / handoff with it, i.e. start on desktop and move to mobile. doesn't work with the cli

I don't think it's about YC specifically, more about VC-driven enshittification of open source software in general.

Oh no, I was using it. What's a good replacement? Guess I'll go back to tmux.

> And while the agent works, I continue taking my notes of what I'm doing, what's next, what the design/architecture needs to evolve to do and so on.

That is still context switching for me. I do the same. While the agent works, which may be 10 minutes, I start planning follow up tasks, etc. But then eventually I still need to come back to the original task to review the agents work.

I don't switch across projects usually. But I can never get into a flow state like that. But I also don't want to sit there for 10 minutes doing nothing

Exactly. It matters that you enjoy working with your tools. Even if they don't make you 10x more productive, focusing on what you're working on and enjoying the process is the most important part.

Especially in the age of LLMs I think a lot of us hate the current tools that force you to context switch between N agents and terminals. Constant context switching takes the joy out of work and prevents you from getting into a flow state. I hope someone figures out the solution.

I'm sorry but I am still confused. What specifically does herdr give me over tmux? "It is designed for agents" is just marketing speech, it means nothing. Ok, it has mouse support, fine, I don't care about that one but I can see how it's useful to some people. That has nothing to do with agents though. zellij also has a better user experience than tmux, so you could just use that or any of the other dozen multiplexers.

I understand it has the agent pane on the left that tells you what is running but... that's it?

I don't understand herdr, can you explain it to me? I've tried it, but I don't understand what it gives me over tmux other than the agent sidebar. But people seem excited about it so I feel like I am missing something.

I don't know what I am looking at but it is fabulous.

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.