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.

What does "fully caught up" mean in the context of an ever evolving technology? I think I'm in support of open weight models (though there are safety implications), but these things aren't cheap to train and run. This fact alone gives no incentive for leading labs to release cutting edge open weight models. Why spend the money then give the product for free?

Now if "fully caught up" means today's level of intelligence is available for free in two years, by then that level of intelligence means very little

It’s never free your shifting costs from paying a company for their api use vs the power costs of running it locally.

Sure, but it’ll be orders of magnitude cheaper in a few years. The consumer industry is already moving in this direction, with Apple leading the pack

Yeah I don't understand it, it's a marathon with three companies perpetually a minute ahead, and people keep saying "I expect the stragglers to catch up".

The only thing I can see them meaning is what you said, "in a minute the stragglers will be where the leaders were a minute ago", which, yeah, sure.

By my estimation, there is a point where these models are "good enough" for the vast vast majority of all appropriate tasks, after which point further investment by the major labs will have diminishing returns. While they might stay ahead by some measure, the open models will be good enough too, and I assume significantly cheaper like they are now.

Or AGI hits and this theory collapses, but that's feeling less likely every day.

It's not a marathon, or any race. There is no a finish line. It doesn't matter that much that someone is a minute ahead.

It makes perfect sense if you think things cannot improve indefinitely

They do approximate any function... within the range they're trained on. And that range is human limited, at least today.

Also, there is a good enough point where improvements for a given use case are on heavy diminishing returns

That's fine. I can afford to wait a minute if it means I pay $10/m instead of $5k/m.

What is the transition state where people start using open source models that you imagine actually happening?

Play out a scenario. An open source model is released that is capable as Mythos. Presumably it requires hardware big enough that running it at home is unfeasible. You are imagining that individuals can run it in the cloud themselves for cheaper than api tokens would cost? Or even small companies? And that Anthropic and OpenAI won't be able to cut costs deeper than their competitors while staying profitable?

If it is fundamentally a commodity, that means "running it yourself" also isn't really interesting as a proposition. Many of the world's biggest companies sell commodities. It's a great business to be in if you can sell them cheaper than anyone else.

The value add here isn't the model, it is "having a bunch of compute and using it more efficiently than anyone else".

Why do people have such faith in "open source" models? There's nothing "open source" about them. No individuals have the ability to train such modules. They are just released by companies to commoditize the models of the competition.

If Mythos is the endgame, companies won't release open-weight equivalents, and no private individuals have the capital to train such models.

The open models cannot be taken away. Anyone with the right hardware can host these. Unlike the API/subscription services where you can be banned from, may have drastic price increases or reduction of their limits.

> There's nothing "open source" about them. No individuals have the ability to train such modules.

I expect that people on subscriptions can be asked to donate 1 query a month towards an open source distillery.

It should be good enough to distill SOTA models over time.

The result won't be perfect, but it will be close.

Think SETI@home, but it'll be model distillation instead.

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.

The autoregressive engine should see whenever the model starts emitting tokens under the user prompt section. In fact it should have stopped before that and waited for new input. If a harness passes assistant output as user message into the conversation prompt, it's not surprising that the model would get confused. But that would be a harness bug, or, if there is no way around it, a limitation of modern prompt formats that only account for one assistant and one user in a conversation. Still, it's very bad practice to put anything as user message that did not actually come from the user. I've seen this in many apps across companies and it always causes these problems.

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.

Like he iterates in the blog post multiple times: It's still MIT licensed, you can fork it to your heart's content. Or keep using the mainline and merge new features to your own fork.

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.

The company is from the person who created Flask which has remained very open and free. It might be a little too early to be cynical about this.

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

For what it's worth, it's pretty straightforward to recreate it I found, at least it's base idea. Readline w/ nice output is a bit of a pain, but still, doable, and if you don't care about that part of it, then the overall agent loop that you'd build on top of? You could build it, I promise.

What you are suggesting might sound difficult to some people, it is possible: in the last week I co-wrote (with Antigravity with Claude as the backend) an Emacs package for agentic coding that also just uses ‘rg’ for finding relevant code for the context, call out to a model, and handle creating a diff with merging tools. I love using my own code that provides inside Emacs vibe coding and I would suggest others try building the same thing.

EDIT: here is a link of what I wrote just for my own use: https://github.com/mark-watson/coding-agent

Exactly! The actual base loop of these agents is remarkably simple.

https://github.com/girvo/girvent this is my silly one :)

looks nice!

How does DMCA apply? The license is MIT.

I have a private Gitea exactly for stuff like that and Gitea can mirror GitHub repos out of the box and keep it in sync (and it's Git so you can always revert).

Or fork it?

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.

I don't want the world, sounds like people would start bothering me all the time. I'd never get a moment to relax. But if I had the world, I'd do better with it than what we got.

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

This title might have triggered something in those bots; most of them have sneaky AI SaaS links in their bio.

Honestly, I never expected this post to become so popular. It was just the outcome of a weekend practice session.

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.

They see those hovers as attention. And they likely calculate how long you linger. The lingering tells them a lot when you are infinite scrolling on other platforms.

They would love to have full on eye tracking. So the next best thing is a cursor. (Even though I’d agreed with anyone who says it’s a poor signal.)

I’m suddenly grateful that I use my iPad, so they can’t use that signal on me.

They can see how long you stop scrolling, too.

Why does opening the youtube app sometimes just go straight into a short?

I mean, i know why it does that. Engagement. But I mean what governs it? Just a percent chance? If you havent watched any shorts in a while?

Definitely another example of the shit UI/UX you speak of.