eldenring

It says rookie contract + endorsements, and I don't think it seems particularly small, just doing the math. Also I doubt that ex: Thrive LPs would care about this kind of thing. This whole comment just seems wrong.

Huh? of course its enough. Transformers immediately started destroying every single baseline out there. The authors definitely knew it was a very significant discovery beforehand.

Hmmmm I don't think that would make sense. Closest analogy is working with humans. The easiest way to work with a human isn't a thin, limited api, but rather to give them context and work together (employment). I think the future of software will look more like Claude Code. Lets the model work in a similar space as us, where it can intelligently seek out information and use tools as a human would.

I recently found out my website has been blocked by AI agents, when I had never asked for it. It seems to be opt-out by default, but in an obscure way. Very frustrating. I think some of these companies (one in particular) are risking burning a lot of goodwill, although I think they have been on that path for a while now.

I'm not American, I think its sad to see my country dismiss AI and continue to fall behind.

I don't think its controversial to say that in net, the US has had a good streak so far with new technologies. Yeah there have been some pretty big misses, but obviously the wins more than make up for them.

This is how America ends up being ahead of the rest of world with every new technology breakthrough. They spend a lot of money, lose a lot of money, take risks, and then end up being too far for others to catch up.

Trying to claim victory against AI/US Companies this early is a dangerous move.

What do you mean by that list? They pretty much have every single big tech company in their portfolio.

I don't think something like this works when you can change the model, retrain it, etc. Or at least its much more difficult to do.

Cash out is a bit of a negative word here. They've shown the ability to build categorically better tooling, so I'm sure a lot of companies would be happy to pay them to fix even more of their problems.

Its interesting because the value is definitely there. Every single python developer you meet (many of who are highly paid) has a story about wasting a bunch of time on these things. The question is how much of this value can Astral capture.

I think based on the quality of their work, there's also an important component which is trust. I'd trust and pay for a product from them much more readily than an open source solution with flaky maintainers.

Serving a model efficiently at 1M context is difficult and could be much more expensive/numerically tricky. I'm guessing they were working on serving it properly, since its the same "model" in scores and such.

A more powerful ASI, the market, is keeping everything in check. Meta's 10 figure offers are an example of this.

yep, also doscoverability is not an issue with Slack. You can find most things with a search, people typically don't go scrolling through a channel to find something.

Yep this article is self centered and perfectly represents the type of ego Sutton was referencing. Maybe in a year or two general methods will improve the author's workflow significantly once again (eg. better models) and they would still add a bit of human logic on top and claim victory.

I don't understand why the current setup for rate limits wouldn't be sufficient to stop this kind of thing.

> Borrowchecker frustration is like being brokenhearted - you can't easily demonstrate it, you have to suffer it yourself to understand what people are talking about. Real borrowchecker pain is not felt when your small, 20-line demonstration snippet fails to compile.

As someone who writes Rust professionally this sentence is sus. Typically, the borrow checker is somewhere between 10th and 100th in the list with regards to things I think about when programming. At the end of the day, you could in theory just wrap something in a reference counter if needed, but even that hasn't happened to me yet.

It's really not that bad.

I'm guessing now that it is GA this won't be a problem.

I mean its useful to the customers who get lower latency too.

> The context window can be compared to working memory in humans: it’s fast, efficient but gets rapidly overloaded. Humans manage this limitation by offloading previously learned information into other memory forms, whereas LLMs can only mimic this process superficially at best.

This is just silly. Humans forget things all the time! If I want to remember something I write it down.

> The nature of hallucination is very different between AR models and humans, as one has a world model and the other doesn’t.

I stopped reading at this point. There's not much signal here, just basic facts about LLMs and then leaps to very bold statements.

Here is an interesting experiment I use to help people understand next token prediction. Think of a simple math problem in your head, maybe 3 digit by 2 digit multiplication. Then speak out every single thought you have while solving it.

Its not the exact same since you can still finetune it, you can modify the weights, serve it with different engines, etc.

This kind of purity test mindset doesn't help anyone. They are shipping the most modifiable form of their model.

They're not new in the same way Attention wasn't new when the transformer paper was written.

No one (publically) had really pushed any of these techniques far, especially not for such a big run.

Why don't you think its possible?

Its a matter of degree. If 90% of the cost savings are from a new, smarter architecture, it doesn't make sense to point to the API terms as the reason for it being so cheap.

This comment is misleading. There is a "free lunch" here in the sense that serving this model is far cheaper than worse, open source models at scale.

Yes they probably are more willing to go down in price due to this, but the architecture is open, and they are charging similarly to a 30B-50B dense model, which is about how many active params deepseek-v3 has.

He mentions this in the video, but the talk is specifically tailored for the "Test of Time" award. This being his 3rd year in a row recieving the award, I think he's earned permission to speak prophetically.

It is probably the same base model as Llama 3.0.

They mention postraining improvements.

of course, running a script from start to finish is better when it's fast, but if your code is slow to run, each cell acts like an in memory cache of your previous work.

In the model card they say they dont train on any user generated data