redox99

Images from their X account

They realized Astra having the same score as Sol was silly so they rushed to update the index so it fits what people expect.

The old index was clearly bad (Astra is way better than Sol) but it's also unscientific to tweak it like this.

Those two you mentioned completely demolish opus 4.5. It's not even close. I'd say they are between opus 4.8 and opus 5. And better in some tasks.

4.5 became useful for one shotting large features.

LLMs were useful for coding ever since GPT3 (copilot), and sonnet 3.5 for agentic coding.

Yeah 5 was very underwhelming.

The jump from 3.5 to 4 felt gigantic to me back then.

GPT 5.0 did feel underwhelming though.

I hate the term "AGI" but IMO Fable, 5.6 Sol, et al. were already AGI.

Definitely not Google, countless horror stories and infamous for killing stuff. OpenAI is still serving GPT 3.5 turbo as far as I remember.

You can ask 100 people and they'll all give you a different list. It's subjective.

I think a less personal ranking would be, as a business owner, which of those providers is more dependable? As in, you don't care about evil, just your stuff working. I think maybe OpenAI?

Google, Zuck, Sama, Elon, Amodei (in no particular order).

They all suck. Pick your poison.

No because the frontier keeps advancing very fast.

Yeah for sure. Easily 10x that. And obviously we're talking just output tokens. I was mostly just pointing out a theoretical lower bound.

> I think it's roughly possible for any piece of software.

Still not possible for games (many people are attempting it on X).

It's definitely getting better with every new model and it's just a matter of time before they can actually do it.

Fable 5.1 is quite strong

https://x.com/loktar00/status/2095192155086057978

(that's just a main menu, other people are attempting entire games which is way more difficult of course and results much weaker)

Obviously the agent would need to build a very extensive test setup for Paint.NET, a lot of it with computer use or something similar.

Opening the same files, performing the same clicks, etc should output the same on paint.net and the clone.

700k LoC of human written code is probably like 2M LoC of LLM "slop". That's probably around 8M tokens. Assuming a single fable agent that one shots the new source code at 30 tok/s, that's 74 hours. Of course it needs wayyy more tokens because it wouldn't one shot it, plus the reasoning tokens. Generation speed would be faster because you'd have many simultaneous agents though.

Bun rewrite took 6 days of coding and was a similar number of LoC, so I think a couple of weeks is definitely reasonable, although you probably need more than 5 Claude subs.

How much in tokens/Claude subs would it cost to make a paint.net open source clone?

Seems like the kind of software that with enough tokens an agent would be able to create from scratch autonomously (because you have a clear goal and reference)

I'd guess with 5 max subs running fable 5.1 almost non stop it should be doable in a couple of weeks.

Qwen 3.8 27B absolutely demolishes Mistral best offerings at coding, and you only need a 5090 or 2x3090 to run it.

Just wait and see

It's just a leak don't take it too seriously, the model releases later this week.

Am I 100% sure the author is telling the truth? No

Does it matter? No

It's not my image

Yes

Yes

https://x.com/lyraxana/status/2093960706051727723

>massive PD disaggregated cluster of B300s connected via NVLink.

So a headless server.

Macs were mentioned because that's what the post is about. It could be a PC (I use a 2x3090 PC). The point is that it's a better experience to have a box dedicated to the LLM than running it in your system. Obviously in your home, so local.

Only people who live in SF are capable of programming a browser?

Firefox has 1.4 billion in reserves and no debt.

If they stop paying tens of millions to their management, they could easily afford paying devs in perpetuity just from investing that money.

You can, you just need a beefier PC, and it's more annoying in terms of noise and heat vs throwing something on your server closet. Plus you don't need to worry about other software stealing resources and whatnot.

Qwen 27B runs very comfortably on a 5090. You need to use Q4 quants and Q8 KV cache. Here's the math

https://news.ycombinator.com/item?id=49514141

32GB of fast unified memory is enough for Qwen 3.8 27B.

- 16GB for the weights at Q4

- 9GB for the full 256K context at Q8

- 7GB spare for overhead and system.

The problem is that these Macs have 32GB of slow unified memory.

Edit: I'm thinking of a headless Mac mini, if you meant running it on the same machine you're using of course you'll need more memory, but LLMs are best served from a headless server so that's what I'd recommend.

The speed at which hy4 usage increased on openrouter, especially considering its not a cheap model, doesn't seem organic to me.

Its already serving as much tokens/day as the incredibly cheap and good GLM 5.3 flash, which had a crazy marketing campaign as ox alpha?

Also those top 5 apps are just 1.58B tokens out of 1.54T tokens from yesterday. Negligible.