All of the "dunks" people are making here are so idiotic. He is saying this about the current state of things, he never said we will never do local inference. Of course there will be better architectures and better local models. But now it's not worth wasting compute on local models because they are weak and the inference is being done in a vastly more inefficient way that's wasteful considering the memory shortage.
sam: local inference is a terrible allocation of scarce resources
“a lot of people want local inference, and somehow they assume that the problem is solved now because we have open-source models, we're going to be able to run them on our computers, and so we're not going to need all the stuff in the data centre
but from a global allocation perspective, actually it's quite terrible to have everybody, like, have a bunch of the world's limited supply of DRAM sitting in their own computer and being used five percent of the time
you want maximum utilisation in the cloud if you're trying to deal with this very supply-constrained environment
somehow many people's intuition is the opposite of that”
You can tell who has a banger model and who doesn't based on the amount of marketing slop they release on Twitter. Case in point, Anthropic just made one post for opus 5.5 and was done with it. Google bought the entire shill industry and had every employee to promote a model not even available to anyone.
The bar for a new model release is so insanely high now after Opus 5.5. Not only do you have to beat it in benchmarks, but you also need to have the same real-world impact and somehow provide almost unlimited usage like Claude pro/max. Not sure even Anthropic can top themselves.
Anthropic's distillation measures really seem to be working; now we are only getting benchmaxxed models. Artificial Analysis' credibility has gone down the drain with the Kimi 3.1 and Muse release.
Here's a wild guess: Ox Alpha is an on-device model either from OpenAI or Google. For a regular model, its performance is quite mid. But if this is an on-device model, it's absolutely game-changing.
It's so annoying to see every new model provider and fanboys claiming their model is Fable level. Like, bro, stop lying. Your model is nowhere close to Fable level and will never be. You simply don't have a hard enough task set or the taste to know the difference.
For some reason, people still believe a high AA index, arena elo or some other benchmark makes a model frontier. The frontier has moved to solving open problems and doing something interesting instead of benchmaxxing public evals. Afaik, there are only two public models that can do this consistently.
This is huge! BTW, this is where the frontier is. Before you claim your model is "Fable-class" show me something like this. No one cares about benchmaxxing.
hello there the jacobian conjecture is false thanx to my close friend akhil for asking about it and my other close friend fable for working during the world cup final
((1+xy)^3 z + y^2 (1+xy) (4+3xy), y + 3 x (1+xy)^2 z + 3 x y^2 (4+3xy), 2 x - 3 x^2 y - x^3 z): \C^3\to \C^3, has jacobian determinant -2, and sends (0, 0, -1/4), (1, -3/2, 13/2), and (-1, 3/2, 13/2) to (-1/4, 0, 0)
Kimi K3 release feels so eerily similar to Gemini 3.1 Pro, everyone posting pointless games and benchmark scores, but it does absolutely nothing interesting like Fable. I would be very interested in the ARC-AGI scores. There are only a few benchmarks that capture benchmarking.
The LLMs used here were GPT-4o, Llama 3 and Command R+. There, I saved you some time. It's amazing that this kind of trash gets published and shared.
A randomized study of 3 LLMs finds they are not ready for public use @NatureMedicine
nature.com/articles/s41591-0…
LLM psychosis is real considering how many people Anthropic has psyoped into thinking they are somehow the good guys.
Every month or so there are people on my timeline with "OpenAI is done for/screwed etc.". How intense can cope be lmao?
It's hilarious to see all these takes on the Physics Nobel prize from people with strong Dunning-Kruger. Like, stop, you're never getting near to even the shortlist of 100k people for the prize and you're acting like they personally stole the prize from you. The prize [1/2]