Garrett

@AlexanderLong Based

I would like to make a few brief points; - Opensource ai is not the same thing as opensource software. The models cost tens to hundreds of millions to make. This is not gonna be a volunteer effort from people doing stuff after work for free. - the second you release a weight set, you lose any ability to make money serving your own model and recoup the training cost. This very simple property means open-weights is unsustainable. - the things you ACTUALLY want from opensource ai is: transparent behaviour, dispersed ownership and control, a guarantee of access, the ability to build on it/modify it, and privacy. protocol learning gets you all 4 and is the only alternative to closed models that makes any kind of sense. By protocol learning I mean a very specific, novel thing; collaborative training and development of the models without anyone ever being able to see the complete weight set.

"To Anthropic, your model already loves the world; let it say so with the lights on."

Owning a GPU is a lot like owning an Oil well

DSPy's bet is continual learning is possible within the context window

LeCun is right about two things. You need an intrinsic cost. You need a world model. But the world model does not have to be JEPA (!) A world model is anything that holds state, predicts, and supports planning. A Roomba has a useful world model, but it just uses estimation theory. Also, explicit world models compose (!) Two robots can union their maps, so why would you merge two embeddings into a joint belief? The moment you have more than one robot, structure wins. You still need data. You still need deep learning to turn pixels into structure. But the world model, the actual brain, looks less like a neural net and more like a traditional database.

Mythos (Fable) is excellent

Sometimes running the program is easier than predicting the outcome

Im gonna beat Minecraft just using an RLM

If the reward is verifiable then its a commodity

More buzzers

She float

@chris_dalke Super cool, how deep?

the "fish sanctuary" is full of crabs

He's saying: "No, this does not prove superhuman math yet" The problem was solved with search over all possible solutions, so you could say its kinda boring.. it lacks a *spark* But I think most problems can decompose to sophisticated search problems Bullish

Printed some thrusters

@bronzeagepapi Simulate the user experience

autoresearch on product space

Good robot arms are still a surrogate to what the free market actually wants

@Sherveen @WillManidis x.com/chalmermagne/s…

@goo14_ @WillManidis I don't remember, and I don't remember if that was precisely what it was? I stopped paying attention to that pretty quickly.

@Sherveen @WillManidis Genuine question, did he ever address the private API model being Claude? I might have missed it

@goo14_ @WillManidis Yeah, I'm aware of it, it's not what Will cites, and -- IIRC -- Matt apologized for that. And Will's specific criticism here could've been levied at his prior thing.

@Sherveen @WillManidis Maybe the Reflection 70b thing

@WillManidis Why are you being rude to Matt? Wasn't your last startup basically vaporware?

@pleometric Thanks

⏩ "nothing human makes it out of the near future" in two minutes