The fact that they are not working on having LLMs consume streams of video and audio boggles my mind. It's such a low hanging fruit. It's not a hardware issue because you can downsam     See More and audio. It's hard to write any real time application with LLMs since they can't access streams of I/O except for text. EDIT: I would just like to say something about sample efficiency. Human beings actually learn from a pretty huge amount of data too. When we give models millions of hours of video, they are fixed and 2D. Human beings can see and interact with the data in 3D with all our senses which gives us much, much more experience. Just imagine how much data and interaction a baby must consume before it's able to do anything of usefulness. Human beings can be quite slow to learn from mistakes and require a lot of trial and error. I believe it's not so much about sample efficiency as it is about generalization and transferring of knowledge. Once you learn to play guitar it's easier to learn bass because you can utilize previous aquired knowledge.    See Less

После разработки модели, требующей в 100 раз меньше данных, компания создаст ИИ-агента для автоматизации     See More ботки собственных продуктов.

Туарек хочет разработать модели, требующие меньше данных и вычислительных мощностей.

То что нужно. То о чём говорит Ян Лекун. Оч круто.    See Less

Funding is crazy these days when you consider that Facebook probably started with $19K and then $500K.     See Less

All of them are leaving to grift. Promise something stupid and earn 10x. How many of them have left already started companies and NOTHING!     See Less

Figured continental learning was the next big step for a while now...     See Less

i have way to do it but i'm poor, i have the algorithm, code and even the math to achieve general symbolic reasoning but lack gpus for it...     See Less

I mean, this is all just sci-fiction, dude coming up with high flown ideas where he wants to go. He hasn't replaced the transformer yet. Sounds like you hype up that your going to build     See More t building in the world but you don't even have a blueprint but maybe you can convince investors to give you money. Dynamic weights in a model is a very stupid idea. You cannot control well enough the direction the model learns. People troll the model, people try to convert the model to their religion, people jell break the model, people teach the model to be racists, people fill the model with propaganda, etc. etc. etc. Take less than two week for user to drive a dynamic weights model completely insane. At some point these idiots will realize what a clean dataset is. That's the total solution to AI data safety and alignment by the way, you just don't put anything in the dataset you don't want to the model to do or say. If there is no erotica in the dataset, the model can't create erotica.    See Less

we've been terraforming earth for decades: chemtrails     See Less

We're actively terraforming our planet, as I type, into a biosphere inhospitable to the current lifeforms living on it. We can do it, clearly, but we're going the wrong direction wi     See More planet.    See Less

LLMs are just huge databases. That’s not thinking, it’s just a db. Perhaps this continuous leaning will actually net some results     See Less