GPT-4 is listening to us now | Joscha Bach and Lex Fridman

again so i'm using your tweets as if this is like plato all right as if this is well thought out novels that you've written uh you tweeted gpt4 is listening to us now um this is one way of asking what are the limitations of gpt3 when it scales so what do you think will be the capabilities of gpt4 gpt5 and so on what are the limits of this approach so uh obviously when we are writing things right now uh everything that we are writing now is going to be training data for the next generation of machine learning models so yes of course tbt4 is listening to us and i think the tweet is already a little bit older and the uh we now have wu dao and we have a number of other systems that basically are placeholders for gpt4 don't know what open ais plans are in the city i read that tweet in several ways so one is obviously everything you put on the internet is used as training data but in the second way i read it is in a uh we talked about agency i read it as almost like gpt4 is intelligent enough to be choosing to listen so not only like did a programmer tell it to collect this data and use it for training i almost saw the humorous angle which is like it has achieved agi kind of thing well the thing is um could we be already be living in gpt5 [Laughter] so gpt4 is listening and dpt5 actually constructed the entirety of the so the reality works we in some sense the what everybody is trying to do right now in ai is to extend the transformer to be able to deal with video and there are very promising extensions where there's a work by google that is called perceiver and that is uh overcoming some of the limitations of the transformer by letting it learn the topology of the different modalities separately and by training it to find better input features so the basic feature abstractions that are being used by uh this successor or to gpt3 are chosen such a way that it's able to deal with video input and there is more to be done so i one of the limitations of gpt three is that it's uh amnesiac so it forgets everything beyond the two pages that it currently reads also during generation not just during learning do you think that's fixable within the space of deep learning can you just make a bigger bigger bigger input no uh i don't think that our own uh working memory is infinitely large it's probably also just a few thousand bits but uh what you can do is you can structure this working memory so instead of just force feeding this thing a certain thing that it has to focus on and it's not allowed to focus on anything else with its network you allow it to construct its own working memory as we do right when we are reading a book it's not that we are focusing our attention in such a way that we can only remember the current page we will also try to remember other pages and try to undo what we learned from them or modify what we learned from them we might get up and take another book from the shelf we might go out and ask somebody and we can edit our working memory in any way that is useful to put a context together allows us to draw the right inferences and to learn the right things so this ability to perform experiments on the world based on an attempt to become fully coherent and to achieve causal closure to achieve a certain aesthetic of your modeling that is something that eventually needs to be done and at the moment we are skirting this in some sense by building systems that are larger and faster so they can use dramatically larger resources and human beings can do much more training data to get to models that in some sense are already very superhuman and in other ways are laughingly incoherent so do you think uh sort of making um the systems like what would you say multi-resolutional so like some uh some of the language models are focused on two pages some are focused on uh two books some are focused on two years of reading some are focused on a lifetime like so it's like stacks of it's the gpt threes all the way down and you want to have gaps in between them so it's not necessarily two years there's no gaps it's things out of two years or out of twenty years or two thousand years or two billion years yeah where you are just selecting those bits that are predicted to be the most useful ones to understand what you're currently doing and this prediction itself requires a very complicated model that's the actual model that you need to be making it's not just that you are trying to understand the relationships between things but what you need to make relationships or discover relationships over i wonder what that thing looks like with the architecture for that for the thing that's able to have that kind of model that i think it needs more degrees of freedom than the current models have so it starts out with the fact that you possibly don't just want to have a feed forward model but you want it to be fully recurrent and to make it fully recurrent you probably need to loop it back into itself and allow it to skip connections once you do this right when you are predicting the next frame and your internal next frame in every moment and you are able to skip connection it means that signals can travel from the output of the uh network into the middle of the network faster than the inputs do you think it can still be differentiable do you think it still can be a neural network uh sometimes it can and sometimes it cannot so it uh it can still be a neural network but not a fully differentiable one and when you want to deal with non-differentiable ones you need to have an attention system that is discrete and dual dimensional and can perform grammatical operations you need to be able to perform program synthesis you need to be able to backtrack in this operations that you perform on this thing this thing needs a model of what it's currently doing and i think this is exactly the purpose of our own consciousness yeah the program things are tricking well in your networks you

Channel: Lex Clips

Views: 19,463

Rating: 4.9565892 out of 5

Keywords: ai, ai clips, ai podcast, ai podcast clips, artificial intelligence, artificial intelligence podcast, computer science, consciousness, deep learning, einstein, elon musk, engineering, friedman, joe rogan, joscha bach, lex ai, lex clips, lex fridman, lex fridman podcast, lex friedman, lex mit, lex podcast, machine learning, math, math podcast, mathematics, mit ai, philosophy, physics, physics podcast, science, tech, tech podcast, technology, turing

Id: fxQ_JsXLMJ0

Length: 6min 47sec (407 seconds)

Published: Tue Aug 24 2021