Generative AI After The OpenAI Crisis?
Channel: Alex Kantrowitz
Published at: 2023-11-28
YouTube video id: jHE_3GJyjYk
Source: https://www.youtube.com/watch?v=jHE_3GJyjYk
hello YouTube how are we doing we're going to continue our coverage of the open AI aftermath and also talk a little bit about where the AI field is going after this so you know as opposed to Bringing on you know a commentator on this I thought it might be worthwhile to actually speak with some people in the field who can talk a little bit about some of the things that we're hearing in the headlines and also give their thoughts about where this field is going to go next um we're doing it with open stream and we'll tell you a little bit more about what open stream is as we go but we have folks from the very top here the CEO Raj tumur is here welcome Raj hello hey thanks for being here and we also have Magnus rang who is the chief product officer at open stream and a former Gartner analyst welcome Agnes thank you pleasure to be here all right folks as we go um if you have any questions feel free to drop them in the chat and we will definitely get to them um I just kind of want to start with qar right which is something that open AI has uh apparently come through it's this big breakthrough that they're talking about um you guys are working on reasoning and that is what qar is also apparently supposed to be an advance in just from the very you know high level what is this advance that openi is apparently uh you know touting behind the scenes and you know is it actually a step forward in the development of Art icial intelligence um well you know without actually having the facts on hand I can speculate based on the news that I have read um and and of course my our own knowledge of where the field is advancing um I'm afraid uh we're not in for um a lot of celebration at as far as reasoning and other things go uh it is extremely unlikely that AGI system can actually have the planning and reasoning based on where things were just earlier this year there was a report um by Professor Rob at Arizona State University that was funded by JP Morgan that actually found out including gbd4 they all performed very poorly almost like 2% score on a planning and reasoning uh capabilities of the of the Genera a systems that they have between 2% to you know whatever breakthrough that people are talking about now we really doubt um whether that is really true okay but you're talking about a previous model and this is a new thing so what what makes you doubt that the new thing is um the fundamental approach itself um because uh planning and reasoning capabilities uh require a lot of um active domain knowledge that is contained and you can't really do that across all kinds of domains um no matter how many how much of hardware and systems that you throw at it um which is one of the reasons why you know not to speak about ourselves but much of our you know industry is focused on the neuros symbolic approach which is you know the symbolic approach is pretty good at reasoning and planning capabilities and the and and the uh neuros neural systems are pretty good at scaling and doing that in an unsupervised way so unless you have a mixture of neuros symbolic approach approach uh you are very unlikely to have a breakthrough in terms of providing reasoning um I mean beyond the epistemic reasoning I'm talking about deontic reasoning um across you know you can talk to few documents here and there but but nothing beyond that when it comes to planning and reasoning and explainability comes from as an offshoot of that so so I I can add something and that is that qar is supposed to have according to rumors this is unsubstantiated facts um supposed to have uh or be based on reinforcement learning um to do reinforcement learning in the way that people have been hyping this qar thing uh is that you would need to have something called a universal feedback loop uh right because basically um you evaluate with reinforcement learning you evaluate how well the answer was that the model was giving right so that it can self-improve uh but to do that in the Universal fashion is is far off of the capabilities that we're capable of we can do it in in localized systems like chess computer games stuff like that uh but bringing that into the real world um is you know uh is is not something I see uh being done in in a sort of a a re Revelation with a single model you would see multiple models working toward that ability of universal feedback loop uh even wouldn't suddenly sort of somebody wakes up overnight and and releases it what do you mean by Universal feedback loop it a universal function that can evaluate the quality of the output uh in reinforcement learning you need to evaluate the quality of the output in order to improve the model and to do that in a universal fashion uh like like we evaluate around outputs as as humans right is was that good was it bad um to do that Universal fashion is is something you know in in Academia and in papers you know very few have have even tried to uh address that problem and and there's just simply a lack of uh papers leading up to that breakthrough right so let me just make sure that I get that right so that's instead of trying to um do one specific task right un you know basically being able to like universally do Universal reinforcement learning means you have a generalized system and it just gets better as it goes yeah is that is that and that's kind of what this qar thing was supposed to be well well it's some of the rumors going in that direction right uh talking about as like a math so how does that relate to the math you know being able to do math that's like the roters and open and uh the information were harping on yeah uh so um so so if you look at the generative AI models today they're really poor at math right because what they're doing is essentially next word prediction uh so they're they're predicting the next word in the sequence and uh they use a lot of computing power and they're really good at at doing it uh but it means it has some trouble with numbers and and uh and and understanding math a planning and reasoning system uh could can turn math problems into symbols and actually do logical deduction and logical uh uh reasoning on top of math and then turn that into you know an output again um and you know that that's not uh that uh unusual uh uh to do um so so so if it does that in sort of like a in speculated uh in in the same kind of architecture vector-based representations like like they have on their other models um I mean it's it's just a it's a linear appro you know thing from even even to add to what Magnus is saying since you asked about math even simple transposition of of the things like you know a plus b is not same as B plus a you get different answers with generate UI as you know um so you know it's it's not so to really uh seek a reasoning system across different domains on the Fly is is really very hard I mean it's it's probably doable in my lifetime but not not anytime soon I mean at least that's my having spent my 25 plus years in in this field at least reading through the papers and our chief scientist has Dr philen has has spent his lifetime on and reasoning systems and uh given where we are from our vage Point uh we don't believe such a breakthrough across domains is possible it is possible that you can take a narrow domain and then achieve you know something that U most of the other neuros symbolic systems are able to achieve and that is probably what they're alluding to again you know I don't want to be speculating but uh I would love to see if uh that is indeed the case Okay interesting so where do we head from here now um what do you think are the areas to watch in the AI field you know after we've seen open ai go through this convulsion and now we have this development that you're both dubious about yeah so you know um there is say backdrop to this whole thing which is which is kind of um alluding to the fact that somebody is capable of thinking and reasoning and those of us you know including you who are in this field know that theism and existential crisis are all really U exaggerated to say the least um you know none of us are really genuinely concerned about it although we are concerned about uh the employment of techniques that are an ethical and that do not have you know checks and balances but that said it's the AA systems are not going to take over anything anytime soon um again I don't want to sound like I'm discounting something that I I don't know uh if it is serious but uh the everybody is is including from meta to uh many of the other companies they are focused on bringing that planning and reasoning capabilities because real life conversations are collaborative we understand each other you know for example I'm watching you I'm I'm seeing whether or not I have your gauge attention I know that you're paying attention to what I'm saying and you also read my facial expressions and a lot of communication happens through the non-verbal Behavior as well uh so sometimes even our inferences are more from the intonation and the facial expression than the actual word spoken uh because much of the meaning is conveyed through that so we I mean we're all working towards that which will lead to better engagement Beyond just question answering in terms of having conversations with systems providing helpful assistance to our and customers um you know in an empathetic way uh I think that's where the AI field is head it uh from from what what we are seeing and almost all the companies are at least the advanced ones are focused on bringing that at the same time before you go on how does that how does that um how is a computer going to be able to take in information that's non-verbal I'm gonna have a camera on my face as I speak to chat GB as you speaking yes you know we all having Zoom calls we have talking to the system so instead of you know typing something on the chat bot you're able to take the phone and and talk to a system so that can actually analyze your internation that is better than just typing and then similarly if you have a video call with the with the system you know customer support can be video based and as it is right now in most of the healthcare related applications where patients actually do call in there are there are things that you can actually analyze and assist the uh uh the the the customers uh by observing their facial expressions instead of asking them what their pain level is on a one to scale of one to 10 you actually can deduce that from the Expressions that you are seeing what what their pain level could be and there are many other such uh behavioral markers uh biomarkers that are visible to the system and a is Advanced enough today um on any video call to actually do that analysis and assess whether you know you are suffering from anxiety or depression or things like that so those are pretty much possible and the same could be used in any social conversation too so if somebody is having some kind of inhibition about saying something you can actually uh make them feel at ease by adopting appropriate verb Ag and ination interesting Magnus you seems like you're ready to say something yeah I'm I'm agreeing uh but it is it is the more data sources you have available when you communicate the more modalities you have available the better you can respond using whatever modalities are are available for you to respond in so so so to limit yourself into text only for example uh means that yeah you can be really good at text but what happens if you have other things available because our natural way of communicating is to use gestures and we use other things and when we use it we expect it to be understood so so so if I if I you know do things like this and and talk with a certain tone um I would expect for example if I if I do a voice-based interface I would expect it to be able to detect irony uh or or or things like that because I'm using my voice and I'm expressing irony with my voice but if it this all the system does is turning it into text where it's really hard to see irony you you know it it it will most likely not detect it and answer in a way that it becomes readily apparent that the machine was not understanding what was being said yeah so where does that leave open AI I mean where does is this something that they're working on I mean natur to assume that they are right so it is natural to assume that they are they have all the money and resources at their disposal and obviously they have a dolly and other models as well so it's it's it's only time that people can combine but right now the way these things are happening are in sequential fashion like you know you have one that is doing a generations and ofar generations and another one that is doing the text analysis and it's U but whereas in real life communication conveying of an intent happens through combination of all these modalities for example I would say you know take that pencil out of your mouth so you know so like to magnet you know I'm watching what he's doing and sorry Magus I didn't mean to say that I'm referring to something that is so contextual and assic as it is happening while I'm talking so it's not based on your historic knowledge of something you have to be observant and so whatever I'm saying is is finding it to take that out of your mouth what is that that that is something that I'm seeing I'm showing and I'm gesturing it so the referential resolution happens through combination of these modalities and that is what we spend our time and life on and there are companies that are focused on that and it is natural to assume that open AI included uh is working on that same thing too because that's the the ultimate Holy Grail for uh for this field actually to be able to replicate human behavior uh completely I want to add something to that and that is when you look at individual models capable of doing amazing stuff like you look at an individual generative AI model you look at you know di or or generative sort of image based uh model uh you have to remember that kind of if you look at a human mind which is what open AI say they're going to do like artificial G gener general intelligence if you look at a human mind our human mind employs many many different cognitive strategies to solve problems and one of the things we do is that we use our executive function uh to regulate what kind of of of focus do we have on a problem what part of the brain are we engaging to solve that problem uh and combine different solutions together to what's the ultimately the right solution um and you can have these amazing models but I think we'll all find out that maybe the big problem is to actually combine them together uh and select appropriate way to attack a problem uh using different kinds of models so that replication of the executive function Uh I that can have a toolbox of cognitive strategies to solve stuff and put them to put together the answers afterwards that is likely uh the biggest problem in Ai and I I think looking at the capabilities of individual models like like like the llms and stuff like that and seeing they're super impressive they're going to surpass you know human ability on all cognitive strategies because it surpassed human capability and one cognitive strategy it's kind of somewhat underestimated underestimating the complexity of the human brain um so so yeah I just wanted to put that in there so it sounds like you're both kind of skeptical like we're going to approach anything close to human level intelligence anytime in the near future well on certain fun yes yeah sorry sorry Magnus certain fun we can we can have overum level performance on tasks without having over superhuman intelligence right um because humans are extremely versatile especially in groups um you you can't you know pull up gd4 and ask can you give me the unified theory of physics mhm right and we're very very far from when you're able to do that because it's reiterating things we already know uh which is amazing but still it doesn't generate new knowledge uh in the way a group of dedicated humans can that's the most important thing because we draw inferences uh you know that are probably something that is congenital pre-wired to draw those inference you know there are many researchers that you probably have seen this but you know a kid just who is not trained on anything and if you are struggling to put some books into a shelf and the Shelf is closed you know this is done by one of the leading professors uh this video is shown there and the kid kind of watches it couple of times and intuitively goes and opens the Shelf door MH so which is something it has never seen before and such a thing that you have never seen before is not something that uh the AI models of today are even capable of right so um so therefore you know the skepticism is stemming from that because um it's we know at least you know some of us know how long it takes to to even get there to because the human man has this unique ability to focus on something at microscopic level and micros macroscopic level at the same time so as you choose I can see the screen right now as set of individual pixels if I want want to or I can see the whole thing as a as a face or with the background and I can choose to like to Magnus Point uh ignore all the books that the day one of course I'm not ignoring your book but the day one and other things that are there on the backdrop I can ignore and just focus only on you so this is this is something that we can we develop as a strategy uh during the conversation as we go about doing things and if we know which ones are to be ignored intuitively and that is really very hard to achieve achievable probably in my lifetime yes MH but not uh not anytime soon okay you guys want to take a question or two from YouTube sure okay we got one from bario Gorman who is working to bring these models together who motivated to do this what standards exist to uh lead us towards this happening well excellent question there are many people in fact people don't realize that there are many standards including the the one I co-wrote at w3c the worldwide Web Consortium called multimodel interaction so which is not generally talking about generative AI or some AIS of type if you will but it's like speech recognition and various other gesture recognition ink modality non-verbal Behavior generation and many other things how to piece them together there is some kind of that the often times we have standards that come only after the field has advanced sufficiently this is where the standards kind of came ahead of the the field advancements itself so but yeah many companies are working on um you know you can Google multimodel AI and you will you'll find a lot of companies including ours speaking of Googling multimodal AI what's gonna go what's going on with Gemini you guys have any idea when Google's going to release that model no idea no idea but what could possibly be the holdup well uh you have to say that the Google is more risk averse than open AI which is startup right um they they they uh they have amazing AI researchers at Google uh and and probably the most academic papers of of any Source in the world right so it doesn't come down to um you know from their looking at them it doesn't come down to you know being unable to do the research and create the models and stuff like that it is their uh how much risk they want to take in making those things into products for you know General people and they have an established Market where they are dominant so naturally they are you know looking at not disrupting that with uh when when they release uh release things that's you have to see it through that lens right right it's funny because uh someone talked about I forgot who it was but how you had open AI who's a company that would effectively you know take the risks that Google wouldn't building on Google technology fire at CEO because it wasn't safe enough like yeah this whole and then bring them back like it's obviously clear now in the aftermath kind of where that company is focused does that sound right to you guys I think it shows that uh you know the the the valuation of the company is really important for the employees of that company so do you think that um I mean obviously yeah they didn't want to lose their their options do you think that this was like amazing marketing for them what happened it seems like it was the best possible marketing like people who never are you sure about that though are you sure about that question yeah you you tell me from a marketing perspective perhaps but with the ones that want to buy into this valuation of them would they think twice now or they wouldn't think twice before what do you guys think is happening with that tender offer our CMO will cringe at the possibility of somebody adopting this kind of strategy for gaining popularity leave it at that but but you know I think it's I'm arguing back let me argue back for a second yeah yeah of course now the world knows Sam Alman Greg Brockman you know Mira moradi every Eliot who nobody knew who these people were before last week now they're spending the entire weekend freaking out about whether qar is going to kill the world whereas before they would completely ignore this it's not this not been incredible for awareness for that there is modum of Truth in uh what you're saying it is very much possible like with anything we don't know for certain but it is too risky for a company to adopt something like that as a stratey I'm not saying they did it intentionally after the fact it worked out yeah after the fact yes after the fact it may have worked uh in their favor but but then also please understand you have ruffle the feathers you have shaken the confidence of many companies and not not not talking about big companies that are banking on this but there are many product companies whose products are solely built on this kind of technology and they have you know second thoughts about their own existence and survival and they want to now have some kind of backup plans and that's probably the Fallout of this are you guys building on open aite Tech yourselves or are you completely proprietary no we are agostic sorry go ahead agnostic when it when it comes to uh using uh large language models we do have our own we do have our own but uh we basically let the use case Define uh what kind of model is more appropriate so the platform itself is capable of orchestrating with uh with many type of models what have uh what have you how have you felt about your uh work with open AI now is everything going to change or I think it's the the the customers uhw customers reacting yeah they're more aware of there being Alternatives right um and we find that often um you know you you can take an open source model and you can pre-train it a little bit additionally and you can fine-tune it and get the same performance for specific use cases that you're employing it for uh than you can get from gp4 right um and and that's basically when you work with Enterprises that wants you know they want transparency they want to reduce hallucinations uh they want uh to monitor the use they they have lots of needs that are you know not out of the box um so our job is often to take that the broadness of these providers that are able to do anything and that's what they're targeting right we want to do anything and we're focusing that into the specific tasks that a particular Enterprise needs we found a a portfolio of smaller models uh trained on Enterprise data actually works better from a performance standpoint and cost standpoint and every other standpoint than any of the ubiquitous large model that you can deliver to an Enterprise so you know if you do just based on that that alone most enter es were on the journey even without this crisis that you you're referring to and now they are patting themselves on the back for having had that that kind of a strategy yeah yeah this whole idea of the models commoditizing the cost coming down and then the building of Open Source has sort of been Central to my theories about what's going to happen does that sound right to you absolutely yeah absolutely we are with you on that so you guys must be loving meta these days well I think think their AI strategy is really good and is beneficial for the overall AI Community what they they're doing at meta and I do like that I also do like yeah I do do have to say that uh you know shout out to an Lon for for his unrelenting focus on bringing planning capabilities or you know not using choices words to describe the lack thereof of the current models that's right yeah Yan and I have been talking about about that for like seven years at this point we have been focused on that you know we have been very highly heavily focused on that and that's something that we we stand you know far far ahead of most of the people that you you're probably seeing in the marketplace yeah I like to ask everybody that I speak with about this thing what do you think is going on with this you know e accelerationist and effective altruist battle oh I will I will answer that um so so first you got to which one you are Magnus I'm neither I'm neither um I think that it it is uh worrying um the influence basically what philosophical belief systems are getting into the boardrooms of of AI companies because uh it's nothing wrong with being an efficient altruist or being a accelerationist but it is philosophical beliefs and philosophical you know background to to to what you're you're standing for uh and that should be open right and the consequence of that should be open that yeah I believe that you know like efficient altruism now I would say it's the long term not not efficient yeah effective but it's the long termism portion of it that is is a challenge in AI the belief that explain that yeah yeah so so the belief is is really out of you know Oxford and Stanford philosophic philosophy departments um and it it is the belief that future humans that are not born yet at the same moral value as people living today and that brings you into some thinking which means that if there is an existential risk and the and the chance of that even if it's really really small mhm if it can do anything to avoid it we should because there is an infinite number of unborn people in the future so if there is a chance of of that not happening that means that you know you you should uh uh not deal with uh or or avoid existential risk which means that even you know if you go into the center of a effective altruism and stuff like that you will actually see that they believe the chance of existential risk for AI to be very very low but because of the long- termism view they still act on it today which is you know I feel that people there's nothing wrong in believing that like there's nothing wrong in believing lots of things but you have to be open about it yeah you have to state that this is why I'm doing it not that you know I believe the world's going under no you believe there's 0.1% MH uh uh exactly chance of it that's a pretty clear explanation that's good yeah yeah all right as as we come to a close if anybody has uh any questions on YouTube just drop them in I'm going to ask a couple more and then we're going to wrap but um what what else like looking forward looking ahead you know obviously this open AI story kind of caught Everybody by surprise um and what what is what are going to be the next things that you think will be important to focus on like if you were my position like what would you be exploring I I can go Raj um go ahead I would focus on um neuros symbolic AI um and and I'm explain why as well so I did explain about the executive function and different you know different types of models working together MH um and really gener AI uh when it came out kind of supercharged a lot of of things in Ai and one thing it did was make every old AI approach that hit some snag in terms of too much work too much this too much that suddenly became relevant again because you could use you know large language models in tandem with these other ways of doing AI uh to eliminate those those bottlenecks uh so so when we talk about neuros symbolic it means that you use large language models multi modal um uh models generative AI models to understand a chaotic world and turn that into symbols that you then reason over and plan over and then you turn it back into the chaotic world uh using uh using these generative AI tools again uh So you you're basically taking different approaches to Ai and combining it in a way that keeps the strength of each of these Technologies and eliminate the weaknesses um and that is really uh why I think uh neuros symbolic is going to be uh the next big thing people are working on neuros symbolic for a long time right uh but it's really with large language models and generative AI that that it's it's turned up an arch what what is possible cool Raj yeah uh absolutely because um humans are good at certain things and we have built over the years um or the centuries a kind of logical reasoning capabilities um and if you can and we can plan and collaborate with users to arrive at answers to to help them achieve things and ultimately what is the objective it's not to create a movie or something like that right so this is I'm talking about practical use cases where this is very helpful um with this technology when you combine them there are I don't want to throw the baby with bath water there are a lot of good things about the generative AI as the model stand today and they all can do those things well but to leave the planning to to such model by itself is not something that I would be comfortable doing it for two reasons obviously Hallucination is not a word that I would use lightly uh in in in our industry in the for our customers uh if you hallucinate you are in for a lawsuit you have to close down so you that's not not even an option right so people have to vet what kind of things that you would say and how you say it um so it's okay to do some search and then get wrong results but you know you can't you're you're doing deontic reasoning you are trying to provide answers to People based on applicability conditions and whether or not they can for example file a claim or whether or not they can use a particular medicine and whether it will have a side effect or not and these are the kinds of questions that we hope to see automated uh through the human expert or virtual expert that is kind of mimicking human for something like that to be offered you need to have some kind of determinism thrown in and that cannot come with uh with neural approach alone and therefore a combination of these two would be the ideal use of technology I will call it as augmented intelligence as opposed to artificial intelligence great yeah and where can people learn a little bit more about Open stream sorry where can people learn a little bit more about Open stream well we are featured in all gner reports and Forest reports all over um probably you know we are the sole Visionary in the Gartner magic quadrant for conation a platforms over the last year two years um and of course they can come to openstream do.ai and our people are waiting to hear from you great Rog and Magnus thanks for coming here and speaking with me very interesting stuff fun to have a great discussion with you guys and let's keep in touch and everybody thank you for watching uh we'll be back on the feed relatively soon and uh looking forward to seeing you there all right that'll do it take care bye