# OpenAI suggests AI licenses (US Senate hearing on AI regulation w/ Sam Altman) ## Метаданные - **Канал:** Yannic Kilcher - **YouTube:** https://www.youtube.com/watch?v=I72_GJHzH3Y - **Дата:** 21.05.2023 - **Длительность:** 16:12 - **Просмотры:** 33,819 ## Описание #ai #openai #gpt4 US Senate hearing on AI regulation. MLST video on the hearing: https://www.youtube.com/watch?v=DeSXnESGxr4 Links: Homepage: https://ykilcher.com Merch: https://ykilcher.com/merch YouTube: https://www.youtube.com/c/yannickilcher Twitter: https://twitter.com/ykilcher Discord: https://ykilcher.com/discord LinkedIn: https://www.linkedin.com/in/ykilcher If you want to support me, the best thing to do is to share out the content :) If you want to support me financially (completely optional and voluntary, but a lot of people have asked for this): SubscribeStar: https://www.subscribestar.com/yannickilcher Patreon: https://www.patreon.com/yannickilcher Bitcoin (BTC): bc1q49lsw3q325tr58ygf8sudx2dqfguclvngvy2cq Ethereum (ETH): 0x7ad3513E3B8f66799f507Aa7874b1B0eBC7F85e2 Litecoin (LTC): LQW2TRyKYetVC8WjFkhpPhtpbDM4Vw7r9m Monero (XMR): 4ACL8AGrEo5hAir8A9CeVrW8pEauWvnp1WnSDZxW7tziCDLhZAGsgzhRQABDnFy8yuM9fWJDviJPHKRjV4FWt19CJZN9D4n ## Содержание ### [0:00](https://www.youtube.com/watch?v=I72_GJHzH3Y) Segment 1 (00:00 - 05:00) hello everyone welcome to ml news today I just want to look into one story which is a hearing in front of the U. S Senate essentially gives the Senate an opportunity to interview some people from the field and to make more informed decisions invited to the hearing as far as I can tell where Christina Montgomery of IBM Gary Marcus professor at NYU and Sam Altman of openai I want to just comment on a few things that I think are worth commenting on I'm working off of the summary of a machine learning Street talk I think Tim does an excellent job doing a broader summary of what happened and giving some of his comments so definitely recommend if you go and watch the mlst video on this one it goes into a lot more topics than I will go into here second clearly defining risks there must be clear guidance on AI uses or also will not be commenting much on Christina Montgomery's testimony right here because coming from IBM it just corporate speech as far as I can tell as this technology advances we understand that people are anxious about how it could change the way we live we are too so I think in Broad here Sam Altman makes a you know really nice presentation and his takes are really smart like he clearly notice he's very intelligent person and he's quite optimistic like he manages to convey that optimism about the future of AI and the future of this technology and all and I find this takes for example how it's going to affect the job market and economy and all of that I find them to be very reasonable and very well argued and very well said here so I want to give compliments to that I feel it makes a really good case for the field of AI to be part of the future and not to make people that afraid or anything like this and to make people genuinely excited about it so I have no troubles with that compliments to Sam for all of this I do want to comment on the things obviously where I disagree but don't take that as an indication that I disagree with Sam on All Things again I recommend to watch the mlst video on this to get more of his testimony the biggest disagreement and I think that's for a lot of people goes when Sam goes into the topic of potentially licensing AI companies here is how it sounds like I think that regulatory intervention by governments will be critical to mitigate the risks of increasingly powerful models for example the U. S government might consider a combination of Licensing and testing requirements for development and release of AI models above a threshold of capabilities so notably he says the licensing and testing requirements should be for models above certain capabilities and he elaborates later on that as well new approach any new law does not stop the Innovation from happening with smaller companies open source models researchers that are doing work at a smaller scale that's a wonderful part of this ecosystem in of America we don't want to slow that down there still may need to be some rules there but I think we could draw align at systems that need to be licensed in a very intense way the easiest way to do it I'm not sure if it's the best but the easiest would be to talk about the amount of compute that goes into such a model so we could you know we could Define a threshold of compute and it'll have to go it'll have to change it could go up or down I could down as we discover more efficient algorithms that says above this amount of compute you are in this regime um what I would prefer it's harder to do but I think more accurate is to Define some capability thresholds and say a model that can do things X Y and Z up to all to decide that's now in this licensing regime but models that are less capable you know we don't want to stop our open source Community individual researchers we don't want to stop new startups can proceed you know with a different frame as you can please state which capabilities you'd propose would consider for the purposes of this definition I would love rather than to do that off the cuff to follow up with your office with like a well perhaps openings opine understanding that you're just responding and you're not making law all right in the spirit of just opinion um I think a model that can persuade manipulate influence a person's Behavior or a person's beliefs that would be a good threshold I think a model that could help create novel biological agents would be a great threshold things like that so I think one thing is clear from this clip this person knows what he's talking about which is not the case with all of the witnesses here in front of the Senate this clearly comes from experience and from a technical expertise in the matter even though Sam Altman might not be on the Forefront of actually training these things there is a clear handle on what actually makes the capabilities and so on I find the reasoning of this I find to be sound I don't think the I disagree but I find the reasoning of it to be sound in terms of okay if we regulate something probably compute is sort of the best proxy for sort of capabilities and it ### [5:00](https://www.youtube.com/watch?v=I72_GJHzH3Y&t=300s) Segment 2 (05:00 - 10:00) would be better to measure capabilities but we can't yet I mean that all is reasonable what I have trouble with is in the first place suggesting this regulation because essentially saying you know there should be a line where you need to be licensed and anything below is fine because we don't want to crush the open source ecosystem and so on you know but then there is a line essentially that says you know if you just do your itty bitty thing in your garden right and it's small and it's cute right you know no problem for us but it's a long-standing element in the Playbook of big corporations to ask and to welcome regulatory actions because imposing regulations imposing licenses automatically means it makes it much harder for anyone else to enter that market so the position is as long as you're doing your things down here that's fine but if you really if you want to come up to our level like if the open source Community for some reason actually gets a handle on these things and threatens open ai's virtual monopoly now or Duo Paulina with Google entering the market then we should you know clamp down regulatorily and conforming with regulations is easy if Microsoft chips billions and billions of dollars into you like you can handle it but you very much know that the open source Community for example cannot handle it they cannot handle a lawsuit they cannot handle or easily handle regulatory compliance we've seen time and time again that the open source Community actually manages to reach and surpass the level of commercial offerings think of the Linux operating system which is now powering almost the entire internet as does the Apache web server and things like this open source technology will catch up and that is exactly the effect that licensing and testing requirements and all of this have is to squash that competition and to make it much harder for small businesses and for startups to reach open ai's level and threaten their business and the fact that oh there is a line and the licensing should only apply above this line it is a bit better in that it actually lets the open source Community do some things at their level but it has the same effects right because ultimately we engage in open source we engage in research we engage in competition too displace the market leader because there is going to be an improved offering this there's going to be a more transparent of offering in the future and in all cases that regulation denies that or makes it much harder and I find it to be quite disturbing that Sam Altman here asks for this while couching it in sort of the language of oh it's all safety related but oh the open source Community is so important like no you profit highly from the fact that open source software in the past not only has reached the capabilities of commercial offerings but vastly surpassed them without any regulatory oversight is there an FDA for Linux no there isn't right the source code is open you can see what's in there and that is much more of a safety guarantee and a guarantee of you understand what's going on than any other thing and yes I get it with neural networks it's not like source code where I can read it and flat out sort of make a test for it but within large enough software systems that's the case too like Linux and I think if we were to publish what data went into these big models how exactly they were trained if we were to publish their weights then we would have insane amounts of knowledge of you know what's possible what's not possible and how to deploy them safely but that's exactly what openai doesn't want to do there are more and more secretive about exactly what's in their model and that's the other part there are some people still agreeing with Sam Altman here and saying well what if they really believe these things are dangerous you know doesn't this make sense right to have the agencies and so on and I'll show you a second clip here labor should we consider independent testing labs to provide scorecards and nutrition labels or the equivalent of nutrition labels packaging that indicates to people whether or not the content can be trusted what the ingredients are so the question is should there be nutrition labels that tell people what's in these models listen to Sam's answer right here that people users are pretty sophisticated and understand where the mistakes are that they need or likely to be that they need to be responsible for verifying what the models say that they go off and check it um I worry that as the models get better and better uh the users can have sort of less and less of their own discriminating thought process around it but I think users are more capable than we could often give them credit for in conversations like this I think a lot of disclosures which if you've used chat gbt you'll see about the inaccuracies of the model um are also important and I'm excited for a world where companies ### [10:00](https://www.youtube.com/watch?v=I72_GJHzH3Y&t=600s) Segment 3 (10:00 - 15:00) publish with the models information about how they behave where the inaccuracies are and independent agencies or companies provide that as well I think it's a great idea Sam directs the conversation into again the realm of capabilities and where the inaccuracies are and what the models are capable of but a really easy thing to say which Gary Marcus repeats sometimes in this hearing is how about you just publish the data that went into it how about you just publish how you trained it how long it was trained and so on open AI notoriously no gloriously and with Glee as far as I can tell denied to give the research Community any of that data in their latest technical report purposefully hiding it like it was very common until that point too say here's our model this is the size here is what it was trained on here is how it was trained here is the loss here is how it behaves and so on different things and openai has anything but that last category where they demonstrated they can take a bar test and pass it anything other than that they have deliberately obscured so they have deliberately went completely against that idea of a nutrition label hey let's tell people what's inside so they can make an informed decision and so that they can deploy these things safely so if someone was really first of all as he says you know give the users more credit because they should be able to make their own decisions using this technology and if you actually thought it was a good idea to do these nutrition labels it'd be very much for you know publishing all of the ingredients right here and an effective regulation could be that once you deploy I don't know a model you have to disclose in some way what's inside I'm not per se for that but rather than whatever is proposed right here just saying you have to disclose the dates or you have to publish the weights I don't know we have to make it economically viable but um clearly directs the conversation away from that and back into the regime of woo capabilities and yeah so combined with the sayings before we need some kind of Licensing and testing above a certain threshold and we will deliberately not tell you what's inside there is I don't think that combo makes a lot of sense and I don't think that's a coherent line except under the aspect of hey let's make sure we have some business tomorrow and make it as hard as possible for others to threaten that business lastly I want to make a comment on the statements made by Gary Marcus here I legitimately don't know why Gary Marcus is in this hearing but yes so good for him I guess um but here is how that goes from the inherent unreliability of current systems a law professor for example was accused by a chatbot of sexual harassment untrue and it pointed to a Washington Post article that didn't even exist the more that happens the more that anybody can deny anything as one prominent lawyer told me on Friday defendants are starting to claim that plaintiffs are making up legitimate evidence these sorts of allegations undermine the abilities of juries to decide what or who to believe can contribute to the undermining of democracy no way are you saying stuff that's written isn't always true that is I I'm con okay forget everything I said like that convinces me because I mean so far before we had gpt4 and all of these things everything that was written was true like any allegation you know any person claiming that something was in a news article or there was some study proving something was certainly correct in doing so every allocation made by defendants about evidence like no one has ever lied in a court or made up stuff or made up evidence this is it's a new world man and we I I'm convinced like you know these things can make false statements that just wasn't possible before I tried using this technology called email once I wanted to scam someone I wanted to persuade someone of something but it just wouldn't let me but this is the statement it can make up stuff like what kind of a bar is this for no other technology in the world has been held to the standards of it can make up stuff therefore it's bad or it can be used to make up stuff a keyboard I can find tons of made-up stuff on Google I can generate tons of made up stuff using Photoshop actually Sam Altman has a take on that as well if you can find it in this testimony where he says and I find that to be very reasonable where he says look the world has seen Photoshop and for a while Photoshop has fooled some people but then very quickly people realized oh hey not everything that's a picture is you know real and he says well this is the same but on steroids but you know we've had to deal for a long time with texts may not be true so I failed to see ### [15:00](https://www.youtube.com/watch?v=I72_GJHzH3Y&t=900s) Segment 4 (15:00 - 16:00) that just because it becomes a bit cheaper to hire you know gpt4 rather than I don't know some labor form in a low wage tree like yes it's a bit cheaper to write false things at scale I don't see how that fundamentally changes how we think about these things and I get tons of email each day asking me for some sort of money and I'm perfectly fine even though you know according to Gary Marcus also these systems they are you know have lacks so far they're not as good as humans yet so the humans trying to convince me but send me these emails they should be smarter than that right and so far they haven't managed yet like I only ever sent money to legitimate uncles who died and I had to pay the lawyers fees then there was a problem and I never got them but it was totally fine so that was my take on these things again Watch Tim's video it's really cool it's about 50 minutes long and you'll get the most important parts of that hearing bye thank you --- *Источник: https://ekstraktznaniy.ru/video/12437*