[ML News] OpenAI is in hot waters (GPT-4o, Ilya Leaving, Scarlett Johansson legal action)

we have to talk about open AI there's so much been going on in the entire world of AI and today we're going to have a full plate with just open AI I know there's tons of other stuff going on but just open a has been through ups and downs and ups and downs recently and is finding itself currently in a controversy with Scarlet Johansson we'll get to that at the end but you surely have heard that introd introducing their new model gp4 o stands for Omni and it is a model that breaks out of the text modality and can natively interact with things like images voice and so on now chat GPT previously had a voice mode but it did it in a multi-stage way so it first transcribed voice input into text then it let a text model handle the text and then it used text to voice synthesis in order to make a speech output GPT 40 at least as far as we know natively processes all of these things at the same time in the same model and therefore can provide a much more unified experience in much more real time so if you have seen the demos and the announcements you will have seen people interacting with the app in a quite a realtime fashion so it will talk back to them now this is YouTube I'm not going to uh show any sound I'll show you why later uh always with YouTube copyright things are at the Forefront but as you can see right here the model is able to naturally interact with video at least on a frame by frame basis you can interrupt it like it's real time like that and it just generally shows a great level of multimodal understanding uh has sort of a humaness to it now the humanness in part comes from the fact that it the voice engine has been you know the development in that has been absolutely incredible and getting this to the realtime like state that it is really crazy so GPT 40 has been you know evaluated by people and it's been evaluated in a quite a special way there is this chatbot Arena by LM they will pit two models against each other and then you have to choose which one's better and so on you don't know which one's which so this is a way to just from a human assessment assess different language model performances now GPT 4 has been smashing that what open AI has actually done is they have submitted a sort of uh fake name uh they called they had different iterations so here is a leaderboard with more of the iterations they had different iterations of this so first it appeared as a model called gpt2 chatbot and that got replaced by a model by I'm a good gpt2 chatbot and then I'm a I'm also a good gbt2 chatbot these models quickly shot to the top and people who have evaluated them have noticed that they are aware of current events and things like this so it is reasonably possible these things do have access to the internet in the background now this is fine I think from an lmis evaluation perspective you just care which model is better no matter how it does that uh but open AI essentially used a you know a Arena a leaderboard in order to do some well first some evaluations but I really think it's just a marketing campaign for them so I think this is the new World things that have been infrastructured that has been put in place in order to help us determine you know what's good what's bad and so on is now being hijacked by these companies in order to do some for free guerilla marketing of their new products so that's we'll have to contend with that again it doesn't seem to be against the rules necessarily but it's it is special uh that obviously L companies consider this marketing Avenue but we'll have to Grapple with the fact that we are no longer in the world of researching AI we are productizing and making money off of that some people have evaluated want to highlight this here from from Bindu who has done early evaluations but I see this mirrored across the Spectrum in that generally uh people find GPT 40 to be better uh um of course with internet

access and so on it is able to deliver more current results and so on more relevant results to the user inputs however it does tend to suffer a little bit on the more difficult uh sort of coding tasks and so on these are just first evaluations but I do see that mirrored from several other people now it could be that the prompting is different or something like this which we've seen in the past but it could also be and actually something that is quite likely is that open AI has made a conscious tradeoff here between performance on things like you know coding reasoning whatever you call that trading that off for this more multimodal performance and uh more speed so GPT 40 I believe is faster and cheaper than gp4 turbo and can interact with all of these different modalities right so the tradeoff here is it may not be as good on very particular tasks as the uh previous models although in the evaluations it does look really good in the ones that open AI has performed but you can see that the there is no longer this ultimate push to deliver like orders of magnitude better performance in these benchmarks rather hey can we expand can we make it into a better product so being faster being cheaper is more important nowadays than being better in terms of performance Benchmark and I think that is the direction the field is going that's when you build a product that's what you focus on rather than some numbers in a benchmark nevertheless their evaluations are GPT 40 is pretty good they've also made some other announcements in their spring update uh in improvements to data analysis in chat GPT by the way uh GPT 40 is also supposed to come for free in various aspects into chat GPT so the race is now really on between Google's new announcements which were also pretty crazy Microsoft's new announcements which is kind of the same as open air but not um so and other announcements apparently anthropic also is brewing something new um there's lots of stuff going on so we really see market dynamics kicking in here there's also a new open a and Reddit partnership so people will remember that Reddit uh close down some of their apis um in order to better be able to monetize their data and they're doing so in a partnership with open AI among others Apple also is said to have a deal with open AI to make chat GPT or open ai's models the provider for features like Siri so Siri has been once been revolutionary because of it their series like humor and quippy responses and so on but Siri was a bit ahead of its time in that they put in a ton of effort to really make that thing work and then as far as I know nobody really used Siri or any of the other phone assistance for anything else except like set an alarm or something or what's the weather like or something like this like all of these prophecies have never materialized that's by the way also why I'm somewhat skeptical of the current generation of Agents even though I do think they have a much higher potential being bottom up data driven and generalist rather than the pre previous generation of Agents like that the Google assistant and the uh Siri and so on that essentially had hundreds and hundreds of hardcoded paths that you would have to trigger with some keywords interesting that apple is says it's nearing a deal um with open AI or um whereas just previously uh there were reports of open AI sorry not open AI of Apple and Google nearing some sort of deal so I guess we'll see how that pans out for the next generation of IOS assistants but I think there is there's cool things to come here also not in today's video but Microsoft just announced like co-pilot integration into everything the operating the Windows operating system does which I don't think it's going to work but if it works it's going to be crazy

there controversies started starting I guess so among a few people leaving and we'll get to that namely Yan Le and Ilia sver um there have been reports popping up that open AI has very restrictive practices with respect to people leaving so the way that uh this works and you can read more about this in this Vox article the way that supposedly it works and some people uh xop a have confirmed this is that they give you part of your salary asignificant in participation units which are essentially shares but not really shares so they're something that's tied to the shares so essentially they give you these pieces of paper and the pieces of paper say hey um in the future you may get some money for this depending on how well the company's doing so it's like a share but they have some restrictions like the company only the company can buy them and they will host events and say okay now we're buying X we're buying them for x amount of dollars and then you can choose to sell your participation units so once you leave it is expected that you keep part of them that's called vesting so the day one you join open AI you if you left again you give back all of them if you after one year leave open AI you give back most of them if you leave after 2 years you give back half of them or something like this and then after a couple of years you may have earned a lot of them and then you can keep those however these things had a sort of a legally legal back door in it such that when people left sorry they were forced to sign a very comprehensive non-disclosure non disparagement agreement that would prevent them from ever in their life basically saying anything bad about open AI uh to anyone or you know pretty much talk about anything at all uh and if they did uh they would lose their participation unit or if they didn't sign that they would lose all the participation units and that is there are such things um but that seems a bit overly strict and overly draconic so other leaders have come out for example suum with chintala has come out and saying holding your already earned stock compensation hostage over signing a on disparagement Clause is not normal pretty sketchy and sleazy it's sleazier uh when a super majority of your compensation comes from comes in the form of stocks and yes uh that has been confirmed so Sam Alman has come out and said oh I said essentially hey I wasn't really aware of that and that is not uh not cool and will stop doing that so controversy one averted Sam come out and say hey we'll make it right that is not okay you know I wasn't aware okay then I already announced some people leaving notably Ilia sver after not having been seen for a while I guess it's become a meme but finally announcing that uh Elia will leave open Ai and company has been great and Wishes the best to the people uh to the people staying and that he will share details in due time yeah we're also excited for what comes next obviously Ilia sver has been a massive contributor to the field and probably also to open for sure also to open AI itself so him leaving is a tragedy for open AI there nothing short of a tragedy um for their ability to in the future push what's possible uh and really Envision future breakthroughs it's only because of people like Ilia seeing when nobody else saw what could be done with scale and how to do it and just going for it like not him alone for sure but people like him who made it possible that we're in the situation we're in right now so also Sam Alman making a statement about Elia parting very sad obviously that resonates with all of us the next person leaving Yan Le also announcing yesterday was my last day as a head of alignment super alignment lead and executive as at open AI highlighting the

things that the team around Yan has done uh wishing the best to everyone staying however saying I've been disagreeing with open a leadership about the company's core priorities for quite some time until we finally reached a breaking point I believe much more of our bandwidth should be spent getting ready for the next Generations of model security monitoring preparedness adversarial robustness super alignment confidentiality societal impact and so on uh yada y saying that the team around like has been struggling to get the compute for the necessary experiments so essentially saying inside of open AI safety alignment and so on has been deprioritized probably in favor of more producty stuff hey can we get chat GPT to do data analysis A can we do Reddit data you know integrate into Mac OS and things like this so you can again see this from this perspective we had at the beginning open AI shifting away from this researchy company more and more towards a product company towards a pro towards a company that wants to monetize their current technology and make it more you know appealing to more users push it out to more channels and so on which not the directly but in a way is in opposition to this oh we're building AGI and you know we need to take care of it if it's if it turns out to be too intelligent and so on so in the past I think we've we've covered um Yan like has views a little bit but I don't want to ascribe things to him personally but generally seems to be in the camp of what one would not necessarily the doomers but in that Direction essentially saying this stuff is going to be very dangerous very soon and we need right now to take care of aligning it with human values and you know what if it more intelligent than we and we can't stop it anymore and so on then and yeah so when someone like this formulates this it all sounds quite reasonable in my personal opinion this is all hogwash nothing personally against Jan or anyone on this team and so on and the research is also super cool that they've put out in my personal opinion deprioritizing this is makes open AI a lot stronger and probably the causation is differently right they want to be a lot stronger and therefore they deprioritize this stuff because at the end of the day just again my personal opinion we are not we are nowhere near the danger of AGI being misaligned or anything like this like yeah people can make deep fakes and things like this but super alignment like that's at some point you get into fairy tale territory and then you never you can always say o this is probably going to be soon and we need to act now and it's an unfalsifiable statement you can never prove them wrong because you can always say well if we build something that's so smart it can outsmart all of you and if we do that isn't that terrible shouldn't we do everything in our power like and then they show some curves you know what's a good curve like track the performance of computers versus humans in Long integer division okay you see like and then it surpasses human l skill like human s it surpasses and then just up out the top right or chess like playing chess or something like this like that's a situation we find ourselves in except the axis is like statistically modeling language or something like this you know sure you can do more stuff with it but it is wildly different from like super intelligence and dangerous super line and so on so again cool research I'm sure good people however these people departing are a good thing for open AI in the current climate where it's about who can win the current market my opinion also here uh Sam Alman wishing Yan good luck Greg Brockman coming out iterating what they are doing in order to keep things safe and so on yeah which

is also something right you can like something's going to go wrong at some point like some someone will use one of these language models to do something bad not saying someone has done that already but then everyone will point fingers right but ultimately like how much stuff has happened so far with all the meta open sourcing llama and so on like how many world destructions have we seen but you can never like it's obvious convince those people always be like yeah that those are the models now once we build the true super intelligence like get out all right next controversy final controversy for today the voices so if you've seen the GPT 40 announcement you may have thought well that's kind of a sexy voice that chat GPT has there and people started questioning hm that sounds a lot like Scarlet Johansson even had a role as like a AI um intelligence uh in the movie her now this all has led open AI to release a blog post saying how the voices for chat GPT were chosen that they have received professional submissions for voicing these things they talked about you know they talked with the with these about sorry actors about using the voice and about Ai and so on and saying we believe that AOS should not deliberately mimic a celebrity's distinctive voice Sky which is the particular voice they showed voice is not an imitation of Scarlet Johansson but belongs to a different professional actress using her own natural speak voice to protect their privacy we cannot share the names of our voice talents and again saying how their casting process went and so on each actor flew in recording session yada yada now they are saying they're currently pausing this Sky voice that sounds like Scarlet Johansson the whole world kind of wondered why if for one they're releasing this oh no no we did nothing wrong it's not Scarlet Johansson it's a different one and also they're pausing this Scarlet Johansson at least I believe her legal team has come out with a statement uh telling a story that we hadn't heard from open AI saying last September I received an offer from Sam Alman who wanted to hire me to voice the current chat GPT for Doo system he told me that he felt that by my voicing the system I could bridge the gap between yada yada yada after much consideration and for personal reasons I declined the offer 9 months later my friends family and the general public all noted how much the newest system named Sky sounded like me when I heard the release demo I was shocked angered disbelief prity prity and this tweet from Sam Alman also not really going so after the release of GPT 40 Sam Alman tweeted her um nobody knows if that was like on purpose as a jab or just uh differently so the timeline here the timeline here syncs up um probably Scarlet Johansson got contacted for this call for this roll call for voicing the system however declined and okay there there's like a good interpretation and a bad interpretation we'll get to that in a second so they um Johansson says I was forced to hire a legal counsil who wrote two letters to Mr Alman and open AI setting out what they had done by asking them to detail the exact process by which they created the sky voice consequently openly I reluctantly agreed to take down the sky voice well grappling yada y our own identities and so on so um the good interpretation of this oh sorry there's this paragraph 2 two days before the chat GPD 40 demo was released Mr Alman contacted my agent asking me to reconsider before we could connect the system was out there okay so the good interpretation is Sam Alman really wanted something like the movie her uh thought hey wouldn't it be a great idea

to contact Scarlet Johansson directly that'd be really cool if she did the voice and so on that that' be amazing and then she declined and then he set out and said hey why don't we just you know hire someone who sounds very similar to that because the just the concept of that movie and of having this sort of you know type of voice in the product seemed appealing and that that's how it came to be and then in to avoid legal stuff they did not mention any of that in this um they just listed how they did that roll call and good okay the other interpretation is of course that Sam wman was salty that Johansson wouldn't do that and said well let's do it anyway we can just either take her voice or we can deliberately choose a person that sounds so much like her that it's basically her right or we can even modify the voice and so on uh there's various spectrums of this however I don't know I think one cannot deny that at least they were going for an effect like that right and Johansson declining to do that should have given them a hint that maybe they should at least create some distance between her voice and that so they're saying this is not her voice and so on but I don't think you can make the argument they were not going for an effect like this which to me okay to me there is the hypocrisy right here you have the stewards of safe AI the people who didn't release gpt2 because it was so dangerous the people are so worried about misinformation and deep fakes unless it's them doing the profiting from something then we're fine then we're okay with you know just sort of being very close to something else in order to create a certain effect not for you plebs though we will see that you plebs cannot do that with our technology but we the also ethical open AI have every right to do anything we want as long as we're sort of Legally clear because it makes us money yeah and that's the people who advise your government tell you to listen to them when it comes to safety and safe deployment and who ask you to trust them that when they invent in AGI they are going to use it responsibly and for the benefit of all of humanity right sure okay um no that's not going to happen so we need open source more than ever uh we're all excited for the big llama 3 next iterations of llama and we're all excited for open AI to get some fires under their butts so open AI sailing high on the release of a new model absolutely and it is a cool model right and they do cool things and it is beyond what anyone else can do at the moment sliding Partnerships with all kinds of things and then sliding right into people leaving ultimately controversy around their compensation of former employees and lastly with in a in currently in a tussle with Scarlet Johansson I think that's enough bravel for today um I want to leave you alone now just know that this is only a tiny part of what happened over the last week or so and hopefully we'll get to the rest soon unless something else is happening we'll keep you up to date stay hydrated and bye-bye