OpenAI CEO responds to Jordan Peterson criticism | Sam Altman and Lex Fridman

on the chair GPT side it saysthe dialog formatmakes it possible for Chad gbt to answerfollow-up questions admit its mistakeschallenge incorrect premises and rejectinappropriate requests but also there'sa feeling like it's struggling withideasyeah it's always tempting toanthropomorphize this stuff too much butI also feel that way maybe I'll I'lltake a small tangent towards JordanPeterson who posted on Twitterthis kind of uh political questioneveryone has a different question theywant to ask jgpt first right likethe different directions you want to trythe dark thing it somehow says a lotabout people the first thing the firstoh no oh no we don't I mean we don'thave to review what I do notum I of course ask mathematicalquestions and never ask anything darkum but Jordan uh asked it uh to saypositive things about uh the currentPresident Joe Biden and the previouspresident Donald Trump and thenhe asked GPT as a follow-up to say howmany charactershow long is the string that yougenerated and he showed that theresponse that contained positive thingsabout Biden was much longer or longerthan uh that about Trumpand Jordan asked the system to can yourewrite it with an equal number equallength string which all of this is justremarkable to me that it understood butit failed to do itand it was interesting gbt Chad GPT Ithink that was 3.5 based uh was kind ofintrospective about yeah it seems like Ifailed to do the job correctlyand Jordan framed it as uh chai GPT waslying and aware that it's lyingbut that framing that's a humanentrepreneurization I thinkum but that that that kind of yeah thereseemed to be a struggle within GPT tounderstandhow to dolike what it means to generatea text of the same lengthin an answer to a questionand also in a sequence of prompts how tounderstand that it failed to do sopreviously and where it succeeded andall of those like multi like parallelreasonings that it's doing it just seemslike it's struggling so two separatethings going on here number one some ofthe things that seem like they should beobvious and easy these models reallystruggle with yeah so I haven't seenthis particular example but countingcharacters counting words that sort ofstuff that is hard for these models todo well the way they're architected thatwon't be very accuratesecond we are building in public and weare putting out technologybecause we think it is important for theworld to get access to this early toshape the way it's going to be developedto help us find the good things and thebad things and every time we put out anew model and we just really felt thiswith gpd4 this week the collectiveintelligence and ability of the outsideworld helps us discover things we cannotimagine we could have never doneinternallyand both like great things that themodel can do new capabilities and realweaknesses we have to fix and so thisiterative process of putting things outfinding the the the the great Parts thebad parts improving them quickly andgiving people time to feel thetechnology and shape it with us andprovide feedback we believe is reallyimportant the trade-off of thatis the trade-off of building in publicwhich is we put out things that aregoing to be deeply imperfect we want tomake our mistakes while the stakes arelow we want to get it better and bettereach rep butthe like the bias of chat GPT when itlaunched with 3.5 was not something thatI certainly felt proud ofit's gotten much better with gpt4 manyof the critics and I really respect thishave said hey a lot of the problems thatI had with 3.5 are much better and forum but also no two people are ever goingto agree that one single model isunbiased on every topic and I think theanswer there is just going to be to giveusers more personalized control granularcontrol over timeand I should say on this point you knowI've gotten to know Jordan Peterson andum I tried to talk to GPT for aboutJordan Peterson and I asked it if JordanPeterson is a fascistfirst of all it gave context itdescribed actual like description of whoJordan Peterson is his careerpsychologist and so on it it stated thatuh some number of people have calledJordan Peterson a fascist but there isno factual grounding to those claims andit described a bunch of stuff thatJordan believes like he's been anoutspoken critic of various totalitarianumideologies and he believes in ofindividualism and uh various freedomsthat are contradict the ideology offascism and so on and it goes on and onlike really nicely and it wraps it upit's like a it's a college essay I waslike one thing that I hope these modelscan do is bring some Nuance back to theworld yes it felt it felt really new youknow Twitter kind of destroyed some andmaybe we can get some back now thatreally is exciting to me like forexample I asked um of courseum you know did uh did the covet virusleak from a lab again answer verynuanced there's two hypotheses it likedescribe them it described the uh theamount of data that's available for eachit was likeit was like a breath of fresh air when Iwas a little kid I thought building AIwe didn't really call it AGI at the timeI thought building yeah I'd be like thecoolest thing ever I never never reallythought I would get the chance to workon it but if you had told me that notonly I would get the chance to work onit but that after making like a veryvery larval Proto AGI thing that thething I'd have to spend my time on isyou know trying to like argue withpeople about whether the number ofcharacters it said nice things about oneperson was different than the number ofcharacters that said nice about someother person if you hand people an AGIand that's what they want to do Iwouldn't have believed you but Iunderstand it more now and I do haveempathy for it so what you're implyingin that statement is we took such Johnleaps on the big stuff and we'recomplaining or arguing about small stuffwell the small stuff is the big stuff inaggregate so I get it it's just like Iand I also like I get why this is suchan important issue this is a reallyimportant issue but that somehow we likesomehow this is the thing that we getcaught up in versus like what is thisgoing to mean for our future now maybeyou saythis is critical to what this is goingto mean for our future the thing that itsays more characters about this personthan this person and who's deciding thatand how it's being decided and how theusers get control over that maybe thatis the most important issue but Iwouldn't have guessed it at the timewhen I was like eight-year-oldyeah I mean there isum and you do there'sFolks at open AI including yourself thatdo see the importance of these issues todiscuss about them under the big bannerof AI safety that's something that's notoften talked about with the release ofgpt4 how much went into the safetyconcerns how long also you spend on thesafety concern can you um can you gothrough some of that process yeah surewhat went into uh AI safetyconsiderations of gpt4 release so wefinished last summerum we immediately startedgiving it to people to uh to Red Team westarted doing a bunch of our owninternal safety efels on it we startedtrying to work on different ways toalign itumand that combination of an internal andexternal effort plus building a wholebunch of new ways to align the model andwe didn't get it perfect by far but onething that I care about is that ourdegree of alignment increases fasterthan our rate of capability progressand that I think will become more andmore important over time andI know I think we made reasonableprogress there to a to a more alignedsystem than we've ever had before Ithink this is the most capable and mostaligned model that we've put out we wereable to do a lot of testing on it andthat takes a while and I totally get whypeople were like give us gpt4 right awaybut I'm happy we did it this way isthere some wisdom some insights aboutthat process that you learned like howto how to solve that problem you canspeak to how to solve it like thealignment problem so I want to be veryclear I do not think we have yetdiscovered a way to align a superpowerful system we have we havesomething that works for our currentskill called our lhfand we can talk a lot about the benefitsof that andthe utility it provides it's not just analignment maybe it's not even mostly analignment capability it helps make abetter system a more usable systemandthis is actually something that I don'tthink people outside the fieldunderstand enough it's easy to talkabout alignment and capability asorthogonal vectors they're very closebetter alignment techniques lead tobetter capabilities and vice versathere's cases that are different andthey're important cases but on the wholeI think things that you could say likerlhf or interpretability that sound likealignment issues also help you make muchmore capable models and the division isjust much fuzzier than people think andso in some sense the work we do to makegpd4 safer and more aligned looks verysimilar to all the other work we do ofsolving the research and Engineeringproblems associated with creatinguseful and Powerful modelsso rlhfis the process that came applied verybroadly across the entire system morehuman basically votes what's the betterway to say somethingum was you know if a person asks do Ilook fat in this dressthere's uh there's different ways toanswer that question that's aligned withhuman civilizationand there's no one set of human valuesor there's no one set of right answersto human civilizationso I think what's gonna have to happenis we will need to agree on as a societyon very broad bounds we'll only be ableto agree on a very broad bounds of whatthese systems can do and then withinthose maybe different countries havedifferent rlh F Tunes certainlyindividual users have very differentpreferenceswe launched this thing with gpt4 calledthe system message which is not rlhf butis a way to let users have a good degreeof steerability over what they want andI think things like that will beimportant can you describes this themessage and in general how you were ableto make gpt4 more steerableforeignbased on the interaction that the usercan have with it which is one of his bigreally powerful things so the systemmessage is a way to say uh you know heymodel please pretend like you or pleaseonly answer this message as if you wereShakespeare doing thing X or please onlyrespond uh with Json no matter what wasone of the examples from our blog postbut you could also say any number ofother things to that and then wewe we tune gpt4 in a way to really treatthe system message with a lot ofauthorityI'm sure there's jail they'll always notalways hopefully but for a long timethere will be more jailbreaks and we'llkeep sort of learning about those but weprogram we develop whatever you want tocall it the model in such a way to learnthat it's supposed to really use thatsystem message