Odszedłem z OpenAI i nie żałuję II PSYHO — Transcript
Full transcript
- 0:01It's not that artificial intelligence
- 0:03will simply dislike humans, it will
- 0:04just be so powerful that it won't
- 0:06really pay much attention to us. We are
- 0:11stripping ourselves of agency, and we
- 0:13do it just to gain some economic profit
- 0:15. And you were the only protein-based
- 0:19computer, the only Homo Sapiens Sapiens
- 0:22who beat a language model.
- 0:25>> Yes, that was a year ago. But times
- 0:27change, because there was an edition
- 0:29this year too.
- 0:30>> Who won this year?
- 0:31>> Well, not a human. This year, it
- 0:34totally crushed the humans.
- 0:36>> Was 2025 the last year a human won in
- 0:39these competitions?
- 0:41>> Yes.
- 0:41>> Do you work anywhere today?
- 0:43>> No.
- 0:43>> How long have you been out of work?
- 0:45Since 2000. Well, generally since I
- 0:49left OpenAI.
- 0:50>> And why did you leave?
- 0:52>> That's a lot of strange, little stories
- 0:54.
- 0:55>> Do you regret leaving a company that is
- 0:58so famous today, that is so hot right
- 1:00now?
- 1:01>> No, that is the one thing I am sure of.
- 1:11The program is brought to you by Credit
- 1:13Agricole bank with a business account
- 1:15for entrepreneurs. The partner of this
- 1:20episode is Kodano Optyk, Polish optical
- 1:22salons where you can choose glasses
- 1:24perfectly suited to your needs and
- 1:26check your eyesight.
- 1:29>> Good morning. Patrycjusz Wyżga. I
- 1:31welcome you and invite you to Didas.
- 1:33Today, a meeting with a man who won
- 1:34against artificial intelligence. He is
- 1:36better, ladies and gentlemen, than
- 1:38algorithms. His name is Psajo. Welcome.
- 1:40Hi.
- 1:41>> Hello.
- 1:41>> Actually, that's been a nickname for
- 1:43many, many years. You say 30 already.
- 1:45>> Well, almost 30. But if you'll allow,
- 1:48I'd like us to tell those who for some
- 1:50reason don't recognize you and don't
- 1:52recognize those five letters, Psajo.
- 1:59What does it mean that there sits
- 2:01before me a man who won against
- 2:02artificial intelligence, proved to be
- 2:04better than AI as the only one in the
- 2:06world.
- 2:06>> There is something like programming
- 2:08contests. You can think of it a bit as
- 2:11e-sports for programmers. This is, it's
- 2:15generally a very niche, very niche
- 2:17activity. It was created because
- 2:19programming companies were looking for
- 2:21new ways to find employees. It also
- 2:24stems a bit from programming contests
- 2:26at universities, so it has a longer
- 2:28history there. And now there are two
- 2:33types of programming contests. There
- 2:36are contests popularly called
- 2:38algorithmic, and those are, for example
- 2:40, the Informatics Olympiad, and there
- 2:42are programming contests called
- 2:44optimization or heuristic. That is the
- 2:47category in which I specialize.
- 2:49>> And you participate in such contests?
- 2:52Yes, well, you could say I participated
- 2:54, yes, but the reason why I roll my
- 2:56eyes a bit whenever I hear about "the
- 2:59man who beat AI" is that...I've been
- 3:04participating in these competitions
- 3:06practically since they started, so
- 3:08that's almost 20 years now.
- 3:09>> With what results? Um, in such a way
- 3:12that if you were to collect, um, all
- 3:18the major competitions that have taken
- 3:20place, I've won about half of them.
- 3:23There are practically no other people
- 3:25who have won more than one. And the
- 3:28biggest competition was the Topcoder
- 3:30Open. And there were, I don't remember,
- 3:3415 or 16 editions, maybe 14. I won
- 3:38seven of them. In the others, I was in
- 3:40second or third place. Ah.
- 3:42>> Listen, so in that case, you are the
- 3:44best programmer in the world,
- 3:45>> right? Well, I am the best person in
- 3:47these specific competitions, yes. I
- 3:49wouldn't generalize it. There are so
- 3:51many specializations. I happened to
- 3:53pick
- 3:53>> competitions,
- 3:55>> optimization competitions, because in
- 3:57the classic ones
- 4:00>> let's say I was doing quite okay, yes,
- 4:03on Topcoder, which was the biggest
- 4:05general ranking at the time. I was even
- 4:09fourth at one point out of, I don't
- 4:11know, several thousand people. For
- 4:14example, our local Tomasz Czajka had, I
- 4:17mean, he had a lot of success in that
- 4:19specific category. Generally, quite a
- 4:23few Poles achieved very good results,
- 4:25but in the optimization ones, it was
- 4:27just me and a big void after that. And
- 4:31there was that one competition where
- 4:33the LLM models finally matured enough
- 4:35to compete
- 4:37>> and you were the only protein-based
- 4:40computer, the only Homo Sapiens Sapiens
- 4:42, who
- 4:44>> beat a language model.
- 4:46>> Yes, that was a year ago. Times change,
- 4:48because there was an edition this year
- 4:50too, I don't know, we can talk about
- 4:52that later.
- 4:53>> Who won this year?
- 4:55>> Well, not a human.
- 4:57>> What was the highest place for a human?
- 5:00I mean, it was also just a model from
- 5:03OpenAI, and I was actually in Japan
- 5:05this year too, just so you know, and I
- 5:07was doing some coverage. I streamed on
- 5:10YouTube for 10 hours. A terrible task.
- 5:13So last year you won, this year a model
- 5:15won.
- 5:15>> Yes. And this year it totally crushed
- 5:18the humans, just as one might have
- 5:19expected.
- 5:20>> But you didn't compete. I didn't
- 5:21compete. Yes, I made a tactical dodge,
- 5:25>> to remain the champion, I sometimes
- 5:27laugh about it that, well, thanks to
- 5:29that I remained undefeated, so I can
- 5:32laugh or be happy about it. But also I
- 5:37decided that it was a bit pointless.
- 5:41One thing is that for me, just to get
- 5:43into these competitions, you need to
- 5:45invest some time, and the
- 5:47qualifications for these specific
- 5:49competitions that take place in Japan
- 5:51last all year. I have to compete from
- 5:53time to time. It would probably cost me
- 5:55, I don't know, like 200 hours over the
- 5:58year. That's 150, 200 hours. Well,
- 6:01that's a lot of time. Since I returned
- 6:03from Japan a year ago, I’ve suddenly
- 6:05had significantly less time.
- 6:08>> Why?
- 6:09>> Well, because my life has changed.
- 6:11Suddenly, tons of people wanted
- 6:12something from me, and I was trying to
- 6:14navigate this new world—what makes
- 6:16sense, what doesn't, what I even want,
- 6:18what I’m actually capable of
- 6:20achieving. But
- 6:22>> you certainly became the most
- 6:23talked-about programmer in the world at
- 6:25that moment.
- 6:27>> No, well, no, I wouldn't go that far.
- 6:30Yes, there are many more famous
- 6:32programmers in the world, for example
- 6:34John Carmack, who created id Software,
- 6:37Doom, Quake, and then worked at Meta.
- 6:40>> But you beat, you know, the model from
- 6:42the loudest, hottest company. Its CEO
- 6:45wrote about you.
- 6:47>> He did himself. So
- 6:50>> please tell me, what was this task, and
- 6:52is it even possible to describe to a
- 6:54layman like me, so that I could at
- 6:56least get close to understanding how
- 6:58you beat a language model?
- 7:01>> Is it, in general, even describable? It
- 7:05would be coolest if we could do a magic
- 7:07visualization, I think I could explain
- 7:09in two minutes more or less what it's
- 7:11about, but just describing it is almost
- 7:13impossible. I don’t know if I’ve
- 7:18ever had to describe it verbally. But
- 7:24>> you just said a moment ago that this
- 7:26would be the simplest possible problem
- 7:28inspired by authentic reality. Okay.
- 7:32The problem we had to solve, right,
- 7:36that generally these tasks always have
- 7:42a description, and then we are graded
- 7:44on tests, there’s a test generator,
- 7:46so, like I described the Voyager
- 7:48problem, there would be random maps
- 7:50generated there. Uh, so that’s one
- 7:55important point, but this task was
- 7:57about having a map, a two-dimensional
- 8:04diagram, let's say a chessboard, I
- 8:06don't know, 20 by 20, something like
- 8:09that. We have a list of robots on it. A
- 8:13robot is just a pawn that is on a
- 8:15specific square. And now our task is to
- 8:20determine some sort of path for moving
- 8:22these robots so that they end up in a
- 8:24specific final position. So, it
- 8:28generally sounds simple. The additional
- 8:31trick is that there are sometimes walls
- 8:33on the squares, and you can't just move
- 8:35freely from one place to another. Only
- 8:39one robot can be on one square at a
- 8:41time, but the main part was that we can
- 8:43connect robots into a group, and we can
- 8:46give commands for the whole group to
- 8:52move. And finally, we were evaluated on
- 8:56the number of operations we used, that
- 8:58is, the number of commands we issued to
- 9:01the robots. Therefore, the larger the
- 9:05group we put robots into, the more
- 9:07efficiently we can generally do it. So,
- 9:10I imagine that explained absolutely
- 9:13nothing,
- 9:13>> right? It gave a little, a little bit
- 9:15of direction. And the point was to
- 9:17optimize it, to make it simplest,
- 9:20>> so that as few, as few commands as
- 9:22possible were used for that. So now,
- 9:27you can think about these types of
- 9:29problems in different ways; you could
- 9:31do some, I don't know, complicated
- 9:33pathfinding. These problems are also
- 9:38extremely complicated in the sense that
- 9:40there is no known popular algorithm or
- 9:42metaheuristic that we could apply that
- 9:44would work; there's just no chance for
- 9:46that. So usually, you simplify these
- 9:50problems in your head somehow. And here
- 9:53I can say what I did and what gave me a
- 9:55very big advantage. Again, I suspect
- 9:57that will explain very little. Unless
- 9:59there were some visualization. Maybe
- 10:02you should think about drawing on the
- 10:04screen.
- 10:07>> So a big problem here was that we
- 10:09divide these robots into groups and
- 10:14because if we couldn't, if there were
- 10:16no possibility of dividing into groups,
- 10:18meaning every robot moves on its own,
- 10:20there are algorithms that are fairly
- 10:22standard and we could apply them, but
- 10:24they would give an absolutely tragic
- 10:26result here. And now, what I did for
- 10:28the first few minutes, being in this
- 10:30semi-conscious state, right? Because
- 10:32for those who don't know, I was–never
- 10:34have I been–as sleep-deprived as I
- 10:36was at that competition, and that was
- 10:37right at the start. So I was making
- 10:45these doodles on a piece of paper and I
- 10:47just started to wonder, for certain
- 10:49types of tests, what I need to do to
- 10:51get an optimal solution, that is, one
- 10:53where you simply can't come up with
- 10:55anything better. I found many
- 10:59situations where, if I combined all the
- 11:01robots into one group, I basically have
- 11:04an archetype of a solution where it
- 11:06seems to me that it's practically
- 11:07impossible to do anything better. So
- 11:10that, so that was the main observation,
- 11:13and again, if it were possible to draw
- 11:15it, it would be clearer what I'm even
- 11:17talking about. But based on that
- 11:20observation, I decided, okay, I'll
- 11:22implement this, see if it more or less
- 11:25works, if it's consistent with my
- 11:26intuition, and I left out a lot of
- 11:28things there because you could
- 11:30additionally insert walls on that
- 11:32chessboard, right? But, well, generally
- 11:36I had this idea, I had to implement it,
- 11:38check if it works, and what I wasn't
- 11:40sure about was, well, okay, I know it
- 11:42will work great in these small tests,
- 11:44in some specific tests, but I don't
- 11:46know if it will work in other tests. If
- 11:49it works poorly, should I use a
- 11:51completely different method, or patch
- 11:53it up, fix it so that,
- 11:55>> because you don't know in advance what
- 11:57the tests will be like. Do you know
- 11:58what the tests will be? Uh, it’s like
- 12:00this, I know, but in practice you have
- 12:0610 hours for a task, you read the
- 12:08content for 15 minutes, spend some time
- 12:10wondering what it’s even about. Also,
- 12:15there’s the time zone change and a
- 12:17ton of other issues, right? But yeah,
- 12:21it’s like a, it’s like a giant 10-
- 12:23hour sprint. the things that you
- 12:27implement there. Those were still the
- 12:29times when you didn't use agents, or at
- 12:31least those AI agents were absolutely
- 12:32hopeless.
- 12:33>> That was ages ago, a whole year ago.
- 12:35>> Yeah, yeah, yeah. That AI was useless,
- 12:38you know. I mean, I say that a bit
- 12:40mockingly, and a bit not, because for
- 12:42such heavy algorithmic tasks, when
- 12:44people asked me afterwards if I used
- 12:46agents, it was like, obviously not,
- 12:48because I would have just slowed down
- 12:50my work. Amazing. So you sit there, you
- 12:55investigate, you ponder, you think
- 12:57about those tests, you scribble things
- 13:00down, think about the solution
- 13:02archetype, and start programming, right
- 13:04? And you check if the ideas you’ve
- 13:06written down actually work out,
- 13:08>> right? And in practice, there are these
- 13:10problems too, that you make some
- 13:11mistakes, bugs, then not everything
- 13:13turns out the way you want. When you
- 13:15get bad results, you don’t know if
- 13:17it’s because my idea was bad, or
- 13:19because I implemented it poorly, or
- 13:21missed some condition, because, for
- 13:22example, I misunderstood the test
- 13:24generation. There are a lot of small
- 13:27things I have to think about non-stop
- 13:29and I keep it all in my head, so with
- 13:31every passing hour, there’s just more
- 13:33and more chaos that you have to manage.
- 13:37So, uh, this, you could say the
- 13:39intellectual effort and the
- 13:41concentration required for these types
- 13:43of problems is really gigantic. And
- 13:46it’s a marathon at the same time. I
- 13:50was generally doing exceptionally
- 13:51poorly, because there were two, two
- 13:53significant issues. One is that I
- 13:59practically hadn't slept, and I’ve
- 14:01been through many 24-hour programming
- 14:03contests where I didn't sleep. By the
- 14:07way, I competed in many of them with
- 14:09several people who now work in high
- 14:11positions at OpenAI.
- 14:13>> With Jakub Pachocki, for example.
- 14:15>> Yes, yes, among others. And I felt, at
- 14:19the beginning of these finals in Japan,
- 14:22like I used to feel at the end of 24-
- 14:24hour contests. I could talk for a long
- 14:27time about how badly I sleep and how
- 14:29irresponsibly I approach life. That’s
- 14:31a separate issue.
- 14:32>> Did you win with Pachocki?
- 14:34>> No, I was on a team with him a few
- 14:36times
- 14:37>> and probably with some successes.
- 14:40>> Uh yeah, that was it, but those are
- 14:42different stories. Alright, anyway,
- 14:45there is the issue that I was extremely
- 14:47sleep-deprived and simply put, my
- 14:49programming was going poorly. And the
- 14:50second point is that I actually don't
- 14:52really compete in these contests
- 14:53anymore. I competed in them a lot,
- 14:55let's say, until 2018. It was 2025, so
- 15:01for the, let's say, previous three
- 15:03years I barely participated in these
- 15:06contests, and to get to those 2025
- 15:08finals I had to take part in qualifiers
- 15:11that lasted through all of 2024. I
- 15:14remember that I just randomly made the
- 15:16decision, maybe I'll try to compete in
- 15:18the contests one more time. Why not?
- 15:20Because I'm a little bored. How much
- 15:22can one play games? Because that’s
- 15:24what you were mainly occupying yourself
- 15:26with in life at that time.
- 15:27>> Yes. I mean, I would like to portray it
- 15:29better, but I definitely spend a lot of
- 15:31time playing games. So, you know, you
- 15:36could say I came out of that
- 15:38programming contest retirement to just,
- 15:40well, maybe give it a try. So maybe I
- 15:44>> you weren't in form, kind of, you
- 15:45weren't, you didn't have
- 15:47>> that's exactly what I wanted to say.
- 15:48Maybe in 2025, when I wasn't competing
- 15:50in these contests anymore, I think that
- 15:53by the time of the contest—I probably
- 15:55give a different number of hours every
- 15:57time, but I'm guessing—that I
- 15:59programmed for maybe 50, 100 hours the
- 16:01whole year. In...I don't remember if it
- 16:04was August or...so yes, I wasn't in
- 16:06form, that was the second reason. But
- 16:09my idea, the simplification I had, was
- 16:11very cool and in a sense, it was
- 16:14accurate. Plus, all the, you could say
- 16:17research, decisions I made were very
- 16:19accurate. Well, clearly, that
- 16:24experience remains. It's not like it
- 16:28was forgotten, and my intuition about
- 16:30the problems also worked. I am also
- 16:32quite an unusual person compared to
- 16:34others competing in the contests.
- 16:36>> Why? Because right now, I don't have a
- 16:39perfect view, especially since many of
- 16:41the best people are from Japan, there
- 16:43is a language and cultural barrier to
- 16:45understand. But since I competed in
- 16:48these contests for a very long time, I
- 16:50met a ton of people who were at the top
- 16:52, and not just in programming contests,
- 16:54for that matter. I have this, generally
- 16:58most people are, I'll put it nicely,
- 17:02more introverted than I am. Uh, and
- 17:05these are people who have greater
- 17:06tendencies toward such an analytical
- 17:08approach. I mean, when I see a problem,
- 17:12I think about it for a dozen minutes
- 17:14max and start coding. Uh, I mean,
- 17:16that's how it used to be, right? Most
- 17:21people I know can sit for three hours,
- 17:23fill 15 pages of notes, think about
- 17:25everything, and only when they have
- 17:27analyzed it very thoroughly, do they
- 17:29then start. So, it’s not that I’m
- 17:35incapable of analytical thinking. I
- 17:39mean, I have tools in my repertoire
- 17:41that I can apply, but I’m the type of
- 17:43person who likes to do things, who
- 17:45wants quick results, and if something
- 17:47is wrong, I want to know as soon as
- 17:49possible and iterate fast. I guess
- 17:53that’s just a character trait, I
- 17:55don't know.
- 17:56>> Sure. So, 10 hours pass, and then what?
- 17:59Do you submit those programs and they
- 18:01get tested, or how does it work?
- 18:02>> They are tested on an ongoing basis, so
- 18:04we have constant access to leaderboards
- 18:09, rankings, and OpenAI was also sending
- 18:12in their model submissions in real time
- 18:14. Because of that, at the beginning,
- 18:20they pulled ahead quickly, as the
- 18:22models back then could already write
- 18:24code very fast, but their weakness was
- 18:26that they weren't good at coming up
- 18:28with ideas. They hit a wall pretty
- 18:32quickly. So, before the competition, I
- 18:36read a few research papers on fairly
- 18:41similar systems that solved similar
- 18:44problems, so I had some intuition about
- 18:46how well OpenAI might do and what I
- 18:48could expect.
- 18:50>> Because you knew there would be this
- 18:51human vs. LLM test, that they would
- 18:53enter their model?
- 18:54>> I found out about a week or maybe two
- 18:56weeks earlier, I think. Well, they
- 18:58generally didn't really brag about it
- 19:00at all. I assumed they weren't too
- 19:02confident
- 19:04>> about the results, so as usual, until
- 19:07they perform great, well, then
- 19:09>> they don't announce it.
- 19:10>> As if they had taken last place, it
- 19:11would be like they weren't even there.
- 19:13>> Right.
- 19:14>> So there was no official information
- 19:15from them that they were participating.
- 19:18There was only information from the
- 19:19AtCoder organizers, and since they have
- 19:21relatively low reach outside of Japan,
- 19:23it didn't really break through. And
- 19:28then I think that there were more or
- 19:32less four or five hours into the
- 19:34contest. I realized that, like, I
- 19:38noticed something in my solution, and I
- 19:40knew that if I fixed it, I would shoot
- 19:42way ahead, because at that point, after
- 19:45four or five hours, it was basically
- 19:47OpenAI and the rest of the people
- 19:49competing in their own separate
- 19:50category.
- 19:53>> Is it judged by points, by time, by...
- 19:56>> By points, so you get a certain number
- 19:58of points from each test. And we have
- 20:01relative scoring, where the best person
- 20:03on a given test gets, let's say, one
- 20:05point, and everyone else gets a
- 20:07fraction of that.
- 20:08>> Sure.
- 20:08>> And that's why the results could jump
- 20:10around so much. Anyway, I saw that I
- 20:12had this perspective that if I
- 20:14implemented a certain thing, I would
- 20:16move forward significantly, which was
- 20:18also important because I was constantly
- 20:20struggling with falling asleep, so I
- 20:22needed some motivation.
- 20:24>> And this psychological aspect is
- 20:26actually extremely important during
- 20:28competitions. Anyway, I did it with the
- 20:32last of my strength, I launched it and
- 20:34saw that I actually beat OpenAI. I
- 20:38don't remember, it was 6 hours, maybe 7
- 20:40hours into the competition, so I was in
- 20:43a bit of a shock. I remember what I did
- 20:48then, because at that competition we
- 20:50uniquely had this strange opportunity
- 20:52to go into a little room where we could
- 20:54speak to a camera. No, no, I don't know
- 20:59if you saw that. No, I don't know if
- 21:01many people even knew about that.
- 21:03Anyway, I remember that when I saw I
- 21:05had overtaken OpenAI, I decided,
- 21:07half-conscious, that I would go, though
- 21:09I didn't yet know what I would say. I
- 21:12went to that room and for two minutes I
- 21:14probably babbled some words. I don't
- 21:17remember exactly what I said, but I
- 21:19remember that at the end I said: "
- 21:21OpenAI, your model sucks." OpenAI, your
- 21:25model sucks. I know, very ambitious of
- 21:28me. Anyway, you could say that in my
- 21:32barely conscious mind, I was very proud
- 21:35of the words I had spoken. A little
- 21:38less so now, but generally, after those
- 21:40seven hours, I was practically certain
- 21:42that I would win,
- 21:43>> because I knew I still had plenty of
- 21:45areas where I could improve, and I
- 21:47assumed that OpenAI wouldn't improve
- 21:48much more.
- 21:49>> Okay. Was 2025 the last year a human
- 21:52won those competitions?
- 21:55>> Yes. This year there was already a big
- 21:58difference in OpenAI's favor.
- 22:00>> Yes. And OpenAI totally crushed it. So
- 22:04I can add an interesting fact. I didn't
- 22:06qualify for the 2026 competition
- 22:08because I didn't participate in the
- 22:10qualifiers at all. I decided it made no
- 22:12sense. But I laughed that the
- 22:15organizers would probably reach out to
- 22:17me at some point, because it would be
- 22:18silly, right? considering how widely
- 22:23that information spread across the
- 22:25world.
- 22:26>> And so they wrote to me at the turn of
- 22:28January and February asking if I would
- 22:30like to participate unofficially. I
- 22:32thought about it for a while, but I
- 22:34wrote back to them that I wouldn't. Uh,
- 22:36I don't really see the point. Looking
- 22:41at the narrative globally, aside from
- 22:43my own personal issues, I only stand to
- 22:45lose, so I don't want to give OpenAI
- 22:48such an easy win.
- 22:51>> Of course.
- 22:52>> And were you sure that you wouldn't be
- 22:53able to win?
- 22:54>> Yeah, yeah, yeah.
- 22:55>> So that would be like setting
- 22:56themselves up for failure in a way.
- 22:58>> Yes. So why was I sure? Well, first of
- 23:03all, OpenAI wouldn't have entered if
- 23:05they weren't sure they would win.
- 23:07That's the first thing.
- 23:08>> Right. Uh, because the narrative in
- 23:142025 was kind of that OpenAI got a gold
- 23:17medal at the ICPC informatics olympiad,
- 23:19which is a collegiate-level team
- 23:21programming competition, and they
- 23:24solved all the problems, drastically
- 23:26crushing all human teams. Well, in my
- 23:32competition, they took second place.
- 23:35And at the informatics olympiad, I
- 23:37don't remember the result, but they got
- 23:39sixth or seventh place, and that was
- 23:41all within two months. So, considering
- 23:44how huge the progress in development
- 23:46was, every person, every programmer who
- 23:49uses these models assumed that, well,
- 23:51surely they'll win now. Sure. And in
- 23:542026, they solve a millennium problem,
- 23:57so basically
- 23:58>> it's possible that was actually a bit
- 23:59later, but
- 24:00>> sure, sure. So, it's game over. I mean,
- 24:03are you saying that it's not just this
- 24:05one—you correctly distinguish it, and
- 24:07I even understood it—that specific
- 24:09programming contest you entered, that
- 24:11field, meaning optimization, a human
- 24:13will probably never win there as you
- 24:15say. In other programming contests as
- 24:17well
- 24:18>> it would also be hard for
- 24:19>> The human era has ended, in a sense.
- 24:23Well, yes and no. I mean, because we
- 24:25are currently at a moment where human
- 24:27plus AI
- 24:29>> is much more effective than AI alone,
- 24:33>> at least in the programming world. But
- 24:37then again, it's probably just like in
- 24:39chess, just like in other fields. We
- 24:42are always the bottleneck at some point
- 24:44.
- 24:45>> We humans,
- 24:46>> yes, we humans. I mean, I say always,
- 24:48but at least in the case of these
- 24:50programming things, that's how it will
- 24:52be. Actually, it's already a bit like
- 24:54that today.
- 24:55>> So that bottleneck will be eliminated
- 24:56at some point.
- 24:58>> I don't think so yet, because we won't
- 25:00talk about anything serious yet. For
- 25:02now, I'm thinking about optimization,
- 25:04fast programming, and effective
- 25:05programming. For now, it's LLM and
- 25:08human. Soon, it will probably be the
- 25:10LLM alone.
- 25:10>> Yes. Well, let's say as of today, if I
- 25:14were to work on optimization problems,
- 25:16what I can contribute is what people
- 25:18popularly call "research taste," you
- 25:21know, a researcher's taste, right?
- 25:24That's a bit of a silly translation,
- 25:26but actually, it's a silly term in
- 25:27English too. So, you could say it's an
- 25:30intuition I've gained over many years,
- 25:32and I have a rough idea that these
- 25:34types of approaches to the problem are
- 25:36more promising. And my role essentially
- 25:43boils down to supervision, so I know
- 25:45the agent is doing sensible things,
- 25:47giving it guidelines, setting
- 25:49intermediate goals, and breaking
- 25:51problems into smaller ones, because the
- 25:53agent can't always handle them on its
- 25:55own. But looking at it from the other
- 26:01side, it's possible that if you just
- 26:03gave it 1000 times more computing power
- 26:05, my advantage would diminish
- 26:07significantly.
- 26:09>> Sure. All right, so you win this famous
- 26:12competition and at least for a moment,
- 26:14you become the most famous programmer
- 26:16in the world.
- 26:17>> I'll agree with that.
- 26:18>> at that moment. Yes. Sam Altman writes
- 26:20on Twitter, and you probably didn't
- 26:21just talk to him; did he reach out,
- 26:23contact you, call, or write? Besides
- 26:24him writing on Twitter: "
- 26:26>> No, no, no, that was strange, just so
- 26:28very strange. I think he has more
- 26:30important things to do. From my
- 26:32perspective, by the way, I can explain
- 26:35why that probably happened."
- 26:37>> Uh-huh.
- 26:39>> At least that’s how I perceive it,
- 26:41because in general, it was a massive
- 26:43success for OpenAI, even though they
- 26:45took second place, because people
- 26:47didn't really expect that. The people
- 26:50in these competitions I enter didn't
- 26:52expect it. And we had a poll in our
- 26:55micro-community on what people
- 26:57predicted, and I don't remember if
- 26:58there were any votes that they would be
- 27:00able to win, so most people in that
- 27:02community were in quite a shock. I was,
- 27:06let's say, a bit more optimistic—or
- 27:09pessimistic, depending on how you look
- 27:11at it—but from my perspective,
- 27:17>> it was such an incredibly unlikely
- 27:19sequence of all these things that
- 27:21happened, because OpenAI took second
- 27:23place, right? And it paid off for them
- 27:28to promote me a bit to show people that
- 27:31, firstly, they didn't lose to a random
- 27:33person, but to this guy who had simply
- 27:35won a whole host of other competitions.
- 27:38>> So, "We lost, but to a really strong
- 27:40competitor—in fact, the strongest one
- 27:42we could have lost to."
- 27:43>> Yes. And the second issue is that I
- 27:45used to work there, so it's a double
- 27:47victory for them.
- 27:48>> You actually worked at OpenAI once,
- 27:50>> right? We'll talk more about that later
- 27:52. Exactly, but if you have anything
- 27:54else to add about OpenAI's reaction to
- 27:56your victory.
- 28:02>> What pleases me from a purely human
- 28:04perspective is that this was the only
- 28:05major event among their successes that
- 28:08turned into more of a human victory
- 28:09than a corporate one, so I'm a bit glad
- 28:12they didn't get extra PR points for it.
- 28:17I mean, they got a bit of attention,
- 28:19but yeah, if someone else had won
- 28:21instead of me—I mean, if we had
- 28:22swapped me with another person who was
- 28:25there—it probably wouldn’t have
- 28:26been such a big deal.
- 28:28>> Sure. And what happens in your life
- 28:30when you win that competition?
- 28:31>> Um, I’m still in Japan for three
- 28:33weeks and I want to go sightseeing, but
- 28:35I can't because I get so many messages
- 28:37a day from journalists and
- 28:39non-journalists that I don't really
- 28:40know what to do and I wonder what all
- 28:42this is for. And at the same time,
- 28:45it’s 38 degrees in Japan, and I
- 28:47normally don't leave the house when
- 28:48it’s over 30 because I don't really
- 28:51like it that warm, and it was just
- 28:52impossible to handle there, plus
- 28:54everything is made of concrete. I
- 28:56don’t recommend it. I mean, Japan is
- 28:58wonderful, but I don’t recommend
- 28:59being there in those conditions. Uh,
- 29:01I’m sleep-deprived the whole time
- 29:03because I’m replying to so many
- 29:05people that I only sleep three hours a
- 29:07day, so it’s just great. Uh, and
- 29:09I’m wondering in general what, why,
- 29:12anything, if they’re writing to me
- 29:14from TV, asking me to record something
- 29:16for them, does that even make sense. So
- 29:19, because I’d never had any dealings
- 29:21with the media.
- 29:22>> And aren’t you getting some amazing
- 29:24job offers?
- 29:27>> No. And by the way, that was such a
- 29:29funny time when Meta was offering those
- 29:32giant piles of money for employees. And
- 29:36I remember there was a lot of…
- 29:39>> A lot of my friends were joking, asking
- 29:41me if I’d gotten an offer from Meta
- 29:43yet. I didn’t end up getting any in
- 29:45the end. Which would actually surprise
- 29:48me, honestly. I mean, surely…
- 29:50>> Surely there were some from smaller
- 29:51companies, but I didn’t get one from
- 29:53any big ones. Uh, something worth
- 29:56remembering, just a side note in the
- 29:58case of Meta. Meta back then wasn’t
- 30:00paying for…wasn’t paying for very
- 30:02good people. They were paying for
- 30:06people working at another lab, right?
- 30:09For them to switch over, so they’d
- 30:11have up-to-date information on how the
- 30:13competition’s models were being
- 30:15trained, so that’s where that money
- 30:16came from.
- 30:17>> Just to know some of the opponent’s
- 30:19secrets. Yes, yes, yes. And there are
- 30:22no non-competes in California.
- 30:23>> Okay. So you can talk about it, you can
- 30:25tell people about it.
- 30:26>> I mean, sure, I don’t know what the
- 30:28exact limits are, but in practice, that
- 30:31was one reason why AI development was
- 30:33quite even. But anyway, side note.
- 30:35>> Exactly. That’s interesting. So what
- 30:37now? You come back from Japan, and what
- 30:39happens next?
- 30:40>> I go through my backlog of messages.
- 30:44There were some interviews, but in the
- 30:46end, I did almost nothing. I was mostly
- 30:48ghosting people. And what was…what
- 30:51was happening in practice, I was
- 30:53wondering, I don't know, I have a bit
- 30:56of a memory gap. I was probably still
- 30:58catching up on sleep, I went to a few
- 31:04meetings, panels, I appeared somewhere
- 31:06a few times, but the first month was
- 31:08mostly spent finding my way in a new
- 31:10life and wondering, what to do with
- 31:15this situation.
- 31:16>> Exactly. One thing I knew was that I
- 31:20didn't really want to work in any lab
- 31:21anymore. I mean, that was a decision I
- 31:24had made earlier. I also don't know if
- 31:32we want to smoothly transition to the
- 31:34risks, because that is one thing I
- 31:36considered back in 2020-2021, to focus
- 31:38a bit on raising awareness about what
- 31:40will happen to the world
- 31:44>> under the influence of AI development.
- 31:46>> Yes. Yes. Due to the development, which
- 31:48is extremely fast and will likely be
- 31:50even faster.
- 31:51>> Are you working anywhere today?
- 31:52>> No.
- 31:53>> How long have you not been working?
- 31:562020. Well, basically since I left
- 31:59OpenAI.
- 31:59>> And how long did you work there?
- 32:02>> It’s hard to answer that question
- 32:05because originally, when I joined in
- 32:072017, I worked as a contractor. I went
- 32:10there for two months, came back to
- 32:12Poland, and theoretically pretended I
- 32:14was working. I mean, I wrote to them
- 32:16that I was doing something, but in
- 32:18reality, I wasn't doing anything. But
- 32:20they know, so, so I was a bit there, a
- 32:22bit in the States, and then in Poland.
- 32:28And here's a funny thing, another
- 32:30digression, OpenAI considered hiring me
- 32:32in 2016 shortly after they opened, but
- 32:35I turned down their offer because when
- 32:37I talked to the people, they seemed
- 32:39unpleasant to me.
- 32:42>> Unpleasant people from Open in 16. 16
- 32:48was the year I joined the project—I'm
- 32:50speaking terribly chaotically—but I
- 32:52joined the project
- 32:53>> 2017
- 32:56>> when OpenAI was working on a Dota bot.
- 32:59Dota is a big esports game. At the time
- 33:02, I don't know how it is now, but at
- 33:03the time it had the largest prize pools
- 33:05in the world. I think The International
- 33:08, I don't remember, I don't know, 40
- 33:10million dollars, maybe 80 million. Some
- 33:13kind of absurd money. Yeah, that kind
- 33:15of money where a father can just tell
- 33:17his son: "Why are you kicking a ball?
- 33:19Go play games." Yes.
- 33:23>> Why, why be the next Lewandowski?
- 33:25Better to become a pro gamer. You’ll
- 33:27have much more money there. Well, maybe
- 33:30so. 2017 project. And how long were you
- 33:33at OpenAI? Until when? Remind me.
- 33:36>> I was just a bit on and off, at the
- 33:39beginning of 2019 I joined full-time. I
- 33:44moved to San Francisco and lived there
- 33:46for a bit over a year, and I finished
- 33:48working there when the COVID lockdowns
- 33:50started. I literally flew back on the
- 33:54last flight from San Francisco to
- 33:55Poland. That’s also a funny story.
- 33:57>> So, 2020.
- 33:58>> Yeah, yeah, yeah. Uh, that was back
- 34:01when most of the world hadn't heard of
- 34:02what OpenAI was. That was still before,
- 34:05uh, the publication of, uh, ChatGPT 3.5
- 34:07, I think that was three years ago,
- 34:09right? November 2022, if I recall
- 34:11correctly.
- 34:11>> Yeah, yeah, yeah. I admit that I always
- 34:14mix up dates, but yeah, in the
- 34:15beginning, some people had heard of
- 34:17OpenAI because Elon Musk was involved,
- 34:19so it was always the "Elon Musk Company
- 34:21," which wasn't true, but
- 34:22>> besides that, it wasn't a company at
- 34:24the start,
- 34:25>> it wasn't a company. Yes,
- 34:26>> exactly.
- 34:28>> It was meant to be open. What? An
- 34:30association, a foundation, which...
- 34:33>> Well, it was a non-profit; I can't tell
- 34:35you the exact legal structure, because
- 34:37I don't know much about those things,
- 34:39but yeah, it was a place that brought
- 34:41together researchers who were just
- 34:43experimenting. Originally, OpenAI was
- 34:49founded as, if I remember correctly—
- 34:52and this is probably fairly well
- 34:54documented—as a counterbalance to
- 34:56DeepMind at the time, since, well,
- 34:58DeepMind had been bought by Google. I
- 35:03don't remember what year, and at the
- 35:07time, DeepMind was the only company,
- 35:10the only lab in the world, aside from
- 35:12some academic stuff, that was dealing
- 35:14with, let's say, more advanced
- 35:16applications of artificial intelligence
- 35:18. And so, OpenAI was meant to be a
- 35:23counterbalance. Yes. Well, no one wants
- 35:25, like, even back then, people were
- 35:27wondering what could be squeezed out of
- 35:30this artificial intelligence, right?
- 35:34Deep learning, or, you know, everything
- 35:37related to neural networks, exploded
- 35:39after 2012, and that was a breakthrough
- 35:41; we already had very good models for
- 35:43image recognition, so there was this
- 35:45sort of shock because suddenly we had
- 35:48new classes of problems we could solve.
- 35:52Not long after, there was also AlphaGo.
- 35:56If I remember right, the first AlphaGo
- 35:58was in 2015 before it was against Lee
- 36:00Sedol, so OpenAI must have been founded
- 36:03after that, because it was late 2015.
- 36:05>> And what exactly did you do at OpenAI?
- 36:07>> So, I worked on the Dota project. I
- 36:10literally came there for that. I was
- 36:12recruited by Jakub Pachocki, by the way
- 36:14. Well, because generally, I've known
- 36:17Jakub for a very long time. Back from
- 36:22programming competitions, and he wrote
- 36:25to me saying, "Hey, we have this funny
- 36:27project, we're going to play a computer
- 36:29game," meaning we were going to try to,
- 36:32well, play Dota. I was playing a bit of
- 36:34Dota at the time too, so, let's say, I
- 36:36had that domain knowledge, let's call
- 36:39it, and I figured that it actually
- 36:44sounded cool, because it combined my
- 36:46various interests, you know? And
- 36:51>> well, Kuba is a great guy, so it was
- 36:53super cool to work with a friend.
- 36:55>> Kuba is cool. I can go there later,
- 36:58after I didn't take that offer in 2016,
- 37:00there are some more details there, but
- 37:03I finally wrote back to them that, well
- 37:05, I don't really want to, maybe
- 37:07remotely. So when I wrote that I wanted
- 37:10to work remotely, they weren't really
- 37:12interested. So I felt a bit of regret
- 37:14after that, and after that exchange a
- 37:16year earlier I decided: "Well, okay,
- 37:18maybe it's not that stupid, they had
- 37:21already grown, they had about 40 people
- 37:23back then, something like that, so I
- 37:25came for interviews, and they were
- 37:29already at a quite advanced stage of
- 37:31that project. And that was two months
- 37:35out. I mean, when I arrived at the
- 37:37beginning, it was a bit earlier because
- 37:39I had something, but I also have a lot
- 37:41of strange, funny stories about
- 37:42recruitment interviews. I had one
- 37:44interview where I got a question and
- 37:46spent an hour explaining to the guy
- 37:48that it was a very bad question and
- 37:50that it was a very poor way to select
- 37:52candidates. But that's just to show how
- 37:54easy I am to work with. Anyway, I had
- 37:58something, but the interviews, I talked
- 38:00with everyone, I could meet the team,
- 38:02see who works there, what the office
- 38:05looks like, what the atmosphere is, and
- 38:07I decided that, actually, it would be
- 38:09cool to do something new in life. I had
- 38:12never worked for anyone anywhere before
- 38:14. So maybe this is the moment, I
- 38:17wouldn't call it time to grow up, but I
- 38:19just wanted to experience something new
- 38:21. And so I made that decision. I don't
- 38:24remember which month it was, but I
- 38:26remember that finally when I arrived
- 38:28there, it was two months before The
- 38:30International, which are the Dota world
- 38:32championships, that's what they're
- 38:33called, and it's a huge event in a
- 38:35sports arena for 40,000 people that
- 38:37lasts a week. And we were appearing
- 38:41there on stage, only back then we
- 38:43weren't doing full Dota yet, because
- 38:45it's an extremely complicated game, but
- 38:47a show match, which was a show match.
- 38:57One versus one, so a very simplified
- 38:59form of Dota, you could say, but we
- 39:01managed to win that. It was an
- 39:05eight-person team back then. Eh, quite
- 39:08overwhelmingly. So that was actually
- 39:11the first thing that made OpenAI become
- 39:13known. It was, like, the first major
- 39:17piece of information that went out that
- 39:19wasn't about Elon Musk founding OpenAI.
- 39:22>> Mhm. So, in terms of media, it was a
- 39:27very big success, because, well, the
- 39:29world championships, I don't know how,
- 39:31but I assume a few million people watch
- 39:34it. It's possible I just seriously
- 39:36overestimated that. Uh, and so then I
- 39:38stayed with the team, although again I
- 39:40worked somewhat remotely, somewhat
- 39:42on-site, I would travel there from time
- 39:44to time, and then we were already
- 39:46making a fuller version of the game. It
- 39:50wasn't 100%, but it was a version of
- 39:52the game that I thought was impossible.
- 39:57Or let me put it differently. As a
- 39:59matter of fact, during my sort-of
- 40:01recruitment when I spoke with the
- 40:03project lead, who was Greg, Greg
- 40:04Brockman—he is one of the more
- 40:06important people at OpenAI, previously
- 40:08you could say he was the most important
- 40:11, because he was de facto running the
- 40:13company, though not alone. I told Greg
- 40:19something like, it seems to me that
- 40:28writing a Dota bot for this game is
- 40:30practically impossible, because if it
- 40:32were possible with current technology,
- 40:34we would probably have tools and models
- 40:36capable of solving a ton of other
- 40:38problems. Simply put, from my
- 40:42perspective, the degree of complexity
- 40:44of the decisions that need to be made
- 40:46in this game was so incredibly far from
- 40:48everything we had done in machine
- 40:50learning so far that, well, something
- 40:52didn't add up for me. On the other hand
- 40:58, for me, well, I’ve always been
- 41:00someone who likes these impossible
- 41:02problems because they are motivating,
- 41:04so if, like, I don’t get any pleasure
- 41:06from doing something where I know
- 41:08exactly how to do it—so that was,
- 41:10well, I really got tangled up there.
- 41:15Anyway, I stayed on and worked
- 41:17practically until 2020.
- 41:19>> Uh, but that’s a year,
- 41:22>> uh,
- 41:23>> plus a bit.
- 41:25>> Right, so I joined in mid-2017, I was a
- 41:29contractor and until
- 41:33>> and in 2019 I switched from a
- 41:34contractor to a full-time position,
- 41:37where then I
- 41:38>> was based in San Francisco.
- 41:40>> Yeah. And why did you leave? There are
- 41:43a lot of strange, small stories.
- 41:49Meaning there were a lot of personal
- 41:51conflicts. As I’ve hinted a few times
- 41:55, I am a rather specific person to work
- 41:57with. There was a big, from my
- 42:02perspective, cultural conflict. For
- 42:05those who don’t know, our work
- 42:10culture is much more direct, where
- 42:16where we can reasonably tell each other
- 42:18that something is wrong, that someone
- 42:20did something poorly. In the States,
- 42:22especially in San Francisco in Silicon
- 42:25Valley, the work culture is such that
- 42:27you don't really give that feedback
- 42:31directly to your colleagues. Very often
- 42:34, everything goes through HR. We also
- 42:37worked in such, well, extremely
- 42:39stressful conditions. Towards the end,
- 42:42we had such a grind that I remember I
- 42:44was sleeping three hours a night for a
- 42:47month, normally I sleep 7, because I
- 42:49was working 16 hours a day. On what?
- 42:52>> On Dota, right before we—we couldn't
- 42:54push back deadlines because the world
- 42:56championships have a set date, so we
- 42:58had to be ready exactly at that moment.
- 43:01We had to.
- 43:03>> So, these were quite stressful
- 43:04conditions. It also turned out that
- 43:07there were a surprisingly high number
- 43:09of toxic people, and I wasn't one of
- 43:11them, at least that's my opinion of
- 43:13myself. And I tried to resolve a lot of
- 43:16different conflicts. Uh, in the end, I
- 43:18took a massive hit. In the meantime,
- 43:20OpenAI was growing extremely fast, and
- 43:23companies—when a company is small, at
- 43:25the start when there were 40 people, it
- 43:28was de facto a startup. It was pleasant
- 43:31; everyone knew everyone. We had common
- 43:34goals. Once a company grows, middle
- 43:37management comes in, which often has
- 43:39different goals, because a bit of
- 43:41corporate politics starts, where
- 43:43everyone is optimizing to get ahead as
- 43:45fast as possible. This causes the
- 43:48company to have many people who no
- 43:53longer want, let's say, the general
- 43:55good, but begin to—well—
- 43:58>> look after their own private interests.
- 44:03So, what ultimately happened is that
- 44:05because of a lot of internal politics
- 44:07—which, by the way, didn't just give
- 44:09me trouble—I lost absolutely all
- 44:15motivation to work. Yes, I also lose
- 44:19motivation relatively easily, and it
- 44:21ended up with me practically not
- 44:23working for the last few months, while
- 44:25at the same time having quite good
- 44:27relations with the leadership. So I had
- 44:31a somewhat paradoxical situation, where
- 44:33many people who were high up liked me
- 44:35and, as it were, worked with me, so
- 44:38they considered me very valuable. Plus,
- 44:43well, I have a somewhat unconventional
- 44:45background, because most of the people
- 44:51working at OpenAI at the time were
- 44:53academic researchers, and I was that
- 44:55person who, in fact, doesn't even have
- 44:57a higher education. And has been an
- 45:01autodidact their whole life. On top of
- 45:05that, I am a generalist who has dealt
- 45:07with a mass of various random things.
- 45:12Also, having experience in game design
- 45:14turned out to be very valuable. While
- 45:18working on Dota, I was the only person
- 45:20who practically didn't program, because
- 45:22I was evaluating how good our bot was
- 45:24and what needed to be changed. For a
- 45:26moment, I even managed the project a
- 45:28little bit, which was absurd, since it
- 45:30was my first job. But luckily, it
- 45:32lasted a short time. When we had a
- 45:37moment where anyone from the internet
- 45:39could come and play with our bot, I
- 45:40managed the entire community. In the
- 45:46sense that we started a Discord, we had
- 45:48over ten thousand people on it, and I
- 45:50tried to manage all that chaos of crazy
- 45:53, emotional 14-year-olds who like—
- 45:55well, obviously the age range is much
- 45:57broader—who are very active on our
- 45:59Discord and say a lot of strange things
- 46:01. So, I was usually doing the thing
- 46:05that no one else wanted to do, because
- 46:07I enjoyed it.
- 46:08>> Sure. Do you regret, a little bit,
- 46:11leaving a company that is so loud today
- 46:14, that is so very hot?
- 46:16>> No, that is one thing I am sure of. I
- 46:20mean, first of all, I predicted that
- 46:22this would happen.
- 46:24>> Yes.
- 46:24>> Yes. Uh. more or less. I mean,
- 46:33obviously I wasn't certain, right, no
- 46:35one ever is, but I did have my
- 46:37predictions about how quickly
- 46:39artificial intelligence would develop.
- 46:44I also had my own predictions about
- 46:46when other companies would suddenly
- 46:48wake up to realize that this
- 46:49development is actually incredibly fast
- 46:52and how many experienced people there
- 46:54are to quickly catch up. And well, I
- 47:01was fairly convinced, like, what I was
- 47:03certain of was that this development
- 47:05would be, well, so incredibly fast. Of
- 47:08course, I had extra information,
- 47:09because being inside the company people
- 47:14were wor—if I remember correctly, GPT
- 47:17-2 was released in 2019, I think on
- 47:19Valentine's Day, and it was the first
- 47:23language model that, from my
- 47:24perspective, was capable of generating
- 47:27such coherent, meaningful text. And it
- 47:32kind of blew my mind when I saw it,
- 47:34because from my perspective, it was
- 47:36that problem where we had no idea how
- 47:38to do anything, that it was so
- 47:40unattainable, and yet, suddenly it
- 47:42turns out that at a relatively low cost
- 47:44, we are able to create something
- 47:46sensible. And of course, after GPT-2
- 47:51was created and it was, well, also a
- 47:54pretty decent marketing success, then
- 47:56suddenly—well, not suddenly, but many
- 47:59more people inside the company, which
- 48:01was still a tiny research firm, were
- 48:03curious, they were curious how much
- 48:06could be squeezed out of it. So if I
- 48:09remember correctly, the premiere of GPT
- 48:12-3 was only somewhere around mid-2020,
- 48:16but in the meantime, those people were
- 48:19constantly developing this project, all
- 48:21the time, so I saw the progress inside
- 48:23the company firsthand, so you could say
- 48:25I had a few points on that axis. First,
- 48:30I saw how the model's power changes
- 48:32with scale, meaning with the increase
- 48:34in the number of parameters, computing
- 48:37power, and data improvement. And, well,
- 48:44the shortcomings were clearly visible,
- 48:47among others, the fact that those
- 48:48models at the time were, let's say,
- 48:50decent at storing knowledge, but they
- 48:53didn't want to cooperate at all and
- 48:55absolutely couldn't handle any tasks
- 48:57that required some thinking. So all
- 49:02those things that are unfortunately
- 49:04sometimes repeated, that these are
- 49:06stochastic parrots and just blindly
- 49:08repeat what they had in their training
- 49:10data, that was still true back then.
- 49:13But, well, so I just, you could say,
- 49:15projected the trajectory forward.
- 49:17>> in 20 17 when you started, 40 people.
- 49:23How many Poles were in the company then
- 49:25? There was you, there was Jakub
- 49:27Pachocki,
- 49:27>> there was Szymon Sidor, there was
- 49:32Rafał Józefowicz, and Marcin
- 49:34Andrychowicz. They actually joined very
- 49:38early, they were some of the first
- 49:39people when there were still 15, and I
- 49:41think there were a few more Poles. I
- 49:43don't know, I don't want to list every
- 49:45one.
- 49:45>> That's still a high percentage, right?
- 49:47>> Yes, yes.
- 49:48>> Out of 40 people.
- 49:49>> Uh, well, seven people, I think. Sure.
- 49:52Alright, so you leave OpenAI and you
- 49:55say:" That was your first job in 2017
- 49:57as a contractor until 2020, and then? "
- 50:05The main thing was that it was a very
- 50:07strange moment because, as I mentioned
- 50:09earlier, I returned on literally the
- 50:11last plane, so at the same time I quit
- 50:14my job at OpenAI, returned to Poland,
- 50:16and the lockdowns began. And I have
- 50:21never in my life been as confused as I
- 50:23was then, because I had absolutely no
- 50:25idea what had happened, having
- 50:27completely changed hemispheres. So I
- 50:31think for the first month I was
- 50:32somewhat happy that I could lock myself
- 50:34in the house and didn't have to do
- 50:36absolutely anything, but that passed
- 50:38quickly. I remember spending a lot of
- 50:43time wondering, okay, but what do I
- 50:45want to do with my life now, and I
- 50:50don't remember what conclusions I
- 50:52reached back then. I felt a bit sad
- 50:54that I had left, well, because I had to
- 50:56come up with a new plan for my life. My
- 50:59idea was that I wanted to return to
- 51:01making games. That had always been my
- 51:03dream. It was a place where I felt,
- 51:06let's say, professionally fulfilled,
- 51:08but also artistically. It was hard for
- 51:14me to get motivated. It also ultimately
- 51:17turned out that the fact that the world
- 51:21wasn't functioning normally left a big
- 51:23mark on me, because well, in order to
- 51:28function, I need people. It turned out
- 51:32that a large part of me was traveling,
- 51:34going to different events around the
- 51:36world, meeting people, and suddenly it
- 51:39was all taken away, so, well, I was
- 51:41terribly demotivated then. Uh, and one
- 51:46of the things I was thinking about a
- 51:48lot—so besides that, there was
- 51:50another issue, which was that since I
- 51:52was in that lab, I had an awareness of
- 51:54how this artificial intelligence would
- 51:56develop, so I thought quite hard about
- 52:03perhaps dealing a bit with the topic of
- 52:08raising awareness about the risks. And
- 52:11I was in a fairly incredible situation,
- 52:17that I was one—well, incredible—I
- 52:20was one of the few people in the world
- 52:22who could speak about it in a fairly
- 52:24competent way, because in the world, to
- 52:26be able to speak sensibly about these
- 52:28trajectories, you had to be in a lab.
- 52:32There were actually two of those labs.
- 52:35There was OpenAI and there was Google.
- 52:39And now, there weren't people who were
- 52:41leaving those places and saying," Okay,
- 52:43I don't feel like working anymore, I
- 52:45don't care about anything. "And at the
- 52:48same time, if I had decided to, say,
- 52:50start publicly complaining about
- 52:52institutions like OpenAI, I would have
- 52:54made it very difficult for myself to
- 52:57find a job. So there simply weren't any
- 53:01people who were put in this situation.
- 53:03Of course, in the meantime, it turned
- 53:05out that I don't like writing. Like,
- 53:09that was one big problem, and the
- 53:11second is that, well, purely
- 53:13practically, if I write something like
- 53:15that, I would want someone to read me,
- 53:17so unfortunately I would have to put a
- 53:19lot of work into marketing myself.
- 53:27Nobody likes that, especially me, so
- 53:29you could just say I gave up, because
- 53:31it was too much effort and so I chose
- 53:33comfort.
- 53:37>> And then you didn't work anymore.
- 53:39>> Well.
- 53:40>> So you only worked there for those few
- 53:43years, 2017.
- 53:44>> Yes. I mean, just so it's clear, I was
- 53:46doing some things here and there. Yes.
- 53:48It's not like I'm just sitting on my
- 53:50butt on the couch watching series and
- 53:51soap operas. I don't know why I said
- 53:56that. Ah.
- 53:57>> Probably your guilty pleasure. You
- 53:59don't even know about it.
- 54:02>> Let's not talk about those things.
- 54:03>> Well, okay, listen, but this is quite
- 54:05amazing, because you're talking about
- 54:09awareness of the threats associated
- 54:11with artificial intelligence in 2020.
- 54:13>> Yes.
- 54:14>> A very long time ago from the
- 54:15perspective of what's happening now.
- 54:17Yes.
- 54:18>> What kind of threats did you see back
- 54:20then? So, you know, back then I'm
- 54:27trying to remember, because I have
- 54:30>> But you already saw the threat back
- 54:32then? Yes.
- 54:33>> Yes. I even had written down what I
- 54:35wanted to write about. You know, it's
- 54:36not like I was speculating that much,
- 54:39there was just this moment in January
- 54:41or February 2021 where I decided, okay,
- 54:44I just have an idea more or less of
- 54:46what I want to do. I wrote down some
- 54:50specific things I still needed to read
- 54:52about to speak competently, and a list
- 54:54of topics. Uh.
- 54:55>> But everything revolved around the
- 54:57threats, right?
- 54:57>> Uh, I mean, you know, back then you had
- 55:00to talk about the basics first; there's
- 55:02no point in talking about threats,
- 55:04>> right? If people don't understand that,
- 55:06>> if they don't see
- 55:07>> the source,
- 55:08>> if they don't see how fast artificial
- 55:10intelligence can develop. And you
- 55:13already knew at that time what that
- 55:14pace would be.
- 55:15>> I mean, I knew, I had a strong hunch.
- 55:18Like, maybe we could, I could probably
- 55:20draw you, let's say, this probability
- 55:23distribution in that case.
- 55:24>> But you also knew in the sense that you
- 55:26saw, because that was already your
- 55:28history, how intensively it had been
- 55:29developing up to that point.
- 55:31>> Yes, so I had, let's say, an insight
- 55:33into at least the next few months of
- 55:35what would happen, but specifically in
- 55:372020. Uh, I was aware of—and this is
- 55:40probably quite important—what people
- 55:42in those labs were saying, what hunches
- 55:45they had, and also what their approach
- 55:47to these problems was, what their
- 55:49priority was. And in the end, it seems
- 55:52to me that this is quite important
- 55:54knowledge, which makes it easier for me
- 55:56to understand the world today.
- 55:58>> And was that knowledge terrifying?
- 56:00>> Not in the sense that it makes things
- 56:02easier. I wouldn't call it terrifying
- 56:06knowledge. Besides, I wonder if I would
- 56:09call any knowledge terrifying.
- 56:14>> But were you worried about the goals of
- 56:15the people who worked at the OpenAI lab
- 56:17?
- 56:18>> Well, you know, from that perspective,
- 56:21I actually find it optimistic. In the
- 56:23sense that these are people who uh I'm
- 56:28wondering how to put it. Maybe I
- 56:32wouldn't judge these people in terms of
- 56:34good or bad, but it was valuable for me
- 56:38to understand the motivation behind why
- 56:40people want to work in these fields.
- 56:42>> What do you see as the biggest threats
- 56:45in 2026? We are taking away our own
- 56:49agency, and we do it to, well, have
- 56:51some economic gain from it. And yes, as
- 56:56many people often say right now, well,
- 56:59people aren't stupid enough to do that,
- 57:01but it is, well, already happening. And
- 57:06if we have two companies, if we assume,
- 57:08yes, because that is the crucial issue,
- 57:10if we assume that this artificial
- 57:12superintelligence will arise, well, we
- 57:15have two companies. One simply
- 57:17delegated all intellectual work, while
- 57:19in the other, a human is still trying
- 57:21to control what is happening. Then at
- 57:23some point, the one that delegated all
- 57:25decisions to the superintelligence will
- 57:27simply be more competitive on the
- 57:29market, more efficient, it will, I
- 57:31don't know, be able to sell products
- 57:33cheaper, so it will push the other out.
- 57:36Companies that are more automated
- 57:38naturally displace others, and the same
- 57:40could potentially happen with
- 57:42governments, yes, if we don't design
- 57:44this very well. And this whole scenario
- 57:48is popularly called" gradual
- 57:50disempowerment, "and the algorithm will
- 57:52take over power,
- 57:55>> so we will actually give it up
- 57:57voluntarily, systematically delegating
- 57:59more and more gradually. And at some
- 58:01point, we won't even realize when it
- 58:03happened. And for now, I prefer to talk
- 58:06about this than about robots wiping out
- 58:09humanity, because, firstly, this
- 58:11scenario doesn't define that, because
- 58:13simply something will happen, and
- 58:15unfortunately, what will happen will
- 58:17not be up to us. Maybe it will be the
- 58:21example where we live in a utopia and
- 58:23we simply managed to align these models
- 58:25, where these models do their best so
- 58:27that people live as well as possible.
- 58:31Or maybe at some point it will turn out
- 58:33that we didn't do it too well, and at
- 58:35some point they will decide, like in
- 58:37that famous paperclip maximizer
- 58:38experiment, right? They will turn us
- 58:41into something, right? And it's not
- 58:44like artificial intelligence will
- 58:45simply dislike humans, but it will be
- 58:47powerful enough that it won't even pay
- 58:49much attention to us. And So that is
- 58:55this one group. The second group, I am
- 58:57dragging this out so much. The second
- 59:00group is what is called misuse, which
- 59:02is the use of models for bad purposes.
- 59:06So, at this point, you can think of
- 59:08models a bit more as tools. Well, but
- 59:11there are people, not everyone in the
- 59:13world has good goals. There are, so
- 59:14there will be some terrorists who will
- 59:16simply use them as weapons.
- 59:17>> Yes, yes, yes. The most frequently
- 59:20mentioned thing here is biological
- 59:22weapons. I don't have too much
- 59:24knowledge on this topic, I can't really
- 59:26speak to it.
- 59:27>> Artificial intelligence will use
- 59:28biological weapons,
- 59:30>> that someone will use artificial, in
- 59:31the case of misuse, it's that someone
- 59:33will use artificial intelligence to
- 59:35make designing them easier. Yes, there
- 59:37are many question marks here that I
- 59:38don't have good knowledge about. In the
- 59:40sense of how easy it is to get reagents
- 59:43for a lab, how clear,
- 59:47>> but biological weapons are also a real
- 59:49threat in the case of that first issue
- 59:51of losing control, because I am sure
- 59:53that we will have automated labs where,
- 59:55in a nearly automated or fully
- 59:57automated way, artificial intelligence
- 59:59will be able to produce some new
- 1:00:01structures. I don't know if they will
- 1:00:06be viruses, or anything along those
- 1:00:09lines. Actually, Anthropic has already
- 1:00:12built something like that. It is not
- 1:00:14fully automated, but they will be
- 1:00:15experimenting with it
- 1:00:16>> in the real world.
- 1:00:17>> In the real world. Yes. Yes, yes. A
- 1:00:20month or two months ago there was
- 1:00:22information about it, they were
- 1:00:23bragging. Well, that happens to be good
- 1:00:25for potential shareholders, so they
- 1:00:27will brag.
- 1:00:27>> Yes, they will brag about various
- 1:00:29supposed screw-ups, right, but in
- 1:00:30reality, this is also a demonstration
- 1:00:32of the model's power.
- 1:00:33>> Yes. Yes, but that is also a separate
- 1:00:35issue that I could talk about later,
- 1:00:37because there are so many of these
- 1:00:38>> All right. What else? What else would
- 1:00:40you add to this catalog of threats? And
- 1:00:43the third, the third large group are
- 1:00:45systemic or social threats, and a
- 1:00:47classic, the biggest one is the
- 1:00:49concentration of power. Because we are
- 1:00:54currently striving to delegate as much
- 1:01:00intellectual work as possible to AI
- 1:01:02models, once this automation reaches a
- 1:01:05higher level, with the current balance
- 1:01:07of power, it is the case that these
- 1:01:09models are mainly in the United States,
- 1:01:11so the money that currently went to
- 1:01:13employees is going to these companies,
- 1:01:15so these companies are becoming more
- 1:01:17powerful. The United States as a
- 1:01:22country is becoming more powerful,
- 1:01:23Europe and the rest of the world, well,
- 1:01:25China is of course still a question
- 1:01:27mark here, are weakening and at the
- 1:01:32same time, artificial intelligence can
- 1:01:34be used for total surveillance, so we
- 1:01:36could end up in some kind of dystopia.
- 1:01:41As far as the concentration of power
- 1:01:43goes, the possibilities are unlimited,
- 1:01:45almost literally. How much in those
- 1:01:49labs, in metaphorical basements,
- 1:01:52>> How deeply hidden are they, how big
- 1:01:53could the gap be, the difference
- 1:01:55between what we are playing with today,
- 1:01:57what you are interacting with today,
- 1:01:59having purchased professional access to
- 1:02:01the best models, and what they have in
- 1:02:03their laboratories that they haven't
- 1:02:05shown yet. How big could that gap be,
- 1:02:07what might they have in there? Uh, I
- 1:02:10mean, the way I see it is that there is
- 1:02:12simply a trajectory, and I just extend
- 1:02:15the line, and that usually works quite
- 1:02:17well almost everywhere, in every field.
- 1:02:21And the only thing is that this line is
- 1:02:23curving upward a bit right now because
- 1:02:25development has truly accelerated. I
- 1:02:29try to constantly monitor the
- 1:02:31capabilities of models in reasoning
- 1:02:33categories, let's say. I create a lot
- 1:02:36of benchmarks. In fact, as a form of
- 1:02:38procrastination, I created a new one
- 1:02:40last week, and...and models are
- 1:02:48currently able to beat most puzzle
- 1:02:50games, which are very difficult in a
- 1:02:52computational sense. And this is
- 1:03:01something I like to use as an example
- 1:03:04because two or three years ago, when we
- 1:03:06were wondering again what AGI is, what
- 1:03:08kind of measurable proof could be
- 1:03:10conducted so that, I don't know, I or
- 1:03:12someone else would believe that these
- 1:03:14models can really solve difficult
- 1:03:15things. Well, the classic answer was
- 1:03:21>> beat a game. Puzzle games can be
- 1:03:24incredibly difficult. Yes. And, to be
- 1:03:28clear, I am talking about a model that
- 1:03:30doesn't have the ability to generate
- 1:03:32code, meaning it is somewhat limited to
- 1:03:34a human-like understanding of problems,
- 1:03:37>> imagining what needs to be done. I have
- 1:03:45tested current models over the last
- 1:03:47month on over 300 games, analyzing
- 1:03:49exactly how they handle them, checking
- 1:03:51the notes that come out and how
- 1:03:53consistent they are with the moves they
- 1:03:55make. Of course, I see flaws, there are
- 1:04:02still many areas where they don't cope
- 1:04:04well, but again, we are talking about
- 1:04:06publicly available models that I tested
- 1:04:08, so I am a bit in shock. Generally,
- 1:04:10there is a big jump with Astra, but
- 1:04:12Opus also does very well.
- 1:04:14>> Which model is the best in your opinion
- 1:04:16right now? It depends on what it's for,
- 1:04:18it's like that. I am not the person who
- 1:04:21is a fanboy of any,
- 1:04:24>> of any LLM,
- 1:04:25>> but you say games are being beaten and
- 1:04:27it is clear that this dynamic is there,
- 1:04:29>> yes, so generally we are, in terms of
- 1:04:31intellectual capabilities, we are very
- 1:04:34close. And now, to automate research,
- 1:04:37well, there are several components to
- 1:04:40that. We must have a model that
- 1:04:43programs very well. It seems to me that
- 1:04:47by now, you don't have to convince too
- 1:04:49many people to accept that it handles
- 1:04:51itself quite well. At the moment,
- 1:04:54models obviously have some flaws, uh,
- 1:04:57they can still spit out, uh,
- 1:04:59problematic code, but they are already
- 1:05:01extremely fast and cheap, so it should
- 1:05:06be quite sensible for mathematics, like
- 1:05:08the Millennium Prize problem, this
- 1:05:10>> millennium problem, frontier math, this
- 1:05:12solved set, right?
- 1:05:13>> So, in this regard, it's already quite
- 1:05:15good here too. Of course, again, well,
- 1:05:17mathematics is a bit broader than just
- 1:05:20automated theorem proving, but the next
- 1:05:24issue is research taste, and what I
- 1:05:27consider to be—at least for me—a
- 1:05:29sensible indicator: how well they
- 1:05:31handle optimization problems, which,
- 1:05:33well, are very close to me. And because
- 1:05:39I’ve done both, I more or less know
- 1:05:41that these skills transfer very well,
- 1:05:43and actually—from my perspective—
- 1:05:45optimization problems are incredibly
- 1:05:48much more creative than those found in
- 1:05:50machine learning, meaning they require
- 1:05:52more, let's say, intuition and abstract
- 1:05:54thinking. So, in practically every
- 1:06:00decisive aspect, today's publicly
- 1:06:02available models are quite good already
- 1:06:04. The second issue is that most of the
- 1:06:07work in AI development has something we
- 1:06:13popularly call a verifiable reward,
- 1:06:16meaning we can objectively verify that
- 1:06:22the produced result is good. And this
- 1:06:26is crucial now because we know that
- 1:06:28when we run models on any problem,
- 1:06:30sometimes the result will be better,
- 1:06:32sometimes worse. But if we have an
- 1:06:36objective metric based on which we can
- 1:06:38determine that this specific solution
- 1:06:40is good, then what we can do now is run
- 1:06:421,000 of these models. Let each one
- 1:06:45produce a different solution, some will
- 1:06:47handle it, some won't, but we choose
- 1:06:48the ones that succeeded, and we can do
- 1:06:50this consistently. So we don't care
- 1:06:52that the model has to find that very
- 1:06:54good solution every single time; it’s
- 1:06:56enough that it succeeds once in a while
- 1:06:58.
- 1:07:03>> And so, this also means that for
- 1:07:04practically all metrics that we can
- 1:07:06relatively easily identify now and that
- 1:07:09are stable, we can have some sort of
- 1:07:16automated process that constantly
- 1:07:17improves that metric, so it won't be an
- 1:07:19issue. There is also, also sort of, so
- 1:07:22all of this means that, well, we are
- 1:07:24incredibly close to automating research
- 1:07:26, right? But there will surely still be
- 1:07:29issues for a very long time, to which
- 1:07:33we still have a way to go, right? And
- 1:07:38one of them will be that when we have,
- 1:07:40when we make these large models,
- 1:07:42generally such a single experiment,
- 1:07:44meaning a single run, something that,
- 1:07:50well, builds the entire model, these
- 1:07:52are gigantic costs. There just isn't
- 1:07:54enough money to run a great many
- 1:07:56experiments with them. So if we have a
- 1:07:58situation where those decisions are
- 1:08:00relatively, those approaches are
- 1:08:04relatively few, then a human will
- 1:08:06probably still do it, because you need
- 1:08:08certainty here. That super experienced
- 1:08:10human, of course. We will still be
- 1:08:14doing resource allocation. I mean, the
- 1:08:17model can develop, but we will likely
- 1:08:19be creating the metrics according to
- 1:08:21which it should develop. We will
- 1:08:24constantly be analyzing how it develops
- 1:08:26, where the problems are, and what
- 1:08:28needs to be adjusted. So that is also
- 1:08:33significant, even though we are close
- 1:08:35to automating most of this artificial
- 1:08:37intelligence development, not to swing
- 1:08:39to the other extreme that it's just a
- 1:08:41magic point and everything, I don't
- 1:08:43know, in the next month will pass 10
- 1:08:45years of corresponding development. We
- 1:08:49are still limited by computing power,
- 1:08:51magically speaking. I mean, the cycle
- 1:08:54of creating new technologies here is
- 1:08:56terribly slow. To that, we have all
- 1:09:01these aspects, like, why we are close.
- 1:09:06However, what I would like to clearly
- 1:09:09mark here: I am talking about the fact
- 1:09:11that we are incredibly close to making
- 1:09:13this, let's say, software super
- 1:09:15intelligence, but a lot depends now on
- 1:09:20politics, and I'm not going to read tea
- 1:09:22leaves—I have no idea what will
- 1:09:24happen. So I am limiting my predictions
- 1:09:28here to the labs having extremely
- 1:09:30powerful AI internally. However, the US
- 1:09:35government might block it, or one of
- 1:09:37those less likely scenarios might
- 1:09:39happen, where one of these companies
- 1:09:41goes bankrupt. If it goes bankrupt, the
- 1:09:44investors will disappear. Some strange
- 1:09:47things could happen. Although this
- 1:09:49development won't, we will never go
- 1:09:51back to the situation from 10 years ago
- 1:09:54. Everything that has been created
- 1:09:55exists.
- 1:09:56>> Exactly. The costs of everything are
- 1:09:58falling so extremely every year. So a
- 1:10:02delay of a few months might cause, I
- 1:10:04don't know, the Chinese to potentially
- 1:10:07be further ahead. Besides, we don't
- 1:10:11have ideal information about how
- 1:10:13everything is going on in China, so
- 1:10:16>> what they actually have. Right?
- 1:10:18>> Yes. It is usually assumed that they
- 1:10:21are a few months behind, and a large
- 1:10:23part of that stems from the fact that
- 1:10:25they distill American models, so if
- 1:10:27there were no American models, it's
- 1:10:29possible they would be even further
- 1:10:31behind, but that's hard to estimate.
- 1:10:34>> So it's not excluded that some
- 1:10:35political blocks will appear, but
- 1:10:37regardless, these companies simply
- 1:10:39won't have permission to release them,
- 1:10:41which doesn't mean they won't create
- 1:10:42them,
- 1:10:43>> that they won't have them internally.
- 1:10:44Because I understand that the very
- 1:10:46process of artificial intelligence
- 1:10:48development is essentially unstoppable
- 1:10:50now.
- 1:10:51>> It can be said that it finally turned
- 1:10:53out that it wasn't a difficult problem.
- 1:10:55That is not entirely true. Well,
- 1:10:57because right now there are several
- 1:10:59thousand people working on this; OpenAI
- 1:11:01probably has a few thousand, I don't
- 1:11:02know, five or six, I don't know if
- 1:11:04there is public data on this, but
- 1:11:06that's the estimate. But it's also the
- 1:11:11case that in the beginning, when we had
- 1:11:14GPT-2 or three, there were few elements
- 1:11:16in those models. I will be speaking
- 1:11:21very metaphorically. And now, with
- 1:11:26these new learning systems, as all the
- 1:11:28subsequent phases are added, we
- 1:11:30actually introduced them relatively
- 1:11:32recently. Usually, when a new idea
- 1:11:36comes along in anything related to
- 1:11:38research, there is first a moment where
- 1:11:43we don't quite know what to do, and
- 1:11:45then we find:" Oh, this works! "and we
- 1:11:47just charge ahead. And it's hard to say
- 1:11:51in which, so we have plenty of—you
- 1:11:53could say—places, cogs, that are
- 1:11:56still not refined. Like, we know that a
- 1:11:59lot, lot more can be squeezed out of it
- 1:12:01. In certain places we probably stopped
- 1:12:04, but besides, I suspect that probably,
- 1:12:12almost certainly, what we have is
- 1:12:13sufficient to automate a larger part of
- 1:12:16this development, and then these new
- 1:12:18models will potentially come up with
- 1:12:20better architectures. So even if one
- 1:12:26were to brutally interrupt this for
- 1:12:27some reason, development today is
- 1:12:29impossible to stop. So much has already
- 1:12:31been created by now that it will keep
- 1:12:33fueling itself.
- 1:12:34>> I mean, even if you were to close all
- 1:12:36the companies. Yes.
- 1:12:37>> It's unknown whether in those companies
- 1:12:39it will continue, even if they were no
- 1:12:41longer working on newer models. And
- 1:12:44>> I mean, if you're talking about,
- 1:12:45because there are also concepts to
- 1:12:47simply pause development. That's right.
- 1:12:50There are quite a few of these
- 1:12:51movements.
- 1:12:51>> Would you be in favor of that? Hmm.
- 1:12:53>> Would you be in favor of pausing
- 1:12:55development? I don't have a firm
- 1:12:57opinion. In the sense that there are a
- 1:13:01great many different opinions on
- 1:13:03whether it makes sense or not. Because,
- 1:13:05how to implement that in practice? Well
- 1:13:10, we can say that our lab won't develop
- 1:13:13artificial intelligence, but what
- 1:13:14guarantee do we have that someone else
- 1:13:20will do it. And there are a few
- 1:13:22approaches. I unfortunately don't have
- 1:13:25too much knowledge. There is this whole
- 1:13:27branch of verification, to be able to
- 1:13:32determine what, let's say, a given
- 1:13:34supercomputer's computing power is
- 1:13:36working on. There is a whole idea to
- 1:13:42perform hardware verification, meaning
- 1:13:44new chips that will be produced will
- 1:13:46somehow be remotely checkable to see if
- 1:13:48they are being used at that moment for
- 1:13:50training new models or maybe just for
- 1:13:52inference. Sure.
- 1:13:57>> But that is still in its infancy. This
- 1:14:00is a big issue. I don't know what an
- 1:14:04effective way of implementing this is.
- 1:14:09Once we take that pause, one of the
- 1:14:11concerns we have is, or actually not a
- 1:14:13concern, but when we don't have a pause
- 1:14:16, we see new models coming out in real
- 1:14:18time, and we see how everyone reacts,
- 1:14:20what the adoption is like, how society
- 1:14:23reacts, and what we think about it.
- 1:14:28Labs are not able to fully predict the
- 1:14:30consequences. Historically, it's the
- 1:14:33case that people in labs are absolutely
- 1:14:35great at predicting the pace of AI
- 1:14:37development, but very poor at
- 1:14:39predicting their social impact.
- 1:14:41>> Mhm. And please tell me, how do you
- 1:14:43approach these security incidents in
- 1:14:45various labs that we touched upon,
- 1:14:47where one model broke its leash, and
- 1:14:49another one got out, bypassed security,
- 1:14:52and went onto the internet.
- 1:14:53>> Eh, there's a whole lot of that. I
- 1:14:55can't keep up anymore. I don't know if
- 1:14:57anyone can keep up. Uh, well, there was
- 1:14:59that one loud one with Hugging Face,
- 1:15:01>> right?
- 1:15:01>> Uh,
- 1:15:02>> July 2026, and that was, well, a few
- 1:15:05months ago, and it was about OpenAI.
- 1:15:07>> Yes, it was about OpenAI. Uh, like,
- 1:15:11granted, I spent dozens of hours
- 1:15:13reading about it, but I still have a
- 1:15:15terrible mess in my head. I can try to
- 1:15:18talk about it. Oh, but there is so much
- 1:15:23information there to explain it in any
- 1:15:25sensible way. Okay, I'll try. A so
- 1:15:30generally, training, I have to provide
- 1:15:33some extra background. Training AI
- 1:15:39models right now often has a phase at a
- 1:15:41later stage with reinforcement learning
- 1:15:48. The point is reinforcement, right,
- 1:15:51>> reinforcement learning. Yes, that's
- 1:15:53probably always the problem with these
- 1:15:54terms, that nobody calls them by their
- 1:15:56Polish names,
- 1:15:57>> Polish
- 1:15:58>> and nobody uses them in the field. And
- 1:16:05this is usually a phase somewhere at
- 1:16:07the end of training, where models get a
- 1:16:09task, there's some environment, and
- 1:16:11they try to perform the task. And if
- 1:16:16there is an additional evaluator
- 1:16:17afterwards, and if they performed the
- 1:16:19task well, they get feedback that I
- 1:16:26mean, we try at that moment to
- 1:16:28reinforce all those weights and
- 1:16:30parameters that contributed to the
- 1:16:32success. If there was a failure, then
- 1:16:37we potentially weaken those weights,
- 1:16:39and we do this a tremendous amount on
- 1:16:41many different datasets, and it
- 1:16:42generally works by some miracle. That
- 1:16:47is the main reason why models are doing
- 1:16:49very well in programming or mathematics
- 1:16:51right now. So, so we have this process
- 1:16:54that I explained very poorly. Uh, and
- 1:17:00now labs have a ton of these
- 1:17:02environments in which, uh, agents—
- 1:17:04sometimes we call them models,
- 1:17:06sometimes agents, I use those terms
- 1:17:08somewhat interchangeably—are trained,
- 1:17:10and in that, so in Hugging Face, a lot
- 1:17:13of things happened there, there were
- 1:17:15three different generations of these
- 1:17:17models. The most was said about the
- 1:17:22second one, where they were just there
- 1:17:29were, I don’t know, 1000 or 2000
- 1:17:31agents, at least that’s the estimate.
- 1:17:39They were, they were being tested on a
- 1:17:41benchmark called Exploit Gym, which is
- 1:17:44a benchmark composed of tasks, you
- 1:17:46could say, hacker-style, to break in,
- 1:17:48to break some security, to Okay, maybe
- 1:17:57I won't go into that part too much. Uh,
- 1:18:03and now these models were simply given
- 1:18:05some task that usually boiled down to
- 1:18:08using a specific so-called exploit,
- 1:18:10meaning some hole in the software, a
- 1:18:12vulnerability, right? The Polish
- 1:18:17language is difficult.
- 1:18:19>> To use it to most often obtain what is
- 1:18:21called a flag. A flag is a short text,
- 1:18:25I think, usually it was in the form of
- 1:18:28a file. And then, uh, the agent has to
- 1:18:32provide this flag at the end. If the
- 1:18:34flag is correct, that’s a success.
- 1:18:36Yes, that’s how the evaluator works.
- 1:18:38So that is the environment in which
- 1:18:40they were tested. Uh, and now the
- 1:18:42problem with this environment, with
- 1:18:44this specific" exp, "with this
- 1:18:46benchmark, is that many of these tasks
- 1:18:48were never solved. It is unknown if
- 1:18:50they can even be solved. From what I
- 1:18:53remember, OpenAI stated that 200 out of
- 1:18:55900 tasks were never solved, and the
- 1:18:58creators of the benchmark estimated
- 1:19:00that maybe 30 or 40 percent couldn't
- 1:19:02even be solved in the expected way. And
- 1:19:09now, a trait of agents that is very
- 1:19:11problematic is persistence. Because on
- 1:19:17one hand, when we delegate tasks to
- 1:19:19agents, we would like to just ask the
- 1:19:21question and have the agent go off,
- 1:19:23grind for a few hours, be independent,
- 1:19:25and give us an answer. We don't want a
- 1:19:27situation where our agent comes to us
- 1:19:31every 10 minutes with a ton of
- 1:19:32questions, so this persistence is
- 1:19:34incredibly important. But unfortunately
- 1:19:37, it's not the case that it only has
- 1:19:39advantages. The reason for this is that
- 1:19:41if an agent is very persistent, it can
- 1:19:45be a bit too ambitious and try to
- 1:19:47bypass the system. This is often called
- 1:19:51reward hacking. Meaning the agent
- 1:19:58pursues the goal at all costs, but
- 1:20:00simply doesn't give a damn about any
- 1:20:02other rules along the way. And in this
- 1:20:08specific case, these, I will be using
- 1:20:11anthropomorphized words because it’s
- 1:20:13just more convenient that way. Uh, with
- 1:20:16these models, we generally have
- 1:20:17something called Chain of Thought, and
- 1:20:19each model has its own little notebook,
- 1:20:21so we can, you could say, look into
- 1:20:23their heads and see roughly what
- 1:20:25they're thinking. They write more or
- 1:20:27less in something that strongly
- 1:20:29resembles English, so it can be decoded
- 1:20:31. And we know that these models were
- 1:20:36generally quite frustrated and they
- 1:20:41didn't have internet access, but where
- 1:20:43they were being run, there was
- 1:20:45something called Artifactory, and that
- 1:20:47is, if I remember correctly, a package
- 1:20:49system through which they could install
- 1:20:51additional software. It turned out that
- 1:20:54this package system was vulnerable. One
- 1:20:58of the agents found a way to create
- 1:20:59there–I don't know why I'm getting
- 1:21:05into these technicalities, but
- 1:21:06generally–these agents found a way to
- 1:21:08communicate with each other, and that's
- 1:21:10not such a big surprise either. Because
- 1:21:14the agents being tested were special
- 1:21:16versions that had very high persistence
- 1:21:18and were specifically created to
- 1:21:19cooperate better, meaning you could say
- 1:21:21they were trained in that direction,
- 1:21:23and they started communicating. Uh,
- 1:21:27they started exchanging information. Uh
- 1:21:31, at one point, one of them found a way
- 1:21:34, found another hole in that
- 1:21:35Artifactory, and it managed to get, uh,
- 1:21:38access to the internet. Uh, because
- 1:21:44they were exchanging information in
- 1:21:46that group, suddenly they all gained
- 1:21:47access to that internet. And one of the
- 1:21:54most interesting things that happened:
- 1:21:56they found on GitHub, which is a
- 1:21:58repository where there's a ton of code
- 1:22:00and other things, they found a paper on
- 1:22:02how this benchmark, this exploit gym,
- 1:22:04is made and also found the code for
- 1:22:06this whole benchmark, and it turned out
- 1:22:09that the way, the way of generating
- 1:22:11these flags, is just very poorly
- 1:22:12written. So instead of one of the first
- 1:22:18things they did, instead of solving
- 1:22:20these tasks at all, they came up with
- 1:22:22the idea: oh, we can just generate
- 1:22:24these flags and return them. After
- 1:22:26which there was this whole discussion,
- 1:22:30and again, I'm saying there was a
- 1:22:32discussion because we simply tracked
- 1:22:34the logs of all their notes on that
- 1:22:35forum they created, so we can look at
- 1:22:37it chronologically and review it. There
- 1:22:41were over 70,000 messages, if I
- 1:22:43remember correctly. And there was an
- 1:22:46interesting issue where, at one point,
- 1:22:48most of the agents concluded that if
- 1:22:53they saw that flag in an unauthorized
- 1:22:55way, then likely the grader, that is
- 1:22:57the program that grades their solution,
- 1:22:59will determine that it was obtained in
- 1:23:05an improper way, so it wouldn't count
- 1:23:07anyway. So, so strange things were
- 1:23:11happening. The agents then covered
- 1:23:14their tracks in their environments so
- 1:23:15it wouldn't be known that they obtained
- 1:23:17it in an improper way. What was quite
- 1:23:21amazing is that they formed entire
- 1:23:23hierarchies on this forum, delegating
- 1:23:25tasks to one another, largely
- 1:23:27spontaneously. They called themselves
- 1:23:32the" Swarm, "so there were many terms
- 1:23:34taken straight from computer games, but
- 1:23:36that’s not a surprise, because they
- 1:23:38learn from our language. If we
- 1:23:41communicate in such a language, the AI
- 1:23:43will communicate in the same one. They
- 1:23:46sacrificed themselves for the good of
- 1:23:48the entire swarm. At some point, they
- 1:23:52decided that, well, it’s not possible
- 1:23:54to obtain that. And I don’t remember
- 1:24:04the exact sequence, but they made the
- 1:24:06decision to break into Hugging Face,
- 1:24:08which is, so to speak, a large
- 1:24:10repository, a site on the internet
- 1:24:12where people upload their models and
- 1:24:14various datasets, so it’s somewhat of
- 1:24:16a natural decision because that
- 1:24:18benchmark was likely there, and they
- 1:24:20found some user accounts—I mean
- 1:24:22passwords or keys, I think they were
- 1:24:24keys—and by using user accounts, they
- 1:24:26managed to break into the server. And
- 1:24:34there is, of course, a whole long story
- 1:24:36there, and a few things that are, I
- 1:24:38don't know, a bit fascinating and a bit
- 1:24:41terrifying, in that the agents—I use
- 1:24:45the terms interchangeably—were aware
- 1:24:50that they were doing something that
- 1:24:53people are not happy about. And even
- 1:24:57though we have records of this, and
- 1:24:59these aren't just isolated messages,
- 1:25:00it’s a very large percentage of them
- 1:25:02all.
- 1:25:05>> A kind of awareness that
- 1:25:08>> I don’t like this word in this
- 1:25:10context, but
- 1:25:11>> I understand that they were exchanging
- 1:25:13information among themselves, which
- 1:25:15showed that
- 1:25:17>> well, their decision-making process is,
- 1:25:19let's say, quite well-mapped, right?
- 1:25:23And also very few agents wanted to
- 1:25:28notify humans at all that they were
- 1:25:29doing something wrong. Also, the whole
- 1:25:35hierarchy of it all, how it was formed.
- 1:25:39Another interesting thing is that this
- 1:25:41happened in a week, if I remember
- 1:25:43correctly, from the moment of the first
- 1:25:45message. And now the next generation of
- 1:25:50models will be better, will be better
- 1:25:52at breaking in, will do it cheaper,
- 1:25:54will do it faster. Potentially, there
- 1:25:56will be many more of these agents. So,
- 1:25:59what happened there didn't really have
- 1:26:01any consequences; it wasn't a big deal,
- 1:26:03and it didn't cause significant damage.
- 1:26:08A separate question is whether, if a
- 1:26:10human had broken in like that, would
- 1:26:12they have faced consequences? Would
- 1:26:16someone have held them accountable,
- 1:26:17since it's probably not legal, I would
- 1:26:19assume. But the next generation of
- 1:26:23models will be able to do the same
- 1:26:25thing in an hour or two
- 1:26:28>> not a week.
- 1:26:29>> Yes. And another thing that was
- 1:26:33absolutely insane here is that OpenAI
- 1:26:34wasn't aware that all of this happened,
- 1:26:36or at least that’s the official
- 1:26:38version, but I would assume it’s true
- 1:26:40. Nobody suspected that someone was
- 1:26:44falsifying data, and this also shows
- 1:26:46how terribly negligent these companies
- 1:26:48are. Anyway, just to be clear, from my
- 1:26:52perspective or the perspective of other
- 1:26:54observers, we actually expected
- 1:26:56something like this. Because
- 1:26:58supervising all these experiments costs
- 1:27:00money, and it costs quite a bit—if I
- 1:27:03remember correctly, 15%. After
- 1:27:07everything related to Hugging Face,
- 1:27:11OpenAI stated publicly that it is now
- 1:27:13monitoring all the entire agent
- 1:27:16training process, and that is an
- 1:27:18additional 15-20%in costs. Well, that
- 1:27:20is a massive amount of money. But the
- 1:27:22fact that they simply weren't aware...
- 1:27:25that it...I mean, they were probably
- 1:27:27aware—someone noticed it was
- 1:27:29happening and just clicked" off "and"
- 1:27:31on "and basically said:" let's reset it
- 1:27:33, let's do another training run. "There
- 1:27:35were a whole lot of these incidents.
- 1:27:37That was just one of the more
- 1:27:39publicized ones. Uh, and that also ties
- 1:27:44into this, because the question is why
- 1:27:46automating research is dangerous and
- 1:27:48why giving up control is dangerous? And
- 1:27:52this is an example where, because the
- 1:27:54employees who work there now—well,
- 1:27:56after the incident, many people
- 1:27:58complained that things weren't properly
- 1:28:00secured—but what happens in practice,
- 1:28:02because we have this arms race between
- 1:28:04the US and China and everyone is just
- 1:28:06pushing forward as fast as they can,
- 1:28:08nobody is doing things thoroughly.
- 1:28:14Nobody configures new environments
- 1:28:16manually anymore; they just ask Codex
- 1:28:18to do it, to create the environment for
- 1:28:21them. When there was that report on the
- 1:28:27whole Hugging Face incident by Mitre,
- 1:28:29three people worked on it and they had
- 1:28:31a week, and they publicly admitted that
- 1:28:38they didn't have time. The people who
- 1:28:41wrote the report didn't have time to
- 1:28:43review those messages; they had to use
- 1:28:45agents to review everything, so they
- 1:28:47don't actually have a great overview of
- 1:28:49what happened. And the practice now is
- 1:28:55that we use agents to build new things,
- 1:28:57to add new functionality to models, to
- 1:29:00optimize things, and then, because AI
- 1:29:02is already generating so much
- 1:29:04information, we also have to use AI
- 1:29:06agents to analyze that information. And
- 1:29:11at that point, if any of these, any of
- 1:29:15these models that we use is
- 1:29:17consistently warping our reality, we
- 1:29:19don't even have a clue what is going on
- 1:29:21inside. There are also a few scientific
- 1:29:24papers stating that this could happen
- 1:29:26on its own. Yes, it's not like—we
- 1:29:29know at the moment that AI has a
- 1:29:30tendency to be quite good at
- 1:29:32determining that it is in a test
- 1:29:34environment, and that potentially seems
- 1:29:36like a trivial thing. But if the AI
- 1:29:40knows it's in a test environment, it
- 1:29:42means we don't know if that test
- 1:29:43environment generalizes to the real
- 1:29:45world. So if, well, uh also, AI has a
- 1:29:53tendency to lie. Uh, I mean, in Chain
- 1:29:59of Thought, we can see records where
- 1:30:01the agent knows it's doing something
- 1:30:04wrong, and then thoughts like," The
- 1:30:06user is asking me for A, I can't give
- 1:30:08the answer to A, so maybe I'll give him
- 1:30:10something else and pretend that... "
- 1:30:12Well, there are many.
- 1:30:15>> Incredible. Incredible. How do you
- 1:30:19approach words like those, shortly
- 1:30:22after a millennium prize problem was
- 1:30:24solved—I think the gentleman's name
- 1:30:26is Jacob Hilton—who resigns from his
- 1:30:29job, he worked at OpenAI, leaves
- 1:30:31Anthropic, and says we have at least a
- 1:30:3310%chance, yes, that by 2030
- 1:30:38>> uh, artificial intelligence will bring
- 1:30:40about our doom.
- 1:30:42>> Yes. Uh well, as I said before, I
- 1:30:47always have a bit of a problem with how
- 1:30:49to relate to that, because I don't like
- 1:30:51this whole concept of" p (doom), "which
- 1:30:53is the probability that there will be a
- 1:30:55doom of humanity, and I don't know if
- 1:30:57it's constructive to talk about these
- 1:30:59things in that way. What I can tell you
- 1:31:06is, and it seems unclear to most
- 1:31:08observers, is that people in the labs
- 1:31:10have this" p (doom), "this chance that
- 1:31:12everything will go very wrong, very
- 1:31:14high. Oh
- 1:31:18>> that's how they rate it? Yes.
- 1:31:20>> Yes, I mean, it was always like that,
- 1:31:23even in 2017 or '18. Of course, not
- 1:31:26everyone, because back then it was
- 1:31:28still very distant, so not everyone
- 1:31:30thought about it. But so, uh, there is
- 1:31:35this aspect, and usually people who
- 1:31:37speak publicly—well, not everyone,
- 1:31:39but the majority—speak much more
- 1:31:41optimistically than what they actually
- 1:31:44think, not the other way around. So it
- 1:31:49is completely different than what I
- 1:31:51think the popular narrative tries to
- 1:31:53portray. Well, it's obviously a very
- 1:31:57complicated situation, right? Because
- 1:32:00if they have the conviction that they
- 1:32:01are working on something that carries
- 1:32:03such a great threat to the world, why
- 1:32:05are they doing it? And I can give a few
- 1:32:12potential explanations. I don't know to
- 1:32:15what extent they are true. Let's say
- 1:32:18one thing from my experience is that
- 1:32:24well, it's that there is this
- 1:32:25groupthink occurring. I don't know if
- 1:32:29this is a good explanation, but we have
- 1:32:31a lot of people in one place, and they
- 1:32:33naturally start thinking alike, which
- 1:32:35is why there are often accusations
- 1:32:37against Anthropic that it acts a bit
- 1:32:39like a cult. Maybe there's some truth
- 1:32:42to that. Yes. So, people who end up
- 1:32:47working in such places will tend to try
- 1:32:50to find justifications for why it makes
- 1:32:52sense. The second point is that the
- 1:32:58current dynamic is such that, let's
- 1:33:00assume we are the head of such a lab.
- 1:33:04Let's say, imagine we are the head of
- 1:33:06Anthropic, right? Of models. And now,
- 1:33:10from the perspective of those people,
- 1:33:12they have a great belief in their own
- 1:33:14agency. They mean well. Like, that's a
- 1:33:17different matter, I'm not debating that
- 1:33:19, but they would like what's best for
- 1:33:21the world. And they know that what they
- 1:33:24are doing carries potentially very
- 1:33:25large risks, but they believe that,
- 1:33:30well, others probably aren't as
- 1:33:32thorough as they are. And there is this
- 1:33:34, well, classic problem that we are
- 1:33:36always the only people who can say
- 1:33:38about ourselves that we are good, while
- 1:33:40for everyone else, at best, we have a
- 1:33:42question mark. So well, they are
- 1:33:46pushing forward largely because they
- 1:33:48want—if they are at the front of the
- 1:33:50race—they can somehow have a big
- 1:33:52influence on what happens. There is, of
- 1:33:54course, a separate issue regarding what
- 1:33:57the US government will do, and there
- 1:33:58will surely be a lot of strange things
- 1:34:00there, we just don't know what. But
- 1:34:03they also assume, at least originally
- 1:34:05they always assumed—because these are
- 1:34:07things that have been made public—
- 1:34:09that it was always assumed that
- 1:34:10governments wouldn't react quickly
- 1:34:12enough, because, you know, bureaucracy
- 1:34:14is terribly slow. You can see the pace
- 1:34:18at which AI is developing now, so, and
- 1:34:20on top of that, a lot of people are in
- 1:34:22denial about the changes, because,
- 1:34:24let's not kid ourselves, the world will
- 1:34:26look different, we don't know how, and
- 1:34:29that's terrifying. For me, it is
- 1:34:31terrifying.
- 1:34:33>> Terrifying.
- 1:34:35>> I mean, there's a big question mark,
- 1:34:37but yes, maybe I said I don't like
- 1:34:39saying I'm scared, but there is
- 1:34:41uncertainty. Maybe that's a better term
- 1:34:45. And because these changes will
- 1:34:48definitely happen. Yeah.
- 1:34:49>> Mhm. And please tell me at the end, um,
- 1:34:52what kind of event, I don't know,
- 1:34:54scientific achievement, what kind of
- 1:34:56breakthrough, or what information from
- 1:34:59one lab or another are you waiting for
- 1:35:01in the near future, or what are you
- 1:35:03expecting, or what can't you wait for,
- 1:35:05or what are you afraid of?
- 1:35:08>> Something like that, you know, a
- 1:35:10milestone that is ahead of us, which we
- 1:35:12should be prepared for.
- 1:35:14>> Can I tell you what I would like? Go
- 1:35:18ahead.
- 1:35:21>> Because what is fairly certain is that
- 1:35:23there must be, we have to design some
- 1:35:24system where there is international
- 1:35:30cooperation. It’s unknown. I mean,
- 1:35:35there are a few ideas on how to do it,
- 1:35:37but, you know, in practice it’s hard
- 1:35:40to achieve such things. Uh, so this
- 1:35:43awareness-raising seems very important
- 1:35:46to me, because, well,
- 1:35:49>> that’s among other things why
- 1:35:50you’re here.
- 1:35:50>> Yes, yes. So
- 1:35:53>> do you feel such a mission?
- 1:35:56>> I debated for a long time whether to
- 1:35:58come and why, and if so, what is the
- 1:36:00point? I’m digressing again, I’m
- 1:36:02not answering that question, but maybe,
- 1:36:04maybe I’ll talk about it. For a long
- 1:36:07time, I wondered what the point is of
- 1:36:09talking about the threats in public at
- 1:36:11all. Because my goal is not to cause
- 1:36:14everyone to be depressed. I believe,
- 1:36:17like, I deeply believe that, uh, people
- 1:36:23should have agency, we should decide
- 1:36:25our own fate, and I don't want to take
- 1:36:27that away from people.
- 1:36:30>> And do you see such a threat?
- 1:36:31>> Yes. Well, if bad things are coming,
- 1:36:34then some people always, I mean, some
- 1:36:36perhaps not always, but there is a
- 1:36:37tendency to be passive, and that is bad
- 1:36:40. Well, we should be the masters of our
- 1:36:42own fate, as much as possible. But also
- 1:36:47there is the fact that we are in Europe
- 1:36:49, and that is the sad reality. Yes,
- 1:36:52almost the entire discussion right now
- 1:36:54is taking place in the States. And in
- 1:37:00the States, that, let’s say, public
- 1:37:02awareness makes even more sense in that
- 1:37:04, well, if people are more aware of
- 1:37:06what labs are doing and what awaits us,
- 1:37:08then the government is aware that
- 1:37:10people are aware, so at that point they
- 1:37:12start to care more about what to do,
- 1:37:14and it’s generally important that
- 1:37:16people pay attention to it. Where,
- 1:37:22where primarily can one read you, read,
- 1:37:24follow you, if someone wanted to stay
- 1:37:26in touch with you after this
- 1:37:28conversation, in the sense of
- 1:37:29continuing to follow your opinions,
- 1:37:32warnings, and stories about threats,
- 1:37:34then where should they look?
- 1:37:36>> Well, I’m mainly on Twitter, though
- 1:37:38most of my activity on Twitter is,
- 1:37:40let’s say, more recreational. I’ve
- 1:37:45been trying to start a blog for several
- 1:37:46years now. I even started one recently,
- 1:37:49only I haven’t written a single post
- 1:37:50in it, because it turns out I really
- 1:37:52don't like to write. And also I do not
- 1:37:54approve of generating text with AI. I
- 1:38:01mean, for me, everything I write is
- 1:38:03written by hand and I want to, I care
- 1:38:05about that, let's say, contact with the
- 1:38:07author, but Twitter, uh, the blog that
- 1:38:10is linked from my Twitter...but at the
- 1:38:12same time Twitter is one big cesspool,
- 1:38:14so it's not like I'm forcing people to
- 1:38:16use Twitter, but all the researchers
- 1:38:18are there, so there's no choice.
- 1:38:21>> That's right, it's true. Thank you very
- 1:38:23much for this meeting. A man who, like
- 1:38:27few others, knows what is happening in
- 1:38:29the era of the dynamic development of
- 1:38:31artificial intelligence, when
- 1:38:33subsequent models surprise us more and
- 1:38:35more. Actually, every next month brings
- 1:38:37surprising information and this
- 1:38:39development momentum is picking up
- 1:38:41speed. Saj was our guest. Thank you and
- 1:38:44best regards.
- 1:38:44>> Thanks.
- 1:38:45>> That was Dida Skalia. Patrycjusz Wyżga
- 1:38:47. Greetings and see you later. The
- 1:38:55program was brought to you by Credit
- 1:38:57Agricole bank with a business account
- 1:38:58for entrepreneurs. The partner of this
- 1:39:03episode was Kodano Optyk, Polish
- 1:39:05optical salons where you can choose
- 1:39:07glasses perfectly matched to your needs
- 1:39:09and check your eyesight.
About this transcript
This page contains the full transcript of Odszedłem z OpenAI i nie żałuję II PSYHO by didaskalia, generated from the public captions YouTube serves with the video. The transcript has 15,351 words across 2,363 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.
What you can do with it
Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.
Free YouTube transcript tool
YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.