Latest / Agentic Conversations (formally mlops.community) / Greg Kamradt: Benchmarking Intelligence | ARC Prize

Transcript

Greg Kamradt: Benchmarking Intelligence | ARC Prize · Agentic Conversations (formally mlops.community)

• 0:10
  That's how we know it's legit now.
• 0:12
  Well. I mean, tell you what, I was
• 0:15
  looking forward. To this dude, I can't tell you
• 0:17
  how stoked I am. Tell me exactly how it went down
• 0:20
  that you were on a live stream with Sam Baldman.
• 0:23
  Yeah. So I run Arc Prize right now,
• 0:26
  right? And we run an AI benchmark
• 0:28
  called ARC AGI. Which is.
• 0:31
  We want AGI progress pulled forward like we, we want tech
• 0:34
  progress pulled forward because we believe it's going to be one
• 0:36
  of the best technologies that humanity's ever had, right.
• 0:40
  There's a big question on, well, how the heck do you make
• 0:41
  progress go faster? And so our route that we've
• 0:44
  chosen is from a benchmark and it's by created by Francois
• 0:48
  Chalet in 2019. It takes a very interesting
• 0:52
  approach. There's a lot of benchmarks out
• 0:53
  there. They go PhD plus plus problems.
• 0:54
  And so they'll ask you like the hardest questions and then even
• 0:57
  harder and they'll say this is the last Test we're ever going
• 0:59
  to have to take because we can't come with any harder questions.
• 1:03
  AI ends up solving those. Like it ends up doing it well.
• 1:05
  Like the ceiling on AI is like really high.
• 1:08
  Like it's insane. It's already doing some
• 1:10
  superhuman stuff. So we take a different approach
• 1:13
  on that. We want to know what types of
• 1:16
  problems are easy for humans but hard for AII.
• 1:19
  Love that. And the reason why just getting
• 1:22
  into it like they'll stick behind it is because we have 1
• 1:24
  proof point of general intelligence right now and
• 1:27
  that's the freaking human brain. So these are things like
• 1:29
  strawberry. I would say that that is a class
• 1:32
  of problems where if you can find things like that, it's
• 1:34
  like, Dang, AI can't do that but humans still can.
• 1:37
  We probably don't have AGI if we can convert those problems right
• 1:40
  now. The hard part, the hard part is
• 1:43
  those are one off questions. And so it's, it's easy to find
• 1:45
  one off questions. But if you want to find a domain
• 1:48
  where you can come up with like 200 questions in the same class
• 1:51
  that you can actually quantify this for, then that becomes a
• 1:55
  lot more difficult. And so our theory about AGI, and
• 1:59
  this is more of a working, this is an observational definition
• 2:01
  rather than like a inherent 1, is when we can no longer come up
• 2:05
  with problems that humans can do but AI can't, then we have AGI.
• 2:11
  Wow. OK.
• 2:12
  The the inverse of that though is if we can come up with
• 2:15
  problems that humans can do and AI can't do, then we don't have.
• 2:18
  AGI there yet? And by virtue of Arc AGI one,
• 2:22
  our first, our first version of our benchmark being out there,
• 2:25
  the fact that it's even out there and unsolved, that's a
• 2:28
  class of problems that humans can do.
• 2:29
  We just came out with ARC AGI 2 and we actually went and
• 2:32
  gathered, we gathered up 400 different people and tested them
• 2:35
  on ARC AGI 2 on every single task within there.
• 2:38
  And we made sure, because if we're going to claim that humans
• 2:40
  can do this. Humans better be able to do it.
• 2:42
  Better be able to do it. So we went we got 400 different
• 2:44
  people down in San Diego and we tested on tested them on all
• 2:47
  this and every task that was in there was solved by at least two
• 2:49
  people in under 2 attempts. So humans can do it.
• 2:52
  We have first party data for that, but AI still can't do it,
• 2:56
  so we claim that we don't have AGF.
• 2:58
  But they're kind of hard tasks. They get harder for sure.
• 3:01
  Yeah, well, that's what's crazy is the way that I think about it
• 3:03
  is there's a gap in between what humans can do and what AI can't.
• 3:06
  That gap is narrowing. And so we need to make sure that
• 3:09
  humans can still do it within a reasonable attempt.
• 3:12
  We're not looking at pH DS look, we're not looking at 2 year olds
• 3:14
  to see if they can do this a a competent person give them these
• 3:18
  tasks and see if they can actually do it.
• 3:20
  So just. If you pluck somebody off the
• 3:21
  streets, they have college education type thing.
• 3:24
  More or less. So like when we did our, when we
• 3:26
  did our filtering, we made sure that they could use the Internet
• 3:28
  like like things like that, like we didn't.
• 3:30
  Want my mom is out we. Didn't want to teach them how to
• 3:33
  use a computer when we taught them what art was, You know what
• 3:35
  I mean? And so that that doesn't allow
• 3:39
  us to make the claim about the average human.
• 3:41
  So we're careful about not saying that that's not what
• 3:43
  we're going for. Yeah.
• 3:45
  We're going for a capable human. Yeah.
• 3:47
  Some people like to argue on it, but that's a different
• 3:49
  conversation on it. So run this benchmark RKGI one.
• 3:53
  OK, great. We get an e-mail from one of
• 3:58
  Open the Eyes board members who we have a relationship with in
• 4:02
  early December and it more or less said, hey, we have a new
• 4:04
  model if you want to test it on ARC.
• 4:07
  And back in that day, it was strawberry.
• 4:09
  It was the. Strawberry model.
• 4:10
  There's so many names going around.
• 4:11
  I mean, there's Orion even at that point.
• 4:13
  There was Strawberry. There was what did Ilia see?
• 4:15
  You know, there's so much stuff going around.
• 4:16
  It's like who? Who the heck knows?
• 4:17
  Like what? What refers what rumor refers to
• 4:19
  what production? Version and it hasn't gotten
• 4:22
  better. To be honest, the official names
• 4:24
  are probably worse than the rumors.
• 4:25
  And I think that that tells you don't expect it to get better
• 4:28
  because it it won't get better. So again, the e-mail says we got
• 4:32
  a new model, we want to test it. OK, cool.
• 4:34
  Yeah, sounds great. It's Open AI, they have a new
• 4:36
  model and they claim to have a very good score, but they didn't
• 4:38
  say what the score was in the e-mail and so.
• 4:41
  On the arc. Press on Arc Press because we
• 4:43
  have public data. And so the, the, the way we run
• 4:44
  our benchmark is there's a bunch of public data that you can go
• 4:46
  train on and you can go kind of test yourself on it.
• 4:49
  But then we have a hidden holdout set.
• 4:51
  Nice that we can get into why that's important in the 1st
• 4:55
  place. It's the only way to do it.
• 4:56
  It's the only way to do it. You have a hidden holdout set
• 4:58
  for it. And they said we want to see,
• 4:59
  are we overfit to this because we think we're doing pretty
• 5:01
  good, but we want to try it on your holdout set.
• 5:03
  Will you come and test it for us?
• 5:05
  So we spent the next two weeks testing it, basically working
• 5:08
  with their team to go do it. This is through nerbs 2024 of
• 5:11
  last year too. So I'm like at nerbs like in
• 5:13
  Vancouver, thinking I'm going to relax and just watch talks all
• 5:15
  the time. I'm like literally testing and
• 5:16
  hitting open as API endpoints for that.
• 5:20
  But we get through and it's like, holy shit, this is good.
• 5:25
  I mean freaking soda. It.
• 5:26
  Was multiples better than we had ever seen another model do
• 5:29
  beforehand? And keep in mind that this thing
• 5:32
  had been this thing had been out there for five years so far and
• 5:34
  there hadn't been this type of progress on it.
• 5:35
  So we're like, holy cow. And so I get on a meeting with
• 5:40
  Jerry and Nat to kind of just like kind of like pre brief
• 5:45
  before we go and do the testing and I say, what score do you do
• 5:49
  you claim on RKGI? Because we're going to go and
• 5:51
  verify that because if they claim what it's different and
• 5:53
  that'd be a big story. They claimed 87% and keep in
• 5:57
  mind like the high, the highest on here with a publicly
• 5:59
  available models like in the 20s and then custom built solutions
• 6:03
  purpose built just to try to be arc, we're scoring in the 40s
• 6:07
  and 50s at that time and and they're claiming 87%.
• 6:10
  And so it's like, all right, this is a really big deal for
• 6:11
  us. Anyway, Long story short, we go
• 6:13
  through it and what's interesting is in inference time
• 6:17
  compute world, you can no longer just say, here's our model,
• 6:20
  here's our score. It's here's the model, here's
• 6:22
  how much inference time compute we spent, here's our score.
• 6:25
  So now there's another, there's another variable on it.
• 6:28
  And what we confirmed for them is that on low compute, and we
• 6:31
  can get into what low compute actually means in a second here,
• 6:33
  they scored 75% and on high compute we saw Yep, more or less
• 6:37
  we validated their 87% score, which was like, okay, Yep, it's
• 6:41
  validated. So we write up our blog post and
• 6:44
  just like a one page or Google doc and somehow, not somehow,
• 6:49
  but just like along the lines, Sam ended up getting pulled up
• 6:51
  on the thread, like on the e-mail thread that we had going
• 6:53
  back and forth and we said we have our results here they are.
• 6:56
  We want to discuss them live. He goes, great, I'm free Tuesday
• 6:58
  at 5:30 or whatever. And it's like, let's go, yeah,
• 7:01
  let's go. Like it's, it's wonderful.
• 7:03
  I mean, it's like, it's a huge opportunity.
• 7:04
  It's like we put together A blog post and we get inside the room
• 7:07
  and we show them basically put up on the screen.
• 7:09
  We show a blog post, everyone reads through it and discussion,
• 7:12
  discussion. And he goes, OK, great.
• 7:15
  You guys should join our live stream on Friday.
• 7:22
  And so we're sitting there not even like we hadn't even
• 7:23
  considered they were going to be testing this new model.
• 7:25
  Like, you know, if you and then he says, you know, on Tuesday,
• 7:28
  you guys should come join us on Friday.
• 7:30
  And our one requirement was that we didn't want to just go up
• 7:32
  there and have them tell us what to say.
• 7:34
  Like we didn't want them to write the script for us.
• 7:36
  And so Mike and I, Mike Knoop, who Co founded Arc prize, we
• 7:41
  basically wrote the script that we were happy, we were
• 7:43
  comfortable with. We gave it to him.
• 7:45
  They said yeah, looks great. Yeah.
• 7:47
  All right, cool. That's that and see Friday.
• 7:50
  Exactly. Well, so they they had a very
• 7:53
  big production No, I wouldn't say big production crew, but you
• 7:55
  know, call it, you know, 12 people between marketing comms
• 8:00
  like videographers and events and sound and everything that
• 8:03
  are that are in there. And so we had two rehearsals
• 8:06
  that went through for that. Two rehearsals went through,
• 8:09
  went great, made edits. And then, you know, the live
• 8:12
  stream comes out on Friday or it went out on Friday.
• 8:15
  It was funny because there was a room that wasn't much bigger
• 8:17
  than this, but it was, it was a much bigger room, but there was
• 8:20
  partitioned off on all the sides, you know, with like kind
• 8:22
  of just like, you know. Stage or whatever set design
• 8:27
  meant to look all it was like it was just a table You guys are
• 8:30
  sitting around Yeah just. A table, yeah, they did a really
• 8:32
  good job with that. But so I'm on the other side of
• 8:34
  the partition. I'm hearing Sam and Mark talk
• 8:36
  about it. Mark Tender at the time, their
• 8:38
  SVP of Research and they're like, now we'd like to invite
• 8:41
  Greg and then just walk from behind the wizard, like behind
• 8:43
  the curtain and just go jump on there.
• 8:45
  And what's wild is that you knew how many people were watching on
• 8:47
  the live stream, but it was it was a small room.
• 8:49
  Like it was like, like I said, only like 10 people in there,
• 8:51
  but it was cool. So did that.
• 8:53
  And but this was an unreleased model.
• 8:55
  So now we call it O3 preview because there's a preview model
• 9:00
  for it. And it was more of a
• 9:02
  capabilities demonstration about what if you push this to the
• 9:05
  Max, what could you actually do for it?
• 9:06
  And to put this into perspective on the low compute, it was they
• 9:10
  were spending about $20 per task on this and we verified with 500
• 9:13
  tasks. And so that's about 10,000 bucks
• 9:14
  of compute that they spent on this just for ARC.
• 9:17
  It's like, Dang, that's a lot, right?
• 9:18
  Like, that's a you're not going to find that many people that
• 9:20
  don't want to spend 10 grand on solid arts tasks, right?
• 9:23
  But that was low compute. And then the O3 preview was the
• 9:27
  first reasoning model, no? No, no, it wasn't the first
• 9:29
  reasoning model because you know it depending on when you want to
• 9:33
  call. I believe their first reasoning
• 9:34
  model was when they did O1 preview.
• 9:36
  I don't even know when that was. That must have been early 2024,
• 9:39
  mid 2024, maybe even 2023. But so we tested O3 preview, low
• 9:45
  compute and then we also tested high compute.
• 9:48
  Which is way more money. That used, I forget the exact
• 9:51
  dollar amounts it was in the thousands of dollars per task.
• 9:54
  What per task? And again, remember, we, we, we
• 9:56
  tested, you know, 500 tasks or whatever.
• 9:58
  I think it was 170 times the amount of compute that was used
• 10:01
  on the low compute side. But either way, the TLD, it's
• 10:04
  like, OK, So what? That's a lot of money.
• 10:05
  What, what's the important part? The important part is it just
• 10:10
  reconfirmed that you can pay for better performance, which is
• 10:12
  crazy, right? And there's open questions as to
• 10:15
  where that scaling actually tops out.
• 10:17
  And so we haven't done as thorough analysis on that system
• 10:20
  as we'd like to. There's a big open question on
• 10:22
  does it asymptote or can you get to 100%?
• 10:25
  If you just threw more and more and more, but keep in mind it's
• 10:28
  log what you give it. So you, they, you know, let's
• 10:31
  just say you spent $1,000,000. In order to get a couple
• 10:34
  percentage points upgrade, you need to spend 10 million and
• 10:36
  then you spend 100 million. It's like, well, where did it
• 10:37
  actually end up stopping for us? Yeah.
• 10:39
  What's that trade off? Yeah, yeah.
• 10:40
  And also, are you going to have to wait a year before it gets
• 10:44
  done well? That's a long time.
• 10:45
  So for the high compute, the job took overnight more or less.
• 10:49
  Like more overnight. Maybe even took longer than.
• 10:52
  I forget the exact time on the duration for it, but it was not
• 10:54
  a short amount of time. You're not going to be sitting
• 10:56
  there waiting for the response. Yeah, it's something you do.
• 10:59
  You let it run, go out, have your life, and then come back
• 11:03
  and see if it was able to do it. But actually just today we
• 11:07
  opening out last week they launched their O3 production,
• 11:10
  great O3. And so there's like a big open
• 11:13
  question. It's like, OK, well, how does
• 11:15
  what we tested in December match to what was publicly released
• 11:18
  today? And we asked, opening how we
• 11:20
  asked Jerry for confirmation on this and a bunch of nuance.
• 11:23
  TLDR, it's not the same model exactly.
• 11:26
  It's not the same model and there's less compute being used.
• 11:28
  So we should expect not the same scores.
• 11:30
  And so we just, we tested on it today and yeah, as expected, it
• 11:34
  does really, really well. It doesn't do as well as the.
• 11:36
  Model that we tested at 87%. It's not 87%, but also they
• 11:41
  release O4 Mini. So they're just, they're
• 11:43
  freaking keeping the models coming, right.
• 11:45
  And so a lot of good testing, a lot of good stuff.
• 11:49
  But what's cool is that like RKGI is the tool that we're
• 11:52
  using to evaluate these things. We have RKGI 2 and all these
• 11:55
  models are still going really really low on RKGI 2.
• 11:57
  So basically Ark AGI 2 is meant to be like that next step, like
• 12:04
  order of difficulty more. So I I've been looking for the
• 12:08
  good analogy with it. So apologies, I don't have it
• 12:10
  down yet, but the way I think about it, and this may be an
• 12:13
  incorrect 1 so apologies if I butcher it, But like if ark AGI
• 12:16
  1 measures is really good at measuring car speed between 20
• 12:21
  miles an hour and 40 miles an hour.
• 12:23
  Below 20, it's not very good. Over 40, it's not very good
• 12:26
  because it just maxed out. It's like literally redlining.
• 12:28
  It's over at the top. RKG I2 is measuring cars from 40
• 12:32
  miles an hour to 80 miles an hour.
• 12:34
  So below 40 miles an hour, you're not going to get much
• 12:36
  signal. You're going to get a little
• 12:37
  little bit of, you know, a little bit of stuff, so.
• 12:39
  It's got to be those premium models it.
• 12:41
  It's going to have to be premium models and we're we're not yet
• 12:44
  seeing the that models are making substantial progress on
• 12:47
  it. So I think that the, the best
• 12:48
  open source model right now I think is getting like 3 to 4%.
• 12:51
  I'm sorry, not best open source. Even when we tested O3 like
• 12:54
  medium, it was getting like 3 to 4% on this.
• 12:57
  And down at that range, we're only talking about 120 tasks on
• 13:00
  that. So like down at that range,
• 13:01
  we're talking about noise. It's not until it starts to get
• 13:04
  like 10-15 that you're really going to start to see something
• 13:06
  substantial from it, yeah. How did and how do you go about
• 13:10
  deciding these questions? Yeah.
• 13:13
  And then there's also the other side of it where you can't just
• 13:16
  have it be a super hard task. You have to almost.
• 13:19
  You have to be creative about. It that's not what like again,
• 13:21
  because we hold ourselves to the restriction that humans need to
• 13:24
  be able to do it and that restricts you from just doing
• 13:27
  hard, hard, hard, hard. Yeah, it can't be this PhD plus.
• 13:29
  It can't be the PhD plus, but as long as we can come up with
• 13:32
  those problems that tells you there is a gap between human
• 13:36
  intelligence and AI. And like people argue with me
• 13:39
  and say, oh, you don't need to aim for human intelligence if
• 13:41
  you want to aim for AGI because they're two different things.
• 13:44
  I agree. But like our hypothesis is that
• 13:47
  the fast track towards AGI is understanding how the human
• 13:49
  brain works and understanding where the gaps are.
• 13:51
  Because if we aim for those gaps, that's going to tell
• 13:53
  something interesting from there.
• 13:55
  We can talk about this later, but like it's nowhere.
• 13:57
  The human brain is nowhere near theoretically optimal
• 14:01
  intelligence. Like we got a lot of biological
• 14:02
  baggage. I could tell you that right now,
• 14:04
  man, I my human brain is not working to full capacity exactly
• 14:09
  ever. So by no means am I saying it's
• 14:11
  the best example, but it is our only example of general
• 14:15
  intelligence, and so we see it as a useful model to go after
• 14:18
  anyway. So how do we pick the problems?
• 14:19
  So in 2019, Francois Chalet came out with this paper called on
• 14:22
  the measure of intelligence, which is so fascinating because
• 14:24
  it's like, how do you come up with the problems?
• 14:26
  That's actually the not the question to start with.
• 14:29
  The question to start with is how do you define intelligence?
• 14:32
  Because if you can define it clearly, then you can come up
• 14:34
  with problems for it, which is, which is the freaking
• 14:37
  fascinating. So Francois came out of this
• 14:39
  paper on the measure of intelligence.
• 14:41
  And so his definition of intelligence was what is your
• 14:46
  ability to learn new things? It's not how good you are at
• 14:52
  chess. It's not how good you are at go,
• 14:54
  It's not how good you are at self driving.
• 14:57
  It's if I give you a new task in a new domain and new set of
• 15:00
  skills that you need to learn in order to do it, can you
• 15:03
  successfully learn that thing? Is it how fast you learn that?
• 15:05
  So now that's a great question. So my, my opening definition of
• 15:08
  intelligence is always just binary.
• 15:09
  Can you learn new things or can you not?
• 15:11
  But his actual definition of intelligence is your efficiency
• 15:15
  of learning new things. So just for example, I like to
• 15:18
  do efficiency in terms of 2 axis.
• 15:21
  Number one is the amount of energy required to learn new
• 15:25
  things, and we'll get into that in a second.
• 15:28
  But like the second dimension is the amount of training data that
• 15:32
  you need to learn that new thing.
• 15:34
  So basically how many times you need to do it before you learned
• 15:36
  it? Exactly.
• 15:37
  So a crude, crude, crude example is if I'm going to teach you how
• 15:42
  to play go, we might need like 6 hours, I'll teach you the rules
• 15:46
  and you'll become you'll become like basic at it.
• 15:48
  We can at least have a conversation around it.
• 15:51
  Think about how much training data went into the the system
• 15:53
  that ended up beating go a lot, a lot, right?
• 15:56
  And so of course that was better skill for it, but there's almost
• 15:58
  out outsized training that went into it.
• 16:01
  So another way to do this is do humans have an internet's worth
• 16:05
  of training data in their head to output the intelligence that
• 16:08
  you see from us right now? And the answer is no, No, it
• 16:10
  doesn't. Language models do.
• 16:12
  And so on the recent podcast, it was the internal one with open
• 16:15
  AI, Sam Allman, and I believe the name the, the fellow's name
• 16:18
  was Daniel, but he was talking about the efficiency of language
• 16:22
  and what is an LLM efficiency of language versus a humans
• 16:25
  efficiency of language. And he said by his estimate, I
• 16:28
  think this might be a little low, but he said that humans are
• 16:31
  100,000 times more efficient with language than with current
• 16:34
  LLMS, which speaks to, and one of the underlying things that
• 16:37
  they kept on talking in the pod was that look, compute isn't
• 16:41
  what's blocking us anymore. We have a shit ton of compute.
• 16:43
  Like we have a lot of compute like Stargate, all of NVIDIA.
• 16:47
  We have so much freaking compute.
• 16:49
  What's blocking us right now is more on the data side.
• 16:52
  But underlying all that, what's blocking us more is also on the
• 16:54
  algorithmic side. It's like we just literally need
• 16:57
  new. We just need new algorithms,
• 16:59
  basically breakthroughs in order to get to the human levels of
• 17:02
  efficiency on it. And just the random point that
• 17:04
  really drive this home is like the other reason why I love
• 17:06
  using the human brain has a benchmark for it is because you
• 17:09
  know how much energy the human brain takes, like literally
• 17:11
  calories. Like how many calories does a
• 17:13
  human brain consume? And you convert calories into
• 17:15
  energy. And then you compare that to
• 17:17
  what is the inference energy used to solve ARC.
• 17:21
  Like you can already tell you're you're miles ahead, you're miles
• 17:24
  and miles ahead. So human brains is a good
• 17:27
  benchmark for us. Also we should note like did
• 17:31
  this all just start you down this path from the needle in the
• 17:35
  haystack? Was that like what blew up, you
• 17:38
  know, needle? In a haystack was is is a fun
• 17:43
  bullet point on my journey like that I've been on so far.
• 17:46
  I I wouldn't call it the thing that did it right.
• 17:48
  I mean, it was cool, but it wasn't like it like it didn't
• 17:51
  make me rich like it didn't like blow me up.
• 17:53
  You know what I'm saying? It's like it was a small little
• 17:55
  thing like on like. You got a small little.
• 17:57
  Retweets. I got a few retweets from, you
• 17:58
  know, it's like, yeah, I got a few likes on Twitter, but it
• 18:01
  wasn't like it wasn't much from. So no.
• 18:05
  But like the inherent thing, like what whatever it is about
• 18:09
  what drives me and like whatever it is about me that makes me put
• 18:14
  my energy where I do needle in the haste that came out of that
• 18:16
  spot, Like other stuff comes out of that spot and like, you know,
• 18:19
  everything. And so like I would say that all
• 18:21
  like all the activities that happened were symptoms of where
• 18:24
  I choose to put my energy and consequences of it and those
• 18:27
  consequences line themselves up to put myself on the path of
• 18:30
  where I. AM and then it opens doors.
• 18:31
  It's like hey this happened cuz for those listeners also they
• 18:36
  should know that you were doing amazing tutorials.
• 18:39
  That was like I. Was doing YouTube work.
• 18:40
  Yeah, that's how I found you is back in the day when you were
• 18:43
  doing the YouTube tutorials, you were like the first guy making
• 18:46
  Ling chain. Tutorials.
• 18:47
  So that's another wild story. It's a super brief on that.
• 18:50
  I remember the first Ling chain. Well I saw I was scrolling
• 18:53
  hacker and it was just trolling or whatever and I saw it was
• 18:56
  this was October like 22. Yeah, it's like right when
• 19:00
  ChatGPT came out. Right when chat GB maybe even a
• 19:02
  hair before and it said show Hacker news Lang chain.
• 19:05
  So it's like literally like there's like the like launch
• 19:07
  blog post the Lang chain and I'm looking at this and I'm like,
• 19:10
  holy shit, this is solving a lot of the problems that I had
• 19:13
  building with the raw API at the time.
• 19:15
  Because keep keep in mind at the time there wasn't a chat model.
• 19:17
  It was just Da Vincio 3. It's like trying to work with
• 19:20
  that thing was obnoxious. Like you had to go.
• 19:23
  There's a lot of friction to get the value out of it.
• 19:24
  Anyway, Lang chain helped out with that a little bit more and
• 19:27
  I was like, this is so cool. And I had had a previous history
• 19:31
  of doing pandas tutorials on YouTube that went nowhere.
• 19:34
  They freaking sucked. Like, like it was talking about
• 19:36
  like me and like my mom's basement, like in my underwear,
• 19:38
  like making pandas at least. You're on.
• 19:40
  It, it wasn't exactly that, but it was along those lines.
• 19:43
  So I'd made, I think like some, I'm gonna make it call it like
• 19:46
  80 pandas tutorials or something like that, because that was my
• 19:48
  craft. Like data analysis was my craft
• 19:50
  at that time. That's what I pride myself on.
• 19:53
  And I saw, I went to YouTube and I typed in Lang Chain and there
• 19:56
  was 1 tutorial by a guy who ended up knowing a little bit
• 19:59
  later on. Super awesome.
• 20:00
  His name's James Briggs. And there was one Lang Chain
• 20:02
  tutorial and I kind of had just like one of those small little
• 20:05
  like little light bulb moments. And I was like, dude Greg, you
• 20:07
  should do what you did for Pandas, but you should do it for
• 20:10
  Lang Chain. And all I did for Pandas was
• 20:12
  just see what I was curious in, go and make a bunch of like
• 20:15
  tutorials and functions. And so at the time, just based
• 20:19
  off just riding my pandas, kind of like success or ripples.
• 20:21
  And it wasn't success. I just mean like, whatever was
• 20:23
  coming from it, I was getting like 3 or 4 new YouTube
• 20:25
  subscribers a day. And I did my first language
• 20:27
  tutorial and I got 16 new subscribers after that one.
• 20:30
  I was like, that's 4X success. That's 4X of where I was.
• 20:33
  Anyway, I did #2 and that next day I got 25.
• 20:36
  And then I did #3 and the next day I got 50.
• 20:38
  And keep in mind, like that's 10X what I was doing beforehand.
• 20:40
  I pulled my wife in the room. I'm like, holy shit, Eliza,
• 20:44
  there's something. Here like like.
• 20:45
  I've, I've retold the story a few times, but like there's a
• 20:48
  few times in life when you notice that the ROI in your
• 20:51
  energy that you get, that you get sometimes, often in life,
• 20:55
  it's like you put out one unit of energy, you're getting like
• 20:57
  20% back. Like it's really not much.
• 21:00
  Like you might get some money, but you're not getting like
• 21:02
  fulfillment, you know, blah, blah, blah.
• 21:03
  At that moment in life, I was putting on one unit of energy
• 21:06
  and I was getting like two or three times back.
• 21:08
  Because you were getting energy I.
• 21:09
  Was getting energy like I couldn't sleep.
• 21:11
  Like I was just like I got to wake up.
• 21:13
  What am I doing today? Like what tutorial am I making
• 21:15
  today? Upgraded my setup.
• 21:17
  Like I was so freaking jazzed on it.
• 21:18
  Like I know I was like met Harrison, did all this other
• 21:21
  stuff. And just through that, just
• 21:23
  natural questions came around, like, how do you do better
• 21:25
  retrieval? All these business questions
• 21:27
  that I had beforehand, how do you do better on that?
• 21:29
  And one of them was for Needle in the Haystack, which is
• 21:32
  everybody was talking about long context.
• 21:34
  Oh, it's longer, longer, longer, longer.
• 21:36
  And I'd seen some tweets that were like, yeah, but it's
• 21:38
  actually not that good at long context.
• 21:40
  And I was like, you guys are idiots.
• 21:42
  Let's just go and test this thing.
• 21:43
  Yeah, there's a process we. Can follow like I was like,
• 21:45
  remember, I'm a data dude, so that's my craft.
• 21:47
  And all I saw in my head was a heat map and I was like the
• 21:50
  length. And then there was that whole
• 21:51
  question around like if the position of where your needle
• 21:54
  was had a factor into it, I was like, might as well throw a two
• 21:56
  by two because it's going to look pretty if if nothing else.
• 21:59
  And so I ended up doing that, and that's where Needle and the
• 22:01
  haystack came around. That's so wild, man.
• 22:03
  So now we were talking before about the reasoning models and
• 22:07
  just this test time compute and you have thoughts?
• 22:10
  Here's the undisputable fact, you spend more money at
• 22:12
  inference time, you get better performance.
• 22:15
  The open questions are, and this is where people argue with me,
• 22:17
  but I still believe it's open is does like for top, top frontier
• 22:21
  models, does asymptote sub 100% or can you get to 100%?
• 22:25
  I think there's too much money that you need to go figure out
• 22:27
  to go try to answer that question.
• 22:29
  That's a big one. So it's it's so high risk that
• 22:32
  why even try? Well, not high risk you it, the
• 22:36
  cost is guaranteed. You're gonna spend a shit ton of
• 22:39
  money. What you, what you get returned
• 22:41
  TBD on where it is. It's not it's not worth it right
• 22:45
  now. But here, here's the other
• 22:46
  thing. It's like I harp a lot on AGI
• 22:47
  and a lot of that stuff we're gonna get.
• 22:49
  We have really, really useful, economically useful models right
• 22:52
  now without having AGI. That's cool.
• 22:55
  Like that's great. I love it.
• 22:55
  That's value to the world. I'm a capitalist at the heart.
• 22:57
  Like I want good tools to be used for the good of humanity.
• 23:01
  LLMSO, 3O4, Mini, all that stuff are great tools that are going
• 23:04
  to bring us really, really good progress.
• 23:07
  The AGI conversation is a separate conversation and that's
• 23:09
  more of a theoretical, philosophical, scientific one
• 23:11
  around, well, what is AGI? How do you actually define and
• 23:13
  how are we going for that? Yeah, what is intelligence?
• 23:15
  What is intelligence and what's wild?
• 23:17
  What blows my mind is I'm I'm you're freaking get me going,
• 23:19
  man. I'm like ranting.
• 23:21
  Well, what's wild? Revving them up.
• 23:22
  What's wild man is that we don't have a formal definition of
• 23:25
  intelligence that the community relies on.
• 23:27
  It's for something as hot topic as AGI and what we have right
• 23:31
  now. It's making me wonder if it can
• 23:34
  be formally defined, if it hasn't already beforehand.
• 23:37
  There's a few stories I can tell in my head.
• 23:38
  One is it can't be. But that also takes a very
• 23:41
  humans are really smart approach.
• 23:43
  And we've seen many times over and over again that like humans
• 23:45
  are not as smart as we think we are.
• 23:47
  So the alternative is that maybe we just don't have a sensitive
• 23:50
  enough understanding about like the actual tools about what we
• 23:52
  need for it. But then the other story is the
• 23:55
  other story that you play in your head is like, yeah, it can
• 23:57
  be. We just don't we.
• 23:58
  Just don't know potentially. And we're never going to.
• 24:02
  Know potentially, yeah. And then there's there's a whole
• 24:05
  different subclass of intelligence, which is human
• 24:07
  relevant intelligence. So like there's a certain class
• 24:09
  of intelligence that you need to survive on Earth right here.
• 24:11
  That's what humans have. That's what we have and built up
• 24:13
  there. But if you if you really expand
• 24:15
  out and this is where we get into like more philosophical
• 24:17
  like in the in the grand scheme of things, the Earth is, is a
• 24:21
  pretty small piece, right. So if you're talking about
• 24:23
  universal intelligence and talking about like theoretical
• 24:26
  intelligence, and let's not go here, but I'll just, I'll just
• 24:30
  light the match. But if you, if you jump into
• 24:33
  people are going to think I'm going over the edge of this.
• 24:35
  But if you jump into like simulation theory, what's the
• 24:39
  intelligence that governs that type of thing that would make
• 24:41
  our own world that come from there?
• 24:43
  I guarantee it's not human relevant intelligence.
• 24:44
  And there's a there's a theoretical optimum that's that
• 24:46
  we're not even going to touch. But that's the other thing too.
• 24:49
  You got to walk before you run. We're going to start with human
• 24:51
  intelligence first anyway, reason was, I mean, they're
• 24:57
  great. I mean, you can scale them,
• 24:59
  throw more money at them, you get better performance.
• 25:00
  They take longer thinking for longer.
• 25:02
  There's big open questions on how these reasoning models
• 25:05
  actually work. And so one simple way to do it
• 25:08
  is the very first reasoning model that people ever came up
• 25:11
  with was they told the model, please think out loud 1st and
• 25:14
  then give me your answer that ended up doing better
• 25:17
  performance. Crazy, right?
• 25:19
  And then what you go do is you go train on processes like that
• 25:22
  for much, much longer. And that's another way to scale
• 25:25
  these things up. You say think for longer, think
• 25:26
  for longer, wait, reflect A reflection step.
• 25:29
  You know, and you say you say keep on going.
• 25:31
  Another method to scale these things up is you say, all right,
• 25:34
  I'm going to tell 10 of you, I want you to, I want 10 of you to
• 25:36
  think out loud. And then I'm going to see what
• 25:38
  all 10 of you respond. And I'm going to pick the best
• 25:40
  answer that comes from there. There's even further ways to do
• 25:43
  it, which is like, I want you to think of the first step in your
• 25:46
  process. OK, Now what are 10 potential
• 25:48
  steps that would come after that first step?
• 25:51
  I'm going to pick the best, the best one of those ten.
• 25:53
  OK, now I'm on Step 2. Think of 10 potential step
• 25:56
  threes. I'm pick the best one and then
• 25:58
  boom boom, boom and go all the way down.
• 25:59
  There's always latency and cost trade-offs that come with those
• 26:01
  things. But either way, I it's
• 26:04
  undeniable the performance you're getting from these things
• 26:06
  and how good they are, even through vibe anecdotes and even
• 26:09
  through RKGI performance. So they're very impressive.
• 26:12
  Yeah, so it's almost subjective and objective.
• 26:16
  Don't get me started dude. I mean, this is another one of
• 26:19
  my things. RKGI is a verifiable domain.
• 26:21
  Like you can just go check is the right answer right?
• 26:23
  What blows my mind is like, there's no right answer for how
• 26:26
  good a summary is, right? There's no right.
• 26:28
  There's no right answer for how good an AI took notes on your
• 26:32
  call and then went and put them into sales force.
• 26:34
  Like how good are the notes, right?
• 26:36
  Yeah, and how good are they to who?
• 26:38
  Well, so that's the whole point is you have to keep in mind what
• 26:43
  is the background engine, what is your eval engine in which
• 26:46
  you're evaluating these things from?
• 26:48
  With ARC, we can it's, it's, it's an equality check.
• 26:50
  We can tell we have the right answer.
• 26:52
  It is not right. Much of what drives the economy
• 26:55
  and drives humans and everything, the eval engine is
• 26:57
  human preference. Well.
• 26:58
  That's what I was going to say with ARC.
• 26:59
  Don't you find that the answers can be subjective?
• 27:04
  So if you're just looking at whether or not the task is
• 27:07
  correct, yes, If you're looking at claims as this is human,
• 27:11
  solvable or whatever, then it's a lot more subjective.
• 27:14
  There's a lot more subjective that comes from there.
• 27:16
  But in terms of eval engines, I have a priority order of my
• 27:23
  favorite eval engines back there.
• 27:24
  Number 1 is going to be physics. And what I mean by that is I
• 27:28
  think that the coolest thing that we could have AI do for us
• 27:31
  is discover new knowledge about reality, basically about
• 27:34
  physics. So you think about what is the
• 27:35
  right answer? Well, it's what does the
• 27:38
  scientific process say about physics as the, as the eval
• 27:40
  engine? That's so freaking cool.
• 27:41
  There's no umbrella that encompasses physics.
• 27:43
  Physics is what we're in, right? And so I think that's, that's
• 27:45
  number one. That's super cool #2 capitalism.
• 27:49
  And so you think about capitalism is a human contract
• 27:52
  of a set of rules that we all, we all play by the kids system
• 27:55
  and there's laws and there's how we choose to do things.
• 27:58
  Running a business is an experiment playing in that
• 28:01
  world, right? And so it's like almost like
• 28:03
  capitalism is the eval engine and like, I'm going to go try to
• 28:06
  make a whole bunch of money, but you got to do within the rules,
• 28:08
  right? And so like there's certain
• 28:09
  things you need to go for. So I think capitalism is a
• 28:11
  really interesting eval engine. And then human preference after
• 28:14
  that, which is like, how good is this summary?
• 28:16
  But the wild thing about human preferences, there's no, there's
• 28:19
  no way to like at scale quantify that, which is really tough,
• 28:24
  which is why when you do RLHF, you got to go spin up like date,
• 28:29
  not data centers, but like huge, huge conference rooms of
• 28:32
  hundreds and thousands of people giving you preference
• 28:34
  optimizations on which one's better, right?
• 28:37
  That's how you do it, which that's crazy, but that's what it
• 28:39
  takes in order to do these things.
• 28:40
  So go back real fast to this capitalism one or even the
• 28:44
  physics one, because in a way we are assuming that what is
• 28:51
  happening to us as humans is discoverable or is the engine
• 28:58
  eval engine. But potentially it's not.
• 29:00
  It's just us as humans, yeah. So a big caveat with that is the
• 29:03
  way that I think about it is if it's true, what we see is what
• 29:06
  we get. Like if it's true, that like
• 29:09
  reality appears to be what it is.
• 29:11
  And I know there's going to be like, even as you start to delve
• 29:14
  into like the quantum stuff and we don't know what's on the
• 29:16
  multi world side, Like we don't know what's on that other
• 29:20
  pending something surprising coming out of there, which I
• 29:22
  would love because it's like, I want the truth.
• 29:24
  And if that's the truth and freaking so be it.
• 29:25
  That's freaking awesome. Pending all that, assuming what
• 29:28
  you see is what you get, then I think what I say still hold on.
• 29:32
  Like I still think that the reality is if there's some
• 29:36
  unexplainable thing that like it's just out of our reach to go
• 29:38
  do it. I'm a little I'm AI like answers
• 29:42
  less that we don't have an explanation for at least, but
• 29:45
  like I'm I'm not ruling it out. I'm saying yes, that is a
• 29:48
  caveat. I'm operating looking at this
• 29:50
  way for it though. When I think about it, it's like
• 29:52
  there's something beyond our understanding.
• 29:56
  Yeah, potentially that is what we are going to get helped to
• 30:00
  understand. AI can help us understand it,
• 30:04
  but it's going to be outside of what we are looking at.
• 30:07
  Just like when you have the the chess move that is played and
• 30:11
  then later it's like, Oh yeah, of course.
• 30:14
  What? I never would have thought of
• 30:15
  that or it would have taken us decades to figure that out.
• 30:18
  Now we get to see. But that's a wild.
• 30:23
  I'm with you man. And humans are traditionally
• 30:26
  very poor at forecasting the unknown unknowns.
• 30:29
  And right now that's all I know. No notes and countless examples.
• 30:32
  Go and ask somebody about something in the 1800s.
• 30:34
  What would today be like? They have to have no freaking
• 30:36
  idea. They just had no idea what came
• 30:37
  from so that will that will happen to us like whatever
• 30:40
  happens. And you know, even with how
• 30:42
  accelerated these timelines people talk about, I mean even
• 30:45
  called 10 years from now like. You know what, I I had a great
• 30:48
  dinner 2 nights ago with a friend and he was saying I had
• 30:52
  as a thought experiment to come up with headlines for what 20-30
• 31:00
  would be saying in different magazines.
• 31:02
  So he was saying I created one headline for Wired and it was
• 31:08
  that teen 3D prints microchip in their basement type thing.
• 31:14
  So that was one. And then another one was, and
• 31:17
  this is a complete tangent, but it's trying to think forward on
• 31:20
  like, oh, what could be possible.
• 31:22
  He was saying data center on the moon opened or second data
• 31:29
  center on the moon is opened by the US type thing.
• 31:32
  And so you're like, well, maybe that's not too far off.
• 31:35
  Yeah, I'm both. Those are very intractable to me
• 31:38
  because the path to do those is you could lay that out like it's
• 31:42
  straightforward. If you said something that
• 31:45
  didn't have a clear obvious like lineage to get towards that,
• 31:48
  then I would start to think about it a little bit more.
• 31:50
  But yeah, I'm over the David Deutsch philosophy that all
• 31:53
  problems are solvable. And that's the argument for
• 31:55
  optimism, is if you believe all problems are solvable, then
• 31:57
  there's nothing out there that like should really worry you
• 31:59
  that much, because you can go figure it out, Go do it all.
• 32:02
  After he told me that, I was trying to think what would my
• 32:06
  headline be for 2030? Where would I go with that?
• 32:09
  Only 5. Five years from now?
• 32:11
  Four years and 3/4. We're going to be specific.
• 32:14
  Yeah. I mean, I mean, you kind of got
• 32:16
  to be with these things. It's like that.
• 32:20
  Quarter could be a big. Because I've been, I've been
• 32:22
  listening, I've been going deep, really deep on there's a big
• 32:25
  conversation around intelligence explosion, 30% growth or GDP
• 32:30
  growth rates and all that. And one of the criticisms I have
• 32:33
  with some of the more outlandish ideas is that they're not as
• 32:36
  tactical and they're not as concrete as I really wish that
• 32:40
  some, some of these projections were.
• 32:42
  So getting concrete saying four years and three quarters, it's
• 32:44
  like, well, damn, opening eye just came out with O4 mini this
• 32:48
  past week. When is O4 coming out?
• 32:50
  When is O4 Pro coming out? Could it be like at the
• 32:52
  beginning of 20252026? If so, you only got 3 years left
• 32:56
  for, for with those types of things to go for it.
• 32:58
  And so concretely, like, how are the how is the GDP going to go
• 33:02
  30%? How is that data center going to
• 33:05
  get up to Mars? How many like windows at launch
• 33:07
  windows are there left or even up to the moon or whatever it
• 33:09
  may be? So it's like what what I'm
• 33:12
  thinking about one thing that's caught my that's nerds on me a
• 33:15
  little bit is like, so Elon wants to go.
• 33:16
  He has a Mars window that he wants to go shoot for.
• 33:18
  Humans are not going to be the first.
• 33:21
  Yeah, why would? We Why would you?
• 33:23
  We already have the Mars Rover. We have the Mars Rover, and so
• 33:25
  humans aren't going to be the first one.
• 33:26
  So that means they're going to send optimists out there.
• 33:30
  Could you? Are we going to have AGI on
• 33:32
  Earth before that window? If so, then you pretty much have
• 33:36
  AGI on optimists because like you just go send a bunch of
• 33:39
  commands and so next thing you know, I feel a little bit like
• 33:41
  almost insecure, but then I need to remind myself not to be so
• 33:44
  emotional. But it's like, damn, humans
• 33:45
  weren't the first on Mars, so. You missed that one.
• 33:50
  You know, I mean, I'm sort of it sounds so lame, but to think
• 33:52
  about it, But that's that was my first reaction.
• 33:54
  I was like, damn, there's gonna be this robot that is
• 33:56
  intelligent, that's its own human being, but it's not a
• 33:58
  human. And then I think it's like,
• 33:59
  damn, am I just specious? And I just love like the human
• 34:02
  race so much. Then I need to open.
• 34:03
  It to claim this I wanted to plant that US flag on Mars.
• 34:06
  You know, I don't know, even if it's just like humanity's flag
• 34:10
  or whatever, maybe. But if you think about it that
• 34:12
  way, there's already been the Mars Rover, so how is it
• 34:14
  different than the Mars Rover and?
• 34:16
  That's where my biological baggage is bringing me down, you
• 34:18
  know, It's just. Because it's like a humanoid
• 34:20
  shape. Humanoid.
• 34:21
  I think it's less the humanoid shape for me and it's more just
• 34:25
  a generally intelligent being. That can do.
• 34:28
  It that's art, that's artificial, that can do it need
• 34:30
  to be. But isn't the Mars Rover?
• 34:32
  The Mars Rover is not being controlled.
• 34:33
  I think it is. I think it is.
• 34:35
  Don't they send it instructions and tell it to go do?
• 34:38
  That's a good question. I I should figure that out?
• 34:40
  Because it moves pretty slow then probably like weights like
• 34:42
  for like. We got it.
• 34:45
  Fact checked that one. Just goes and takes actions,
• 34:48
  yeah. It's like, what next?
• 34:49
  And you're like 3 minutes later or 4 minutes later, OK, turn
• 34:53
  right or pick up the raw. I mean, whatever.
• 34:55
  I don't think it's that far off. I feel sure it's like that.
• 34:57
  That's funny. Yeah.
• 34:58
  I thought it was a bit more autonomous.
• 35:00
  Yeah. And or maybe they send 3 or 4
• 35:03
  instructions at once and if it fails, then resend them or
• 35:07
  figure out where we're at now. Something like that.
• 35:10
  Somebody will have to give us that one.
• 35:11
  Because that is that is hilarious.
• 35:13
  What else you've been thinking about?
• 35:15
  Yeah, well, in terms of headlines, I still haven't given
• 35:17
  you a headline I'm thinking about here.
• 35:19
  So headline 2030 Wired says it I, I don't think it's outside of
• 35:26
  the question that there could be a headline that says humans are
• 35:29
  no longer able to come up with questions that AI can't answer,
• 35:34
  which is, isn't that sensationalist.
• 35:36
  Like it's kind of muted from a sensation standpoint.
• 35:38
  Like if you use our definition, like the, the observational
• 35:41
  definition of AGI, that what what are the problems are there
• 35:45
  right? But I.
• 35:46
  Still wonder if there's a world where you have run out of
• 35:50
  questions but you're still not seeing it, where every once in a
• 35:53
  while you'll find that question. It's not that you can find 100
• 35:56
  of them, but there's still those stupid questions where it is
• 36:00
  like the strawberry or the 9.11, Yeah.
• 36:02
  And here's the thing, I don't want to give the viewer the
• 36:05
  impression that I'm relying on this as a formal definition.
• 36:07
  It's I think it's a pretty good working definition for sure.
• 36:10
  It's easy to communicate and it's easy for us to go against.
• 36:13
  I think we'll come up with a formal definition.
• 36:14
  But to your point, how often do you ask a human a question?
• 36:18
  And it's like, what were you thinking?
• 36:23
  You know what I mean? So like as efficiency is such a
• 36:26
  good big, big piece of this here.
• 36:28
  It's like that Will Smith I robot meme that keeps on going
• 36:30
  around. You know, it's like you ask me a
• 36:31
  question. Can you right.
• 36:33
  So. Yeah.
• 36:33
  So then I I could see that though, yeah, we can't come up
• 36:36
  with more questions or we have to have AI come up with the
• 36:40
  questions that it can't answer. Potentially and you know that
• 36:42
  that's a whole another I think that's a very under explored it
• 36:46
  it it's talked about using AI to help build AI to help align AI
• 36:51
  to help test it, you know, all that other stuff and that that
• 36:53
  will happen because like, again, definitions are important, but
• 36:56
  like, look at all the people using cursor right now to go
• 36:58
  build AI models. It's like, is that using AI to
• 37:00
  help you build AI? It's like, yeah, it is.
• 37:04
  So it just depends about how direct you want to use AI for.
• 37:06
  But yeah, So what we're thinking about for RKGI 3, because we're
• 37:10
  coming out with RKGI 2, it's going to get beaten one day,
• 37:12
  right? We know that RKGI 2 could be
• 37:14
  brute forced. So if you give a data center's
• 37:18
  worth of compute and energy and time like a month worth of data
• 37:21
  center, yeah, go brute force it and literally try all random
• 37:23
  permutations using one of the DSLS to like try to solve ARC,
• 37:27
  Yeah, you're going to figure it out.
• 37:29
  But that's why efficiency is a big piece of this.
• 37:31
  Then the energy and the cash that you need to do go do that,
• 37:33
  isn't it? Doesn't make us interested
• 37:35
  because it's a verifiable domain.
• 37:36
  Do you consider Arc 1 beaten because of that 87%?
• 37:41
  Like is 87% a pass grade? That's like AB?
• 37:44
  Plus for a long time we talked about 85% being basically human
• 37:49
  human threshold on RKGI 1. I think that much like the
• 37:54
  battle of MMLU where people were like we got 88.8, well we got
• 37:58
  88.9, well we got 90.1. It's like at that point you're
• 38:02
  redlining where your signal is actually telling me for you and
• 38:05
  you're actually losing signal and that you know, you get
• 38:07
  diminishing returns on the signal that comes from it.
• 38:09
  So I think the RKGI one for anything between like 5 to
• 38:15
  probably 90%, I think it gives you really good signal on where
• 38:18
  something is anything outside those bounds, it is not going
• 38:21
  to, it isn't giving you a ton of it for it.
• 38:22
  So I think it's still a really useful tool today for it.
• 38:26
  It'll eventually go out of vogue, though, once models get
• 38:29
  so good. At it, it's getting closer and
• 38:31
  closer, so it's almost like you see the end of this total
• 38:35
  lifespan. And like, here's the deal.
• 38:36
  It's like benchmark. There isn't 1 benchmark to rule
• 38:39
  them all. Like even if you wanted to
• 38:41
  understand a model's capabilities, you need a
• 38:43
  portfolio. Not only that, but look how many
• 38:45
  benchmarks like had their place and then we're phased out
• 38:49
  because they did their job. Like just even for example, look
• 38:52
  at Imagenet. What happened with that 2012, a
• 38:55
  big data set of images that had a huge impact on the industry
• 38:59
  and it did its job. Would anybody go and do they
• 39:01
  report? They didn't report on Imagenet
• 39:03
  today and that's OK. They have other types of
• 39:05
  benchmarks where they need to go deeper into like image vision
• 39:08
  capabilities in order to get a better job about it.
• 39:11
  So that's where RKG I2 sits. But like I said, we're not
• 39:13
  seeing meaningful, meaningful performance on it yet to to have
• 39:16
  it give us a ton of signal. Brute force it if you want, but
• 39:20
  then you're kind of defeating the purpose totally, and so even
• 39:24
  though you can do it, should you?
• 39:28
  So we run an, we run a Calico competition to try to beat our
• 39:30
  KGI too. The incentives there are to beat
• 39:33
  it at any means necessary because we have money on the
• 39:35
  line and within the competition rules.
• 39:37
  There's no type of solution requirements.
• 39:40
  So people brute force the crap out of that all the time.
• 39:42
  Like that's that's a whole other part of my life.
• 39:44
  It's like if anyone was talking to me about benchmarks, great.
• 39:46
  I love, I can talk about all day long.
• 39:47
  If anybody wants to talk to me about running an AI competition,
• 39:50
  talk to me about it all the time.
• 39:52
  We did it all last year. We put $1,000,000 up to anybody
• 39:55
  who could beat Arc AGI on Cagle and nobody was nobody was able
• 39:58
  to. But like we saw leaderboard
• 40:00
  probing, we saw people getting around the rules.
• 40:02
  We saw where our incentives were.
• 40:04
  We made assumptions about competitors, participants
• 40:07
  incentives. We're not in line with our
• 40:09
  assumptions. Wait, how?
• 40:11
  So? Yeah, basically, if somebody, if
• 40:15
  you wanted to win prize money, you needed to open source your
• 40:18
  solution. We thought that the money being
• 40:21
  a monetary incentive would be enough to make people open
• 40:23
  source their solution. There was one group out there
• 40:26
  who had a really strong solution, really, really
• 40:27
  awesome, and they made the choice.
• 40:30
  I'm not actually, I'm not exactly sure to the exact reason
• 40:33
  it was, it was one of two. It was either we think that we
• 40:35
  have a better chance at not open sourcing our solution and
• 40:38
  competing next year for the grand prize to do really good at
• 40:43
  it. And so they wanted the $700,000
• 40:45
  instead of just the the the yearly $100,000.
• 40:48
  Or it was because it was so close to their startups
• 40:51
  proprietary information that they didn't want to open source
• 40:54
  it. Both of which are not
• 40:57
  necessarily in spirit of what we were.
• 40:59
  Look, we were aiming for as a competition, but we did not
• 41:02
  properly construct the incentives enough or communicate
• 41:06
  early enough that this was an issue.
• 41:08
  And so we basically did what we could, which is we took him off
• 41:11
  the leaderboard because you're not, you're not placing if you
• 41:14
  don't open source. And then this year we made a lot
• 41:17
  more. We're being much more clear
• 41:18
  about our intentions with. How did?
• 41:20
  How are you aligning incentives though?
• 41:23
  Through better communication. And then not only that, this
• 41:26
  year we have a public and private leaderboard.
• 41:27
  So the leaderboard that's that's seen right now is all just based
• 41:30
  off of public data. But the final leaderboard that
• 41:33
  says whether or not you've you've even placed their done
• 41:35
  well is all in hidden data. And if you want to get your
• 41:37
  private score, you need to. Open source, OK, yeah.
• 41:41
  So we're, we're hoping that that does it either way.
• 41:44
  We're not talking about hundreds of thousands of teams here.
• 41:46
  We're talking about maybe 10 teams that are in the running.
• 41:49
  I can go and I can go and have conversations with each one of
• 41:51
  those 10's and make sure that they're seeing it the same way.
• 41:55
  And also the gaming of the leaderboard, you saw that?
• 41:58
  Yeah, I mean, so people get creative because like.
• 42:01
  Money's on the. Line money's on the line.
• 42:03
  And Caglers are professional competition people.
• 42:06
  They're really good at data science stuff and they're really
• 42:07
  good at playing competitions. And so one thing that I saw is
• 42:10
  that they will try to suss out attributes about ARC tasks one
• 42:15
  at a time. And what they'll do is, is
• 42:18
  they'll put in a wait statement in their script that says if you
• 42:21
  see this current task attribute wait 50 seconds.
• 42:24
  And then when they submit their solution to Cagle, they say, did
• 42:26
  it run instantly or did it wait 50 seconds?
• 42:28
  And then that's way that that's a way you can tease out some
• 42:31
  more information about it. Because the only other
• 42:33
  information you get is you get a score that you only get.
• 42:36
  You get a single integer, which is your score out the other end.
• 42:38
  You can't really tell that much information about that.
• 42:41
  Cagle tries to prevent that a little bit more with some
• 42:43
  obfuscation about how long it actually took.
• 42:45
  But people, people get creative like that.
• 42:48
  And now with the ARC prize two, yeah.
• 42:54
  Do you have to create a variety of different tasks or is it very
• 43:01
  much in all right? We're in this one field trying
• 43:04
  to. Do it well.
• 43:04
  So you brought up a question earlier which was good and I
• 43:06
  didn't answer it fully, which is does it take a lot of research
• 43:09
  and like deep thinking to build these things?
• 43:11
  So I would say for Francois's paper in 2019, that's what a lot
• 43:14
  of work to put that, that hypothesis together, that formal
• 43:17
  definite or that that definition of intelligence right out of
• 43:20
  that came, OK, using this definition of intelligence, what
• 43:24
  would a problem look like that would actually go and test these
• 43:26
  things? And that's where the arc
• 43:27
  paradigm came in. So what it is, is basically you
• 43:31
  have an input and you have an output grid and it's just, it
• 43:33
  looks like a checkerboard and you see, OK, the input turns
• 43:38
  into the output. Some way I need to figure out
• 43:40
  how to how do you transform the input into the output?
• 43:42
  You get a few examples and then you get a test.
• 43:45
  And on that test, you only have the input and what your goal is.
• 43:47
  You have to go sell by sell and type out what the output would
• 43:50
  be. The important part is that each
• 43:54
  separate problem on ARC AGI requires a different rule or a
• 43:59
  different transformation to actually solve.
• 44:01
  So what I mean by that is let's super variety, super variety.
• 44:06
  And it's almost like it's a meta thing.
• 44:07
  I'll get why this is important in a second.
• 44:09
  The the way that why this is important is let's just say 1
• 44:12
  arc task has a square on it and on the input outputs, all you're
• 44:15
  doing is you're just adding a border to the square.
• 44:17
  OK, now we're on the test input, we're going to give you a
• 44:19
  square, we're going to ask you and on output, you just need to
• 44:21
  put a border. OK, cool.
• 44:24
  That border transformation rule will only be asked once on
• 44:28
  another task. What we might ask you to do is
• 44:31
  fill in the corner of every single shape and go fill in the
• 44:33
  corners of all those different shapes.
• 44:35
  And So what we're what we're forcing the tester to do is
• 44:39
  learn the new mini skill in each one of those questions.
• 44:43
  And then we're forcing you to demonstrate that you've learned
• 44:45
  that skill on the test. By doing.
• 44:47
  It by doing it which goes back to the definition of Francois's
• 44:50
  Francois's definition of intelligence, which is learning
• 44:52
  new skills that. Is so simple for us, right?
• 44:58
  Like write a border on a square. But it is.
• 45:00
  But it's exactly that and the reason.
• 45:02
  Why it's so hard for machines is because humans are very good at
• 45:04
  abstraction and reasoning. It's like, oh, duh, just put a
• 45:07
  border. Yeah, OK, but that's actually
• 45:09
  really hard for AI to go do. Now arc one, people are like,
• 45:14
  oh, it's so simple. It's not a good test of AI.
• 45:15
  Well, keep in mind for five years it was unbeaten, right?
• 45:18
  And it actually pinpointed the moment like right when models
• 45:20
  started to get good was the exact moment that reasoning
• 45:23
  models took off. OK, that's just something really
• 45:26
  interesting about reasoning the models and using Arc 1 as a
• 45:29
  capabilities assertion. You can actually tell something
• 45:31
  about reasoning the models that there's a non zero level of
• 45:34
  fluid intelligence that actually comes from that, which is very
• 45:36
  cool. Arc 2 is a simple extension of
• 45:38
  the Arc One domain. We still have input output.
• 45:41
  We still ask you to do rules that the the difference with it
• 45:44
  is that the rules are much deeper and they require a bit
• 45:46
  more thought from a human perspective.
• 45:48
  Go to go do it. So instead of just doing a
• 45:50
  border, we might ask you to do a border and do the corners.
• 45:53
  So there's or put next and now there's 2 rules.
• 45:56
  I won't go to the details on it. We actually have a full Francois
• 45:59
  put it together. A we hosted a private preview of
• 46:02
  RKG I2 for donors for for our price because we're a nonprofit.
• 46:05
  I should have said that earlier nonprofit.
• 46:07
  And he gave a 30 minute presentation on RKG I2.
• 46:10
  Wow. But where I want to talk about
• 46:12
  RKGI 3. Of course, RKGI 3 is going to be
• 46:17
  departing from the RKGI 1 and RKGI 2 framework.
• 46:20
  Style of doing it. Style of doing it.
• 46:22
  So it's very scoped and narrow domain.
• 46:24
  If you just have matrices input, output, you know, if it's very
• 46:27
  scoped, you don't have very many axes of freedom for that.
• 46:30
  So we are taking inspiration from simulations and games.
• 46:38
  So back in 2000, I think 2017 DeepMind, they put together an
• 46:42
  exploration, they called it Agent 57.
• 46:44
  So they tried to get an agent, more or less an RL agent, to go
• 46:47
  and try to beat a bunch of different Atari games, right?
• 46:50
  There's like 4 that they didn't solve, which is super
• 46:52
  fascinating. What arc AGI one and two don't
• 46:57
  allow you to don't make you do is they don't make you figure
• 46:59
  out what is the goal. They don't make you figure out
• 47:01
  the rules of the environment. They don't make you have long
• 47:03
  term memory with hidden states. So like you learned something
• 47:06
  early on in the game and you have to remember that that thing
• 47:08
  still applies later on in the game.
• 47:10
  And So what I tell people is if, if you can make an AI that beats
• 47:13
  one game, well, we've done that a bunch.
• 47:15
  We've made AIB chess, we made AIB go.
• 47:17
  OK, cool. If you can make an AI that beats
• 47:20
  50 games, that's much more interesting.
• 47:25
  But the problem is that those 50 games are all public, and you
• 47:27
  can have developer intelligence and developer intuition as to
• 47:30
  how to go beat those 50 games. What arc AGI 3 is going to be is
• 47:34
  we're going to make AI beat 50 games it has never seen
• 47:37
  beforehand, and they're each novel from each other.
• 47:40
  And that is a much further extension and actually is a
• 47:45
  freedom about where we're taking this.
• 47:46
  What you can assert about the model that beats it is it will
• 47:49
  have had no choice but to interact with this environment.
• 47:52
  Learn the rules of the game in 50 different novel situations.
• 47:55
  But you're not letting it just simulate for hours and hours.
• 47:59
  Or maybe you are. That's the test time.
• 48:01
  Compute. Type thing we will, and that's
• 48:03
  where efficiency comes into it. And So what we're going to do is
• 48:05
  we're going to go test 400 humans on those 50 games and
• 48:08
  we're going to see how many actions does it take for a human
• 48:10
  to actually solve this and how many actions does it take for AI
• 48:13
  to go solve it. So that's where we get our
• 48:14
  efficiency that comes from it in addition to cost and energy that
• 48:16
  comes from that. I think we got to go.
• 48:19
  I just saw beautiful man boss man.
• 48:20
  That's a great way of ending it. We'll cut it there.
• 48:23
  This is too good. You were awesome, man.
• 48:26
  This is great. But it was perfect.
• 48:29
  It was like.