Is AI Slowing Down? Nathan Labenz Says We’re Asking the Wrong Question

0:00:02 AI is not synonymous with language models.
0:00:04 AI is being developed with pretty similar architectures
0:00:07 for a wide range of different modalities.
0:00:08 And there’s a lot more data there.
0:00:11 Feedback is starting to come from reality.
0:00:13 Maybe we’re running out of problems we’ve already solved
0:00:16 when we start to give the next generation of the model
0:00:19 these power tools and they start to solve
0:00:21 previously unsolved engineering problems.
0:00:22 I think you start to have something that looks
0:00:24 kind of like superintelligence.
0:00:28 There’s a growing debate about whether AI progress has plateaued
0:00:30 or if our expectations have simply caught up
0:00:31 to the pace or change.
0:00:34 On this episode, I’m joined by Nathan LeBenz,
0:00:36 host of The Cognitive Revolution,
0:00:38 to unpack whether AI innovation is actually slowing.
0:00:41 We break down the case for slowdown
0:00:43 from Cal Newport’s argument that students are using AI
0:00:46 to get lazier to the claims that GPT-5
0:00:48 wasn’t a leap over GPT-4.
0:00:50 Nathan and I look at what’s really happening
0:00:51 under the hood of AI,
0:00:53 from new reasoning and math capabilities
0:00:56 to real scientific discoveries and multimodal systems
0:00:58 that go far beyond chatbots.
0:01:00 We also discuss agents, automation,
0:01:03 and how quickly work itself is starting to change.
0:01:05 And we’ll end on the big question.
0:01:07 If progress isn’t slowing down,
0:01:09 how should we shape it toward a future we actually want?
0:01:11 Let’s get started.
0:01:16 Nathan, I’m stoked to have you on the A&Z podcast
0:01:17 for the first time.
0:01:19 Obviously, we’ve been podcast partners for a long time
0:01:20 with you leading Cognitive Revolution.
0:01:21 Welcome.
0:01:22 It’s great to be here.
0:01:22 Thank you.
0:01:26 So we were talking about Cal Newport’s podcast appearance
0:01:27 on Lost Debates.
0:01:28 And we thought it was a good opportunity
0:01:30 to just have this broad conversation
0:01:31 and really entertain this sort of question of,
0:01:33 is AI slowing down?
0:01:35 So why don’t you sort of steel man some of the arguments
0:01:37 that you’ve heard on that side,
0:01:38 either from him or more broadly,
0:01:40 and then we could sort of have this broader conversation.
0:01:42 Yeah, I mean, I think for one thing,
0:01:43 it’s really important to separate
0:01:45 a couple of different questions,
0:01:46 I think with respect to AI.
0:01:47 One would be,
0:01:51 is it good for us right now, even?
0:01:54 And is it going to be good for us in the big picture?
0:01:57 And then I think that is a very distinct question
0:02:00 from are the capabilities that we’re seeing
0:02:03 continuing to advance and at a pretty healthy clip.
0:02:05 So I actually found a lot of agreement
0:02:09 with the Cal Newport podcast that you shared with me
0:02:12 when it comes to some of the worries
0:02:14 about the impact that AI might be having
0:02:16 even already on people.
0:02:17 You know, he looks over students’ shoulders
0:02:18 and watches how they’re working
0:02:21 and finds that basically he thinks
0:02:23 that they are using AI to be lazy,
0:02:24 which is no big revelation.
0:02:26 I think a lot of teachers would tell you that.
0:02:26 Shocker.
0:02:28 He puts that in, yeah,
0:02:30 puts that in maybe more dressed up terms
0:02:34 that people are not even necessarily moving faster,
0:02:37 but they’re able to reduce the strain
0:02:38 that the work that they’re doing
0:02:40 places on their own brains
0:02:43 by kind of trying to get AI to do it.
0:02:45 And, you know, if that continues,
0:02:46 and I think he’s been,
0:02:48 I think a very valuable commenter
0:02:50 on the impact of social media,
0:02:52 certainly I think we all should be mindful of
0:02:55 how is my attention span evolving over time
0:03:00 and am I getting weak or averse to hard work?
0:03:01 Those are not good trends
0:03:04 if they are showing up in oneself.
0:03:05 So I think he’s really right
0:03:06 to watch out for that sort of stuff.
0:03:08 And then as we’ve covered
0:03:09 in many conversations in the past,
0:03:10 I’ve got a lot of questions
0:03:13 about what the ultimate impact of AI is going to be.
0:03:15 And I think he probably does too.
0:03:17 But then when it comes to,
0:03:19 it’s a strange move from my perspective
0:03:20 to go from,
0:03:23 there’s all these sort of problems today
0:03:24 and maybe in the big picture
0:03:28 to, but don’t worry,
0:03:28 it’s flatlining,
0:03:29 like kind of worry,
0:03:29 but don’t worry
0:03:31 because it’s not really going anywhere further than this
0:03:33 or it’s scaling is kind of petered out
0:03:35 or we’re not going to get better AI
0:03:37 than we have right now.
0:03:40 Or even maybe the most easily refutable claim
0:03:41 from my perspective is
0:03:44 GPT-5 wasn’t that much better than GPT-4.
0:03:47 And that I think is where I really was like,
0:03:47 what, wait a second.
0:03:49 I was with you on a lot of things
0:03:50 and some of the behaviors
0:03:52 that he observes in the students,
0:03:54 I would cop to having exhibited myself.
0:03:56 When I’m trying to code something these days,
0:03:57 a lot of times I’m like,
0:03:59 oh man, can’t the AI just figure it out?
0:04:00 I really don’t want to have to sit here
0:04:02 and read this code
0:04:04 and figure out what’s going on.
0:04:05 It’s not even about typing the code anymore.
0:04:07 You know, I’m way too lazy for that.
0:04:09 But it’s even about figuring out how the code is working.
0:04:10 Can’t you just make it work?
0:04:11 Try again, you know,
0:04:12 and just try again.
0:04:14 And I do find myself at times
0:04:16 falling into those traps.
0:04:17 But I would say a big part of the reason
0:04:18 I can fall into those traps
0:04:20 is because the AIs are getting better and better.
0:04:22 And increasingly,
0:04:23 it’s not crazy for me to think
0:04:24 that they might be able to figure it out.
0:04:28 So, that’s my kind of first slice
0:04:30 at the takes that I’m hearing.
0:04:32 There’s almost like a two-by-two matrix
0:04:34 maybe that one could draw up
0:04:34 where it’s,
0:04:36 do you think AI is good or bad
0:04:37 now and in the future?
0:04:38 And do you think it’s like
0:04:39 not a big deal or a big deal?
0:04:42 And I think it’s both on the good and bad side.
0:04:43 I definitely think it’s a big deal.
0:04:46 The thing that I struggle to understand the most
0:04:48 is the people who kind of don’t see
0:04:49 the big deal
0:04:50 that it seems pretty obvious to me
0:04:51 and the,
0:04:52 especially when it comes again
0:04:54 to the leap from GBD-4 to GBD-5,
0:04:56 maybe one reason that’s happened
0:04:57 a little bit is that
0:04:59 there were just a lot more releases
0:05:02 between GBD-4 and 5.
0:05:04 So, what people are comparing to
0:05:07 is, you know,
0:05:08 something that just came out a few months ago,
0:05:09 03, right,
0:05:11 that only came out a few months before GBD-5.
0:05:12 Whereas with GBD-4,
0:05:15 it was shortly after ChatGPT
0:05:16 and it was all kind of this moment of,
0:05:17 whoa, this thing is like
0:05:19 exploding onto the scene.
0:05:20 A lot of people were seeing it
0:05:20 for the first time.
0:05:23 And if you look back to GBD-3,
0:05:24 there’s a huge leap.
0:05:25 I would contend that the leap
0:05:27 is similar from GBD-4 to 5.
0:05:28 These things are hard to score.
0:05:29 There’s no single number
0:05:30 that you could put on it.
0:05:30 Well, there’s loss.
0:05:31 But of course,
0:05:32 one of the big challenges
0:05:33 is that what exactly
0:05:34 does a loss number
0:05:35 translate into
0:05:36 in terms of capabilities?
0:05:40 So, it’s very hard to describe
0:05:43 what exactly has changed.
0:05:44 But we could go through
0:05:45 some of the dimensions of change
0:05:45 if you want to
0:05:46 and enumerate some of the things
0:05:47 that I think people
0:05:49 maybe are starting to
0:05:50 or have come to take for granted
0:05:51 and kind of forget
0:05:52 like that GBD-4
0:05:53 didn’t have a lot of the things
0:05:55 that now were sort of
0:05:57 expected in the GBD-5 release
0:05:58 because we’d seen them in
0:06:00 4.0 and 0.1
0:06:01 and 0.3
0:06:02 and all those things
0:06:03 sort of maybe
0:06:04 boiled the frog a little bit
0:06:05 when it comes to
0:06:06 how much progress
0:06:07 people perceived
0:06:09 in this last release.
0:06:10 Yeah, a couple reactions.
0:06:11 So, one is
0:06:12 and even to complicate
0:06:12 your 2×2
0:06:14 even further in the sense of
0:06:16 is it bad now
0:06:17 versus is it bad later?
0:06:18 Cal is not really
0:06:19 who we both admire
0:06:20 by the way a lot.
0:06:21 Cal’s a great guy
0:06:22 and a valuable contributor
0:06:23 to the thought space
0:06:24 but he’s not as concerned
0:06:25 about sort of this
0:06:27 sort of future AI concerns
0:06:28 that sort of the
0:06:30 AI safety folks
0:06:30 and many others
0:06:31 are concerned about.
0:06:32 He’s more concerned about
0:06:33 what it means to life
0:06:35 for cognitive performance
0:06:37 and development now
0:06:37 in the same way
0:06:38 that he’s worried about
0:06:40 social media’s impact
0:06:41 and you think that’s a concern
0:06:42 but nowhere near
0:06:42 as big a concern
0:06:43 as what to expect
0:06:44 in the future
0:06:45 and then also
0:06:46 he presents sort of
0:06:47 this theory of why
0:06:48 we shouldn’t worry
0:06:48 about the future
0:06:49 because it’s slowing down
0:06:50 and why don’t we just share
0:06:51 how we interpreted
0:06:52 kind of his history
0:06:53 which as I interpreted
0:06:54 it was this idea of
0:06:55 hey, the simplistic version
0:06:56 is we’ve figured out
0:06:57 this way such that
0:06:57 if you throw
0:06:58 a bunch of data
0:06:59 into the model
0:07:00 it gets better
0:07:01 in sort of order
0:07:01 of magnitude
0:07:02 and so the difference
0:07:04 between GPT-2 and GPT-3
0:07:05 and then GPT-3 and GPT-4
0:07:07 but then that sort of
0:07:07 significant
0:07:08 the difference
0:07:09 but then achieved
0:07:10 sort of diminishing returns
0:07:11 significantly
0:07:12 and we’re not seeing
0:07:13 it at GPT-5
0:07:13 and thus
0:07:15 we don’t have to worry anymore.
0:07:15 How would you edit
0:07:16 the characterization
0:07:17 of his view
0:07:17 of the history
0:07:18 and then we can get
0:07:19 into the differences
0:07:19 between 4 and 5?
0:07:21 the scaling law idea
0:07:22 which is definitely
0:07:24 worth agreeing
0:07:24 taking a moment
0:07:25 to note that
0:07:26 it is not a law of nature
0:07:27 we do not have
0:07:28 a principled reason
0:07:29 to believe
0:07:30 that scaling
0:07:32 is some law
0:07:32 that will go
0:07:33 indefinitely
0:07:34 all we really know
0:07:35 is that it has held
0:07:36 through quite a few
0:07:37 orders of magnitude
0:07:37 so far
0:07:39 I think that
0:07:41 it’s really not clear
0:07:42 yet to me
0:07:43 whether or not
0:07:44 the scaling laws
0:07:45 have
0:07:46 petered out
0:07:48 or whether
0:07:49 we have just found
0:07:50 a steeper gradient
0:07:51 of improvement
0:07:52 that is giving us
0:07:53 better ROI
0:07:54 on another
0:07:55 front that we can
0:07:55 push on
0:07:56 so they did train
0:07:58 a much bigger model
0:07:59 which was GPT-4.5
0:08:01 and that did get released
0:08:03 and
0:08:04 there are
0:08:05 a number of interesting
0:08:06 of course
0:08:07 there’s a million
0:08:07 benchmarks
0:08:07 whatever
0:08:08 but the one
0:08:08 that I zero in
0:08:09 on the most
0:08:09 in terms of
0:08:10 understanding
0:08:12 how GPT-4.5
0:08:13 relates to
0:08:14 both
0:08:15 O3
0:08:16 and GPT-5
0:08:17 and
0:08:17 OpenAI
0:08:18 obviously
0:08:19 famously
0:08:19 terrible at naming
0:08:20 we can all agree
0:08:21 on that
0:08:21 I think a decent
0:08:22 amount of this
0:08:23 confusion
0:08:23 and sort of
0:08:23 disagreement
0:08:24 actually does
0:08:25 stem from
0:08:27 unsuccessful
0:08:28 naming decisions
0:08:29 4.5
0:08:30 on this
0:08:31 one benchmark
0:08:31 called
0:08:32 SimpleQA
0:08:33 which is really
0:08:34 just a
0:08:35 super long tail
0:08:36 trivia benchmark
0:08:37 it really
0:08:38 just measures
0:08:39 do you know
0:08:39 a ton
0:08:40 of esoteric
0:08:40 facts
0:08:41 and they’re
0:08:42 not things
0:08:42 you can really
0:08:43 reason about
0:08:43 you either
0:08:43 just have to
0:08:44 know or
0:08:44 don’t know
0:08:45 these particular
0:08:45 facts
0:08:47 the
0:08:48 O3
0:08:50 class of
0:08:50 models
0:08:51 got about
0:08:51 a 50%
0:08:52 on that
0:08:53 benchmark
0:08:54 and GPT-4.5
0:08:55 popped up
0:08:56 to like
0:08:56 65%
0:08:58 so in
0:08:58 other words
0:08:58 it
0:08:59 basically
0:09:00 of the
0:09:00 things that
0:09:01 were not
0:09:01 known
0:09:02 to
0:09:03 the
0:09:03 previous
0:09:03 generation
0:09:03 of
0:09:03 models
0:09:04 it
0:09:04 picked
0:09:04 up
0:09:04 a
0:09:04 third
0:09:05 of
0:09:05 them
0:09:06 now
0:09:07 there’s
0:09:07 obviously
0:09:07 still
0:09:07 two
0:09:07 thirds
0:09:08 more
0:09:08 to
0:09:08 go
0:09:08 but
0:09:08 I
0:09:08 would
0:09:08 say
0:09:09 that’s
0:09:09 a
0:09:09 pretty
0:09:09 significant
0:09:10 leap
0:09:10 right
0:09:10 these
0:09:10 are
0:09:10 super
0:09:11 long
0:09:11 tail
0:09:11 questions
0:09:12 I
0:09:12 would
0:09:12 say
0:09:12 most
0:09:12 people
0:09:13 would
0:09:13 get
0:09:13 close
0:09:13 to
0:09:14 a
0:09:14 zero
0:09:14 you’d
0:09:14 be
0:09:14 like
0:09:14 the
0:09:15 person
0:09:15 sitting
0:09:15 there
0:09:15 at
0:09:15 the
0:09:15 trivia
0:09:16 night
0:09:16 who
0:09:17 maybe
0:09:17 gets
0:09:17 one
0:09:17 a
0:09:18 night
0:09:18 is
0:09:18 kind
0:09:18 of
0:09:18 what
0:09:18 I
0:09:41 might
0:09:42 was
0:09:42 was
0:09:42 never
0:09:43 trained
0:09:44 with
0:09:44 the
0:09:44 same
0:09:46 power
0:09:46 of
0:09:46 post
0:09:46 training
0:09:47 that
0:09:47 GPT-5
0:09:48 has
0:09:48 had
0:09:49 and
0:09:49 so
0:09:49 we
0:09:49 don’t
0:09:49 really
0:09:50 have
0:09:50 an
0:09:50 apples
0:09:50 to
0:09:50 apples
0:09:51 comparison
0:09:51 but
0:09:51 people
0:09:52 did
0:09:52 still
0:09:52 find
0:09:53 some
0:09:53 utility
0:09:53 in
0:09:53 it
0:09:54 I
0:09:54 think
0:09:55 maybe
0:09:55 the
0:09:55 way
0:09:55 to
0:09:56 understand
0:09:56 why
0:09:56 they’ve
0:09:57 taken
0:09:57 that
0:09:57 offline
0:09:57 and
0:09:58 gone
0:09:58 all
0:09:58 in
0:09:58 on
0:09:59 GPT-5
0:09:59 is
0:10:00 just
0:10:00 that
0:10:00 model
0:10:00 is
0:10:00 really
0:10:00 big
0:10:01 it’s
0:10:02 expensive
0:10:02 to
0:10:03 run
0:10:03 the price
0:10:03 was
0:10:03 way
0:10:04 higher
0:10:04 was
0:10:12 it
0:10:12 for
0:10:12 them
0:10:13 to
0:10:13 consume
0:10:13 all
0:10:13 the
0:10:14 compute
0:10:14 that
0:10:14 it
0:10:14 would
0:10:14 take
0:10:15 to
0:10:15 serve
0:10:16 that
0:10:16 and
0:10:16 maybe
0:10:17 they
0:10:17 just
0:10:17 find
0:10:17 that
0:10:17 people
0:10:17 are
0:10:18 happy
0:10:18 enough
0:10:18 with
0:10:19 the
0:10:19 somewhat
0:10:19 smaller
0:10:20 models
0:10:20 for
0:10:20 now
0:10:21 I
0:10:21 don’t
0:10:21 think
0:10:21 that
0:10:22 means
0:10:22 that
0:10:22 we
0:10:22 will
0:10:22 never
0:10:23 see
0:10:23 a
0:10:24 bigger
0:10:25 GPT-4.5
0:10:25 model
0:10:26 with all
0:10:26 that
0:10:26 reasoning
0:10:27 ability
0:10:27 and I
0:10:27 would
0:10:28 expect
0:10:28 that
0:10:28 that
0:10:28 would
0:10:29 deliver
0:10:29 more
0:10:30 value
0:10:30 especially
0:10:31 if
0:10:31 you’re
0:10:31 really
0:10:31 going
0:10:31 out
0:10:31 and
0:10:31 trying
0:10:32 to
0:10:32 do
0:10:33 esoteric
0:10:34 stuff
0:10:34 that’s
0:10:35 pushing
0:10:35 the
0:10:35 frontier
0:10:35 of
0:10:36 science
0:10:36 or
0:10:36 what
0:10:36 have
0:10:36 you
0:10:38 but
0:10:38 in
0:10:38 the
0:10:38 meantime
0:10:39 the
0:10:40 current
0:10:40 models
0:10:40 are
0:10:40 really
0:10:40 smart
0:10:40 and
0:10:41 you
0:10:41 can
0:10:41 feed
0:10:41 them
0:10:41 a lot
0:10:41 of
0:10:42 context
0:10:42 that’s
0:10:42 one
0:10:42 of
0:10:42 the
0:10:42 big
0:10:43 things
0:10:43 that
0:10:43 has
0:10:44 improved
0:10:44 so
0:10:44 much
0:10:44 over
0:10:44 the
0:10:45 last
0:10:45 generation
0:10:46 when
0:10:47 GPT-4
0:10:47 came
0:10:47 out
0:10:48 at least
0:10:48 the
0:10:48 version
0:10:48 that
0:10:49 we
0:10:49 had
0:10:49 as
0:10:49 public
0:10:50 users
0:10:50 was
0:10:50 only
0:10:51 8,000
0:10:51 tokens
0:10:51 of
0:10:52 context
0:10:53 which
0:10:53 is
0:10:53 like
0:10:53 15
0:10:54 pages
0:10:54 of
0:10:55 text
0:10:55 so
0:10:56 you
0:10:56 were
0:10:56 limited
0:10:56 you
0:10:57 couldn’t
0:10:57 even
0:10:57 put
0:10:57 in
0:10:57 like
0:10:58 a
0:10:58 couple
0:10:59 papers
0:11:00 you
0:11:00 would
0:11:00 be
0:11:01 overflowing
0:11:01 the
0:11:02 context
0:11:02 and
0:11:02 this
0:11:02 is
0:11:03 where
0:11:03 prompt
0:11:03 engineering
0:11:04 initially
0:11:04 kind of
0:11:05 became
0:11:05 a thing
0:11:05 it was
0:11:05 like
0:11:05 man
0:11:05 I’ve
0:11:06 really
0:11:06 only
0:11:06 got
0:11:07 such
0:11:07 a
0:11:07 little
0:11:08 bit
0:11:08 of
0:11:09 information
0:11:09 that
0:11:09 I
0:11:09 can
0:11:10 provide
0:11:10 I
0:11:10 gotta
0:11:10 be
0:11:10 really
0:11:11 careful
0:11:11 about
0:11:12 what
0:11:13 information
0:11:13 to
0:11:14 provide
0:11:14 lest
0:11:15 I
0:11:15 overflow
0:11:15 the
0:11:15 thing
0:11:15 and
0:11:16 it
0:11:16 just
0:11:16 can’t
0:11:16 handle
0:11:17 it
0:11:18 there
0:11:18 were
0:11:18 also
0:11:19 as
0:11:19 context
0:11:19 windows
0:11:20 got
0:11:20 extended
0:11:20 there
0:11:20 were
0:11:20 also
0:11:22 versions
0:11:22 of
0:11:22 models
0:11:23 where
0:11:23 they
0:11:23 could
0:11:24 nominally
0:11:25 accept
0:11:25 a lot
0:11:25 more
0:11:25 but
0:11:25 they
0:11:25 couldn’t
0:11:26 really
0:11:27 functionally
0:11:28 use
0:11:28 them
0:11:28 they
0:11:29 could
0:11:30 fit
0:11:30 them
0:11:31 at
0:11:31 the
0:11:31 API
0:11:32 call
0:11:32 level
0:11:32 but
0:11:33 the
0:11:33 models
0:11:33 would
0:11:33 lose
0:11:34 recall
0:11:34 they
0:11:35 sort
0:11:35 of
0:11:35 unravel
0:11:35 as
0:11:35 they
0:11:36 got
0:11:36 into
0:11:36 longer
0:11:36 and
0:11:37 longer
0:11:37 context
0:11:38 now
0:11:38 you
0:11:38 have
0:11:38 obviously
0:11:39 much
0:11:39 longer
0:11:39 context
0:11:40 and
0:11:40 the
0:11:40 command
0:11:40 of
0:11:40 it
0:11:41 is
0:11:41 really
0:11:41 good
0:11:42 so
0:11:42 you
0:11:42 can
0:11:48 accept
0:11:49 them
0:11:49 but
0:11:49 it
0:11:49 will
0:11:50 do
0:11:50 pretty
0:11:50 intensive
0:11:51 reasoning
0:11:51 over
0:11:51 them
0:11:52 and
0:11:52 with
0:11:52 really
0:11:52 high
0:11:53 fidelity
0:11:54 to
0:11:54 those
0:11:55 inputs
0:11:55 so
0:11:55 that
0:11:56 skill
0:11:56 I
0:11:56 think
0:11:57 does
0:11:57 kind
0:11:57 of
0:11:58 substitute
0:11:58 for
0:11:58 the
0:11:58 model
0:11:59 knowing
0:11:59 facts
0:12:00 itself
0:12:00 you
0:12:00 could
0:12:00 say
0:12:01 let’s
0:12:01 try
0:12:01 to
0:12:02 train
0:12:02 all
0:12:03 these
0:12:03 facts
0:12:03 into
0:12:03 the
0:12:03 model
0:12:04 we’re
0:12:04 going
0:12:04 to
0:12:04 need
0:12:05 a
0:12:05 trillion
0:12:05 or
0:12:06 five
0:12:06 trillion
0:12:07 parameters
0:12:08 to
0:12:08 fit
0:12:08 all
0:12:08 these
0:12:09 super
0:12:09 long
0:12:10 tail
0:12:10 facts
0:12:11 or
0:12:11 you
0:12:11 could
0:12:11 say
0:12:11 well
0:12:12 a
0:12:13 smaller
0:12:13 thing
0:12:13 that’s
0:12:13 really
0:12:13 good
0:12:14 at
0:12:14 working
0:12:14 over
0:12:15 provided
0:12:16 context
0:12:17 can
0:12:18 if
0:12:19 people
0:12:19 take
0:12:19 the
0:12:19 time
0:12:20 or
0:12:20 you
0:12:20 know
0:12:20 go
0:12:20 to
0:12:20 the
0:12:20 trouble
0:12:21 of
0:12:21 providing
0:12:21 the
0:12:22 necessary
0:12:22 information
0:12:23 I
0:12:23 can
0:12:23 kind
0:12:23 of
0:12:24 access
0:12:24 the
0:12:24 same
0:12:24 facts
0:12:25 that
0:12:25 way
0:12:26 so
0:12:27 you
0:12:27 have
0:12:27 a
0:12:27 kind
0:12:27 of
0:12:28 do
0:12:28 I
0:12:28 want
0:12:29 to
0:12:29 push
0:12:29 on
0:12:29 this
0:12:30 size
0:12:30 and
0:12:30 do
0:12:30 I
0:12:30 want
0:12:30 to
0:12:30 bake
0:12:31 everything
0:12:31 into
0:12:31 the
0:12:31 model
0:12:31 or
0:12:32 do
0:12:32 I
0:12:32 want
0:12:32 to
0:12:32 just
0:12:33 try
0:12:33 to
0:12:33 get
0:12:34 as
0:12:34 much
0:12:34 performance
0:12:43 the
0:12:43 same
0:12:44 way
0:12:44 that
0:12:44 the
0:12:45 models
0:12:45 themselves
0:12:45 are
0:12:45 always
0:12:46 kind
0:12:46 of
0:12:46 in
0:12:46 the
0:12:46 training
0:12:47 process
0:12:48 taking
0:12:48 a little
0:12:49 step
0:12:49 toward
0:12:49 improvement
0:12:50 the
0:12:51 outer
0:12:51 loop
0:12:51 of
0:12:51 the
0:12:52 model
0:12:52 architecture
0:12:52 and
0:12:53 the
0:12:53 nature
0:12:53 of
0:12:53 the
0:12:53 training
0:12:54 runs
0:12:54 and
0:12:54 where
0:12:55 they’re
0:12:55 going
0:12:55 to
0:12:55 invest
0:12:55 their
0:12:55 compute
0:12:56 is
0:12:56 also
0:12:57 kind
0:12:57 of
0:12:57 going
0:12:57 that
0:12:57 direction
0:12:58 and
0:12:58 they’re
0:12:58 always
0:12:58 looking
0:12:58 at
0:12:59 like
0:12:59 well
0:12:59 we
0:12:59 could
0:13:00 scale
0:13:00 up
0:13:00 over
0:13:00 here
0:13:00 maybe
0:13:01 get
0:13:01 this
0:13:01 kind
0:13:01 of
0:13:01 benefit
0:13:01 a
0:13:01 little
0:13:02 bit
0:13:02 or
0:13:02 we
0:13:02 could
0:13:02 do
0:13:03 more
0:13:03 post
0:13:03 training
0:13:03 here
0:13:04 and
0:13:05 get
0:13:05 this
0:13:05 kind
0:13:05 of
0:13:05 benefit
0:13:06 and
0:13:06 it
0:13:06 just
0:13:06 seems
0:13:06 like
0:13:06 we’re
0:13:06 getting
0:13:07 more
0:13:07 benefit
0:13:08 from
0:13:09 the
0:13:09 post
0:13:09 training
0:13:10 and
0:13:10 the
0:13:10 reasoning
0:13:11 paradigm
0:13:11 than
0:13:12 scaling
0:13:12 but I
0:13:12 don’t
0:13:12 think
0:13:12 either
0:13:13 one
0:13:13 is
0:13:16 dead
0:13:17 we
0:13:17 haven’t
0:13:17 seen
0:13:17 yet
0:13:18 what
0:13:19 4.5
0:13:19 with
0:13:19 all
0:13:19 that
0:13:19 post
0:13:20 training
0:13:20 would
0:13:20 look
0:13:21 like
0:13:21 yeah
0:13:22 and
0:13:23 so
0:13:24 one
0:13:24 of
0:13:24 the
0:13:24 things
0:13:24 that
0:13:24 you
0:13:24 mentioned
0:13:25 that
0:13:25 Cal
0:13:27 analysis
0:13:27 missed
0:13:27 was
0:13:28 that
0:13:28 it
0:13:28 underestimated
0:13:32 right
0:13:33 and
0:13:33 so
0:13:34 what
0:13:34 would
0:13:34 it
0:13:34 mean
0:13:35 to
0:13:35 fully
0:13:36 appreciate
0:13:36 that
0:13:37 well
0:13:38 I
0:13:38 mean
0:13:38 a
0:13:38 big
0:13:38 one
0:13:39 from
0:13:39 just
0:13:39 the
0:13:39 last
0:13:39 few
0:13:39 weeks
0:13:40 was
0:13:40 that
0:13:40 we
0:13:40 had
0:13:40 an
0:13:40 IMO
0:13:41 gold
0:13:41 medal
0:13:41 with
0:13:42 pure
0:13:42 reasoning
0:13:43 models
0:13:44 with
0:13:44 no
0:13:45 access
0:13:45 to
0:13:45 tools
0:13:46 from
0:13:46 multiple
0:13:47 companies
0:13:48 and
0:13:49 you know
0:13:49 that
0:13:49 is
0:13:50 night
0:13:50 and
0:13:50 day
0:13:50 compared
0:13:50 to
0:13:51 what
0:13:51 GPT-4
0:13:51 could
0:13:52 do
0:13:52 with
0:13:52 math
0:13:52 right
0:13:53 and
0:13:53 these
0:13:53 things
0:13:53 are
0:13:54 really
0:13:54 weird
0:13:54 like
0:13:55 nothing
0:13:55 I
0:13:55 say
0:13:55 here
0:13:56 should
0:13:56 be
0:13:57 intended
0:13:57 to
0:13:58 suggest
0:13:58 that
0:13:58 people
0:13:59 won’t
0:13:59 be able
0:14:00 to
0:14:00 find
0:14:00 weaknesses
0:14:00 in
0:14:00 the
0:14:01 models
0:14:01 I
0:14:01 still
0:14:01 use
0:14:02 a
0:14:02 tic-tac-toe
0:14:02 puzzle
0:14:03 to this
0:14:04 day
0:14:04 where
0:14:04 I
0:14:05 take
0:14:05 a
0:14:05 picture
0:14:05 of
0:14:05 a
0:14:05 tic-tac-toe
0:14:06 board
0:14:06 where
0:14:07 one of
0:14:08 the
0:14:08 players
0:14:08 has
0:14:08 made
0:14:08 a
0:14:09 wrong
0:14:09 move
0:14:10 that
0:14:10 is
0:14:11 not
0:14:11 optimal
0:14:11 and
0:14:11 thus
0:14:12 allows
0:14:12 the
0:14:12 other
0:14:12 player
0:14:12 to
0:14:13 force
0:14:13 a
0:14:13 win
0:14:14 and
0:14:14 I
0:14:14 ask
0:14:14 the
0:14:15 models
0:14:15 if
0:14:17 somebody
0:14:17 can
0:14:17 force
0:14:17 a
0:14:17 win
0:14:18 from
0:14:18 this
0:14:18 position
0:14:19 only
0:14:19 very
0:14:19 recently
0:14:20 only
0:14:20 the
0:14:20 last
0:14:21 generation
0:14:21 of
0:14:21 models
0:14:21 are
0:14:22 starting
0:14:22 to
0:14:22 get
0:14:22 that
0:14:22 right
0:14:22 some
0:14:23 of
0:14:23 the
0:14:23 time
0:14:23 almost
0:14:24 always
0:14:24 before
0:14:24 they
0:14:25 were
0:14:25 like
0:14:26 tic-tac-toe
0:14:26 is a
0:14:27 solved
0:14:27 game
0:14:27 you know
0:14:28 you can
0:14:28 always
0:14:28 get a
0:14:29 draw
0:14:29 there’s
0:14:30 and
0:14:30 they
0:14:30 would
0:14:31 wrongly
0:14:32 assess
0:14:32 my
0:14:32 board
0:14:33 position
0:14:33 as
0:14:34 player
0:14:34 can
0:14:34 still
0:14:35 get
0:14:35 a
0:14:35 draw
0:14:37 so
0:14:37 there’s
0:14:38 a lot
0:14:38 of
0:14:38 weird
0:14:38 stuff
0:14:38 right
0:14:39 the
0:14:39 jagged
0:14:40 capabilities
0:14:40 frontier
0:14:41 remains
0:14:41 a
0:14:41 real
0:14:42 issue
0:14:42 and
0:14:42 people
0:14:49 approaching
0:14:50 IMO
0:14:50 gold
0:14:51 problems
0:14:51 it
0:14:51 was
0:14:52 still
0:14:52 struggling
0:14:52 on
0:14:52 like
0:14:52 high
0:14:53 school
0:14:53 math
0:14:54 and
0:14:54 since
0:14:54 then
0:14:55 we’ve
0:14:55 seen
0:14:55 this
0:14:55 high
0:14:55 school
0:14:56 math
0:14:56 progression
0:14:57 all the
0:14:57 way up
0:14:57 through
0:14:57 the
0:14:57 IMO
0:14:58 gold
0:14:58 now
0:14:58 we’ve
0:14:58 got
0:14:58 the
0:14:58 frontier
0:14:59 math
0:14:59 benchmark
0:15:00 that
0:15:00 is
0:15:01 I
0:15:01 think
0:15:01 now
0:15:01 like
0:15:01 up
0:15:01 to
0:15:02 25%
0:15:03 it
0:15:03 was
0:15:03 2%
0:15:04 about
0:15:04 a
0:15:04 year
0:15:04 ago
0:15:04 or
0:15:05 even
0:15:05 a
0:15:05 little
0:15:05 less
0:15:05 than
0:15:05 a
0:15:05 year
0:15:05 ago
0:15:06 I
0:15:06 think
0:15:07 and
0:15:08 we
0:15:08 also
0:15:08 just
0:15:08 today
0:15:08 saw
0:15:09 something
0:15:09 where
0:15:10 and
0:15:10 I
0:15:11 haven’t
0:15:11 absorbed
0:15:11 this
0:15:11 one
0:15:12 yet
0:15:13 but
0:15:13 somebody
0:15:13 just
0:15:13 came
0:15:13 out
0:15:13 and
0:15:14 said
0:15:14 that
0:15:14 they
0:15:14 had
0:15:14 solved
0:15:15 a
0:15:16 canonical
0:15:17 super
0:15:17 challenging
0:15:17 problem
0:15:17 that
0:15:18 no less
0:15:18 than
0:15:18 Terrence
0:15:18 Tao
0:15:19 had
0:15:19 put
0:15:19 out
0:15:20 and
0:15:20 it
0:15:20 was
0:15:21 like
0:15:21 this
0:15:21 thing
0:15:22 happened
0:15:22 in
0:15:23 days
0:15:23 or
0:15:23 weeks
0:15:23 of
0:15:24 the
0:15:24 model
0:15:24 running
0:15:24 versus
0:15:25 it
0:15:39 that
0:15:39 that
0:15:40 really
0:15:40 hard
0:15:42 jumping
0:15:42 capabilities
0:15:42 to
0:15:43 miss
0:15:43 I
0:15:43 also
0:15:43 think
0:15:44 a lot
0:15:44 about
0:15:44 the
0:15:45 Google
0:15:46 AI
0:15:47 co-scientist
0:15:47 which
0:15:48 we did
0:15:48 an episode
0:15:48 with
0:15:49 you can
0:15:50 check out
0:15:50 the full
0:15:50 story
0:15:50 on that
0:15:50 if you
0:15:51 want
0:15:51 to
0:15:51 but
0:15:52 they
0:15:53 basically
0:15:54 just
0:15:55 broke
0:15:55 down
0:15:55 the
0:15:55 scientific
0:15:56 method
0:15:56 into
0:15:57 a
0:15:57 schematic
0:15:58 and this
0:15:59 is a lot
0:15:59 of what
0:15:59 happens
0:16:00 when
0:16:00 people
0:16:00 there’s
0:16:00 one
0:16:01 thing
0:16:01 to
0:16:01 say
0:16:01 the
0:16:01 model
0:16:02 will
0:16:03 respond
0:16:03 with
0:16:04 thinking
0:16:04 and it
0:16:04 will
0:16:04 go
0:16:04 through
0:16:04 reasoning
0:16:05 process
0:16:05 and
0:16:06 the more
0:16:06 tokens
0:16:09 that’s
0:16:09 true
0:16:10 then
0:16:10 you
0:16:10 can
0:16:10 also
0:16:10 build
0:16:11 this
0:16:11 scaffolding
0:16:11 on top
0:16:11 of
0:16:12 that
0:16:12 and
0:16:12 say
0:16:12 okay
0:16:13 well
0:16:13 let me
0:16:14 take
0:16:14 something
0:16:14 as
0:16:15 broad
0:16:15 and
0:16:16 you know
0:16:16 aspirational
0:16:17 as
0:16:17 the
0:16:17 scientific
0:16:18 method
0:16:19 and let
0:16:19 me
0:16:19 break
0:16:19 that
0:16:19 down
0:16:20 into
0:16:20 parts
0:16:21 okay
0:16:21 there’s
0:16:22 hypothesis
0:16:22 generation
0:16:23 then there’s
0:16:23 hypothesis
0:16:24 evaluation
0:16:25 then there’s
0:16:25 you know
0:16:25 experiment
0:16:26 design
0:16:26 there’s
0:16:26 literature
0:16:27 review
0:16:27 there’s
0:16:27 all
0:16:27 these
0:16:27 parts
0:16:27 to
0:16:28 the
0:16:28 scientific
0:16:28 method
0:16:29 what
0:16:29 the
0:16:29 team
0:16:29 at
0:16:29 Google
0:16:30 did
0:16:30 is
0:16:30 created
0:16:30 a
0:16:31 pretty
0:16:31 elaborate
0:16:32 schematic
0:16:33 that
0:16:33 represented
0:16:34 their
0:16:34 best
0:16:34 breakdown
0:16:35 of
0:16:35 the
0:16:35 scientific
0:16:35 method
0:16:36 optimized
0:16:37 prompts
0:16:37 for
0:16:37 each
0:16:38 of
0:16:38 those
0:16:38 steps
0:16:39 and
0:16:39 then
0:16:41 gave
0:16:41 this
0:16:42 resulting
0:16:42 system
0:16:42 which
0:16:43 is
0:16:43 scaling
0:16:44 inference
0:16:44 now
0:16:44 kind of
0:16:44 two
0:16:44 ways
0:16:45 it’s
0:16:45 both
0:16:45 the
0:16:45 chain
0:16:45 of
0:16:45 thought
0:16:45 but
0:16:46 it’s
0:16:46 also
0:16:46 all
0:16:47 these
0:16:47 different
0:16:47 angles
0:16:47 of
0:16:48 attack
0:16:48 structured
0:16:49 by
0:16:49 the
0:16:49 team
0:16:50 and
0:16:50 they
0:16:50 gave
0:16:51 it
0:16:51 legitimately
0:16:52 unsolved
0:16:53 problems
0:16:53 in
0:16:53 science
0:16:54 and
0:16:55 in
0:16:55 one
0:16:56 particularly
0:16:57 famous
0:16:57 kind of
0:16:57 notorious
0:16:58 case
0:16:59 it
0:16:59 came up
0:17:00 with a
0:17:01 hypothesis
0:17:02 which
0:17:02 it
0:17:02 wasn’t
0:17:02 able
0:17:02 to
0:17:02 verify
0:17:03 because
0:17:03 it
0:17:03 doesn’t
0:17:04 have
0:17:04 direct
0:17:05 access
0:17:05 to
0:17:05 actually
0:17:05 run
0:17:05 the
0:17:06 experiments
0:17:06 in
0:17:06 the
0:17:06 lab
0:17:06 but
0:17:07 it
0:17:07 came up
0:17:07 with
0:17:07 a
0:17:07 hypothesis
0:17:08 to
0:17:08 some
0:17:08 open
0:17:09 problem
0:17:10 in
0:17:10 virology
0:17:11 that
0:17:12 had
0:17:13 stumped
0:17:13 scientists
0:17:14 for years
0:17:15 and it
0:17:15 just so
0:17:15 happened
0:17:15 that
0:17:15 they
0:17:15 had
0:17:16 also
0:17:16 recently
0:17:16 figured
0:17:17 out
0:17:17 the
0:17:17 answer
0:17:17 but
0:17:18 not
0:17:18 yet
0:17:18 published
0:17:18 their
0:17:19 results
0:17:19 and
0:17:19 so
0:17:20 there
0:17:20 was
0:17:20 this
0:17:20 confluence
0:17:20 where
0:17:21 the
0:17:21 scientists
0:17:21 had
0:17:22 experimentally
0:17:23 verified
0:17:23 and
0:17:24 Gemini
0:17:25 in the
0:17:25 form of
0:17:25 this
0:17:25 AI
0:17:26 co-scientist
0:17:27 came up
0:17:27 with
0:17:27 exactly
0:17:28 the
0:17:28 right
0:17:29 answer
0:17:29 and
0:17:29 these
0:17:30 are
0:17:30 things
0:17:30 that
0:17:30 literally
0:17:31 nobody
0:17:31 knew
0:17:32 before
0:17:33 and
0:17:34 GPT-4
0:17:34 just
0:17:34 wasn’t
0:17:34 doing
0:17:35 that
0:17:35 these
0:17:35 are
0:17:37 qualitatively
0:17:37 new
0:17:38 capabilities
0:17:39 that
0:17:39 thing
0:17:40 I think
0:17:40 ran
0:17:40 for
0:17:40 days
0:17:41 you know
0:17:41 it
0:17:41 probably
0:17:42 cost
0:17:43 hundreds
0:17:43 of
0:17:43 dollars
0:17:43 maybe
0:17:43 into
0:17:43 the
0:17:44 thousands
0:17:44 of
0:17:44 dollars
0:17:45 to
0:17:45 run
0:17:45 the
0:17:46 inference
0:17:47 you know
0:17:48 that’s
0:17:48 not
0:17:48 nothing
0:17:49 but
0:17:49 it’s
0:17:49 also
0:17:50 like
0:17:50 very
0:17:51 much
0:17:51 cheaper
0:17:51 than
0:17:52 years
0:17:52 of
0:17:52 grad
0:17:52 students
0:17:53 and
0:17:54 if
0:17:54 you
0:17:54 can
0:17:55 get
0:17:55 to
0:17:55 those
0:17:56 caliber
0:17:56 of
0:17:56 problems
0:17:56 and
0:17:57 actually
0:17:57 get
0:17:57 good
0:17:57 solutions
0:17:57 to
0:17:58 them
0:17:58 like
0:17:59 you
0:17:59 know
0:17:59 what
0:17:59 would
0:17:59 you
0:17:59 be
0:17:59 willing
0:17:59 to
0:18:00 pay
0:18:00 right
0:18:00 for
0:18:00 that
0:18:01 kind
0:18:01 of
0:18:01 thing
0:18:02 so
0:18:03 yeah
0:18:04 I don’t
0:18:04 know
0:18:04 that’s
0:18:04 probably
0:18:04 not a
0:18:04 full
0:18:05 appreciation
0:18:05 we could
0:18:05 go on
0:18:06 for a
0:18:06 long
0:18:06 time
0:18:06 but
0:18:07 I
0:18:07 would
0:18:08 say
0:18:08 in
0:18:09 summary
0:18:10 GPT-4
0:18:10 was
0:18:11 not
0:18:12 able
0:18:12 to
0:18:13 push
0:18:13 the
0:18:14 actual
0:18:14 frontier
0:18:14 of
0:18:14 human
0:18:15 knowledge
0:18:15 to my
0:18:16 knowledge
0:18:16 I don’t
0:18:16 know
0:18:16 that
0:18:16 ever
0:18:17 discovered
0:18:17 anything
0:18:18 new
0:18:19 it’s
0:18:19 still
0:18:19 not
0:18:19 easy
0:18:19 to
0:18:20 get
0:18:20 that
0:18:20 kind
0:18:21 of
0:18:21 output
0:18:22 from
0:18:23 a
0:18:23 GPT-5
0:18:24 or a
0:18:24 Gemini
0:18:24 2.5
0:18:25 or
0:18:25 a
0:18:25 Claude
0:18:26 Opus
0:18:26 4
0:18:26 or
0:18:26 whatever
0:18:27 but
0:18:28 it’s
0:18:28 starting
0:18:28 to
0:18:28 happen
0:18:29 sometimes
0:18:29 and
0:18:30 that
0:18:30 in
0:18:30 and of
0:18:30 itself
0:18:30 is
0:18:31 a
0:18:31 huge
0:18:31 deal
0:18:32 how do
0:18:32 we
0:18:33 explain
0:18:33 the
0:18:35 bearishness
0:18:35 or the
0:18:35 kind
0:18:35 of
0:18:36 vibe
0:18:36 shift
0:18:36 around
0:18:37 GPT-5
0:18:37 then
0:18:38 one
0:18:38 potential
0:18:40 contributor
0:18:40 is this
0:18:41 idea
0:18:41 that
0:18:41 if
0:18:42 the
0:18:43 improvements
0:18:43 are
0:18:43 at
0:18:43 the
0:18:44 frontier
0:18:44 not
0:18:44 everyone
0:18:45 is
0:18:45 working
0:18:45 with
0:18:46 advanced
0:18:47 math
0:18:47 and
0:18:47 physics
0:18:48 and
0:18:48 day-to-day
0:18:48 so
0:18:48 maybe
0:18:48 they
0:18:48 don’t
0:18:49 see
0:18:49 the
0:18:50 benefits
0:18:50 in
0:18:50 their
0:18:51 day-to-day
0:18:51 lives
0:18:51 in
0:18:51 the
0:18:51 same
0:18:51 way
0:18:52 that
0:18:53 the
0:18:54 jumps
0:18:54 in
0:18:54 chat
0:18:54 GPT
0:18:54 were
0:18:55 obvious
0:18:56 and
0:18:56 shaped
0:18:57 day-to-day
0:18:58 I think
0:18:58 a decent
0:18:59 amount
0:18:59 of it
0:19:00 was
0:19:01 that
0:19:01 they
0:19:02 kind
0:19:02 of
0:19:02 fucked
0:19:02 up
0:19:02 the
0:19:02 launch
0:19:03 simply
0:19:03 put
0:19:04 they
0:19:05 were
0:19:05 tweeting
0:19:06 Death Star
0:19:06 images
0:19:08 came back
0:19:08 and said
0:19:08 no
0:19:09 you’re
0:19:09 the
0:19:09 Death Star
0:19:09 I’m
0:19:10 not
0:19:10 the
0:19:10 Death Star
0:19:11 but
0:19:11 I
0:19:11 think
0:19:11 people
0:19:12 thought
0:19:12 that
0:19:12 the
0:19:12 Death Star
0:19:12 was
0:19:12 supposed
0:19:13 to
0:19:13 be
0:19:13 the
0:19:13 model
0:19:14 that
0:19:14 was
0:19:14 generally
0:19:15 the
0:19:16 expectations
0:19:16 were
0:19:16 set
0:19:17 extremely
0:19:17 high
0:19:19 the
0:19:20 actual
0:19:20 launch
0:19:21 itself
0:19:22 was
0:19:22 just
0:19:23 technically
0:19:23 broken
0:19:23 so
0:19:24 a lot
0:19:24 of
0:19:24 people’s
0:19:24 first
0:19:25 experiences
0:19:26 of
0:19:26 GPT-5
0:19:27 they’ve
0:19:27 got this
0:19:28 model
0:19:28 router
0:19:28 concept
0:19:29 now
0:19:29 where
0:19:31 I think
0:19:32 another way
0:19:32 to understand
0:19:32 what they’re
0:19:33 doing here
0:19:33 is
0:19:33 they’re
0:19:34 trying
0:19:34 to
0:19:35 own
0:19:35 the
0:19:36 consumer
0:19:36 use
0:19:37 case
0:19:38 and
0:19:38 to
0:19:39 own
0:19:39 that
0:19:39 they
0:19:40 need
0:19:40 to
0:19:40 simplify
0:19:41 the
0:19:41 product
0:19:42 experience
0:19:43 relative
0:19:43 to
0:19:43 what
0:19:43 we
0:19:43 had
0:19:43 in
0:19:43 the
0:19:44 past
0:19:44 which
0:19:44 was
0:19:44 like
0:19:44 okay
0:19:44 you
0:19:44 got
0:19:45 GPT-4
0:19:46 4-0
0:19:46 and 4-0
0:19:46 mini
0:19:47 and
0:19:47 4-3
0:19:47 and
0:19:48 4-4
0:19:48 mini
0:19:49 and
0:19:49 other
0:19:50 things
0:19:50 4-5
0:19:51 was in
0:19:51 there
0:19:51 at one
0:19:51 point
0:19:52 you got
0:19:52 all these
0:19:52 different
0:19:53 models
0:19:53 which one
0:19:53 should I
0:19:53 use
0:19:54 for which
0:19:54 it’s
0:19:54 like
0:19:54 very
0:19:55 confusing
0:19:55 to most
0:19:55 people
0:19:56 who
0:19:56 aren’t
0:19:56 obsessed
0:19:56 with
0:19:57 this
0:19:57 and
0:19:58 so
0:19:58 one
0:19:58 of
0:19:58 the
0:19:58 big
0:19:58 things
0:19:59 they
0:19:59 wanted
0:19:59 to do
0:19:59 was
0:20:00 just
0:20:01 shrink
0:20:01 that
0:20:01 down
0:20:01 to
0:20:02 just
0:20:03 ask
0:20:03 your
0:20:03 question
0:20:04 and
0:20:04 you’ll
0:20:05 get
0:20:05 a
0:20:05 good
0:20:05 answer
0:20:05 and
0:20:06 we’ll
0:20:06 take
0:20:06 that
0:20:06 complexity
0:20:07 on
0:20:07 our
0:20:07 side
0:20:07 as
0:20:08 the
0:20:08 product
0:20:09 owners
0:20:10 to do
0:20:10 that
0:20:11 interestingly
0:20:12 and I
0:20:12 don’t
0:20:12 have a
0:20:12 great
0:20:12 account
0:20:13 of
0:20:13 this
0:20:14 but
0:20:15 one
0:20:15 thing
0:20:15 you
0:20:15 might
0:20:15 want
0:20:16 to do
0:20:16 is
0:20:16 kind
0:20:16 of
0:20:17 merge
0:20:17 the
0:20:18 models
0:20:18 and
0:20:18 figure
0:20:18 out
0:20:19 just
0:20:19 have
0:20:20 the
0:20:20 model
0:20:20 itself
0:20:20 decide
0:20:21 how
0:20:21 much
0:20:21 to
0:20:21 think
0:20:22 or
0:20:22 maybe
0:20:22 even
0:20:23 have
0:20:23 the
0:20:23 model
0:20:23 itself
0:20:24 decide
0:20:24 how
0:20:24 many
0:20:24 of
0:20:24 its
0:20:25 experts
0:20:25 if it’s
0:20:25 a mixture
0:20:25 of
0:20:26 experts
0:20:26 architecture
0:20:27 it needs
0:20:27 to
0:20:28 use
0:20:28 or
0:20:28 maybe
0:20:29 there’s
0:20:29 been
0:20:30 a bunch
0:20:30 of
0:20:30 different
0:20:31 research
0:20:31 projects
0:20:31 on
0:20:32 skipping
0:20:32 layers
0:20:32 of
0:20:32 the
0:20:33 model
0:20:33 if
0:20:33 the
0:20:34 task
0:20:34 is
0:20:34 easy
0:20:34 enough
0:20:34 you
0:20:34 could
0:20:35 skip
0:20:35 a
0:20:35 bunch
0:20:35 of
0:20:35 layers
0:20:36 so
0:20:36 you
0:20:36 might
0:20:37 have
0:20:37 hoped
0:20:37 that
0:20:38 could
0:20:39 genuinely
0:20:39 on
0:20:39 the
0:20:39 back
0:20:40 end
0:20:41 merge
0:20:42 all
0:20:42 these
0:20:42 different
0:20:42 models
0:20:42 into
0:20:43 one
0:20:43 model
0:20:43 that
0:20:44 would
0:20:44 dynamically
0:20:45 use
0:20:45 the
0:20:45 right
0:20:46 amount
0:20:46 of
0:20:46 compute
0:20:47 for
0:20:47 the
0:20:48 level
0:20:49 of
0:20:49 challenge
0:20:49 that
0:20:50 a
0:20:50 given
0:20:50 user
0:20:51 query
0:20:51 presented
0:20:52 it
0:20:52 seems
0:20:52 like
0:20:52 they
0:20:53 found
0:20:53 that
0:20:54 harder
0:20:54 to
0:20:54 do
0:20:55 than
0:20:55 they
0:20:56 expected
0:20:56 and
0:20:56 so
0:20:57 the
0:20:57 solution
0:20:57 that
0:20:57 they
0:20:57 came
0:20:57 up
0:20:58 with
0:20:58 instead
0:20:58 was
0:20:58 to
0:20:58 have
0:20:59 a
0:20:59 router
0:21:00 where
0:21:00 the
0:21:01 router’s
0:21:01 job
0:21:01 is
0:21:02 to
0:21:02 pick
0:21:02 is
0:21:02 this
0:21:02 an
0:21:03 easy
0:21:03 query
0:21:15 with
0:21:15 and
0:21:15 without
0:21:16 thinking
0:21:18 the
0:21:19 problem
0:21:19 at
0:21:19 launch
0:21:19 was
0:21:19 that
0:21:20 that
0:21:20 router
0:21:20 was
0:21:20 broken
0:21:21 so
0:21:21 all
0:21:22 of
0:21:22 the
0:21:22 queries
0:21:22 were
0:21:23 going
0:21:23 to
0:21:23 the
0:21:24 dumb
0:21:24 model
0:21:24 and
0:21:24 so
0:21:25 a lot
0:21:25 of
0:21:25 people
0:21:25 literally
0:21:25 just
0:21:26 got
0:21:27 bad
0:21:28 outputs
0:21:28 which
0:21:28 were
0:21:29 worse
0:21:29 than
0:21:30 03
0:21:30 because
0:21:31 they
0:21:31 were
0:21:31 getting
0:21:31 non
0:21:31 thinking
0:21:32 responses
0:21:33 and
0:21:33 so
0:21:33 the
0:21:33 initial
0:21:34 reaction
0:21:34 of
0:21:34 like
0:21:34 okay
0:21:34 this
0:21:35 is
0:21:35 dumb
0:21:35 and
0:21:35 that
0:21:35 sort
0:21:35 of
0:21:35 you
0:21:36 know
0:21:37 traveled
0:21:37 really
0:21:38 fast
0:21:38 I
0:21:38 think
0:21:38 that
0:21:39 kind
0:21:39 of
0:21:39 set
0:21:40 the
0:21:40 tone
0:21:41 my
0:21:41 sense
0:21:41 now
0:21:42 is
0:21:42 that
0:21:42 as
0:21:42 the
0:21:42 dust
0:21:43 has
0:21:43 settled
0:21:44 most
0:21:44 people
0:21:44 do
0:21:45 think
0:21:45 that
0:21:45 it
0:21:46 is
0:21:46 the
0:21:46 best
0:21:46 model
0:21:47 available
0:21:48 and
0:21:49 you know
0:21:50 things like
0:21:50 the
0:21:51 meter
0:21:52 the infamous
0:21:52 meter
0:21:53 task
0:21:53 length
0:21:53 chart
0:21:54 it
0:21:55 is
0:21:55 the
0:21:56 best
0:21:56 you know
0:21:56 we’re
0:21:56 now
0:21:57 over
0:21:57 two
0:21:57 hours
0:21:58 and
0:21:59 it
0:21:59 is
0:21:59 still
0:21:59 above
0:21:59 the
0:22:00 trend
0:22:00 line
0:22:01 so
0:22:01 if
0:22:01 you
0:22:02 just
0:22:02 said
0:22:02 you know
0:22:03 do I
0:22:03 believe
0:22:03 in
0:22:03 straight
0:22:04 lines
0:22:04 on
0:22:12 power
0:22:14 logarithmic
0:22:15 scale
0:22:15 graphs
0:22:17 it
0:22:17 shouldn’t
0:22:17 really
0:22:18 change
0:22:18 your
0:22:18 mind
0:22:18 too
0:22:19 much
0:22:19 it’s
0:22:19 still
0:22:19 above
0:22:20 the
0:22:20 trend
0:22:20 line
0:22:21 I
0:22:21 talked
0:22:21 to
0:22:21 Zvi
0:22:21 about
0:22:22 this
0:22:24 legendary
0:22:25 infovore
0:22:25 and
0:22:26 AI
0:22:26 industry
0:22:27 analyst
0:22:27 on a
0:22:27 recent
0:22:28 podcast
0:22:28 too
0:22:28 and
0:22:28 kind
0:22:28 of
0:22:28 asked
0:22:28 him
0:22:29 the
0:22:29 same
0:22:29 question
0:22:29 like
0:22:30 why
0:22:30 do
0:22:30 you
0:22:30 think
0:22:31 the
0:22:31 you know
0:22:32 even
0:22:32 some
0:22:32 of
0:22:33 the
0:22:33 most
0:22:35 plugged
0:22:35 in
0:22:36 you know
0:22:36 sharp
0:22:37 minds
0:22:37 in the
0:22:37 space
0:22:38 have
0:22:38 seemingly
0:22:39 pushed
0:22:39 timelines
0:22:39 out
0:22:40 a bit
0:22:40 as a
0:22:40 result
0:22:40 of
0:22:41 this
0:22:41 and
0:22:41 his
0:22:41 answer
0:22:42 was
0:22:42 basically
0:22:42 just
0:22:43 it
0:22:43 resolved
0:22:43 some
0:22:44 amount
0:22:44 of
0:22:44 uncertainty
0:22:45 you know
0:22:45 you had
0:22:46 open
0:22:47 question
0:22:47 of
0:22:47 maybe
0:22:47 they
0:22:47 do
0:22:47 have
0:22:48 another
0:22:48 breakthrough
0:22:48 you know
0:22:48 maybe
0:22:49 it really
0:22:49 is
0:22:49 the
0:22:49 death
0:22:49 star
0:22:51 you know
0:22:51 if they
0:22:52 surprise us
0:22:52 on the
0:22:52 upside
0:22:53 then all
0:22:53 these
0:22:53 short
0:22:54 timelines
0:22:54 you know
0:22:54 we could
0:22:54 have
0:22:55 expected
0:22:55 a
0:22:58 I guess
0:22:58 one way
0:22:58 to think
0:22:59 about it
0:22:59 is like
0:22:59 the
0:23:00 distribution
0:23:01 was sort
0:23:01 of
0:23:02 broad
0:23:02 in terms
0:23:03 of
0:23:03 timelines
0:23:04 and if
0:23:04 they
0:23:05 had
0:23:05 surprised
0:23:05 on the
0:23:05 upside
0:23:06 it
0:23:06 might
0:23:06 have
0:23:07 narrowed
0:23:08 and
0:23:08 narrowed
0:23:08 in
0:23:09 toward
0:23:09 the
0:23:09 front
0:23:09 end
0:23:09 of
0:23:09 the
0:23:10 distribution
0:23:10 and
0:23:10 if
0:23:11 they
0:23:11 surprised
0:23:11 on
0:23:11 the
0:23:12 downside
0:23:12 or
0:23:12 even
0:23:12 just
0:23:12 were
0:23:13 purely
0:23:13 on
0:23:13 trend
0:23:14 then
0:23:14 you
0:23:14 would
0:23:14 take
0:23:14 some
0:23:14 of
0:23:14 your
0:23:15 distribution
0:23:15 from
0:23:15 the
0:23:15 very
0:23:16 short
0:23:16 end
0:23:16 of
0:23:16 the
0:23:16 timelines
0:23:17 and
0:23:17 kind
0:23:17 of
0:23:17 push
0:23:17 them
0:23:17 back
0:23:18 toward
0:23:18 the
0:23:18 middle
0:23:18 or
0:23:18 the
0:23:19 end
0:23:19 and
0:23:19 so
0:23:20 his
0:23:20 answer
0:23:20 was
0:23:22 AI
0:23:23 2027
0:23:23 seems
0:23:23 less
0:23:24 likely
0:23:25 but
0:23:25 AI
0:23:26 2030
0:23:27 seems
0:23:27 basically
0:23:28 no less
0:23:28 likely
0:23:28 maybe
0:23:28 even
0:23:28 a little
0:23:28 more
0:23:29 likely
0:23:29 because
0:23:29 some
0:23:29 of
0:23:30 the
0:23:30 probability
0:23:31 mass
0:23:31 from
0:23:31 the
0:23:31 early
0:23:32 years
0:23:32 is
0:23:32 now
0:23:33 sitting
0:23:33 there
0:23:34 so
0:23:34 it’s
0:23:35 not
0:23:35 that
0:23:36 I
0:23:36 don’t
0:23:36 think
0:23:36 people
0:23:37 are
0:23:37 moving
0:23:39 the
0:23:39 whole
0:23:40 distribution
0:23:40 out
0:23:40 super
0:23:40 much
0:23:40 I
0:23:41 think
0:23:41 there
0:23:41 may
0:23:41 be
0:23:41 more
0:23:41 just
0:23:41 kind
0:23:42 of
0:23:42 shrinking
0:23:43 the
0:23:44 you know
0:23:44 it’s
0:23:44 getting
0:23:44 a little
0:23:45 tighter
0:23:45 because
0:23:45 it’s
0:23:46 maybe
0:23:46 not
0:23:46 happening
0:23:46 quite
0:23:47 as
0:23:47 soon
0:23:47 as
0:23:47 it
0:23:48 seemed
0:23:48 like
0:23:48 it
0:23:49 might
0:23:49 have
0:23:49 been
0:23:50 but
0:23:50 I
0:23:50 don’t
0:23:50 think
0:23:50 too
0:23:50 many
0:23:51 people
0:23:51 at
0:23:51 least
0:23:52 are
0:23:53 really
0:23:53 plugged
0:23:53 in
0:23:53 on
0:23:53 this
0:23:54 are
0:23:54 pushing
0:23:54 out
0:23:55 too
0:23:55 much
0:23:55 past
0:23:55 2030
0:24:04 Dario
0:24:04 says
0:24:05 2027
0:24:05 Demis
0:24:06 says
0:24:06 2030
0:24:06 I’ll
0:24:06 take
0:24:07 that
0:24:07 as
0:24:07 my
0:24:07 range
0:24:08 so
0:24:08 coming
0:24:08 into
0:24:09 GPT-5
0:24:10 I
0:24:10 was
0:24:11 kind
0:24:11 of
0:24:11 in
0:24:11 that
0:24:12 space
0:24:12 and
0:24:12 now
0:24:12 I’d
0:24:12 say
0:24:13 well
0:24:13 I
0:24:13 don’t
0:24:13 know
0:24:14 Dario’s
0:24:14 got
0:24:15 what
0:24:15 cards
0:24:15 does
0:24:15 he
0:24:16 have
0:24:16 up
0:24:16 his
0:24:16 sleeve
0:24:16 you know
0:24:17 they
0:24:17 just
0:24:17 put
0:24:17 out
0:24:18 4.1
0:24:19 opus
0:24:19 and
0:24:19 in
0:24:19 that
0:24:20 blog
0:24:20 post
0:24:20 they
0:24:20 said
0:24:21 we
0:24:21 will
0:24:21 be
0:24:21 releasing
0:24:22 more
0:24:23 powerful
0:24:24 updates
0:24:24 to
0:24:24 our
0:24:24 models
0:24:25 in
0:24:25 the
0:24:25 coming
0:24:25 weeks
0:24:26 so
0:24:26 they’re
0:24:26 due
0:24:26 for
0:24:27 something
0:24:27 pretty
0:24:27 soon
0:24:28 you know
0:24:28 maybe
0:24:28 they’ll
0:24:28 be
0:24:29 the
0:24:29 ones
0:24:29 to
0:24:29 surprise
0:24:30 on
0:24:30 the
0:24:30 upside
0:24:30 this
0:24:30 time
0:24:30 or
0:24:31 maybe
0:24:31 Google
0:24:31 will
0:24:31 be
0:24:32 I
0:24:33 wouldn’t
0:24:33 say
0:24:33 2027
0:24:33 is
0:24:34 out of
0:24:34 the
0:24:34 question
0:24:35 but
0:24:36 yeah
0:24:36 I
0:24:36 would
0:24:36 say
0:24:37 2030
0:24:37 still
0:24:38 looks
0:24:38 just
0:24:38 as
0:24:39 likely
0:24:39 as
0:24:39 before
0:24:40 and
0:24:40 again
0:24:41 from my
0:24:41 standpoint
0:24:41 it’s
0:24:41 like
0:24:43 that’s
0:24:43 still
0:24:44 really
0:24:44 soon
0:24:45 you know
0:24:45 so
0:24:45 if
0:24:45 we’re
0:24:46 on
0:24:46 track
0:24:46 whether
0:24:46 it’s
0:24:46 28
0:24:47 29
0:24:47 30
0:24:48 I
0:24:48 don’t
0:24:49 really
0:24:49 care
0:24:49 I
0:24:50 try
0:24:50 to
0:24:51 frame
0:24:51 my
0:24:52 own
0:24:52 work
0:24:52 so
0:24:52 that
0:24:52 I’m
0:24:52 kind
0:24:53 of
0:24:53 preparing
0:24:54 myself
0:24:54 and
0:24:54 helping
0:24:54 other
0:24:54 people
0:24:55 prepare
0:24:55 for
0:24:56 what
0:24:56 might
0:24:56 be
0:24:56 the
0:24:57 most
0:24:57 extreme
0:24:57 scenarios
0:24:58 and
0:24:58 kind
0:24:59 of
0:24:59 you know
0:25:00 one of
0:25:00 these
0:25:00 things
0:25:00 where
0:25:00 if
0:25:00 we
0:25:00 aim
0:25:01 high
0:25:01 and
0:25:01 we
0:25:01 miss
0:25:01 a
0:25:01 little
0:25:01 bit
0:25:01 and
0:25:02 we
0:25:02 have
0:25:02 a
0:25:02 little
0:25:02 more
0:25:02 time
0:25:03 great
0:25:03 I’m
0:25:03 sure
0:25:03 we’ll
0:25:03 have
0:25:03 plenty
0:25:04 of
0:25:04 things
0:25:04 to
0:25:04 do
0:25:05 to
0:25:05 use
0:25:05 that
0:25:06 extra
0:25:06 time
0:25:06 to
0:25:07 be
0:25:07 ready
0:25:07 for
0:25:08 whatever
0:25:08 powerful
0:25:08 AI
0:25:09 does
0:25:09 come
0:25:10 online
0:25:13 but
0:25:13 yeah
0:25:13 I guess
0:25:14 I don’t
0:25:16 my worldview
0:25:16 hasn’t
0:25:16 changed
0:25:16 all that
0:25:16 much
0:25:17 as a
0:25:17 result
0:25:17 of
0:25:17 these
0:25:18 summers
0:25:19 developments
0:25:20 anecdotally
0:25:21 I don’t
0:25:21 hear as
0:25:21 much
0:25:21 about
0:25:22 AI
0:25:22 2027
0:25:22 or
0:25:23 situational
0:25:23 awareness
0:25:24 to the
0:25:25 same
0:25:25 degree
0:25:26 I do
0:25:26 talk to
0:25:26 some
0:25:26 people
0:25:26 who’ve
0:25:26 just
0:25:27 moved
0:25:27 it
0:25:27 a
0:25:28 few
0:25:28 years
0:25:28 back
0:25:29 to
0:25:29 your
0:25:30 point
0:25:31 but
0:25:32 yeah
0:25:33 Darkash had
0:25:35 his
0:25:35 believes
0:25:35 in
0:25:36 it
0:25:36 but
0:25:37 maybe
0:25:37 because
0:25:38 this
0:25:38 gap
0:25:39 in
0:25:39 continual
0:25:39 learning
0:25:40 or
0:25:41 something
0:25:41 to
0:25:41 the
0:25:42 effect
0:25:42 that
0:25:43 maybe
0:25:43 it’s
0:25:44 going
0:25:44 to
0:25:44 be
0:25:44 a bit
0:25:44 slower
0:25:45 to
0:25:46 diffuse
0:25:48 and
0:25:48 meters
0:25:49 paper
0:25:49 as you
0:25:49 mentioned
0:25:50 showed
0:25:50 that
0:25:51 engineers
0:25:51 are less
0:25:52 productive
0:25:52 and so
0:25:52 maybe
0:25:53 there’s
0:25:53 less
0:25:54 of a
0:25:55 concern
0:25:56 around
0:25:57 people
0:25:57 being
0:25:57 replaced
0:25:58 to
0:25:58 the
0:25:58 next
0:25:59 few
0:25:59 years
0:26:00 in
0:26:01 mass
0:26:01 I think
0:26:02 when we
0:26:02 spoke
0:26:02 maybe a
0:26:02 year
0:26:03 go
0:26:03 about
0:26:03 this
0:26:03 or
0:26:04 I think
0:26:04 you said
0:26:04 something
0:26:04 like
0:26:05 50%
0:26:05 of
0:26:06 jobs
0:26:08 I’m
0:26:08 curious
0:26:08 if
0:26:08 that’s
0:26:08 still
0:26:09 your
0:26:10 litmus
0:26:10 test
0:26:11 or
0:26:11 how
0:26:11 do
0:26:11 you
0:26:11 think
0:26:11 about
0:26:12 it
0:26:12 well
0:26:13 for
0:26:13 one
0:26:13 thing
0:26:13 I think
0:26:13 that
0:26:14 meter
0:26:15 paper
0:26:16 is
0:26:16 worth
0:26:17 unpacking
0:26:17 a little
0:26:17 bit
0:26:18 more
0:26:18 because
0:26:19 this
0:26:19 was
0:26:19 one
0:26:19 of
0:26:20 those
0:26:20 things
0:26:20 that
0:26:20 was
0:26:20 and
0:26:21 I
0:26:22 I’m
0:26:22 a
0:26:22 big
0:26:22 fan
0:26:22 of
0:26:22 meter
0:26:22 and
0:26:23 I
0:26:23 have
0:26:23 no
0:26:25 shade
0:26:26 on
0:26:26 them
0:26:26 because
0:26:26 I
0:26:26 do
0:26:26 think
0:26:27 do
0:26:28 science
0:26:33 result
0:26:34 and
0:26:34 everything
0:26:34 you
0:26:34 put
0:26:34 out
0:26:35 conform
0:26:35 to
0:26:35 a
0:26:35 narrative
0:26:36 but
0:26:36 I
0:26:36 do
0:26:36 think
0:26:37 it
0:26:37 was
0:26:38 a
0:26:38 little
0:26:39 bit
0:26:43 too
0:26:44 easy
0:26:44 for
0:26:44 people
0:26:44 who
0:26:44 wanted
0:26:45 to
0:26:45 say
0:26:46 that
0:26:46 oh
0:26:46 this
0:26:46 is
0:26:46 all
0:26:47 nonsense
0:26:47 to
0:26:48 latch
0:26:48 on
0:26:48 to
0:26:48 that
0:26:50 and
0:26:50 you
0:26:50 know
0:26:50 again
0:26:50 there’s
0:26:51 something
0:26:51 there
0:26:51 that I
0:26:51 would
0:26:52 kind of
0:26:52 put
0:26:52 in the
0:26:52 Cal
0:26:52 Newport
0:26:53 category
0:26:53 too
0:26:54 where
0:26:55 for me
0:26:55 maybe
0:26:55 the
0:26:55 most
0:26:56 interesting
0:26:56 thing
0:26:56 was
0:26:56 the
0:26:58 users
0:26:59 thought
0:26:59 that
0:26:59 they
0:26:59 were
0:27:00 faster
0:27:00 when
0:27:01 in fact
0:27:01 they
0:27:01 seemed
0:27:01 to
0:27:01 be
0:27:02 slower
0:27:02 so
0:27:02 that
0:27:02 sort
0:27:03 of
0:27:04 misperception
0:27:04 of
0:27:04 oneself
0:27:05 I
0:27:05 think
0:27:05 is
0:27:06 really
0:27:07 interesting
0:27:08 personally
0:27:08 I
0:27:08 think
0:27:08 there’s
0:27:08 some
0:27:16 thing
0:27:16 might
0:27:16 have
0:27:16 been
0:27:17 done
0:27:17 for
0:27:17 quite
0:27:17 a
0:27:17 while
0:27:17 by
0:27:18 the
0:27:18 time
0:27:18 I
0:27:18 get
0:27:18 back
0:27:19 so
0:27:19 honestly
0:27:19 one
0:27:20 really
0:27:20 simple
0:27:24 thing
0:27:24 that
0:27:25 the
0:27:25 products
0:27:25 can
0:27:25 do
0:27:26 to
0:27:27 address
0:27:28 those
0:27:28 concerns
0:27:28 is
0:27:28 just
0:27:28 provide
0:27:29 notifications
0:27:29 like
0:27:29 the
0:27:30 thing
0:27:30 is
0:27:30 done
0:27:30 now
0:27:31 so
0:27:31 stop
0:27:32 scrolling
0:27:32 and
0:27:32 come
0:27:33 back
0:27:33 and
0:27:33 check
0:27:33 its
0:27:33 work
0:27:35 that
0:27:35 in
0:27:35 terms
0:27:35 of
0:27:36 just
0:27:36 clock
0:27:37 time
0:27:37 it
0:27:37 would
0:27:37 be
0:27:37 interesting
0:27:38 to
0:27:38 know
0:27:38 what
0:27:39 applications
0:27:39 did
0:27:39 they
0:27:40 have
0:27:40 open
0:27:40 maybe
0:27:41 they
0:27:41 took
0:27:41 a
0:27:41 little
0:27:41 longer
0:27:42 with
0:27:43 cursor
0:27:43 than
0:27:43 doing
0:27:43 it
0:27:43 on
0:27:44 their
0:27:44 own
0:27:45 but
0:27:45 how
0:27:46 much
0:27:46 of
0:27:46 the
0:27:46 time
0:27:46 was
0:27:47 cursor
0:27:47 the
0:27:47 active
0:27:47 window
0:27:47 and
0:27:48 how
0:27:48 much
0:27:48 of
0:27:48 it
0:27:48 was
0:27:48 some
0:27:49 other
0:27:49 random
0:27:50 distraction
0:27:50 while
0:27:50 they
0:27:50 were
0:27:51 waiting
0:27:52 but
0:27:52 I
0:27:52 think
0:27:52 a
0:28:11 product
0:28:12 cursor
0:28:12 in
0:28:13 the
0:28:13 area
0:28:13 where
0:28:14 it
0:28:14 was
0:28:14 known
0:28:14 to
0:28:14 be
0:28:15 least
0:28:15 able
0:28:15 to
0:28:15 help
0:28:16 this
0:28:17 study
0:28:17 was
0:28:17 done
0:28:17 early
0:28:17 this
0:28:18 year
0:28:18 so
0:28:18 it
0:28:19 was
0:28:19 done
0:28:19 with
0:28:21 one
0:28:21 depending
0:28:21 on
0:28:21 how
0:28:21 you
0:28:22 want
0:28:22 to
0:28:22 count
0:28:22 a
0:28:22 couple
0:28:23 releases
0:28:24 ago
0:28:25 with
0:28:27 codebases
0:28:27 that
0:28:28 are
0:28:28 large
0:28:29 which
0:28:29 again
0:28:29 strains
0:28:30 the
0:28:31 context
0:28:31 window
0:28:32 and
0:28:32 that’s
0:28:33 one
0:28:33 of
0:28:33 the
0:28:34 frontiers
0:28:34 that
0:28:34 has
0:28:34 been
0:28:35 moving
0:28:38 very
0:28:38 mature
0:28:39 codebases
0:28:39 with
0:28:40 high
0:28:40 standards
0:28:40 for
0:28:41 coding
0:28:42 and
0:28:42 developers
0:28:43 who
0:28:43 really
0:28:43 know
0:28:43 their
0:28:44 codebases
0:28:44 super
0:28:44 well
0:28:45 who’ve
0:28:45 made
0:28:46 a lot
0:28:46 of
0:28:46 commits
0:28:47 to
0:28:47 these
0:28:48 particular
0:28:48 codebases
0:28:49 so
0:28:49 I
0:28:49 would
0:28:49 say
0:28:50 that’s
0:28:50 basically
0:28:50 the
0:28:51 hardest
0:28:51 situation
0:28:52 that
0:28:52 you
0:28:52 could
0:28:52 set
0:28:52 up
0:28:53 for
0:28:54 an
0:28:54 AI
0:28:54 because
0:28:55 the
0:28:55 people
0:28:55 know
0:28:56 their
0:28:56 stuff
0:28:56 really
0:28:56 well
0:28:56 the
0:28:57 AI
0:28:57 doesn’t
0:28:58 the
0:28:59 context
0:28:59 is
0:29:00 huge
0:29:00 people
0:29:00 have
0:29:00 already
0:29:01 absorbed
0:29:02 that
0:29:02 through
0:29:02 working
0:29:03 on
0:29:03 it
0:29:03 for
0:29:03 a
0:29:03 long
0:29:03 time
0:29:04 the
0:29:04 AI
0:29:04 doesn’t
0:29:04 have
0:29:05 that
0:29:06 that
0:29:06 knowledge
0:29:07 and
0:29:08 again
0:29:09 a couple
0:29:09 generations
0:29:10 ago
0:29:10 models
0:29:12 and
0:29:12 then
0:29:12 a big
0:29:12 thing
0:29:13 too
0:29:13 is
0:29:13 that
0:29:13 the
0:29:14 people
0:29:14 were
0:29:14 not
0:29:15 very
0:29:15 well
0:29:15 versed
0:29:16 in
0:29:16 the
0:29:16 tools
0:29:17 why
0:29:18 because
0:29:18 the
0:29:19 tools
0:29:19 weren’t
0:29:20 really
0:29:20 able
0:29:20 to
0:29:20 help
0:29:20 them
0:29:20 yet
0:29:20 I
0:29:21 think
0:29:22 the
0:29:22 sort
0:29:22 of
0:29:23 mindset
0:29:23 of
0:29:23 the
0:29:24 people
0:29:24 that
0:29:24 came
0:29:24 into
0:29:24 the
0:29:24 study
0:29:25 in
0:29:25 many
0:29:25 cases
0:29:25 was
0:29:25 like
0:29:26 well
0:29:27 I
0:29:27 haven’t
0:29:27 used
0:29:27 this
0:29:27 all
0:29:27 that
0:29:28 much
0:29:28 because
0:29:29 it
0:29:29 hasn’t
0:29:29 really
0:29:29 seemed
0:29:29 to
0:29:30 be
0:29:30 super
0:29:31 helpful
0:29:31 they
0:29:31 weren’t
0:29:32 wrong
0:29:32 in
0:29:32 that
0:29:33 assessment
0:29:33 given
0:29:33 the
0:29:34 limitations
0:29:36 and
0:29:37 you
0:29:37 could
0:29:37 see
0:29:37 that
0:29:37 in
0:29:38 terms
0:29:38 of
0:29:38 the
0:29:39 some
0:29:39 of
0:29:39 the
0:29:39 instructions
0:29:39 and
0:29:40 the
0:29:40 help
0:29:40 that
0:29:40 the
0:29:40 meter
0:29:41 team
0:29:41 gave
0:29:41 to
0:29:41 people
0:29:42 one
0:29:42 of
0:29:42 the
0:29:42 things
0:29:43 that
0:29:43 is
0:29:43 in
0:29:43 the
0:29:43 paper
0:29:43 that
0:29:43 they
0:29:44 would
0:29:45 if
0:29:45 they
0:29:46 noticed
0:29:46 that
0:29:46 you
0:29:46 weren’t
0:29:47 using
0:29:47 cursor
0:29:48 super
0:29:48 well
0:29:48 they
0:29:48 would
0:29:48 give
0:29:48 you
0:29:49 some
0:29:49 feedback
0:29:49 on
0:29:49 how
0:29:49 to
0:29:49 use
0:29:49 it
0:29:50 better
0:29:50 one
0:29:50 of
0:29:51 the
0:29:51 things
0:29:51 that
0:29:51 they
0:29:51 were
0:29:51 telling
0:29:52 people
0:29:52 to
0:29:52 do
0:29:52 is
0:29:53 make
0:29:53 sure
0:29:54 you
0:29:55 tag
0:29:55 a
0:29:55 particular
0:29:56 file
0:29:57 to
0:29:57 bring
0:29:57 that
0:29:57 into
0:29:58 context
0:29:58 for
0:29:59 the
0:29:59 model
0:29:59 so
0:30:00 that
0:30:00 the
0:30:00 model
0:30:00 has
0:30:01 the
0:30:01 right
0:30:02 context
0:30:02 and
0:30:02 that’s
0:30:02 literally
0:30:03 the
0:30:03 most
0:30:03 basic
0:30:04 thing
0:30:04 that
0:30:04 you
0:30:04 would
0:30:05 do
0:30:05 in
0:30:05 cursor
0:30:06 that’s
0:30:06 the
0:30:06 thing
0:30:06 you
0:30:06 would
0:30:06 learn
0:30:07 in
0:30:07 your
0:30:07 first
0:30:08 hour
0:30:08 your
0:30:08 first
0:30:08 day
0:30:08 of
0:30:09 using
0:30:09 it
0:30:09 so
0:30:09 it
0:30:10 really
0:30:10 does
0:30:10 suggest
0:30:10 that
0:30:10 these
0:30:11 were
0:30:11 while
0:30:12 very
0:30:13 capable
0:30:13 programmers
0:30:13 like
0:30:14 basically
0:30:15 mostly
0:30:15 novices
0:30:16 when
0:30:16 it
0:30:16 came
0:30:16 to
0:30:17 using
0:30:17 the
0:30:18 AI
0:30:18 tools
0:30:19 so
0:30:19 I
0:30:19 think
0:30:20 the
0:30:20 result
0:30:20 is
0:30:20 real
0:30:22 but
0:30:22 I
0:30:22 would
0:30:23 be
0:30:23 very
0:30:23 cautious
0:30:24 about
0:30:25 generalizing
0:30:25 too
0:30:26 much
0:30:26 there
0:30:28 in
0:30:28 terms
0:30:29 of
0:30:29 I
0:30:29 guess
0:30:29 what
0:30:30 else
0:30:30 was
0:30:31 the
0:30:31 other
0:30:31 question
0:30:31 it
0:30:32 what
0:30:32 is
0:30:32 the
0:30:32 expectation
0:30:33 for
0:30:33 jobs
0:30:33 I
0:30:33 mean
0:30:34 we’re
0:30:35 starting
0:30:35 to
0:30:35 see
0:30:35 some
0:30:35 of
0:30:35 this
0:30:36 right
0:30:36 we
0:30:36 are
0:30:36 definitely
0:30:37 seeing
0:30:38 no
0:30:38 less
0:30:38 than
0:30:38 like
0:30:38 Mark
0:30:39 Benioff
0:30:39 has
0:30:39 said
0:30:39 that
0:30:39 they
0:30:40 have
0:30:40 been
0:30:40 able
0:30:40 to
0:30:40 cut
0:30:40 a
0:30:40 bunch
0:30:41 of
0:30:41 head
0:30:41 count
0:30:42 because
0:30:42 they’ve
0:30:42 got
0:30:42 AI
0:30:43 agents
0:30:43 now
0:30:43 they’re
0:30:44 responding
0:30:44 to
0:30:44 every
0:30:44 lead
0:30:46 Klarna
0:30:47 of
0:30:47 course
0:30:47 has
0:30:48 said
0:30:50 very
0:30:50 similar
0:30:50 things
0:30:51 for a
0:30:51 while
0:30:51 now
0:30:52 they
0:30:53 also
0:30:53 I
0:30:54 think
0:30:54 have
0:30:54 been
0:30:54 a
0:30:54 little
0:30:54 bit
0:30:55 misreported
0:30:55 in
0:30:55 terms
0:30:56 of
0:30:56 like
0:30:56 oh
0:30:56 they’re
0:30:57 backtracking
0:30:57 off
0:30:57 of
0:30:57 that
0:30:58 because
0:30:58 they’re
0:30:58 actually
0:30:58 going
0:30:58 to
0:30:58 keep
0:30:59 some
0:30:59 customer
0:30:59 service
0:31:00 people
0:31:00 not
0:31:00 none
0:31:02 and
0:31:03 I
0:31:03 think
0:31:03 that’s
0:31:03 a bit
0:31:03 of
0:31:03 an
0:31:04 over
0:31:04 reaction
0:31:04 like
0:31:05 they
0:31:05 may
0:31:05 have
0:31:05 some
0:31:06 people
0:31:06 who
0:31:06 are
0:31:06 just
0:31:07 insistent
0:31:07 on
0:31:07 having
0:31:08 a
0:31:08 certain
0:31:08 experience
0:31:08 and
0:31:08 maybe
0:31:08 they
0:31:09 want
0:31:09 to
0:31:09 provide
0:31:09 that
0:31:10 and
0:31:10 that
0:31:10 makes
0:31:10 sense
0:31:11 you
0:31:11 know
0:31:11 it
0:31:11 doesn’t
0:31:12 I
0:31:12 think
0:31:12 you
0:31:12 can
0:31:12 have
0:31:13 a
0:31:15 spectrum
0:31:15 of
0:31:16 service
0:31:16 offerings
0:31:17 to
0:31:17 your
0:31:17 customers
0:31:18 I
0:31:18 once
0:31:19 coded
0:31:19 up
0:31:19 a
0:31:19 pricing
0:31:19 page
0:31:20 for
0:31:20 a
0:31:20 set
0:31:20 and
0:31:20 I
0:31:20 actually
0:31:21 just
0:31:21 vibe
0:31:21 coded
0:31:22 up
0:31:22 a
0:31:23 pricing
0:31:23 page
0:31:23 for
0:31:24 a
0:31:24 SaaS
0:31:24 company
0:31:25 that
0:31:25 was
0:31:25 like
0:31:27 basic
0:31:28 level
0:31:28 with
0:31:28 AI
0:31:29 sales
0:31:29 and
0:31:29 service
0:31:29 is
0:31:30 one
0:31:30 price
0:31:30 if
0:31:30 you
0:31:31 want
0:31:31 to
0:31:31 talk
0:31:31 to
0:31:31 human
0:31:32 sales
0:31:32 that’s
0:31:32 a
0:31:33 higher
0:31:33 price
0:31:33 and
0:31:33 if
0:31:33 you
0:31:33 want
0:31:33 to
0:31:33 talk
0:31:34 to
0:31:34 human
0:31:35 sales
0:31:35 and
0:31:35 support
0:31:36 that’s
0:31:36 a
0:31:36 third
0:31:36 higher
0:31:37 price
0:31:38 and
0:31:38 so
0:31:39 literally
0:31:39 that
0:31:39 might
0:31:39 be
0:31:39 what’s
0:31:40 going
0:31:40 on
0:31:40 in
0:31:40 some
0:31:40 of
0:31:41 these
0:31:41 cases
0:31:41 and
0:31:41 it
0:31:42 could
0:31:42 very
0:31:43 well
0:31:43 be
0:31:43 a
0:31:43 sensible
0:31:44 option
0:31:44 for
0:31:44 people
0:31:45 but
0:31:46 I
0:31:46 just
0:31:47 I
0:31:47 do
0:31:47 see
0:31:49 the
0:31:52 intercom
0:31:52 I’ve
0:31:52 got an
0:31:52 episode
0:31:53 coming up
0:31:53 with
0:31:54 they
0:31:54 now
0:31:54 have
0:31:54 this
0:31:55 fin
0:31:55 agent
0:31:56 that
0:31:56 is
0:31:56 solving
0:31:56 like
0:31:57 65%
0:31:58 of
0:31:58 customer
0:31:59 service
0:32:00 tickets
0:32:00 that
0:32:00 come
0:32:01 in
0:32:02 so
0:32:02 you
0:32:02 know
0:32:02 what’s
0:32:02 that
0:32:02 going
0:32:03 to
0:32:03 do
0:32:03 to
0:32:03 jobs
0:32:03 are
0:32:03 there
0:32:03 really
0:32:04 like
0:32:04 three
0:32:04 times
0:32:04 as
0:32:05 many
0:32:05 customer
0:32:05 service
0:32:06 tickets
0:32:06 to
0:32:07 be
0:32:09 handled
0:32:09 like
0:32:09 I
0:32:09 don’t
0:32:09 know
0:32:09 I
0:32:09 think
0:32:10 there’s
0:32:10 kind
0:32:10 of
0:32:10 a
0:32:11 relatively
0:32:11 inelastic
0:32:14 that
0:32:14 they’re
0:32:14 going
0:32:14 to
0:32:14 get
0:32:14 better
0:32:15 faster
0:32:15 answers
0:32:16 but
0:32:16 I
0:32:16 don’t
0:32:16 think
0:32:16 we’re
0:32:16 going
0:32:16 to
0:32:16 see
0:32:17 like
0:32:17 three
0:32:18 times
0:32:18 more
0:32:18 tickets
0:32:19 by the
0:32:19 way
0:32:19 that
0:32:20 number
0:32:20 was
0:32:20 like
0:32:21 55%
0:32:22 three or
0:32:22 four
0:32:22 months
0:32:22 ago
0:32:23 so
0:32:24 you know
0:32:24 as they
0:32:25 ratchet
0:32:25 that
0:32:25 up
0:32:25 the
0:32:26 ratios
0:32:26 get
0:32:26 really
0:32:26 hard
0:32:27 right
0:32:27 at
0:32:28 half
0:32:29 ticket
0:32:30 resolution
0:32:32 in
0:32:32 theory
0:32:32 maybe
0:32:33 you get
0:32:33 some
0:32:33 more
0:32:33 tickets
0:32:33 maybe
0:32:34 you
0:32:34 don’t
0:32:34 need
0:32:34 to
0:32:34 adjust
0:32:34 head
0:32:35 count
0:32:35 too
0:32:35 much
0:32:35 but
0:32:35 when
0:32:35 you
0:32:35 get
0:32:35 to
0:32:44 really
0:32:45 hard
0:32:45 to
0:32:46 imagine
0:32:46 that
0:32:47 so
0:32:47 I
0:32:47 don’t
0:32:47 think
0:32:48 these
0:32:48 things
0:32:48 go
0:32:48 to
0:32:49 zero
0:32:49 probably
0:32:49 in
0:32:50 a lot
0:32:50 of
0:32:50 environments
0:32:51 but
0:32:52 I
0:32:52 do
0:32:53 expect
0:32:54 that
0:32:54 you
0:32:54 will
0:32:54 see
0:32:56 significant
0:32:56 head
0:32:56 count
0:32:57 reduction
0:32:57 in
0:32:57 a lot
0:32:57 of
0:32:58 these
0:32:59 places
0:32:59 and
0:32:59 the
0:33:00 software
0:33:00 one
0:33:00 is
0:33:00 really
0:33:00 interesting
0:33:01 because
0:33:01 the
0:33:02 elasticities
0:33:02 are
0:33:03 really
0:33:03 unknown
0:33:04 you
0:33:05 can
0:33:05 potentially
0:33:06 produce
0:33:07 X
0:33:07 times
0:33:07 more
0:33:08 software
0:33:08 per
0:33:09 user
0:33:09 per
0:33:10 cursor
0:33:10 user
0:33:10 or
0:33:10 per
0:33:11 developer
0:33:11 at
0:33:12 your
0:33:12 company
0:33:12 whatever
0:33:14 but
0:33:14 maybe
0:33:14 you
0:33:14 want
0:33:14 that
0:33:15 you
0:33:15 know
0:33:15 maybe
0:33:15 there
0:33:15 is
0:33:15 no
0:33:16 limit
0:33:16 or
0:33:16 no
0:33:16 you
0:33:16 know
0:33:17 maybe
0:33:17 the
0:33:18 regime
0:33:18 that
0:33:19 we’re
0:33:19 in
0:33:19 is
0:33:20 such
0:33:20 that
0:33:21 if
0:33:21 there’s
0:33:22 you
0:33:22 know
0:33:22 10
0:33:23 times
0:33:23 more
0:33:23 productivity
0:33:24 that’s
0:33:24 all
0:33:24 to
0:33:24 the
0:33:24 good
0:33:24 and
0:33:24 you
0:33:24 know
0:33:25 we
0:33:25 still
0:33:25 have
0:33:25 just
0:33:25 as
0:33:25 many
0:33:27 jobs
0:33:27 because
0:33:27 we
0:33:27 want
0:33:28 10
0:33:28 times
0:33:28 more
0:33:28 software
0:33:29 I
0:33:29 don’t
0:33:29 know
0:33:29 how
0:33:29 long
0:33:29 that
0:33:30 lasts
0:33:31 again
0:33:31 the
0:33:31 ratios
0:33:32 start
0:33:32 to get
0:33:33 challenging
0:33:33 at
0:33:33 some
0:33:33 point
0:33:37 but
0:33:37 yeah
0:33:37 I
0:33:37 think
0:33:37 the
0:33:38 old
0:33:38 Tyler
0:33:38 Cowan
0:33:39 thing
0:33:39 comes
0:33:39 to
0:33:39 mind
0:33:39 you
0:33:40 are
0:33:40 a
0:33:40 bottleneck
0:33:40 you
0:33:40 are
0:33:41 a
0:33:41 bottleneck
0:33:42 I
0:33:42 think
0:33:42 more
0:33:43 often
0:33:43 it
0:33:43 is
0:33:44 are
0:33:44 people
0:33:45 really
0:33:45 trying
0:33:46 to
0:33:47 get
0:33:47 the
0:33:47 most
0:33:47 out
0:33:47 of
0:33:47 these
0:33:48 things
0:33:48 and
0:33:48 are
0:33:48 they
0:33:49 using
0:33:50 best
0:33:50 practices
0:33:50 and
0:33:50 have
0:33:51 they
0:33:52 really
0:33:52 put
0:33:52 their
0:33:52 minds
0:33:53 to
0:33:53 it
0:33:54 or
0:33:54 not
0:33:55 often
0:33:56 the
0:33:57 real
0:33:57 barrier
0:33:57 is
0:33:57 there
0:33:58 I
0:33:58 I’ve
0:33:58 been
0:33:58 working
0:33:59 a
0:33:59 little
0:33:59 bit
0:33:59 with
0:33:59 a
0:33:59 company
0:34:00 that
0:34:00 is
0:34:01 doing
0:34:04 basically
0:34:05 government
0:34:05 doc
0:34:06 review
0:34:06 I’ll
0:34:06 obstruct
0:34:06 a little
0:34:06 bit
0:34:07 away
0:34:07 from
0:34:07 the
0:34:07 details
0:34:09 really
0:34:09 gnarly
0:34:10 stuff
0:34:10 like
0:34:11 scanned
0:34:12 documents
0:34:13 you know
0:34:13 handwritten
0:34:15 filling
0:34:15 out of
0:34:16 forms
0:34:17 and
0:34:17 they’ve
0:34:18 created
0:34:18 this
0:34:18 auditor
0:34:19 AI
0:34:20 agent
0:34:21 that
0:34:21 just
0:34:22 won
0:34:22 a
0:34:22 state
0:34:22 level
0:34:23 contract
0:34:24 to
0:34:24 do
0:34:25 the
0:34:25 audits
0:34:26 on
0:34:26 like a
0:34:27 million
0:34:27 transactions
0:34:28 a
0:34:28 year
0:34:28 of
0:34:29 these
0:34:31 packets
0:34:31 of
0:34:31 documents
0:34:32 scanned
0:34:32 handwritten
0:34:33 all this
0:34:33 kind
0:34:33 of
0:34:33 crap
0:34:34 and
0:34:34 they
0:34:35 just
0:34:35 blew
0:34:35 away
0:34:36 the
0:34:36 human
0:34:39 workers
0:34:39 that
0:34:39 were
0:34:39 doing
0:34:39 the
0:34:40 job
0:34:40 before
0:34:41 so
0:34:41 where
0:34:41 are
0:34:41 those
0:34:41 workers
0:34:41 going
0:34:42 to
0:34:42 go
0:34:42 I
0:34:42 don’t
0:34:42 know
0:34:43 they’re
0:34:43 not
0:34:43 going to
0:34:43 have
0:34:43 10
0:34:44 times as
0:34:44 many
0:34:44 transactions
0:34:45 I
0:34:45 can
0:34:45 be
0:34:45 confident
0:34:46 in
0:34:46 that
0:34:47 are
0:34:47 there
0:34:47 going
0:34:47 to
0:34:48 be
0:34:48 a
0:34:48 few
0:34:48 still
0:34:48 that
0:34:49 are
0:34:49 there
0:34:49 to
0:34:50 supervise
0:34:50 the
0:34:50 AIs
0:34:50 and
0:34:51 handle
0:34:51 the
0:34:51 weird
0:34:51 cases
0:34:52 and
0:34:52 answer
0:34:53 the
0:34:53 phones
0:34:54 sure
0:34:55 maybe
0:34:56 they won’t
0:34:56 go
0:34:56 anywhere
0:34:57 the
0:34:57 state
0:34:58 may
0:34:58 do
0:34:58 a
0:34:58 strange
0:34:59 thing
0:34:59 and
0:35:00 just
0:35:00 have
0:35:00 all
0:35:00 those
0:35:00 people
0:35:00 sit
0:35:00 around
0:35:01 because
0:35:01 they
0:35:01 can’t
0:35:01 bear
0:35:01 to
0:35:02 fire
0:35:02 them
0:35:02 like
0:35:02 who
0:35:02 knows
0:35:02 what
0:35:03 the
0:35:04 ultimate
0:35:04 decision
0:35:04 will
0:35:04 be
0:35:05 but
0:35:05 I
0:35:05 do
0:35:05 see
0:35:06 a lot
0:35:06 of
0:35:06 these
0:35:06 things
0:35:06 where
0:35:07 I’m
0:35:07 just
0:35:07 like
0:35:08 when
0:35:08 you
0:35:08 really
0:35:08 put
0:35:08 your
0:35:09 mind
0:35:09 to
0:35:09 it
0:35:10 and
0:35:10 you
0:35:10 identify
0:35:12 what
0:35:12 would
0:35:12 create
0:35:13 real
0:35:13 leverage
0:35:13 for
0:35:13 us
0:35:14 can
0:35:14 the
0:35:14 AI
0:35:15 do
0:35:15 that
0:35:15 can
0:35:15 we
0:35:15 make
0:35:15 it
0:35:16 work
0:35:16 you
0:35:16 can
0:35:16 take
0:35:17 a
0:35:17 pretty
0:35:17 large
0:35:18 chunk
0:35:18 out
0:35:18 of
0:35:19 high
0:35:19 volume
0:35:20 tasks
0:35:21 very
0:35:22 reliably
0:35:23 in
0:35:23 today’s
0:35:23 world
0:35:24 and so
0:35:24 the
0:35:26 impacts
0:35:26 I think
0:35:26 are starting
0:35:27 to be
0:35:27 seen
0:35:27 there
0:35:27 on
0:35:28 a lot
0:35:28 of
0:35:29 jobs
0:35:31 humans
0:35:31 I think
0:35:31 are
0:35:32 you know
0:35:32 the
0:35:32 leadership
0:35:33 is
0:35:33 maybe
0:35:33 the
0:35:33 bottleneck
0:35:33 or
0:35:34 the
0:35:34 will
0:35:34 in
0:35:35 a lot
0:35:35 of
0:35:35 places
0:35:35 might
0:35:35 be
0:35:36 the
0:35:36 bottleneck
0:35:37 and
0:35:37 software
0:35:37 might
0:35:37 be
0:35:37 an
0:35:38 interesting
0:35:38 case
0:35:38 where
0:35:39 there
0:35:40 is
0:35:40 just
0:35:40 so
0:35:40 much
0:35:40 pent
0:35:40 up
0:35:41 demand
0:35:41 perhaps
0:35:42 that
0:35:43 it
0:35:43 may
0:35:43 take
0:35:43 a
0:35:43 little
0:35:44 longer
0:35:44 to
0:35:45 see
0:35:45 those
0:35:45 impacts
0:35:46 because
0:35:47 you
0:35:47 really
0:35:47 do
0:35:47 want
0:35:47 you
0:35:48 know
0:35:48 10
0:35:48 or
0:35:48 100
0:35:49 times
0:35:49 as
0:35:49 much
0:35:49 software
0:35:51 let’s
0:35:51 talk
0:35:51 about
0:35:51 code
0:35:52 because
0:35:53 it’s
0:35:53 where
0:35:53 Anthropic
0:35:53 made
0:35:54 a
0:35:54 big
0:35:54 bet
0:35:55 early
0:35:55 on
0:35:56 perhaps
0:35:57 inspired
0:35:57 by
0:35:57 this
0:35:57 sort
0:35:57 of
0:35:58 automated
0:35:59 researcher
0:35:59 recursive
0:35:59 self
0:36:00 improvement
0:36:02 sort
0:36:02 of
0:36:04 desired
0:36:04 future
0:36:05 and
0:36:05 we
0:36:05 saw
0:36:06 open
0:36:06 AI
0:36:07 make
0:36:07 moves
0:36:08 there
0:36:09 as well
0:36:10 why don’t
0:36:10 we
0:36:11 flesh that
0:36:11 out
0:36:11 or
0:36:12 talk
0:36:12 a little
0:36:12 about
0:36:13 what
0:36:14 inspired
0:36:14 that
0:36:14 and
0:36:14 where
0:36:14 you
0:36:14 see
0:36:15 that
0:36:15 going
0:36:17 utopia
0:36:17 or dystopia
0:36:18 is really
0:36:18 the big
0:36:18 question
0:36:19 there
0:36:19 I
0:36:19 think
0:36:19 right
0:36:19 I
0:36:19 mean
0:36:20 is
0:36:21 maybe
0:36:22 one
0:36:22 part
0:36:22 technical
0:36:22 two
0:36:23 parts
0:36:23 social
0:36:23 in
0:36:23 terms
0:36:24 of
0:36:24 why
0:36:24 code
0:36:24 has
0:36:24 been
0:36:24 so
0:36:25 focal
0:36:26 the
0:36:26 technical
0:36:27 part
0:36:27 is
0:36:27 that
0:36:27 it’s
0:36:27 really
0:36:28 easy
0:36:28 to
0:36:29 validate
0:36:30 code
0:36:30 you
0:36:31 generate
0:36:31 it
0:36:31 you
0:36:31 can
0:36:32 run
0:36:32 it
0:36:32 if
0:36:32 you
0:36:32 get
0:36:32 a
0:36:33 runtime
0:36:33 error
0:36:33 you
0:36:33 can
0:36:33 get
0:36:34 the
0:36:34 feedback
0:36:34 immediately
0:36:36 it’s
0:36:36 somewhat
0:36:37 harder
0:36:37 to do
0:36:37 functional
0:36:38 testing
0:36:39 replet
0:36:39 recently
0:36:40 just in
0:36:40 the last
0:36:40 48
0:36:41 hours
0:36:41 released
0:36:41 their
0:36:42 v3
0:36:42 of
0:36:43 their
0:36:43 agent
0:36:44 and
0:36:44 it
0:36:49 do
0:36:49 that
0:36:49 and
0:36:49 it
0:36:50 could
0:36:50 go
0:36:50 for
0:36:50 minutes
0:36:50 and
0:36:51 you
0:36:51 know
0:36:51 in
0:36:51 some
0:36:51 cases
0:36:52 generate
0:36:52 dozens
0:36:53 of
0:36:53 files
0:36:53 and
0:36:53 I’ve
0:36:53 had
0:36:54 some
0:36:54 magical
0:36:55 experiences
0:36:55 with that
0:36:55 where I was like
0:36:58 wow you just did that whole thing in one prompt and it like worked amazing
0:37:11 other times it will sort of code for a while and hand it off to you and say okay does it look good is it working and you’re like no it’s not I’m not sure why you get into a back and forth with it but the difference between v2 and v3 is that instead of handing the baton
0:37:41 back to you it now uses a browser and the vision aspect of the models to go try to do the qa itself so it doesn’t just say okay hey I tried my best wrote a bunch of code like let me know if it’s working or not it takes that first pass at figuring out if it’s working and you know again that really improves the flywheel just how much you can do how much you can validate how quickly you can validate it the speed of that loop is really key to
0:38:11 the pace of improvement so it’s a problem space that’s pretty amenable to the sorts of you know rapid flywheel techniques second of course they they’re all coders right at these places so they want to you know solve their own problems that’s like very natural and third I do think on the you know sort of social vision competition who knows where this is all going they do want to create the automated AI researcher that’s another data point by the way from
0:38:41 this was from the o3 system card they showed a jump from like low to mid single digits to roughly 40 percent of prs actually checked in by research engineers at open ai that the model could do so prior to o3 not much at all you know low to mid single digits as of o3 40 percent I’m sure those are the easier 40 percent or whatever again there will be
0:39:11 you know caveats you know caveats you know caveats to that but that’s you’re entering maybe the steep part of the s curve there and that’s presumably pretty high end you know I don’t know how many easy problems they have at open ai but presumably you know not that many relative to the rest of us that are out here making generic web apps all the time so you know at 40 percent you got to be starting to I would think get into some pretty hard tasks some pretty high value stuff you know at that at what point does
0:39:41 that ratio really start to tip where the AI is like doing the bulk of the work gbd5 notably wasn’t a big update over o3 on that particular measure I mean it also wasn’t going back to the simple qa thing gbd5 is generally understood to not be a scale up relative to 40 and o3 and you can see that in the simple qa measure it basically scores the same on these long tail trivia questions it’s not a bigger model that has absorbed like lots more
0:40:11 world knowledge um it’s not a bigger world knowledge um it is it is you know cal is right i think it is analysis that it’s it’s post training um but that post training you know is potentially entering the steep part of the s curve when it comes to the ability to do even the kind of hard problems that are happening at uh at open ai other on the research engineering front and you know yikes so i i’m a little worried about that honestly the um the idea that we could go from these companies having a few hundred
0:40:41 research engineer people to having you know unlimited overnight and like what would that mean in terms of how much things could change and also just our ability to steer that overall process um i’m not super comfortable with the idea of the companies tipping into a recursive self-improvement regime especially given the the level of control and the level of unpredictability that we currently see in the models but that does seem to be what they are going for so in terms of like why um
0:40:57 i think this has been the plan for quite some time even you remember that leaked anthropic uh fundraising deck from maybe two years ago where they said that in 2025 and 2026 the companies that train the best models will get so far ahead that nobody else will be able to catch up
0:41:09 um i think that in the 2526 timeframe they’d get this like automated researcher and once you have that how’s anybody you know who doesn’t have that going to catch up with you um
0:41:17 obviously some of that remains to be validated but um i do think they have been pretty intent on that for a long time
0:41:21 five years from now are there more engineers or fewer engineers
0:41:29 i tend to think less um you know already if i just think about my own
0:41:34 life and work i’m like would i rather have
0:41:38 a model or would i rather have like a junior marketer
0:41:40 i’m pretty sure i’d rather have the model
0:41:46 would i rather have the models or a junior engineer
0:41:53 i think i’d probably rather have the models in a lot of cases i mean it obviously depends on you know the exact person you’re talking about
0:42:00 um but truly forced choice today now that and then you’ve got
0:42:05 cost adjustment as well right i’m not spending nearly as much on my
0:42:11 cursor subscription as i would be on a you know an actual human engineer so even if they have some advantages
0:42:18 you know and i also have not scaffolded um i haven’t gone full co-scientist right on my
0:42:21 uh cursor problems i think that that’s another interesting
0:42:24 you start to see why
0:42:30 folks like sam altman are so focused on questions like energy and the seven trillion dollar build out
0:42:34 because these power law things are weird
0:42:35 and
0:42:37 you know to get
0:42:40 incremental performance for 10x the cost
0:42:45 is weird it’s a it’s definitely not the kind of thing that we’re used to dealing with
0:42:49 but for many things it might be worth it and it still might be cheaper than
0:42:51 the human alternative you know if it’s like
0:42:52 well cursor costs me
0:42:55 whatever 40 bucks a month or something
0:42:58 uh would i pay 400 for you know however much better
0:43:01 yeah probably would i pay 4 000 for however much better
0:43:03 well it’s still you know a lot less than
0:43:05 a full-time human engineer
0:43:07 uh and the costs are obviously coming down dramatically
0:43:10 too right that’s another huge thing
0:43:13 gpt4 was way more expensive it’s like 90 uh
0:43:18 it’s like a 95 discount from gpt4 to gpt5
0:43:22 that’s you know no small thing right i mean it’s
0:43:26 apple staff was a little bit hard because the chain of thought does
0:43:30 spit out a lot more tokens and so that you get you give back a little
0:43:32 on a per token basis it’s dramatically cheaper
0:43:33 more tokens generated
0:43:37 um you know just does eat back into some of that savings
0:43:41 but everybody seems to expect the trends will continue in terms of prices
0:43:42 continuing to fall
0:43:45 and so you know how many more of these like
0:43:46 price reductions
0:43:49 do you have to to then be able to
0:43:51 you know do the power law thing a few more times
0:43:56 i guess i think i think i i think less
0:43:59 um and i i think that’s probably true even if
0:44:00 we don’t get like
0:44:04 full-blown agi that’s you know better than humans at everything
0:44:06 i think you could easily imagine a situation where
0:44:10 of however many million people are currently employed as
0:44:11 professional software developers
0:44:15 some top tier of them that do the hardest things
0:44:16 can’t be replaced
0:44:19 but there’s not that many of those you know they and the
0:44:21 the real like
0:44:25 rank and file you know the people that over the last 20 years were told
0:44:27 learn to code you know that’ll be your thing
0:44:29 like the people that are the really top top people
0:44:31 didn’t need to be told to learn to code right they just
0:44:34 it was their thing they had a passion for it they were amazing at it
0:44:39 um we may not it wouldn’t wouldn’t shock me if we like still can’t replace those
0:44:44 people in three four five years time but i would be very surprised if you can’t get
0:44:54 your nuts and bolts web app mobile app type things spit out for you for far less
0:45:01 um and far faster than and probably honestly with significantly higher quality and less back and
0:45:07 forth um with an ai system than you know with your kind of middle of the pack developer
0:45:13 um in that time frame one thing i do want to call out you know there are definitely people have
0:45:18 concerns about progress moving too fast but there’s also concern and maybe it’s rising about progress
0:45:24 not moving fast enough in the sense that um you know a third of the stock market is is mag seven
0:45:31 um you know ai capex is you know over one percent of gdp and so we are kind of relying on some of this
0:45:38 progress in order to uh sort of sustain our sustain our economy yeah and with the um you know another
0:45:45 thing that i would say has been slower to materialize than i would have expected are ai culture wars or
0:45:54 you know sort of the the ramping up of protectionism of various industries we just saw um josh holly
0:45:59 i don’t know if he introduced a bill or just said he intends to introduce a bill to ban self-driving
0:46:09 cars nationwide um you know god help me uh i’ve dreamed of self-driving cars since i was a little
0:46:14 kid truly like sitting at red lights i used to be like there’s got to be a way dude we took a way
0:46:21 together yeah and it’s it’s so good um and the safety you know no i think whatever people want to
0:46:26 argue about jobs it’s going to be pretty hard to say right 30 000 americans should die every year
0:46:34 uh so that people’s incomes don’t get disrupted it seems like you have to be able to get over that hump
0:46:40 and say like the you know saving all these lives if nothing else is just really hard to uh to argue
0:46:48 against but we’ll see you know i mean he’s uh not uh without influence obviously so yeah i mean i am uh
0:46:55 very much on team abundance and you know my old mantra i’ve been saying this less lately but
0:47:05 adoption accelerationist hyper scaling pauser the tech that we have you know could do so so much for
0:47:11 us even as is i think if if progress stopped today i still think we could get to 50 to 80 percent of
0:47:17 work automated over the next like five to ten years it would be a real slog you’d have a lot of
0:47:24 you know co-scientist type breakdowns of complicated tasks to do you have a lot of work to do to go sit
0:47:28 and watch people and say why are you doing it this way what’s going on here what’s this you’d handled
0:47:33 this one differently why did you handle that one differently all this uh tacit knowledge that people
0:47:39 have and the kind of know-how procedural um you know just instincts that they’ve developed over time
0:47:43 those are not documented anywhere they’re not in the training data so the ais haven’t had a chance to
0:47:48 learn them but again no if i when i say like no breakthroughs i still am allowing there for like
0:47:53 you know fine tuning of things to just like capabilities that we have that haven’t been
0:47:58 applied to particular problems yet um so just going through the economy and and just sitting with
0:48:01 people and being like why are you doing this you know let’s let’s document this let’s get the
0:48:08 you know the model to learn your particular niche thing um that would be a real slog and in some ways
0:48:14 i kind of wish that were the future that we were going to get um because it would be a methodical you
0:48:20 know kind of one step one foot in front of the other you know no quantum leaps like it would probably
0:48:25 feel pretty manageable i would think in terms of the pace of change hopefully society could you know could
0:48:30 absorb that and kind of adapt to it as we go without you know one day to the next like oh my god you
0:48:34 know all the drivers so you know are are getting replaced or that one would be a little slower
0:48:38 because you’d have to have the actual physical build out but in some of these things you know
0:48:44 customer service could get ramped down real fast right like if a call center has something that they
0:48:49 can just drop in and it’s like this thing now answers the phones and talks like a human and has a higher
0:48:55 success rate and scales up and down um one thing we’ve seen at waymark small company right we’ve
0:49:00 always prided ourselves on customer service we do a really good job with it our customers really love
0:49:09 our customer success team but i looked at our intercom data and it takes us like half an hour to resolve
0:49:16 tickets uh we respond really fast we respond in like under two minutes most of the time but when we respond
0:49:20 you know two minutes is still long enough that the person has gone on to do something else right it’s
0:49:24 the same thing as with the cursor thing that we were talking about earlier right they’ve tabbed over to
0:49:29 something else so now we get the response back in two minutes but they are doing something else so then
0:49:33 they come back at you know minute six or whatever then they respond but now our person has gone and done
0:49:39 something else so the resolution time even for like simple stuff can be easily a half an hour
0:49:45 and the ai you know it just responds instantly right so you don’t have to have that kind of
0:49:50 back and forth you’re just in and out so i do think some of these categories could be
0:49:55 really fast changes um others will be slower
0:50:12 but yeah i mean i kind of wish we had that um i kind of wish we had that slower path in front of us
0:50:15 my best guess though is that we will probably continue to see things that will be
0:50:21 significant leaps and that there will be like actual disruption another one that’s come to mind
0:50:24 recently you know maybe we can get the abundance department on these new antibiotics have you seen
0:50:31 this uh development no tell us about it i mean it’s not a language model i think that’s another
0:50:36 thing people really underappreciate or that you could kind of look back at gpt4 to 5 and then
0:50:43 imagine a pretty easy extension of that so gpt4 initially when it launched the we didn’t have image
0:50:49 um understanding capability they did demo it at the time of the launch but it wasn’t released for some
0:50:54 months later the first version that we had could understand images could do a pretty good job of
0:50:59 understanding images still with like jagged capabilities and whatever um now with the new
0:51:06 nano banana from google you have this like basically photoshop level ability to just say hey take this
0:51:12 thumbnail like we could take our two uh feeds right now you know take a snapshot of you a snapshot of me
0:51:18 put them both into nano banana and say generate the thumbnail for the youtube preview featuring these
0:51:23 two guys put them in the same place same background whatever it’ll mash that up you can even have it
0:51:30 you know put text on top progress since gpt4 whatever we want to call it um gpt5 is not a bust
0:51:37 uh and it’ll spit that out and you see that it has this deeply integrated understanding that bridges
0:51:42 language and image and that’s something that it can take in but now it’s also something can put out
0:51:51 as all as part of one core model with like a single unified intelligence that i think is going to come
0:51:58 to a lot of other things um we’re at the point now with these biology models and material science models
0:52:04 where they’re kind of like the image generation models of a couple years ago they can take a real
0:52:09 simple prompt and they can do a generation but they’re not deeply integrated where you can have like a
0:52:17 true conversation back and forth um and have that kind of unified understanding that bridges language
0:52:24 and these other modalities but even so it’s been enough for this group at mit to use some of these
0:52:33 relatively you know narrow purpose-built biology models and create totally new antibiotics new in the
0:52:39 sense that they have a new mechanism of action like they’re they’re affecting the bacteria in a new way
0:52:49 and uh notably they they do work on um antibiotic resistant bacteria this is some of the first new
0:52:53 antibiotics we’ve had in a long time now they’re gonna have to go through you know when i say that
0:52:57 get the abundance department on it it’s like where’s my operation warp speed for these new
0:53:03 antibiotics right like we’ve got people dying in hospitals from drug resistant strains all the time
0:53:09 um why is nobody you know crying about this i think one of the things that’s happening to our society in
0:53:14 general is just so many things are happening at once it’s kind of the it’s like the flood the zone
0:53:18 thing except like there’s so many ai developments flooding the zone that nobody can even keep up with
0:53:23 all of those and that’s that’s come from me by the way too i would say two years ago i was like
0:53:26 pretty in command of all the news and a year ago i was starting to lose it and now i’m like
0:53:32 wait a second there was new antibiotics developed you know i’m kind of uh missing things you know just
0:53:40 like everybody else despite my best efforts but key point there is ai is not synonymous with language
0:53:46 models there are ais being developed with pretty similar architectures for a wide range of different
0:53:54 modalities we have seen this play out with text and image where you had your text only models and you had
0:53:58 your image only models and then they started to come together and now they’ve come really deeply together
0:54:03 and so i think you’re going to see that across a lot of other modalities over time as well
0:54:07 and there’s a lot more data there you know we might i don’t know what it means to like run out of data
0:54:12 um in the reinforcement learning paradigm there’s always more problems right there’s always some
0:54:17 something to go figure out there’s always something to go engineer the feedback is starting to come from
0:54:23 reality right that was one of the things elon talked about on the croc 4 launch was like maybe we’re
0:54:30 running out of problems we’ve already solved and you know we only have so much of those sitting around
0:54:35 in inventory you only have one internet you know we only have so much of that stuff but over at tesla
0:54:39 over at spacex like we’re solving hard engineering problems on a daily basis and they seem to be
0:54:45 never ending so when we start to give the next generation of the model these power tools the
0:54:49 same power tools that the professional engineers are using at those companies to solve those problems
0:54:55 and the ai start to learn those tools and they start to solve previously unsolved engineering problems
0:54:59 like that’s going to be a really powerful signal that they will be able to learn from
0:55:07 and now again fold in those other modalities right the ability to have sort of a sixth sense for the
0:55:13 space of small molecules the space of proteins you know the space of material science possibilities
0:55:20 when you can bridge or or unify the understanding of language and those other things i think you start
0:55:25 to have something that looks kind of like super intelligence even if it’s like not able to you know write
0:55:32 poetry at a superhuman level necessarily its ability to see in these other spaces is going to be truly a
0:55:37 superhuman uh thing that i think will be pretty hard to miss you said that that was one thing that cal’s
0:55:42 analysis missed is just the lack of appreciation for non-language modalities and and how they drive in
0:55:48 some of the innovations that you’re talking about yeah i think people are often just kind of equating the
0:55:58 chatbot experience with ai broadly yeah and you know the that that conflation will not last probably too much
0:56:03 longer because we are going to see self-driving cars unless they get banned um and that’s a you know very
0:56:07 different kind of thing and talk about your impact on jobs too right it’s like what four or five million
0:56:15 professional drivers in the united states um that is a big that is a big deal i don’t think most of those folks
0:56:20 are going to be super keen to learn to code and even if they do learn to code you know i’m not sure how
0:56:27 long that’s going to last so that’s going to be a disruption and then general robotics is like not
0:56:32 that far behind you know the and this is one area where i do think china might be actually ahead of the
0:56:38 united states right now but regardless of whether that’s true or not you know these robots are getting
0:56:42 really quite good right they can like walk over all these obstacles and these are things that a few years
0:56:48 ago they just couldn’t do it all you know they they could barely balance themselves and walk a few steps
0:56:54 under ideal conditions now you’ve got things that you can like literally do a flying kick and it’ll like
0:57:01 absorb your kick and shrug it off and just keep going uh you know right itself and and uh continue on its
0:57:07 way super rocky you know uneven terrain all these sorts of things are getting quite good um
0:57:12 you know the same thing is working everywhere i think one of the other thing that’s kind of
0:57:20 there’s always a lot of detail to the work so it’s it’s a sort of inside view outside view right inside
0:57:25 view you’re like there’s always this minutiae there’s always you know these problems that we had and things
0:57:31 we had to solve but you zoom out and it looks to me like the same basic pattern is working everywhere
0:57:39 and that is like if we can just gather enough data to do some pre-training you know some kind of raw
0:57:48 rough you know not very useful but just enough at least to kind of get us going then we’re in the game
0:57:55 and then once we’re in the game now we can do this flywheel thing of like you know rejection sampling
0:58:00 like have it try a bunch of times take the ones where it succeeded you know re-fine-tune on that
0:58:06 the rlhf you know feedback that the sort of preference take two which one was better fine you know fine
0:58:11 tune on that the reinforcement learning all these techniques that have been developed over the last
0:58:16 few years it seems to me they’re absolutely going to apply to a problem like a humanoid robot as well
0:58:20 and that’s not to say there won’t be a you know a lot of work to figure out exactly how to do that
0:58:24 but i think the big difference between language and robotics
0:58:33 is really mostly that there just wasn’t a huge repository of data to train the robots on at first
0:58:39 and so you had to do a lot of hard engineering to make it work at all you know to even stand up right
0:58:40 you had to have all these control systems and whatever
0:58:45 because there was nothing for them to learn from in the way that the language models could learn from the internet
0:58:50 but now that they’re working at least a little bit you know i think all these kind of refinement techniques
0:58:55 are going to work it’ll be interesting to see if they can get the error rate low enough that i’ll actually
0:59:02 like allow one in my house around my kids um you know that they’ll probably be um better deployed in like
0:59:06 factory settings first more controlled environments than uh the chaos of my house as you
0:59:10 you know have seen in this uh in this recording but
0:59:13 i do think they’re gonna they’re gonna work
0:59:19 what’s the state of agents more broadly uh at the moment where do you see things playing out
0:59:23 where do you see it going well broadly i think you know we’re it’s the the task length story from meter
0:59:29 of the you know every seven months or every four months doubling time we’re at two hours
0:59:37 ish with gbt5 repli just said their new agent v3 can go 200 minutes that if that’s true that would
0:59:44 even be a new you know high point on the um on that graph again it’s a little bit sort of apples to
0:59:47 oranges because they’ve done a lot of scaffolding how much have they broken it down like how much
0:59:52 scaffolding are you allowed to do you know with these things before you sort of are off of their
0:59:58 chart and onto maybe a different chart but if you extrapolate that out a bit and you’re like okay
1:00:05 take take the four month case just to be a little aggressive um that’s three doublings a year that’s
1:00:14 eight x task length increase per year that would mean you go from two hours now to two days in one year
1:00:19 from now and then if you do another eight x on top of that you’re looking at basically say two days to
1:00:27 two weeks of work in two years that would be a big deal you know to say the least if you could
1:00:32 delegate an ai two weeks worth of work and have it do a you know even half the time right the meter
1:00:37 thing is that they will succeed half the time on tasks of that size but if you could take a two-week
1:00:43 task and have a 50 chance that an ai would be able to do it even if it did cost you a couple hundred bucks
1:00:48 right it’s like well that’s again a lot less than it would cost to hire a human to do it um and it’s
1:00:53 all on demand it’s kind of you know it’s immediately available um if i’m not using it i’m not paying
1:00:58 anything the transaction costs are just like a lot lower the whole you know the many many other aspects
1:01:06 are favorable for the ai there so you know that would suggest that you’ll see a huge amount of
1:01:12 automation in in all kinds of different places the other thing that i’m watching though is the
1:01:18 reinforcement learning does seem to bring about a lot of bad behaviors re and um reward hacking being
1:01:25 one you know the the any sort of gap between what you are rewarding the model for and what you really
1:01:32 want can become a big issue um we’ve seen this in coding in many cases where the ai will claude is
1:01:38 like notorious for this will put out a unit test that always passes you know that just has like return
1:01:44 true in the unit test why is it doing that like well uh it must have learned that what we want
1:01:49 is for unit tests that pass you know we want it to pass unit tests but we didn’t mean to write fake
1:01:55 unit tests that always pass but that technically did you know satisfy the reward condition and so we’re
1:02:01 seeing those kind of weird behaviors with that comes this like scheming kind of stuff we we don’t
1:02:08 really have a great handle on that yet there is also situational awareness that seems to be on the
1:02:13 rise right where the models are like increasingly in their chain of thought you’re seeing things like
1:02:20 this seems like i’m being tested um you know maybe i should be conscious of what my tester is really
1:02:26 looking for here and that makes it hard to evaluate models in tests because you don’t know if they’re
1:02:33 actually going to behave the same way when they’re out in the real world so those you know i wouldn’t
1:02:40 say this is a high level or high confidence prediction but like one model of the future i’ve been playing
1:02:50 with is the task length keeps doubling while at the same time these weird behaviors pop up and then are
1:02:57 suppressed and we have seen in the cloud four and in the gpt5 system cards cloud four reported i think a
1:03:07 two-thirds reduction in um reward hacking and you know in gpt5 they reported a few different dimensions
1:03:15 but you know say something similar um reduction in deceptive behavior but those behaviors kind of just
1:03:19 emerged so it’s sort of like weird behavior emerges then they sort of figure out how to tamp it down
1:03:23 but not entirely presumably in the next generation they’ll like tamp it down some more but maybe some
1:03:27 new additional weird behavior could emerge and then we’ll have to kind of tamp that one down
1:03:33 all the while the the tasks are expanding in scope you know with a every four months doubling
1:03:42 so you could end up in a world where you can delegate really like major things to ais but there’s some
1:03:53 small but not necessarily totally vanishing chance that it like actively screws you over in the way
1:03:58 that it is trying to do that task and so you have a you know maybe it’s like okay i think here i’m about
1:04:03 to get two weeks worth of work done for a hundred bucks okay that’s amazing but there’s also a one in
1:04:10 10 000 chance that it like legitimately you know attacks me in a you know like meaningful way some
1:04:15 of the things that we have seen these are like fairly famous at this point but in the cloud four
1:04:22 system card they reported blackmailing of the human the the setup was that the ai had access to the
1:04:28 engineer’s email and they told the ai that it was going to be like replaced with a you know a less
1:04:34 ethical version or something like that it didn’t want that and it found in the engineer’s email that
1:04:41 the engineer was having an affair so it started to blackmail the engineer to so as to avoid being
1:04:50 replaced with a less ethical version people i think are way too quick in my view to move past these uh
1:04:55 anecdotes people are sort of often like well you know they set it up that way and you know that’s not
1:05:01 really realistic but another one was whistleblowing you know there was another thing where they sort
1:05:07 of set up this dynamic where there was some you know unethical illegal behavior going on and again
1:05:13 the model had access to this data and it decided to just email the fbi and and tell uh the fbi about it
1:05:19 so first of all i don’t think we really know what we want you know to some degree maybe you do want
1:05:25 ais to report certain things to authorities um that could be one way to think about the
1:05:31 bioweapon risk you know it’s like not only should the models refuse but maybe they should report you
1:05:35 to the authorities if you’re actively trying to create a bioweapon um i certainly don’t want them
1:05:43 to be doing that too much i don’t want to live under the you know surveillance of um claude five that’s
1:05:49 always going to be you know threatening to turn me in but i do sort of want some people to be turned
1:05:54 in if they’re doing sufficiently bad things we don’t have a good resolution society wide on you
1:06:02 know what we want the models to even do in those situations um and i think it’s also you know it’s
1:06:07 like yes it was set up yes it was research but it’s a big world out there right we got a billion users
1:06:13 already on these things and we’re plugging them in to our email so they’re going to have very deep
1:06:17 access to information about us you know i don’t know what you’ve been doing in your email i don’t
1:06:21 know i hope there’s nothing too crazy in mine but like now i gotta think about it a little bit right
1:06:27 what what did i have i ever done anything that i you know geez i don’t know um or even that it could
1:06:32 misconstrue right like it’s obviously not um maybe i didn’t even really do anything that bad but it just
1:06:38 misunderstands what exactly was going on so that could be a weird you know if there’s one thing that could
1:06:49 kind of stop the agent momentum in my view it could be like the one in 10 000 or whatever you know we
1:06:57 ultimately kind of push the the really bad behaviors down to is maybe still just so spooky to people that
1:07:06 they’re like i can’t deal with that you know and that might be hard to resolve so well you know what
1:07:11 happens then um you know it’s hard to check two weeks worth of work every couple hours or whatever
1:07:15 right like that’s part of where the where the whole then you bring another ai in to check it you know
1:07:20 that’s again where you start to get to the now i see why we need more electricity and and seven trillion
1:07:27 dollars of build out is yikes you know they’re going to be producing so much stuff i can’t possibly
1:07:33 even review it all i need to rely on another uh ai to help me do the review of the first ai
1:07:37 to make sure that if it is trying to screw me over you know somebody’s catching it i can’t
1:07:41 monitor that myself i think redwood research is doing some really interesting stuff like this where
1:07:46 they are trying to get systematic on like okay let’s just assume this is quite a different quite a
1:07:53 departure from the traditional ai safety work where the you know the big idea traditionally was let’s
1:07:58 figure out how to align the models make them safe you know make them not do bad things great redwood
1:08:03 research has taken the other angle which is let’s assume that they’re going to do bad stuff they’re
1:08:11 going to be out to get us at times um how can we still work with them and get productive output and
1:08:19 you know get value without you know uh fixing all those problems and that involves like again all these
1:08:25 sort of ais supervising other ais and um crypto might have a place to to a role to play in this
1:08:32 another episode coming out soon with ilia pulisukhin who’s the founder of near really fascinating guy
1:08:38 because he was one of the eight authors of the attention is all you need paper and then he started
1:08:45 this near company it was originally an ai company they took a huge detour into crypto because they were
1:08:49 trying to hire task workers around the world and couldn’t figure out how to pay them so they were like
1:08:54 this sucks so bad to pay these task workers in all these different countries that we’re trying to
1:09:01 get data from that we’re going to pivot into a whole blockchain uh side quest now they’re coming back
1:09:08 to the ai thing and their tagline is the blockchain for ai and so you might be able to get you know a
1:09:17 certain amount of control from you know the the sort of crypto security that the the blockchain type
1:09:24 technology can provide but i could see a scenario where the these the bad behaviors just become so
1:09:34 costly when they do happen that people kind of get spooked away from using the frontier capabilities in
1:09:41 terms of just like how much you know work the the ais can do but that wouldn’t be a that wouldn’t be a
1:09:47 pure capability stall out it would be a we can’t solve you know some of the long tail safety issues
1:09:54 challenge and you know that if that is the case then you know that’ll be um that’ll be an important
1:10:01 fact about the world too i i always nobody ever seems to solve any of these things like 100 right they
1:10:06 always every every generation it’s like well we reduced hallucinations by 70 oh we reduced deception
1:10:13 by two-thirds we reduced um you know scheming or or whatever by however much but it’s always still
1:10:19 there you know and it’s and if you take the even you know lower rate and you multiply it by a billion
1:10:24 users and thousands of queries a month and agents running in the background and processing all your
1:10:27 emails and you know all the deep access that people sort of envision them happening
1:10:34 it could be a pretty weird world where there’s just this sort of negative lottery of like ai accidents
1:10:39 um another episode coming up is with the ai underwriting company and they are trying to bring
1:10:44 the insurance industry and all the you know the wherewithal that’s been developed there to
1:10:50 price risk figure out how to you know create standards you know what can we allow what sort of
1:10:53 guardrails do we have to have to be able to insure this kind of thing in the first place
1:10:58 um so that’d be another really interesting area to watch is like can we sort of financialize those
1:11:04 risks um in the same way we have you know with car accidents and all these other mundane things but the
1:11:10 the space of car accidents is only so big the space of weird things that ais might do to you um
1:11:17 you know as they have weeks worth of runway is much bigger and so it’s it’s going to be a hard
1:11:21 challenge but you know people are people are working we got some of our best people working
1:11:29 on it what do you make the claim that 80 percent of ai startups have chinese open models um and what
1:11:34 do you make of the claim and and the implications i think that maybe that probably is true with the
1:11:41 one caveat that it is only measuring companies that are using open source models at all i think
1:11:49 most companies are not using open source models and i would guess you know the vast majority of tokens
1:11:59 being processed by american ai startups are their their api calls right to to the usual suspects um
1:12:07 so weighted by actual usage i would say still the majority as far as i could tell would be going to
1:12:14 commercial models um for those that are using open source i do think it’s true that the chinese models
1:12:18 have become the best um you know the american
1:12:27 bench there was always kind of thin right it was basically meta that was willing to put in huge amounts
1:12:35 of money and resources and then open source it you’ve got you know um paul allen funded group the allen
1:12:40 institute for ai ai too um you know they’re they’re doing good stuff too but they don’t have pre-training
1:12:44 resources so they do you know really good post training and and open source their recipes and all
1:12:52 that kind of stuff so it’s not like american and open source is bad you know and again it’s a time
1:12:59 this is another way in which i think you can really validate that things are moving quickly because if you
1:13:04 take the best american open source models and you take them back a year they are probably as good if not a
1:13:10 a little better than anything that we had commercially available at the time if you
1:13:18 compare to chinese you know they have i think uh surpassed so there’s been like pretty clear
1:13:22 change at the frontier i think that means that the best chinese models are like
1:13:27 pretty clearly better than anything we had a year ago um commercial or otherwise so
1:13:32 yeah i mean that just means like things are moving i think that’s like hopefully i’ve uh
1:13:36 made that case compellingly but that’s another data point that i think makes it hard to you i don’t
1:13:42 think you can believe both that um the chinese models are now the best open source models and that ai
1:13:46 has stalled out and we haven’t seen much progress since gpt4 like those seem to be kind of contradictory
1:13:55 notions um i believe the the one that is wrong is the lack of progress in terms of what it means i mean
1:14:00 i don’t really know it’s uh
1:14:08 we’re not going to stop china yeah the the whole i’ve always been a skeptic of the no selling chips
1:14:17 to china thing the notion originally was like we’re going to prevent them from doing you know some super
1:14:24 cutting-edge military applications and it was like well we can’t really stop that um but we can at least
1:14:28 stop them from training frontier models and then it was like man well we can’t necessarily really stop
1:14:33 that but now we can you know at least keep them from like having tons of ai agents well we’ll have
1:14:40 like way more ai agents than they do and i don’t love that line of thinking really at all um but one
1:14:48 upshot of it potentially is they just don’t have enough compute available to provide inference as a
1:14:52 service you know to the rest of the world so instead the best they can do is just say okay well
1:14:56 we’ll train these things and you know you can figure it out here here you go like have at it
1:15:05 um it’s kind of a soft power play presumably um i did an episode with uh anjane from a16z who i i thought
1:15:10 really did a great job of providing the perspective
1:15:17 of what i when i started calling countries three through 193 if the u.s and china are one and two
1:15:22 three through there’s a big gap you know there’s like i think the u.s is still ahead but not by
1:15:31 that much in terms of research and you know ideas relative to china we do have this compute advantage
1:15:37 and that does seem like it matters one of the upshots may be that they’re open sourcing and countries three
1:15:43 three through 93 are like or three through one 93 are significantly behind um so for them it’s a way
1:15:51 to you know try to bring more countries over to the chinese camp potentially in the u.s china rivalry
1:15:56 it seems like the model everybody and i don’t like this at all i i don’t like technology decoupling
1:16:02 as somebody who worries about you know who’s the real other here i always say the the real other
1:16:09 are the ai’s not the chinese so if we do end up in a situation where yikes like you know we’re seeing
1:16:15 some crazy things it would be really nice if we were on basically the same technology paradigm to the
1:16:20 degree that we really decouple and you know not just the chips are different but maybe the ideas start to
1:16:25 become very different publishing gets shut down you know tech tech trees evolve and kind of grow apart
1:16:31 um that to me seems like a recipe for
1:16:39 you know it’s harder to know what the other side has it’s harder to trust one another it seems to feed
1:16:45 into the arms race dynamic which i do think would you know is is a real uh existential risk factor i would
1:16:51 hate to see us you know create another sort of mad type dynamic where we all live under the threat of ai
1:16:55 destruction um but that very well could happen
1:17:02 and so yeah i don’t know i i i do kind of um
1:17:10 have some sympathy for the recent decision that the administration made to be willing to sell the
1:17:15 h20s to china and then it was funny that they turned around and rejected them which to me seemed like a
1:17:21 mistake i don’t know why they would be rejecting them if i were them i would buy them um and i would maybe
1:17:26 be i would maybe sell inference on the models that i’ve just been uh creating and i would try to make
1:17:32 my money back doing that but in the meantime they can at least you know demonstrate the greatness of
1:17:38 the chinese nation by showing that they’re not uh far behind the frontier and they can also make a
1:17:44 pretty powerful appeal to countries three through 193 and say like you know what do you really want to
1:17:51 you see how the us is acting uh in general you know you really want to they cut us off from chips uh
1:17:56 they had a even a long you know the last administration had an even longer list of countries that couldn’t
1:18:02 get chips this administration is doing all kinds of crazy stuff you know you get 50 tariffs here there
1:18:08 whatever um how do you know you can really rely on them to continue to provide you ai into the future
1:18:14 well you can rely on us we open source the model you can have it um so you know come work with us
1:18:18 and buy our chips because by the way our models will you know as we mature they’ll be optimized to run
1:18:26 on our chips so i don’t know that’s a complicated stuff a complicated situation i do think it’s true
1:18:31 i i don’t think the adoption is as high as that 80 i think that is you know within that subset of
1:18:35 companies that are doing stuff with open source we’re going to experiment with that at waymark but
1:18:42 we to be honest we have never done anything with an open source model in our product to present
1:18:49 everything we’ve ever done has been through commercial um at this point we are going to try
1:18:53 doing some reinforcement fine-tuning we are going to do that on a quinn model i think first
1:18:59 um so you know that’ll put us in that 80 but i’m guessing that at the end of the day
1:19:03 we’ll take that quinn model we’ll do the reinforcement fine-tuning and we’ll probably get
1:19:10 roughly up to as good as you know gpd5 or cloud4 or whatever and then we’ll say okay do we really
1:19:16 want to have to manage inference ourself how much are we really going to save and at the end of the
1:19:19 day i would guess we probably are still going to end up just being like yeah we’ll pay a little
1:19:24 bit more on a monthly bill basis for one of these frontier models or a little bit better maybe still
1:19:30 and you know it’s operationally a lot easier and they’ll have upgrades you know um
1:19:35 so yeah i mean of course there’s regulated industries there’s all there’s a lot of places
1:19:41 where you know you you have hard constraints you just can’t get around and that forces you to do those
1:19:45 chinese thing chinese models then there’s also going to be the question of like are there backdoors in
1:19:47 them um you know people have seen the sleeper agents
1:19:56 project where a model was trained to be good up until a certain point of time and you know people
1:20:01 put the today’s date in the system prompt all the time right today’s date is this you are clawed you
1:20:06 know here you go so then that’s gonna be another kind of thing for people to worry about um
1:20:13 and we don’t really have great there there have been some studies anthropic did a thing where they
1:20:19 trained models to have some hidden objectives and then challenged teams to figure out what those
1:20:24 hidden objectives were and with certain interpretability techniques they were able to
1:20:30 figure that stuff out relatively quickly so you might be able to get enough confidence that you take this
1:20:36 open source thing you know created by some chinese company whatever and then put it through
1:20:41 you know some sort of not exactly audit because you can’t trace exactly exactly what’s happening
1:20:47 but some sort of examination you know to see can we detect any hidden goals or any you know secret
1:20:53 backdoor bad behavior or whatever’s and maybe with enough of that kind of work you could be confident
1:20:57 that you don’t have it um but the more and more critical this stuff gets us but you know again going
1:21:02 back to that task length doubling weird behavior now you got to add into the mix what if they
1:21:11 intentionally programmed it to do certain bad things under certain you know rare circumstances um we’re
1:21:17 just headed for a really weird future you know that we’ve got all these there’s there’s no limit to
1:21:26 it you know all these things are valid concerns they often are in direct tension with each other um i don’t i i’m
1:21:31 not one who uh you know wants to see one tech company take over the world by any means so i
1:21:37 definitely think we would do really well to have some sort of broader more buffered ecological like
1:21:43 system where you know all the ais are kind of in some sort of competition you know mutual coexistence
1:21:49 with each other but we don’t really know what that looks like and we don’t really know um
1:21:53 you know we don’t really know what an invasive species might look like you know when it gets introduced into
1:22:00 that very you know nascent and as yet like not battle-tested uh ecology so yeah i don’t know
1:22:05 bottom line i think the future is gonna be really really weird yeah well i uh i do want to close on a
1:22:10 on a uplifting note so maybe maybe as a as a gearing towards closing question we could get into some
1:22:15 areas where we’re already seeing some some exciting capabilities emerge and sort of transform the
1:22:20 experience maybe maybe around education or health care or any other areas you want to you want to
1:22:28 highlight yeah it’s boy it’s all over um one of my mantras is that there’s never been a time better
1:22:33 time to be a motivated learner yeah so i think a lot of these things do have kind of you know two sides
1:22:39 of the coin there’s the worry that the students are taking the shortcuts and they’re you know losing the
1:22:47 ability to sustain focus and endure cognitive strain flip side of that is as somebody who’s fascinated by
1:22:53 the intersection of ai and biology sometimes i want to read a biology paper and i really don’t have the
1:23:03 background an amazing thing to do is turn on voice mode and share your screen with chat gpt and just go
1:23:07 through the paper reading it’s you don’t even have to talk to it most of the time you’re doing your
1:23:12 reading it’s watching over your shoulder and then at any random point you have a question you can
1:23:17 verbally say what’s this why why are they talking about that what’s going on with this what is the
1:23:22 role of this particular protein that they’re referring to or whatever and it will have the
1:23:30 answers for you so if you really want to learn in a sincere way you know the the things are unbelievably
1:23:37 good at helping you do that flip side is you can take a lot of shortcuts and you know maybe never
1:23:42 have to learn stuff on the biology front you know again like we’ve got
1:23:51 multiple of these sort of discovery things happening the antibiotics one we covered there was another one
1:23:56 that i did another episode on with a stanford professor named james zhao who created something
1:24:04 called the virtual lab and basically this was an ai agent that could spin up other ai agents
1:24:11 depending on what kind of problem it was given then they would go through a deliberative process
1:24:15 where you’d have you know one expert in one thing would give its take and they’d you know bat it back
1:24:19 and forth there was a critic in there that would criticize you know the ideas that had been given
1:24:24 eventually they’d synthesize then they were also given some of these narrow specialist tools so you
1:24:32 have agents using the alpha fold type um not just alpha fold you know there’s a whole whole wide wide
1:24:37 array of those at this point but using that type of thing to say okay well can we simulate you know how
1:24:45 this would interact with that um agents are running that loop and they were able to get this language
1:24:53 model agent with specialized tool system to generate new treatments for novel strains of covet that had
1:25:01 you know kind of escaped um the previous treatments amazing stuff right i mean the flip side of that of
1:25:05 course is you know you got the bioweapon risk so all these things do seem like they’re going to be
1:25:14 even even on just the abundance front itself right like we may have a world of unlimited professional
1:25:20 private drivers but we don’t really have a great plan for what to do with the five million people that
1:25:25 are currently doing that work we may have infinite software but you know what especially once the five
1:25:29 million drivers pile into all the coding boot camps and you know get coding jobs i don’t know what we’re
1:25:33 going to do with the 10 million people that we’re coding when you know nine million of them become
1:25:38 superfluous so yeah i don’t know i think we’re we’re headed for a weird world nobody really knows
1:25:45 what it’s going to look like in five years there was a great moment at at um google’s io where they
1:25:52 brought up some journalist i know you we uh we’re skeptical of journalists this is a great moment to uh
1:25:57 we’re going direct right this is a great reason or example of why one would want to do that they brought
1:26:05 up this person to interview demis and uh sergey brandon they the guy asked like what is search
1:26:10 going to look like in five years and sergey brandon like almost spit out his coffee on the on the stage
1:26:16 and was like search we don’t know what the world is going to look like in five years so i think that’s
1:26:23 really true like the biggest risk i think for so many of us and i you know include myself here is
1:26:29 thinking too small you know the the worst thing i think we could do would be to underestimate how far
1:26:37 this thing could go i would much rather be i would much rather be mocked for things happening on twice
1:26:45 the time scale that i thought than to find myself unprepared when they do happen so whether it’s 27 29
1:26:51 31 i’ll take that extra buffer honestly where we can get it my thinking is just
1:26:58 get you know get ready as as much and as fast as possible and again if we do have a little grace time
1:27:08 to uh you know to do extra thinking then great um but i would the i think the worst mistake we could
1:27:14 make would be to dismiss and and not feel like we need to get ready for big changes should we wrap
1:27:19 directly on that or is there any other last note you want to make sure to get across regarding anything
1:27:24 we we said today one of my uh other mantras these days is the scarcest resource is a positive vision
1:27:30 for the future yeah i do think it’s always really striking whether it’s sergey or you know or sam
1:27:37 altman or dario like dario probably has the best positive vision of the frontier developer ceos with
1:27:43 machines of love and grace but it’s always striking to me how little detail there is on these
1:27:50 things and when they launched gbt4o which was the voice mode they were pretty upfront about
1:27:57 saying yeah this was kind of inspired by the movie her and so i do think like even if you are not a
1:28:06 researcher you know not great at math not somebody who codes um i think that this technology wave really
1:28:12 rewards play it really rewards imagination i think literally writing fiction might be one of the
1:28:18 highest value things you could do especially if you could write aspirational fiction that would get
1:28:22 people at the frontier companies to think geez maybe we could steer the world in that direction
1:28:28 like wouldn’t that be great if you could plant that kind of seed um in people’s minds it could come from
1:28:37 a totally non-technical place and potentially be really impactful play fiction um i had one other
1:28:41 dimension to that but yeah play fiction positive vision for the future anything that you could do to offer a
1:28:49 a positive oh behavioral too it’s like these days because you can get the ais to code so well i’m
1:28:55 starting to see people who have never coded before i’m working with one guy right now who’s never coded
1:29:02 before but does have a sort of behavioral science background and he’s starting to do legitimate
1:29:09 frontier research on how are ais going to behave under various kind of esoteric circumstances so
1:29:19 i think nobody should count themselves out from the ability to contribute to figuring this out and
1:29:27 even to shaping this phenomenon um it is not just something that the you know the technical minds can
1:29:32 contribute to at this point literally philosophers fiction writers uh people literally just messing
1:29:41 around um pliny the jailbreaker you know there’s there are almost unlimited cognitive profiles that
1:29:46 would be really valuable to add to the mix of people trying to figure out what’s going on with ai so
1:29:52 come one come all is kind of my attitude on that that’s a great place to to wrap nathan
1:29:55 thank you so much for coming on the podcast thank you eric it’s been fun
1:30:03 thanks for listening to this episode of the a16z podcast if you like this episode be sure to like
1:30:09 comment subscribe leave us a rating or review and share it with your friends and family for more
1:30:15 episodes go to youtube apple podcast and spotify follow us on x a16z and subscribe to our
1:30:21 substack at a16z.substack.com thanks again for listening and i’ll see you in the next episode
1:30:27 as a reminder the content here is for informational purposes only should not be taken as legal business
1:30:32 tax or investment advice or be used to evaluate any investment or security and is not directed at
1:30:38 any investors or potential investors in any a16z fund please note that a16z and its affiliates may also
1:30:43 maintain investments in the companies discussed in this podcast for more details including a link to our
1:30:59 our investments please see a16z.com forward slash disclosures