Latest / Agentic Conversations (formally mlops.community) / The Myth of AI Breakthroughs // Jonathan Frankle // #205

Transcript

The Myth of AI Breakthroughs // Jonathan Frankle // #205 · Agentic Conversations (formally mlops.community)

• 0:00
  I am Jonathan Frankel. I am Chief Scientist of Neural
• 0:03
  networks at Databricks and Mosaic.
• 0:06
  Mel and I don't do coffee, but I am an avid Chipotle connoisseur
• 0:11
  Chipotle, which is an excellent Databricks customer.
• 0:14
  Welcome back to the Envelops Community Podcast.
• 0:17
  I am your host, Dimitrios, and today I am joined by none other
• 0:20
  than my man, Danny Lee. What's going on bro?
• 0:24
  How's it going, Dimitrios? Glad to be here.
• 0:26
  Glad to be of some use occasionally.
• 0:29
  Well, I'm super excited about this because we got Jonathan on
• 0:32
  here, the man of many talents. We talked about so many
• 0:36
  different pieces on what he's done in the past, what he has
• 0:42
  basically how he joined the Databricks team where you're
• 0:45
  currently working. So for those who do not know,
• 0:47
  your developer advocate at Databricks, you are doing some
• 0:51
  incredible work. We met I think almost a year ago
• 0:54
  and we had a lot of fun organising all kinds of
• 0:57
  different conferences and doing all kinds of stuff together.
• 1:01
  We could say which we get into in this podcast a bit and we got
• 1:05
  to do. We did a little bit of the
• 1:07
  censored version, but the. We can't reveal everything.
• 1:11
  That's true. Yeah.
• 1:12
  We don't want to give away any plot twists that come up, but
• 1:16
  Jonathan was a man that I met, I think, when we organised the
• 1:21
  Avalanche in June. He sat on one of the panels that
• 1:24
  I was hosting and in that moment I knew that this guy is a bit of
• 1:31
  a curveball on panels. We talked about that.
• 1:34
  He mentioned how in some of the panels in New York he is not
• 1:38
  invited back to. And so we had a good laugh about
• 1:42
  that. But then we talked about his
• 1:44
  work in law and what he did with policymakers.
• 1:47
  I found that fascinating. Did you have any huge takeaways?
• 1:50
  Because at the end, we also talked about Mosaic and what
• 1:52
  they're they've been doing, how they're utterly focused on
• 1:55
  making sure training and customers are able to do exactly
• 2:01
  what they need to do when making large language models.
• 2:04
  I know you have all kinds of extra context on this, but in
• 2:07
  the little box of this conversation, where were some of
• 2:11
  your huge takeaways? I mean I think the key and the
• 2:15
  reason why you know, #1 thankful for always joining you for these
• 2:18
  type of sessions, but I think the key thing that you really
• 2:22
  want to learn from Jonathan here unequivocally is it's through
• 2:26
  the science, right? Yeah, you have to be able to cut
• 2:28
  through the hype through what is often BS and let the scientific
• 2:33
  process take its course. There are plenty of people who
• 2:40
  this community is very diverse, very variety, very different
• 2:43
  opinions. That's a great thing, not a bad
• 2:45
  thing. But there's also this obsession
• 2:48
  for takedowns. And again, as much as we make
• 2:51
  fun of those things right there, there's the there's the big, the
• 2:56
  big companies that you you can absolutely do that.
• 2:59
  In fact, it's fun in a lot of ways.
• 3:01
  By the same token, there's a lot of really, really good science
• 3:05
  being done by really, really smart people that don't have
• 3:08
  nearly as many resources. And it's important that we go
• 3:11
  out of our way to support science, even if it's not
• 3:15
  exactly right. Because guess what?
• 3:17
  It helps us learn. And this is the key thing,
• 3:19
  right? This is this field everybody
• 3:21
  wants to make a big deal about. Ohh yeah, we're going to go
• 3:25
  through the next breakthrough breakthrough tomorrow.
• 3:28
  And again, Jonathan does a wonderful job explaining how to
• 3:32
  sort of level set your expectations and how to look at
• 3:35
  this these problems in a better way.
• 3:37
  And again, just to reiterate what Jonathan said, it's about
• 3:40
  science. Yeah, he really has his feet on
• 3:43
  the ground. I love how he talked about when
• 3:45
  they're iterating, when they're running with projects, it takes
• 3:48
  months for them to figure out a if the project is worth
• 3:53
  investing more time into. And he also is very pragmatic in
• 3:58
  the way that he looks at it. Like, hey, if I'm about to train
• 4:01
  a $10 million model, yeah, I need to have the confidence that
• 4:07
  that model is going to come out good and usable.
• 4:10
  And so I love that pragmatic approach.
• 4:13
  I also really appreciated him talking about and and we joked
• 4:17
  around at the beginning on the panels and how he is a bit of a
• 4:21
  curveball. But then he was very supportive
• 4:23
  of the scientific community. And like you said, even if
• 4:29
  people are doing work and they're putting out papers that
• 4:31
  are really cool and they get a lot of hype, but when you try
• 4:36
  and put them to the test in the industry, they don't necessarily
• 4:40
  hold up. He is not the guy that is going
• 4:43
  out there and bashing them for that, right?
• 4:45
  And that was really nice to hear and how much of an advocate he
• 4:49
  is for the scientists that are out there that are finishing
• 4:53
  their PhDs. And he he started off the
• 4:56
  conversation was saying like PhDs are hard.
• 4:58
  It's not yet. Yeah nothing to be taken
• 5:01
  lightly. So I appreciate you being on
• 5:03
  here. I appreciate Jonathan coming on
• 5:04
  here and of course I appreciate Databricks Mosaic for being what
• 5:09
  I like to call the premium brand partner of the ML OPS community.
• 5:14
  Hope you like that new term that I just.
• 5:16
  Coined. I love it.
• 5:17
  Absolutely enjoy it. This is awesome.
• 5:19
  All right. We'll see you all.
• 5:20
  On the flip side, let's get into this conversation. right, man.
• 5:29
  So it's good to have you on here.
• 5:31
  I'm excited to talk to you. I haven't actually chatted with
• 5:35
  you since we were live in person, and that was almost six
• 5:40
  months ago, I feel like when we were in California.
• 5:44
  And yeah, that was a, I mean, if I remember right, that was a fun
• 5:49
  day. I had not slept for a week prior
• 5:51
  to that. I was, I was at my least
• 5:55
  inhibited, shall we say, and I also love.
• 6:00
  Sorry, are you meeting? We Are you meeting?
• 6:02
  We liquored you up. Is that where you said?
• 6:03
  Ohh, you did you you liquored me up with sleep deprivation and
• 6:08
  negotiating an acquisition until the very last moment and filling
• 6:12
  out paperwork until the very last moment.
• 6:14
  And so by that point, I was ready to have some fun.
• 6:18
  And I love panels in general and all that good stuff.
• 6:22
  Like, you know, it's great fun because you can ask hard
• 6:27
  questions in front of an audience to your friends.
• 6:29
  And I love getting to, you know, what's the right word?
• 6:36
  I love getting to kind of steal moderation a little bit and kind
• 6:39
  of turn the questions around other people who are there
• 6:41
  because it just makes things a little more fun.
• 6:44
  Yeah, you made my job. Needs to eat or or more
• 6:47
  complicated, I should say. I I was hoping for more
• 6:49
  complicated. So I you know, in the New York
• 6:53
  AI startup community, I think I've been banned from doing
• 6:56
  panels informally. Now I keep getting invited to
• 6:58
  fireside chats before panels and I don't know if anyone wants to
• 7:02
  be on panels with me anymore. All right, so.
• 7:06
  We'll throw some for from this upcoming year.
• 7:09
  Well, we'll definitely throw some new ones for you, OK?
• 7:11
  Just just special just for. You it it has to be new Co
• 7:14
  panellists and new moderators. They don't think anybody else
• 7:16
  has ever gonna invite me back. And so, yeah, there were points
• 7:21
  where I was ready to take your microphone from you and I do
• 7:26
  not, I'm not at all surprised that you've been banned from
• 7:30
  panels in New York. There was and we'll we'll lay
• 7:33
  out the premise for everyone because we just jumped into it.
• 7:36
  But for those who do not know, in June this year 2023, so this
• 7:41
  last year, by the time this comes out, we.
• 7:44
  Oh, oh, now we've lost his oh oh Oh dear.
• 7:48
  Well, I guess it's just you and me.
• 7:50
  Ohh. OK, well, so I, I, I, I I guess
• 7:53
  I could provide the context. Yeah, we.
• 7:54
  Can run the podcast, yeah. There we go.
• 7:56
  Yeah, exactly. That's fair.
• 7:57
  So the context that Demetrius was actually saying was that
• 8:00
  because we're actually recording live anyway.
• 8:02
  So it's perfectly fine. It was that last June in 2023
• 8:06
  and I think Dimitrios was trying to say that by 2024 when this
• 8:10
  podcast comes out, it'll be last year.
• 8:12
  But the context is that we had this event as part of the Data
• 8:15
  and AI Summit 2023. The the conference or sorry, the
• 8:20
  meet up, excuse me, was called the Elam Avalanche with a
• 8:24
  Demetrius Alexey, myself as the organisers.
• 8:28
  And so it was a little bit of fun and we put together the
• 8:33
  event in less than three weeks. It was a little crazy, but we
• 8:36
  were all at the Contemporary Jewish Museum as our venue.
• 8:39
  And then Jonathan, our esteemed speaker here is was one of the
• 8:44
  key panellists for within LM Avalanche and he proceeded to
• 8:51
  ask questions from the audience when normally or the other
• 8:55
  panels when normally the moderator is actually supposed
• 8:57
  to be the one asking those questions.
• 8:59
  And so, yes, that's actually I think where we went.
• 9:01
  No, nobody told me that. Like when you got to tell me in
• 9:04
  advance of the moderators. Lottery.
• 9:05
  No, no, no. It's OK.
• 9:07
  I purposely left it that way. I wasn't actually what was
• 9:12
  actually trying to stop anything from that, but actually let's
• 9:15
  talk a little bit about that context, because I actually
• 9:18
  forgot which panel was it that we put you on, actually.
• 9:20
  I think it was about cost or about something along those
• 9:24
  lines. I remember I had, you know, or
• 9:25
  maybe even been about evaluation.
• 9:27
  There was someone from Anthropic there.
• 9:28
  There was someone from Google Deep Mind there.
• 9:30
  It was an amazing, right. I think someone from cohere was
• 9:32
  there potentially. That's right, that's.
• 9:34
  Fantastic group of people and honestly I had a lot of
• 9:36
  questions I wanted to ask them. So I appreciate the opportunity
• 9:39
  to have gotten to you know, find out stuff.
• 9:41
  Ohh, yes, I think that was the performance panel that we
• 9:43
  actually. Had set up.
• 9:44
  That sounds right. Yeah, I think that's.
• 9:45
  Yeah, yeah. So yeah, because we were
• 9:47
  actually, ohh, we lost Demetrius again.
• 9:49
  So because actually I did want to actually ask a little bit
• 9:52
  about the from a security panel because we were debating on, on,
• 9:55
  on when it came to the Security and Safety Panel whether that
• 9:58
  actually made sense for you to jump on that.
• 9:59
  But saying that, hey, you must be having you back.
• 10:03
  You know. Yeah, we can hear you,
• 10:05
  absolutely. Well, let's talk a little bit
• 10:08
  about your how you are an adjunct professor of law, but
• 10:15
  you have no legal background. Well, what's going on there?
• 10:18
  Yeah, there's a lot of I have some weird things on my resume,
• 10:22
  but this is by far the weirdest. Prior to my PhD, I'd finished up
• 10:27
  my undergrad and my masters and was supposed to start a PhD,
• 10:30
  Lasted exactly a week and dropped out because I was not
• 10:34
  happy and it just wasn't working for me.
• 10:36
  So for anyone who's having a hard time during your their PhD,
• 10:39
  then they're done. That dropped out once and almost
• 10:41
  dropped out a second time after going through 4 advisors.
• 10:44
  In my first two years of the MIT PhDs are hard.
• 10:47
  It's OK for them to be art. And I was just sitting at home
• 10:50
  looking for something to do. And through a bunch of weird
• 10:54
  coincidences and circumstances, the folks at Georgetown Law at
• 10:57
  this brand new centre called the Centre for Privacy and
• 11:00
  Technology happened to have received a grant.
• 11:04
  They did not apply for. And also we're looking to hire
• 11:08
  what they call in DC, a technologist.
• 11:10
  A technologist is anyone who knows about anything science or
• 11:12
  technology like me. The chemists, the climate folks,
• 11:15
  they're all technologists, but they're looking to hire someone
• 11:18
  who knew their technology really well and could come in and just
• 11:22
  kind of provide input to a lot of the work they were doing.
• 11:25
  And so I showed up and they put me to work, you know, helping
• 11:28
  out for to understand the technical aspects of police use
• 11:31
  of facial recognition. And this was back in 2015.
• 11:33
  So back then, this was today. It looks like an AI question
• 11:36
  back then was a privacy and surveillance question.
• 11:39
  And I think the work which eventually, you know, revealed
• 11:42
  that I think at that point like, you know, some huge percentage
• 11:45
  of all American drivers were actually in police facial
• 11:47
  recognition databases or American adults were in police
• 11:49
  facial recognition databases. Now it's basically everyone and,
• 11:52
  you know, even other kinds of biometric information like DNA
• 11:55
  samples, you know, a large portion of Americans are now in
• 11:59
  those databases. It was the first to really break
• 12:01
  that story for state and local police.
• 12:03
  And I got to learn a lot about facial recognition.
• 12:06
  And you know just how. You know how much of our data is
• 12:10
  available to police every day and what can go wrong?
• 12:14
  How was it available just from Facebook?
• 12:17
  Ohh from Like so, a lot of state drivers licence databases were
• 12:20
  being made available to the FBI or to police agencies.
• 12:24
  Other databases of mugshots, even immigration data were
• 12:28
  circulating around. And keep in mind this is almost
• 12:30
  a decade ago and all these databases have gotten a hell of
• 12:33
  a lot better and more sophisticated since then.
• 12:36
  So if you're an American driver, the police can probably look up
• 12:39
  your face whenever they have security camera footage of
• 12:42
  someone robbing a store or anything like that.
• 12:44
  And if you happen to be unlucky on that day, and there have been
• 12:46
  some unlucky people, you may get to spend the night in gaol.
• 12:50
  And you know, for someone who may miss work the next day or
• 12:54
  may have kids or anything, this can completely turn your life
• 12:57
  upside down. Yeah, well, this goes into the
• 13:00
  ethics question of it all. I imagine back in those days the
• 13:05
  accuracy was not that high. So the accuracy was not that
• 13:10
  high. But also I, and this is an
• 13:12
  argument you know, many of my friends who are now at the
• 13:14
  centre will make. I don't think bad accuracy is
• 13:18
  the right argument for we shouldn't use AI because in the
• 13:22
  limit accuracy of pretty much any AI system is going to be
• 13:25
  great. Like back at that time, we
• 13:27
  surface some potential concerns about bias and facial
• 13:29
  recognition that the other folks, Joy Buolamwini most
• 13:32
  prominently, took to the next level and showed we're really a
• 13:36
  huge problem. But bias is also not an excuse
• 13:41
  or not a kind of a basis on which to make policy decisions
• 13:44
  about AI. Because in the limit, you know,
• 13:46
  facial recognition is actually a really easy task for deep
• 13:49
  learning. With enough data and with good
• 13:51
  data kind of with what we have in 2023, we could solve facial
• 13:55
  recognition to A to a decent extent.
• 13:57
  And I could probably do that as a summer project if a Police
• 13:59
  Department hired me as an intern.
• 14:01
  I'm not available for internships right now, just for
• 14:03
  the record. So if you want to make a policy
• 14:06
  argument about AI and whether we should use it or when we should
• 14:10
  use it, it's not about the quality, it's about whether
• 14:13
  automated decision making should be a part of the process and
• 14:16
  whether things should be so frictionless.
• 14:18
  That's really where the argument should lie, in my view.
• 14:22
  Right. And break that down a little bit
• 14:24
  more automated decision making, full on.
• 14:27
  So there's no human in the loop or where the human touches the
• 14:30
  loop. Is that kind of the argument
• 14:32
  that you're saying? So there are there are two ways
• 14:35
  of you know, AI impacting the world in some sense like an AI
• 14:39
  system itself cannot impact the world unless either one we
• 14:43
  delegate decision making authority to it or to we use it
• 14:48
  to provide input to some other decision making process.
• 14:50
  Otherwise it's literally just an AI system and of that and you
• 14:53
  know it can't do anything to touch anyone or change anything
• 14:56
  about the environment. So, you know, you can imagine,
• 15:00
  you know, A, I don't know, let's say that we have a hiring
• 15:02
  algorithm that just determines, you know, which resumes get
• 15:05
  screened by a human. That's authority we've delegated
• 15:09
  to an AI algorithm. You know, alternatively we have
• 15:13
  an A I algorithm that provides a score for every resume and then
• 15:16
  that goes to a human that is in some sense using an AI system
• 15:20
  and then the human makes the decision.
• 15:22
  Now then, people look for the easy out and say, well, humans
• 15:25
  should check every facial recognition match before they go
• 15:28
  and bust down someone's door. And you know, so we looked into
• 15:32
  that a little bit and there's shocking no one, a whole
• 15:34
  literature on the fact that humans are actually quite bad at
• 15:37
  doing facial recognition ourselves.
• 15:39
  There is a well documented other race effect going back decades.
• 15:42
  We're really good at recognising people who look like us, not
• 15:44
  people who don't. There is a whole pseudoscience
• 15:49
  of how to do facial recognition properly and I emphasise
• 15:51
  pseudoscience because there's not a lot of science there.
• 15:54
  There is training that makes people better at this.
• 15:56
  There are people who are super recognizers.
• 15:59
  There's a whole science of the psychology of doing facial
• 16:01
  recognition, but suffice it to say that having a human in the
• 16:04
  loop is not a solution either. Necessarily, humans are flawed
• 16:07
  too. That is fascinating.
• 16:10
  I guess it just comes down to giving up our power and how we
• 16:13
  think we can always do it better.
• 16:15
  Or at least we have the we're willing to take on the
• 16:22
  repercussions if something does go wrong because we've done it
• 16:27
  as opposed to outsourcing it and not really having any hutch it.
• 16:33
  It means that it's like it's kind of a random crapshoot, even
• 16:37
  though it's not it. It still feels like, well,
• 16:40
  whatever, anything can happen now.
• 16:43
  Yeah, I kind of. I think of this two ways.
• 16:46
  Some people are much more comfortable with the human
• 16:48
  making the decision because at least there's someone who can
• 16:51
  explain themselves, you know? Yeah, neural networks are not
• 16:54
  explainable. Even if you ask ChatGPT to
• 16:56
  explain itself, it will make up a rationale.
• 16:58
  Rationalising and explaining are two very different things.
• 17:01
  The explanation for why Chat TPT said something is that you did
• 17:04
  gradient descent a large number of times and a lot of training
• 17:07
  data. The weights were set to this.
• 17:08
  That's the explanation. On the other hand, there is
• 17:11
  something very attractive about looking to some process that
• 17:15
  doesn't have the human biases, at least in an idealised sense.
• 17:18
  Like the The Seattle Police Department for a while had this
• 17:21
  FAQ on their website prior to our report where it was, you
• 17:24
  know, isn't this system biassed? And the answer is no.
• 17:27
  It's an automated algorithm. It can't see race or gender.
• 17:30
  And you know, in 2015 I think that wasn't a crazy thing for
• 17:34
  someone to, you know, naively say in 2023 the whole different
• 17:38
  story. But you know, you getting the
• 17:41
  same thing with autonomous vehicles.
• 17:42
  Would you rather, you know, find out that a pedestrian was
• 17:45
  injured by a human driver who was tired or intoxicated or
• 17:49
  would you rather find out that it was because a self driving
• 17:52
  car thought the human was a cloud?
• 17:55
  You know which which one is more comforting to us as people?
• 17:58
  Not entirely sure. Yeah, me neither.
• 18:02
  That and that gets into. It's a serious ethical debate
• 18:06
  and it becomes very nuanced, especially when you look at the
• 18:13
  the attachments that we have on this type of stuff.
• 18:16
  But I also wanted to ask you about this perpetual line Ucom
• 18:21
  because the. Suspense is killing.
• 18:23
  Emphasis on dot org, its.org. That's why.
• 18:27
  That's why it's not up, because I've been going to dot com ohh
• 18:32
  that's why it's not working for me.
• 18:35
  OK, So what is it? Break it down for me and the
• 18:40
  audience? Well, I am looking at it.
• 18:42
  Yeah, this is the study. This is a years worth of work on
• 18:47
  the part of, you know, two incredible lawyers, Claire
• 18:51
  Garvey, who led the whole study and worked tirelessly for a
• 18:54
  year. Alvaro Bedoya, who supervised it
• 18:57
  and is now an FTC commissioner who can affect these issues from
• 19:00
  government. And you know me as the third
• 19:03
  author and the least important, but the technical contributor,
• 19:07
  and this is our study as of 2015 of you know, who was using
• 19:12
  facial recognition for what Claire filed Freedom of
• 19:15
  Information Act requests on. I believe it was over 100
• 19:19
  different state and local police departments to get all the
• 19:22
  information about whether they were using facial recognition or
• 19:24
  not and in what capacity. And, you know, it turned up
• 19:29
  thousands or 10s of thousands of pages of of stuff that she and
• 19:33
  Alvaro and I went through. And we tried to figure out where
• 19:36
  is this being used? How is it being used?
• 19:38
  What are the safeguards, if any? What are the companies that
• 19:40
  we're providing this at that time?
• 19:42
  How does the technology work? What are the potential risks and
• 19:45
  a lot of policy suggestions? This is kind of, you know, if
• 19:48
  you were to ask me what is the work that you've done that
• 19:51
  you're proudest of, this is it. It's not my lottery ticket work.
• 19:55
  It's not Mosaic, Mel, it's this. This has made a difference in
• 20:00
  people's lives and affected everyday people.
• 20:04
  That to me is what meaning and impact looks like.
• 20:07
  And for anyone in the Mill community, we always need to
• 20:10
  take a step back and ask, you know, why are we here?
• 20:12
  What are we doing? What's the point?
• 20:14
  And for me, this is the point. This is why I got my PhD in
• 20:18
  machine learning. It was to provide better
• 20:20
  information to policy folks and help us make better decisions
• 20:24
  about the world. And I mean, I literally the call
• 20:27
  before this one, I do a bunch of policy work and I was just
• 20:30
  hanging out on a call, you know, helping to advise an
• 20:32
  international body on how they do policy.
• 20:35
  I don't tend to try to be in the spotlight for it, but I think
• 20:37
  it's just important that we're in the room and we're trying to
• 20:41
  be useful. Being useful doesn't mean that
• 20:43
  we tell people what to do. It means that when people ask us
• 20:45
  for information, we provide it to the best of our ability.
• 20:48
  One thing I actually would love to add, like ask, excuse me from
• 20:52
  that context is it seems to be that there is this need from
• 20:57
  what you're talking about for lawyers and like the
• 21:03
  technologists actually work Better Together to enact or
• 21:06
  create better policies. I mean, because again, you start
• 21:10
  off saying that, you know, it seems weird that you're you're a
• 21:13
  professor at law school, yet that it seems like this is
• 21:16
  exactly what's extremely helpful, especially when you're
• 21:18
  looking at the perpetuallineup.org.
• 21:21
  Yeah, I think it's, I would frame it differently.
• 21:24
  There's a need for those with technical expertise to be of
• 21:27
  service. That's a different framing.
• 21:30
  We're not equal partners here, nor should we be equal partners.
• 21:33
  The policy folks are trained in taking into account societies
• 21:37
  and needs and how to accomplish different ends that are better
• 21:40
  for society. We as technical people, that is
• 21:43
  not our job. We are not here to have amateur
• 21:45
  hour where we think about ethics and think about, you know, how
• 21:49
  we would run society if we were dictators or anything.
• 21:52
  About podcasts, about it, all that fun stuff.
• 21:55
  I I find it to be so stupid that we as technical people in
• 22:01
  general, we're big on disruption.
• 22:02
  We're big on reinventing the wheel because we think we can do
• 22:05
  it better for some reason. Policy is not a place to disrupt
• 22:08
  and it is not a place to reinvent.
• 22:10
  The wheel is a place to come and say how can I be of service And
• 22:14
  to be, you know, to be useful to the people who actually know how
• 22:18
  to take all this into account and to be useful to our society.
• 22:21
  Not to be useful to ourselves or come in and say I can do it
• 22:24
  better than you people. And yet, that's what I typically
• 22:26
  encounter, I think, the technical world.
• 22:29
  No policymaker needs an armchair.
• 22:31
  Quarterback No, I We can say what we want about the political
• 22:35
  process. Politics is different than
• 22:37
  policy. Policy is what we want to
• 22:40
  accomplish. Politics is you know how we
• 22:43
  decide who our leaders are, and you know who gets to choose
• 22:46
  which policies we pursue. But when it comes to actually
• 22:49
  making the policy and trying to provide options about what we
• 22:53
  think is best for the world, far be it from me to make that
• 22:58
  decision for all of society. My use is in knowing a lot about
• 23:01
  AI, given that Databricks and Mosaic are very open about what
• 23:05
  we do. And we don't have some big
• 23:06
  secret model talking openly about the process by which we
• 23:10
  build AI, because that insight is really important to
• 23:12
  policymakers. And then being useful in just
• 23:15
  answering questions when we're asked, speaking when spoken to,
• 23:19
  and not speaking out of turn. Yeah, leave me policy making to
• 23:26
  the professionals. Yes, is what it comes down to.
• 23:29
  Unless you'd like to invite them to train your model, in which
• 23:32
  case you know. Yeah, you do a little bring your
• 23:37
  friend to work day and they can do the same little swap.
• 23:41
  One thing that's interesting that jumped out at me from the
• 23:45
  website perpetuallineup.org is that major face recognition
• 23:51
  systems are not audited for misuse.
• 23:54
  And that is pretty wild. Also, how evasive some of these
• 23:59
  different police departments were and sharing their data with
• 24:02
  you and you, you stay here a few times.
• 24:05
  It's like NYPD just straight up ghosted us.
• 24:08
  You put it in different terms, you put it in more politically
• 24:12
  correct terms. But it's basically like yeah, we
• 24:14
  tried but they denied our records requests entirely.
• 24:19
  I believe that is so. It's it's important to
• 24:21
  emphasise, you know, all of this is a was in 2015.
• 24:25
  If I recall correctly, there was actually legal action between
• 24:28
  Georgetown and the NYPD over the fact that they had publicly
• 24:32
  spoken about facial recognition systems and then said they had
• 24:34
  no responsive records. Now, you know, I think it's
• 24:38
  important to say, you know it's important to speak on both both
• 24:41
  ends here. Like it's easy to dunk on some
• 24:44
  of the law enforcement agencies that were not cooperative or I
• 24:47
  think, you know were straight up misleading to us.
• 24:51
  But I also like we did a site visit to the Michigan State
• 24:54
  Police which has one of the most robust facial recognition
• 24:57
  systems released at that time did in collaboration with
• 25:00
  Michigan State University which has the world's best lab on
• 25:03
  automated biometrics, Professor Anil Jain.
• 25:05
  There is incredible and we visited both and I was actually
• 25:09
  very impressed. So you know it's there, I think
• 25:13
  over 10,000 law enforcement agencies in the US between
• 25:15
  state, local and federal. And you will see anything and
• 25:18
  everything from some things that are quite frankly ridiculous and
• 25:22
  unacceptable to a lot of really, really thoughtful people who are
• 25:26
  actually doing a great job. So again, easy to dunk on law
• 25:29
  enforcement and there's there are always comically bad
• 25:31
  examples that come up. But that's not to say that, you
• 25:35
  know, everything is completely broken and there aren't folks
• 25:38
  who are doing the best work they can.
• 25:40
  They're also a lot of like five person police departments in
• 25:43
  very small places that don't have the time to go and talk to
• 25:46
  a pH. D in computer science or get a
• 25:48
  PhD in computer science to figure out how to use this
• 25:51
  stuff. And so a lot of our work in that
• 25:52
  report was, you know, providing guidance so that somebody who
• 25:55
  just Googled police use of facial recognition would come
• 25:58
  across both a lot of people trying to sell them systems and
• 26:01
  a procurement template for what they what questions they should
• 26:04
  ask so that they can make the right decisions because most
• 26:07
  folks have good intentions. It's just hard to know
• 26:10
  everything about everything when dealing with systems like this.
• 26:15
  Well, that's fascinating stuff, man.
• 26:16
  And then Fast forward. At some point in your life,
• 26:19
  Mosaic did happen. You also talked about some
• 26:22
  lottery stuff there. What was that?
• 26:23
  You did you figure out how to win the lottery?
• 26:26
  A lot of people probably think that based on some of the emails
• 26:29
  I get occasionally from people asking me how to pick the right
• 26:32
  lottery numbers and I So my my dissertation research was on
• 26:35
  this topic called the Lottery ticket hypothesis.
• 26:37
  Best name I've ever come up with by far, and probably the best
• 26:41
  name I ever will come up with by far.
• 26:43
  Which was just a very simple idea about asking, you know, how
• 26:46
  how big of a neural network do you really need to learn
• 26:48
  something when you're using a neural network of any given
• 26:51
  size? Are all parts would actually
• 26:53
  getting used for learning? And you know, really asking are
• 26:57
  we training these neural networks in a smart way?
• 26:59
  How much headroom is there for us to improve the efficiency of
• 27:02
  training? And you know, that work ended up
• 27:05
  being very popular. I will say these days I don't
• 27:08
  think it's the world's most exciting topic to work on.
• 27:10
  So you know, I don't go and advise PhD students to work on
• 27:13
  that necessarily. The field has changed a lot
• 27:14
  since 2018, and you go where the cool problems are, not where the
• 27:18
  cool problems were. I certainly don't advise my PhD
• 27:21
  students to work on that topic today, but it was a really
• 27:25
  exciting topic at the time. I think it opened a lot of
• 27:27
  people's minds to, number one, that maybe there was a better
• 27:30
  way to train networks. Maybe there was a lot of room to
• 27:32
  reduce costs for those of us who were not in GPU rich at that
• 27:36
  time. For those of us who literally
• 27:37
  just had a laptop and a K-80 that we managed to find
• 27:40
  somewhere in the lab that nobody had figured out.
• 27:43
  And you know, all that good stuff.
• 27:45
  And it also, I think, helped to push forward this idea that
• 27:49
  maybe we need to be empirical about how we study AI.
• 27:52
  These systems have gotten so big and complicated and are so, you
• 27:56
  know, are so specific to the real world and the data set you
• 28:01
  use. And you know, there's so many
• 28:03
  ways to train a network badly, but we really only care about
• 28:05
  the few ways to train a network. Well, you know, if you want to,
• 28:08
  you can train GPT 4 and change the learning rate by 10X and
• 28:12
  you'll get garbage. Bring it down by 10X and you get
• 28:15
  GPT 4. But.
• 28:17
  And we don't care about the garbage, we care about GPT 4.
• 28:20
  So we need to be empirical, that being theoretical and proving
• 28:23
  things about neural networks isn't enough and we need to
• 28:25
  study them in the real world. And this was, I think at the
• 28:28
  time, not the most popular view of the world.
• 28:31
  But in my work, because I'm really bad at math, was focused
• 28:35
  exclusively on empiricism. And that is just the one hammer
• 28:38
  that I bring to every problem. And I think that was
• 28:40
  influential. And then that, you know, that
• 28:42
  became Mosaic in a way, the idea that we could make training more
• 28:46
  efficient is the genesis story for Mosaic.
• 28:48
  Naveen Rao, who became our CEO, just called, emailed me and said
• 28:51
  nice paper, I want to do a startup.
• 28:53
  I was like, no. And here we are.
• 28:54
  So you know, Naveen is very persuasive if you ever interact
• 28:58
  with him, Watch out. He will convince you to do
• 28:59
  whatever he wants you to do and you'll be grateful for it.
• 29:03
  But you know Mosaic doesn't apply the lottery ticket ideas,
• 29:06
  correct precisely. But the empirical idea and the
• 29:10
  idea that there's a lot of room to reduce the cost of training
• 29:13
  for the rest of us who are not GPU rich, that is basically the
• 29:17
  entire story of Mosaic Camel. Yeah.
• 29:20
  Break that down a little bit, because for me, on one hand,
• 29:24
  there's just breakdown, what you were doing at Mosaic and then
• 29:28
  how much of what you were doing was giving your expertise to
• 29:33
  clients so that they were able to fully leverage that and get
• 29:38
  better models. Yeah.
• 29:40
  So at Mosaic, our product ended up and you know, I say ended up
• 29:44
  because as with all good things, you wander for a while and you
• 29:46
  try to figure things out. Certainly during my PhD before
• 29:49
  and after with Mosaic, we wandered a bunch.
• 29:52
  But the final product is a platform where you can just come
• 29:57
  and train a model and it will work and you will get good
• 30:00
  results. That means that Gu failures,
• 30:03
  everything from the hardware issues all the way U to lost
• 30:07
  spikes, optimization challenges, good hyperparameters that you
• 30:09
  know will be efficient and get you the best performance per
• 30:12
  dollar. All of that is just a baked end,
• 30:15
  and that is a tonne of engineering and a tonne of
• 30:17
  scientific research. Hey, I'm Vishnu.
• 30:21
  I'm a data scientist at first hand and I definitely think that
• 30:24
  you should subscribe to the ML OPS Coffee Sessions podcast.
• 30:28
  It's the best podcast out there to stay on.
• 30:30
  Top of what? MLP's.
• 30:31
  Actually is to talk to the true thought leaders in the space.
• 30:34
  And oh by the way, Demetrius is absolutely hilarious.
• 30:38
  What a weird guy. You should definitely subscribe
• 30:40
  to the podcast. The way I think of it is that my
• 30:48
  team's job, you know, I run our research and science team.
• 30:51
  My team's job is to push the frontier of our knowledge and
• 30:55
  put to the test knowledge that other people have claimed to
• 30:58
  exist in service of finding the best and cheapest way to train
• 31:02
  state-of-the-art stuff. And then we don't directly
• 31:05
  consult like we're not a consulting company or anything
• 31:07
  like that. We build our expertise into our
• 31:09
  product. So if you use our product, you
• 31:11
  are getting the very best that we have as of that day.
• 31:14
  And if a new paper comes out that week that is getting a lot
• 31:18
  of hype on Twitter, we will put it to the test.
• 31:20
  And I would say like 7 times out of 10, it doesn't hold U in the
• 31:25
  settings that our customers would care about.
• 31:28
  But I think we're still one of the only shops in the world
• 31:30
  where you can just come and train basically an arbitrary
• 31:33
  scale language model from scratch and get it right the
• 31:37
  first time. I remember talking with Naveen
• 31:42
  and he was very clear on basically, and this was I think
• 31:48
  almost a year ago and he was the first person to just tell me
• 31:52
  straight up. Listen, if you have this much
• 31:55
  data, you're gonna want to think about just doing fine tuning.
• 31:59
  There's no need to do anything with an LM.
• 32:02
  It's pointless. You don't have the data for it.
• 32:04
  Once you get above this amount of data, and I can't remember
• 32:07
  the exact amount, you probably can quote me on it, but it was.
• 32:10
  I think it was in the in the range of like 10 billion, but I
• 32:15
  can't remember. It 10 billion is probably what I
• 32:17
  would have said and probably. Yeah.
• 32:21
  So I think it was around there. He was like, then it becomes
• 32:24
  advantageous for you to start training your own stuff.
• 32:27
  What I'm actually super interested to hear about though,
• 32:30
  is how you were taking the expertise and then building that
• 32:34
  into the product. What does that actually look
• 32:36
  like in practise? So you're saying, alright, cool.
• 32:39
  I know that, for example, this this part of training is a real
• 32:45
  pain in the ass. How can we make the product just
• 32:49
  automatically intuitive in that way so you don't even have to
• 32:52
  think about it or deal with it? So some of it is that our
• 32:55
  engineers are just superstars and my team uses everything that
• 33:00
  they build before it goes to customers because we're also
• 33:03
  training our own models. We have to prove.
• 33:05
  If I need to convince you that I can train a really good LLM, I
• 33:07
  have to train one first and open source it to convince you and
• 33:10
  convince the world that we can actually do that thing.
• 33:13
  And so that means we're beating the crap out of all of our tools
• 33:16
  before the customers ever see it.
• 33:17
  And so it has to be good enough for literally me as a manager
• 33:21
  who's in 30 something hours of meetings a week to also train
• 33:24
  models on the side and and not break anything in the process
• 33:28
  for good things to happen. And so we have to like our
• 33:32
  engineering team is top notch and we have eaten all of that
• 33:36
  dog food well before it ever goes to any customer.
• 33:39
  The other piece is then how do we get our expertise in.
• 33:42
  This is a huge challenge for us as a company.
• 33:44
  If you roll back the clock, you know this may feel like a long
• 33:46
  time, but four whole years or three whole years, there were a
• 33:50
  lot of different models people wanted to train.
• 33:52
  There was a different one for image classification versus
• 33:53
  segmentation versus detection versus you know, text
• 33:56
  classification versus sentiment analysis.
• 33:58
  We were starting to converge on Burt as maybe a basis for an
• 34:01
  architecture, you know, before we even had the word foundation
• 34:04
  model. But it was still all really
• 34:06
  different and the methods that made each of those things work
• 34:09
  well was really different. And I was terrified that it was
• 34:11
  going to be death by 1000 cuts. I was going to have to do this
• 34:13
  for every single stack and we picked models we had Resnet 50
• 34:17
  that we can train for like you know, less than $10 at this
• 34:20
  point on Imagenet to you know, really good quality accuracy.
• 34:23
  We had birth that we can train, you know, for less than $20 and
• 34:26
  free trained on whatever custom data you want.
• 34:29
  We did that for image segmentation.
• 34:30
  We did it for a bunch of different things and then Elms
• 34:33
  have kind of taken over at this point and a few kind of vision
• 34:36
  foundation models. And now we just focus on really
• 34:39
  nailing those particular use cases and coming up with great
• 34:41
  recipes. So I think of this as like we
• 34:45
  have our own LLM stack that looks really different from a
• 34:48
  lot of stuff that other people are doing in terms of our
• 34:50
  optimizer, our optimization strategy, our choice of position
• 34:54
  encodings are even. Our choice of model architecture
• 34:57
  is now like it's a transformer, but it's a it's a thing that is
• 35:01
  markedly different from I think what everybody else uses.
• 35:04
  If you look at some of our stuff, they've just kind of
• 35:06
  evolved to the point where they they look like a weird
• 35:08
  Frankenstein kind of version of what other people are using.
• 35:12
  And we we are public about this stuff and you should expect more
• 35:15
  in the not too distant future about what our current
• 35:17
  architecture looks like. It's pretty different from our
• 35:19
  last one and from what anybody else is doing.
• 35:22
  But you know, these things evolve and then we make that
• 35:25
  recipe available to all of our customers.
• 35:27
  And you know, well we tried a version it so that people aren't
• 35:31
  dealing with constant changes. But the important part is that
• 35:34
  we've vetted the hell out of every single one of these
• 35:36
  decisions. If you sit down with me and ask
• 35:38
  Roe or alibi, which position encoding should you use?
• 35:41
  I have pages and pages of really rigorous data on large scale
• 35:46
  models telling you which one you should use and why for both
• 35:49
  performance and model quality reasons.
• 35:52
  When you talk about sequence like extrapolation, there are a
• 35:54
  lot of different strategies out there including just using alibi
• 35:56
  naively. I reams of data on that when you
• 36:00
  tell me about, like how many tokens per parameter should you
• 36:03
  use? I have tonnes of data on that
• 36:06
  and helping you kind of navigate that tradeoff between training
• 36:09
  efficiency and inference efficiency.
• 36:11
  So we have the receipts to back up all of these decisions, but
• 36:16
  we also have a set of best decisions that you can just turn
• 36:19
  the key on and, you know, get going and know that you're doing
• 36:22
  the right thing according to our best practises.
• 36:24
  Know that you're doing literally the thing that I choose to do
• 36:26
  when I'm spending my money to train my models.
• 36:29
  When you're talking about incorporating all of this in
• 36:32
  there, you said something. Is it rope?
• 36:35
  It's rope, isn't it? What is it?
• 36:38
  Ohh, so actually I'm going to be making a decision this week on
• 36:41
  Roe versus alibi. That is a this week decision and
• 36:45
  if people are looking for the TLDR, Mosaic has gone with alibi
• 36:48
  historically. We love alibi.
• 36:51
  It gives you free extrapolation in a way that Rogue does not, or
• 36:54
  at least did not. The biggest problem is that the
• 36:57
  Lama folks chose to use rope, and so there's much better
• 36:59
  support for rope in the ecosystem in general.
• 37:02
  We've had to add our own flash attention kernels for alibi.
• 37:07
  We've had to add our own support for inference for alibi.
• 37:09
  Even though Ali is way simpler, I think I hope due to some of
• 37:13
  our advocacy, there's now parity like flash attention has both
• 37:17
  alibi and rope support, so there isn't really a hard choice.
• 37:19
  But also rope extrapolation has gotten a lot better.
• 37:22
  So you know, rope does come with some slowdowns.
• 37:25
  There are like, there are costs to doing it because alibi is
• 37:28
  just a much simpler thing to do. So I'm going to have to make a
• 37:32
  tough decision this week. But right now the data is
• 37:34
  basically that. I see no huge difference between
• 37:36
  the two of them in any meaningful setting.
• 37:38
  When it comes to just using it naturally using modern sequence
• 37:41
  like extrapolation methods, I don't see a huge difference.
• 37:45
  So honestly, one could flip a coin at this point, and there
• 37:48
  are great choices either way, which is not usually true.
• 37:51
  Usually I've got a very clear, definitive view about, you know,
• 37:54
  UL2. You know, loss has not provided
• 37:58
  any benefit whatsoever for us. Just to pick on my friends at
• 38:00
  Google a bit, because, you know, I don't like to pick on HD
• 38:03
  students who write papers. Haven't seen a benefit from that
• 38:06
  at all. In fact, has been slower and
• 38:07
  worse than just doing next token prediction in any setting we've
• 38:10
  measured. So you know, there are things
• 38:13
  that I have very strong opinions about Rope and Alibi.
• 38:16
  Ironically, not one of them. Do you have strong opinions on?
• 38:20
  I know some people have come through the community and asked
• 38:22
  about slum and the use of SLURM. It feels to me like that's kind
• 38:27
  of default. Do you have strong opinions
• 38:29
  about that? I I do have a strong opinion
• 38:32
  about that, that one should use the Mosaic ML slash Databricks
• 38:35
  orchestration solution because we thought about using SLAM and
• 38:39
  tried using SLAM. I've used SLAM in a lot of
• 38:41
  places and Slum was not built for machine learning workloads.
• 38:45
  There's a lot of stuff about just gang scheduling, higher
• 38:49
  level primitives. Slimmer, is not dockerized.
• 38:53
  There's a lot of stuff that just doesn't work well in SLAM
• 38:57
  because it was really built for a different universe and the
• 38:59
  slum folks are doing a fantastic job trying to keep U with that.
• 39:03
  But they have a tonne of different requirements and a
• 39:05
  tonne of different audiences to support.
• 39:06
  We chose to write our own orchestration stack and that's
• 39:10
  honestly the core of our product is everything is rolled into
• 39:12
  that orchestration stack. You use that and we can
• 39:15
  automatically pull in all of your, you know, API keys and
• 39:18
  secrets to various places you know, log to ML, flow or weights
• 39:22
  and biases, Whatever you choose automatically automatically do
• 39:26
  hardware failure detection. It's, you know it.
• 39:30
  All of that stuff is, you know, pretty straightforward to just
• 39:34
  bake into the product and in slum, you kind of have to roll
• 39:38
  it all on your own. And that gets pretty tricky.
• 39:40
  And I say this as an avid Slurm user in my academic life at
• 39:43
  Harvard. Yeah, it feels like it is one of
• 39:46
  those things that is almost an artefact from the academic life
• 39:50
  that's been brought over into the industry.
• 39:52
  And because most of the people that are that we're doing it in
• 39:58
  academia, they learn how to do it and they were playing with it
• 40:01
  there. They come and get a job and it's
• 40:05
  their job to train a model and they're like, well, I guess I'm
• 40:07
  going to keep using what I know how to use, but potentially
• 40:10
  their lives could be a lot easier.
• 40:14
  I I mean, I wouldn't go so far as to say that like Slim is an
• 40:19
  excellent tool. And if you were to put me in an
• 40:21
  academic setting where, you know, costs were limited and,
• 40:25
  you know, in some sense I just even if I wanted to pay to make
• 40:28
  folks more productive, I didn't have the money to do it.
• 40:30
  I would roll with SLURM. I think what Mosaic offers in
• 40:33
  some sense is that, you know, if you have the money to pay for
• 40:39
  better tools to make people more productive, you know, you go
• 40:42
  with Mosaic orchestration In the same way that, you know, cost is
• 40:45
  an issue, you can go with Tensor Board.
• 40:47
  And Tensor Board is a perfectly reasonable way to do experiment
• 40:49
  tracking. Things like mill flow or weights
• 40:52
  and biases or comment mill or a lot of the paid solutions are
• 40:56
  way better and will make you more productive.
• 40:58
  And it just comes down to, you know, what you need to pay for
• 41:01
  and what you don't. There's open source Spark, which
• 41:04
  is great if you can get it set up.
• 41:07
  Or there's managed Spark from Databricks where it's a hell of
• 41:11
  a lot easier to use and way faster.
• 41:13
  But it comes down to just where can you spend your budget.
• 41:16
  And I know plenty of folks where the budget is getting consumed
• 41:19
  by GPU and getting consumed by staffing.
• 41:22
  And so you got to make tough choices about, you know, some of
• 41:25
  the other products that you use. You can't buy everything as much
• 41:28
  as you'd like to. Yeah, well dude, Speaking of
• 41:32
  GPU's, we gotta talk about it. Seems like you have a special
• 41:36
  affinity towards the GPU poor and what is?
• 41:41
  Give us the story there. I mean, I I grew up GPU poor and
• 41:45
  GP underprivileged as a PhD student.
• 41:49
  And you know, did you know, I I joke.
• 41:51
  But to be completely honest, like when I did the lottery
• 41:55
  ticket work, which made my whole career and is the reason I get
• 41:58
  to have this conversation right now.
• 42:00
  And I did it on a bunch of spare Kads and borrowed GPUs from the
• 42:06
  Vision cluster that I was temporarily given access to at
• 42:08
  MIT before it was taken away from me because I abused it.
• 42:11
  And on a bunch of GPUs that IBM generously gave me, even though
• 42:16
  I had applied for a whole different project from them and
• 42:19
  totally abused it for this and cost them a lot of money in
• 42:22
  stupid ways and took down their transatlantic links.
• 42:25
  I'm forever indebted to IBM. They basically paid for my pH
• 42:28
  D's. But I remember sitting there and
• 42:32
  thinking, why does this have to be so expensive?
• 42:35
  All my friends have GPUs, but I am not in a lab that has GPUs at
• 42:39
  that point. I was in a lab that did
• 42:40
  technology policy. Of course we didn't have GPUs.
• 42:42
  And then I moved to my lab with Michael Carbon, who became my
• 42:45
  advisor collaborator, now my fellow staff member at Mosaic,
• 42:51
  in addition to his job at MIT, and he was running a lab on
• 42:54
  programming language research. He didn't have a GPU.
• 42:57
  So the first set of lottery ticket experiments, if you go to
• 43:00
  Archive and look at Archive V One, which I recommend everybody
• 43:02
  does if you want to see a really naive early draught with paper
• 43:05
  that did have some impact. I did it on my laptop.
• 43:09
  I trained such tiny networks that I could do it all on my
• 43:11
  laptop and then got one KAD and that eventually found the rest
• 43:15
  of the 80s and got up to four K 80s.
• 43:17
  And I have a dead K10 behind me right now, which IBM gave me as
• 43:21
  a token of all the GPUs I killed in the process of writing that
• 43:24
  paper. But I never really had much
• 43:28
  compute during my PhD unless I did internships.
• 43:32
  And honestly, Google generously gave me access to a bunch of
• 43:34
  free Tpus during my PhD. Um, but otherwise I was just
• 43:39
  doing it on a shoestring and doing the best I could.
• 43:43
  It's a lot harder to do that today because I think the nature
• 43:46
  of deep learning changes at scale and we've gotten really
• 43:49
  interested in scale in a way that I don't think we were in
• 43:51
  2018, so that it's harder. But my mission from that moment
• 43:56
  forward was kind of why do these people get to have all those
• 44:00
  GPUs and I don't and can we level the playing field and
• 44:04
  Mosaic like we spent two plus years just working to level the
• 44:08
  playing field to drive down the cost of the workloads that we
• 44:11
  thought were most important for industry.
• 44:12
  And personally I thought our most important for academia.
• 44:15
  When you can pre train Burt for 20 bucks you have no excuse not
• 44:18
  to do that in your paper and you can do Imagenet training for
• 44:20
  like a few dollars. You have no excuse not to do
• 44:23
  everything on Imagenet when you do your science, and I'd like to
• 44:27
  hope that those recipes have helped to just democratise the
• 44:30
  ability to to train these models in academia and do science.
• 44:34
  And now we're doing that with all it's like we'll have a lot
• 44:36
  to say in a little while about some much more efficient ways of
• 44:40
  training alms that I hope will bring down the cost of doing
• 44:44
  this in academia by, I don't know, call it conservatively
• 44:48
  like 3 to 5X and less conservatively 5 to 10 X.
• 44:51
  And I hope that that means we get great science and that just
• 44:56
  any random PhD student, regardless of what lab they're
• 44:58
  in, can go and write an impactful paper and build a
• 45:00
  career in this the same way that I had the opportunity to do.
• 45:04
  You guys have obviously found some breakthroughs with Q *.
• 45:09
  Uh, yeah. Yeah.
• 45:09
  So we have this algorithm we call Qatar.
• 45:12
  We've been working really hard on it for three years and I'm
• 45:14
  not sure how you found out about it, but it's going to
• 45:17
  revolutionise the way that you build AGI.
• 45:19
  It will bring about the Singularity and I think three to
• 45:21
  six months. And you know, in general there
• 45:25
  will be no such thing as money after that.
• 45:27
  No I one of my big. This is a podcast, so please do
• 45:33
  note the sarcasm here. Yeah, emphasis on sarcasm.
• 45:36
  If you're reading the transcript of this, that is sarcasm.
• 45:39
  Just for the record, that is sarcasm.
• 45:41
  Do not quote me on this. Do not call me before Congress
• 45:45
  to testify about this. You know, Do not send me to
• 45:48
  prison for this. Do not pour water on my GPU's
• 45:51
  for fear that they will have sudden drops and loss and, you
• 45:55
  know, become sentient. No.
• 45:56
  I, you know, I joke. But it's also, you know, I I'll
• 46:02
  be serious about kind of, I guess two things that worry me
• 46:06
  about the field. One is that we do tend to get
• 46:12
  caught up in the hype. And the hype can be as simple
• 46:14
  AS2 characters that were posted on Twitter.
• 46:18
  The hype can be as simple as a paper that somebody did a great
• 46:21
  job on at the scale that they could, but requires further
• 46:24
  validation at the scales that folks in industry care about and
• 46:28
  that, you know, will actually impact the field.
• 46:31
  And we get so caught up in that I think everybody's really
• 46:34
  excited right now And we're so used to the idea that just
• 46:39
  something will drop that will completely wreck our research
• 46:43
  agenda and force us to rethink everything that we're all just
• 46:46
  kind of a little shell shocked. And we think that may, you know,
• 46:49
  that could happen at any moment and maybe you know, Q Star, Oh
• 46:52
  my gosh, maybe that's the next chat.
• 46:53
  T which is the next Burt, which is the next, you know,
• 46:56
  Transformer, which is the next what have you.
• 46:58
  And we need to take a step back and pause and say like, look,
• 47:03
  this is a really exciting tweet, but can I reproduce it?
• 47:07
  I would say that my lab probably falsifies one major prominent
• 47:13
  AER on Twitter about once a month, just in the course of
• 47:15
  doing our science to validate the stuff and scale it up.
• 47:18
  And it's not a laughing matter. These are, you know, this is the
• 47:21
  hard work of in many cases PhD students who are doing their
• 47:24
  best science and didn't have the resources or just the scale of
• 47:28
  team that I did to really beat the crap out of these ideas.
• 47:30
  We tend not to publicise it when it happens, because you know, I
• 47:33
  and respect the scientific process and we can be wrong and
• 47:37
  be I've been that PhD student and you know, even sometimes
• 47:41
  when the air isn't perfect, you've done great science and
• 47:43
  you've demonstrated that you deserve to be in this field and
• 47:47
  one jerk in industry can ruin your career.
• 47:49
  I don't want to be that jerk, especially for people who are
• 47:53
  students and who are just learning how to do science for
• 47:54
  the first time. But we do tend to find that a
• 47:58
  lot of popular papers is the reason I picked on Google before
• 48:00
  and not on a bunch of student papers.
• 48:03
  But there are a lot of popular papers we find that just do not
• 48:06
  hold U in practise as far as we can tell in our settings, let
• 48:09
  alone looking at 2 characters like Qatar and then
• 48:12
  extrapolating what it must mean and all that good stuff.
• 48:17
  So you know, that's thing #1, hype and science.
• 48:20
  The other is that we get just so caught U in the like.
• 48:23
  The field is moving so fast, so it must continue to move so fast
• 48:26
  and the field has changed so much so it will continue to
• 48:29
  change and there will be breakthroughs constantly.
• 48:31
  If you would I'm sure, asked a nuclear physicist in the 1940s
• 48:35
  or 50s what the world would look like in 2023, They would say
• 48:39
  that everything would be nuclear powered.
• 48:41
  We would have nuclear fusion. We would have very small safe
• 48:44
  nuclear reactors powering every device, probably down to things
• 48:47
  like cell phones, and electricity would be so cheap
• 48:50
  and plentiful it wouldn't need to be metered.
• 48:52
  Sometimes science doesn't progress in a neat line.
• 48:56
  It progresses in leaps and then consolidation.
• 48:59
  If you told someone who looked at the progress of flight from,
• 49:04
  you know, let's say, 1930 and propeller planes, or 1910 and
• 49:07
  the first powered flight to jets toward the end of World War 2
• 49:11
  to, you know, landing on the moon 1520 years later, you know,
• 49:16
  we'd be exploring other solar systems by now.
• 49:20
  But you know, again, sometimes science doesn't proceed in such
• 49:23
  a neat, clean way. And that's OK too.
• 49:27
  We've got a lot to consolidate here.
• 49:30
  You know, Quantum physics did not lead straight to cell
• 49:32
  phones. There was a lot of work we had
• 49:34
  to do in the process of getting there and making the most of it.
• 49:37
  Even if technology stopped today in terms of developing an AI, we
• 49:40
  have decades more juice to squeeze out of this and make
• 49:44
  really cool systems that are useful in our day-to-day lives,
• 49:46
  useful for businesses, you know, just good for the world.
• 49:51
  One of the things I'd love to just following U with that
• 49:54
  context is how would you then help?
• 49:57
  Like what would you tell folks that are trying to decipher like
• 50:02
  the noise Basically, right you're you've called, I
• 50:05
  absolutely agree with everything you're saying, but then what's
• 50:07
  the advice you would give to folks?
• 50:09
  The how they can sort of like understand or read through all
• 50:14
  the hype per se. Yeah.
• 50:15
  And I'm, I'm actually very surprised that you don't call
• 50:18
  people out or at least like, say, hey, this didn't work for
• 50:21
  us. Like, it's working view,
• 50:22
  especially if there's something that is gaining a lot of
• 50:25
  traction and it can mislead us on the Internet that read it,
• 50:30
  it's like, whoa, this is incredible.
• 50:33
  I mean, quite frankly, I don't care about you on the Internet.
• 50:37
  I care about my customers and I care about the people who
• 50:40
  actually need to go and make really tough choices to make an
• 50:43
  impact. So it's not that I keep this
• 50:45
  stuff secret and just tell my customers, but I think we're
• 50:49
  really, we're really into takedowns in machine learning to
• 50:52
  some extent. I've seen this happen to several
• 50:55
  papers that have come out lately where prominent researchers
• 50:57
  basically in a tweet wanted to go and take down some student
• 51:01
  who had the nerve to post a paper that, you know, they
• 51:04
  believed in and you know, may not have been fully right or
• 51:09
  were they overclaimed a little bit.
• 51:11
  And that's not a good culture either.
• 51:13
  So I do like to correspond privately with authors.
• 51:17
  I do like to share quietly among other folks in the community who
• 51:20
  I think benefit from that knowledge.
• 51:22
  And we do have ways of conveying that.
• 51:24
  If you look at some of our blog posts, we'll talk about what
• 51:27
  worked and what didn't work for us.
• 51:29
  But it's not in a takedown sense.
• 51:31
  It's just a, you know, in a sense of kind of what we found
• 51:34
  and we haven't. So it's I want to be
• 51:37
  constructive and I want to be positive because science is also
• 51:40
  not a straight line. You can look at my lottery
• 51:42
  ticket work. It worked.
• 51:44
  In the settings I tested. It did not work anywhere beyond
• 51:48
  that. I had to go and basically redo
• 51:50
  the whole thing and rewrite the paper to figure out how to scale
• 51:52
  U. And the findings were a lot less
• 51:54
  compelling at scale than they were at small scales.
• 51:57
  I've been there. Like, that's just how science
• 52:00
  works. You do the best you can, and
• 52:02
  then you iterate and other people provide feedback and you
• 52:05
  improve. So it's not about keeping it
• 52:06
  secret. It's more about how do you
• 52:08
  contribute that knowledge in a useful constructive way and how
• 52:12
  do you respect the well-being and the careers of junior
• 52:15
  scientists who are learning how to do this.
• 52:18
  In an incredibly high pressure, high stress environment where
• 52:22
  millions of people in the world are watching what they do and
• 52:26
  want to take them down, that's a really tough place to be.
• 52:30
  And then some lab with a lot of resources and some people maybe
• 52:34
  you respect comes and does a takedown blog post of your ideas
• 52:39
  and makes you look like an idiot.
• 52:41
  That's when you leave the field. That's how you destroy our
• 52:45
  pipeline and, you know, ruin the careers and the happiness of a
• 52:47
  lot of people who will become more productive and become
• 52:52
  better at science over time, given the opportunity.
• 52:54
  So I just won't operate that way.
• 52:56
  That is not how I like to do things.
• 52:59
  There are other ways of making sure that knowledge gets out
• 53:01
  into the field that are a little more respectful of the process
• 53:04
  and of people. I just have to highlight that we
• 53:08
  started this conversation with how you've been banned from
• 53:11
  panels in New York and then you come to this, which is showing a
• 53:16
  whole different side of you. I think that is absolutely
• 53:20
  awesome. You get to see that you are very
• 53:23
  you are not one-dimensional at all.
• 53:25
  And it is cool to know that you are focusing on what works, what
• 53:33
  doesn't work and not focusing on, hey, let's go and get in
• 53:37
  Twitter feeds with people or just figure out how to make
• 53:40
  others look bad. Let's just show people what is
• 53:42
  working for us. And so, kind of getting back to
• 53:45
  Denny's question, cutting through the noise a little bit,
• 53:48
  besides reading the Mosaic blog, which I told you before we hit
• 53:52
  record I love and I know you're you've got stuff in the pipeline
• 53:55
  that's going to be coming out soon on that.
• 53:57
  What are other ways that you make sure to keep up to date and
• 54:02
  know what is real and what's not?
• 54:04
  Is it the lab that you're doing these tests on and that's how
• 54:09
  you're making sure that things are for real, or are there other
• 54:13
  ways? You know, so I I'll say one last
• 54:15
  thing about your you know, you do.
• 54:17
  They need contrast of me, the person who's getting banned from
• 54:19
  panels in New York and and me, the person who cares a lot about
• 54:23
  student well-being. My friends at anthropic and
• 54:27
  Hugging face and open an eye and scale AI can take it.
• 54:32
  And you know, especially on panels, there is so much
• 54:34
  bullshit and so many people who are stuck in their talking
• 54:38
  points, and my usual goal on a panel is to cut through that
• 54:41
  crap and get to the nuts and bolts.
• 54:44
  I don't want to hear you give your marketing pitch for the
• 54:46
  10th time. That's a waste of my time and
• 54:47
  the audience's time. I want to get to the real nuts
• 54:50
  and bolts. I want to know what it costs and
• 54:51
  how it works and what the numbers are and what's useful
• 54:53
  for people. And I will get very combative
• 54:57
  with anyone who does not get straight to that point.
• 55:01
  But a junior scientist who is writing a paper and is getting
• 55:05
  technical, they've already met that standard.
• 55:08
  And the rest is making sure that, you know, I treat them in
• 55:11
  a respectful and productive way and I look out for their
• 55:14
  well-being the same as I love to do with anyone who I have a
• 55:17
  conversation with. I love fireside chats for that
• 55:19
  reason. Because, you know, we don't have
• 55:22
  to stick to talking points. We can get to the nuts and
• 55:24
  bolts. We can talk about what's
• 55:25
  actually hard, what I've done wrong, what I've done right.
• 55:27
  I can talk about how the lottery ticket work like, highly
• 55:30
  imperfect. That first paper, which is what
• 55:32
  a lot of people read in classes these days, is not the paper I
• 55:35
  would recommend anyone read. There's a subsequent paper that
• 55:38
  I think is much better and much more realistic.
• 55:40
  That is my favourite paper I've ever written, but doesn't get
• 55:43
  read anywhere near as much because it's not the first AER
• 55:47
  but for how to cut through the noise.
• 55:49
  I, you know, the first bias I apply is just to basically, you
• 55:56
  know, not look at anything for a while after it comes out.
• 56:00
  You know I will look at a paper like two to three months after
• 56:03
  it comes out if anyone still cares about that paper because I
• 56:08
  think you need to alley a bit of a like you need to smooth things
• 56:12
  out overtime. I don't tend to make big changes
• 56:16
  to model architectures until they've been out and popular for
• 56:19
  a few months. Rope became popular again post
• 56:24
  alibi, but I don't know whether that was going to stick around.
• 56:27
  It was going to get replaced by something else.
• 56:28
  And you see this happen all the time where there's a first paper
• 56:31
  that like shows something is really cool, followed by like 10
• 56:34
  more papers that consolidate. And you get to like the really
• 56:37
  good version of it after a little bit.
• 56:39
  So I think you just have to smooth things out a bit.
• 56:42
  There was a really cool paper that came out on, you know,
• 56:44
  state space models and showing how they scale up yesterday.
• 56:47
  I'll read that paper in a couple of months.
• 56:50
  If it's still relevant to read that paper in a couple of
• 56:52
  months. Maybe it's a big idea and a
• 56:54
  bunch of people will build on it very quickly.
• 56:56
  Maybe, you know, it doesn't actually matter that much and
• 56:58
  it's a bunch of hype, but time tends to smooth things out a
• 57:02
  little bit and the field doesn't move that fast.
• 57:04
  If someone like me, who's feeling a lot of pressure to be
• 57:06
  on the bleeding edge, is willing to wait two or three months to
• 57:09
  look at any paper, that should make you comfortable that you
• 57:12
  can do the same thing and you don't need to keep up
• 57:13
  constantly. And the other is just having a
• 57:16
  good team around me to say dumb things to and bounce ideas off
• 57:19
  of. I love my team.
• 57:21
  I think I work with the most brilliant people in the world
• 57:25
  and the most humble people in the world and anyone's willing
• 57:28
  to tell me I'm wrong or tell me we should be doing something
• 57:31
  better. I am not the king of this team
• 57:33
  or anything like that. I'm not infallible.
• 57:36
  I rely on them to be the people who are on the ground.
• 57:39
  I think there's a huge trend of people who get somewhat mill
• 57:42
  famous and emphasis on ML famous and not actual famous because
• 57:47
  quite frankly asked my mom who Yan Lagoon is and you know she
• 57:52
  doesn't know Yan Lagoon from, you know, don't know anybody
• 57:55
  else. But people then try to turn into
• 57:59
  prophets and predict the future. And I think this is a field
• 58:03
  where the people on the ground were working with this stuff in
• 58:06
  day-to-day have the very best insights into the nature of what
• 58:10
  works, what doesn't, where the technology is, where the
• 58:12
  technology isn't. And those of us who are no
• 58:14
  longer with our feet on the ground every single day are just
• 58:18
  not the best experts. And I'm a manager now.
• 58:21
  I'm an administrator in many ways.
• 58:23
  I am not the person I look to for the best insights on this.
• 58:27
  Typically, when people reach out to me, I will route them to the
• 58:29
  person on my team who is building the thing because
• 58:32
  they've read all the papers, they've done all the
• 58:34
  experiments, they have the best insight.
• 58:37
  I love hiring people straight from PhD onto my team.
• 58:39
  And I don't love hiring people who are like senior and famous
• 58:42
  and their careers, because I think those people are actually
• 58:45
  much worse than people fresh out of PHD's who have grappled with
• 58:48
  the stuff going hands on way and continue to.
• 58:51
  I love hanging out with PhD students for that reason, and I
• 58:53
  think they know a lot more than I do.
• 58:56
  It goes back to the idea of let the policymakers make the
• 59:00
  policies, let the PHD's do that, do that hard work out of what it
• 59:07
  is, and even let the pH. D Be the one to make the
• 59:13
  predictions, not the managers who think they know what they're
• 59:17
  doing or where things are going to go.
• 59:20
  And I do see a lot of, like, you know, for my friends who are at
• 59:22
  any big lab, I do see a lot of management versus people on the
• 59:26
  frontline conflict over which ideas should be pursued and what
• 59:29
  makes sense and what doesn't. And my solution to that is to
• 59:33
  trust the hell out of my team. When someone on my team comes to
• 59:36
  me with a really strongly held view that we should do ex, I'll
• 59:39
  interrogate them and push them a little bit just to make sure
• 59:42
  that that view is really strongly held and they have the
• 59:44
  evidence to back it U. And then I'll say, OK, we'll do
• 59:46
  that because in my experience, they're right many more times
• 59:51
  than I am. So in some sense, you know,
• 59:53
  having me on a podcast, like, yeah, cool, I technically run
• 59:56
  the lab. I am the least interesting
• 59:58
  person to talk to of everyone on my team.
• 1:00:00
  And the next time you invite someone back from Databricks,
• 1:00:04
  you should talk to the most junior person on my team,
• 1:00:06
  because they are the one who has the most interesting and
• 1:00:09
  practical things to say. And how long do you feel like
• 1:00:14
  these iterations or the different ideas that you're
• 1:00:17
  running with, how long do those take to bear fruit or show that
• 1:00:22
  they're not working? And how do you look at like, all
• 1:00:25
  right, we're trying to replicate this paper.
• 1:00:27
  Is it days, weeks, months? And what does that process look
• 1:00:31
  like? It's months.
• 1:00:32
  It's always months. And I think I I run into this
• 1:00:36
  all the time with even folks at Databricks.
• 1:00:38
  So they'll come to me and say, hey, can you just test out this
• 1:00:40
  idea? He'll let me know tomorrow what
• 1:00:41
  you think my answer is. I'll let you know next quarter
• 1:00:44
  or next half or next year what this is to do good science.
• 1:00:49
  It takes time, but you know it's probably one month minimum to
• 1:00:54
  like. Reproduce the AER and test it
• 1:00:56
  carefully. It's six months for ideas that
• 1:00:59
  are worth putting six months into, and that's fine for a
• 1:01:03
  question like rope versus alibi. If you're about to train a model
• 1:01:06
  that will cost you millions of dollars, you sure as hell better
• 1:01:09
  put three months into that. If you're trying to figure out
• 1:01:12
  whether Laura is hype and it just happened to show up at the
• 1:01:15
  right time when a bunch of people were new to the field and
• 1:01:17
  thought that was the only way to do parameter efficient fine
• 1:01:19
  tuning. Whether Laura is genuinely a
• 1:01:21
  better way to do parameter efficient fine tuning, that's a
• 1:01:25
  decision that will have huge implications for all the models
• 1:01:27
  you build for a long time to come.
• 1:01:29
  You sure as hell better put months into that.
• 1:01:31
  When you're making big changes to your model architecture, you
• 1:01:34
  better make sure they scale or you're going to really regret
• 1:01:36
  it. So I think it is like this is an
• 1:01:41
  artisanal process. You know what we do at Mosaic is
• 1:01:46
  farm raised and organic and you know, hand tossed and you know
• 1:01:51
  what have you. But the end result is that every
• 1:01:56
  technical decision we've made about how we train our models
• 1:02:00
  and how we recommend other people train our models, I have
• 1:02:04
  a fact sheet that I can hand you to make the case that that was
• 1:02:07
  the right decision. And I think that there you
• 1:02:10
  can't, well, actually you can put a price on that, go train a
• 1:02:13
  $10 million model and then realise you didn't do it well.
• 1:02:16
  That's the price on that kind of thing.
• 1:02:19
  Yeah, that's the danger, right? That you you're like, oops, all
• 1:02:23
  that GPU, nothing to show for it.
• 1:02:27
  That is a little bit scary to think about.
• 1:02:29
  And so one other piece that I was wondering about, what is on
• 1:02:36
  your radar right now? What is interesting to you?
• 1:02:40
  So there are a bunch of things that are interesting to me.
• 1:02:42
  I'm always on the lookout for tweaks to, you know, model
• 1:02:45
  architectures or tweaks to the way we train that improve
• 1:02:48
  efficiency. I'm not on the lookout for
• 1:02:49
  breakthroughs. I think breakthroughs come once
• 1:02:52
  in a very long while and I can, I can rattle off, if you really
• 1:02:56
  want, a bunch of fun things that have come out in the past
• 1:02:59
  several years that everybody looked at as breakthroughs at
• 1:03:03
  the time or thought, oh, this might be the next big thing.
• 1:03:06
  And now like anybody heard of like NF Nets, the Normalisation
• 1:03:11
  free Nets that were supposed to replace resets and get rid of
• 1:03:13
  batch Norm. We still don't use them.
• 1:03:15
  They're they're really cool architecture and it was a really
• 1:03:18
  great research. Anybody heard of Vision
• 1:03:21
  Transformers? I don't know.
• 1:03:23
  I still train a lot of resets and Resnet seemed to have, I
• 1:03:27
  think, survived the vision transformer push and are still
• 1:03:30
  the thing that I go to first. There are all sorts of new
• 1:03:34
  attention schemes that have come and gone.
• 1:03:37
  So I'm not, you know, the next big thing in machine learning
• 1:03:42
  has come like once in a decade or two.
• 1:03:44
  If you look at convents and runs and LTMS and you know,
• 1:03:50
  Transformers, maybe diffusion models, like, they don't come
• 1:03:54
  around very often. And if you're looking for the
• 1:03:56
  next big thing, you're going to be sorely disappointed or you're
• 1:03:58
  just going to get taken by every LinkedIn post and tweet that you
• 1:04:01
  see. It's possible that yesterday we
• 1:04:03
  just had a huge breakthrough that's going to change
• 1:04:04
  everything. Or it's possible that we had
• 1:04:06
  another very exciting piece of science that doesn't end up
• 1:04:10
  replacing anything that we're doing right now could go either
• 1:04:13
  way. So I look for the incremental
• 1:04:15
  things I look for. I also look for kind of new
• 1:04:19
  questions people are asking. Like there are a lot of
• 1:04:22
  questions right now about how we measure the performance of
• 1:04:24
  models in long contexts. That's a really interesting
• 1:04:27
  question. Anthropic has great work on 200K
• 1:04:31
  context windows. They can fit it into memory.
• 1:04:34
  Can you actually take advantage of it in a meaningful way?
• 1:04:37
  Can the model actually like, you know, are they compromising on
• 1:04:40
  so much that the information gets lost or the model can't
• 1:04:43
  holistically take advantage of it?
• 1:04:44
  I don't know because I know how to measure that.
• 1:04:46
  I look for things on a valuation right now in better ways to
• 1:04:48
  evaluate models. I think that's a mess.
• 1:04:50
  I'm really interested in synthetic data right now because
• 1:04:53
  the textbooks are all you need papers from Microsoft.
• 1:04:56
  Those are really cool and those are compelling to me.
• 1:04:59
  The Llama Two folks kind of indicate in the paper that they
• 1:05:01
  did a lot of automatic generation of supervised fine
• 1:05:04
  tuning data once they bootstrap their model to a certain point.
• 1:05:06
  That's pretty intriguing because supervised tuning data is
• 1:05:09
  expensive and hard to come by. You know, I think a lot about
• 1:05:13
  debug ability of neural networks.
• 1:05:14
  Like how do you change a neural network once it's in flight to
• 1:05:16
  fix a problem. I think a lot about how end
• 1:05:20
  users can create new tasks. If a Databricks user comes to me
• 1:05:23
  and they say I want to fine tune your model to do this new
• 1:05:25
  business process, OK, they need a training set, they need a test
• 1:05:29
  set, and they need a metric to figure out whether they're doing
• 1:05:33
  well on that test set. Those are really nontrivial, and
• 1:05:36
  there's a lot of expertise that goes into developing any
• 1:05:38
  training or test set that's publicly available.
• 1:05:41
  How do you help them to do that quickly and with very little
• 1:05:44
  expertise so that they can just get back to work on solving
• 1:05:46
  their problems? Those are the kinds of things
• 1:05:49
  where I'm like, on the lookout for interesting ideas because
• 1:05:51
  those are real problems I have. And if I were talking to someone
• 1:05:54
  in academia right now, those are the problems that make a
• 1:05:57
  difference. Not like can you build the next
• 1:05:59
  transformer like that's a problem where the answer is no
• 1:06:03
  you probably can't because it took a lot of people trying and
• 1:06:06
  someone who got, you know, both very talented and also right
• 1:06:09
  place right time to actually do that.
• 1:06:12
  O yeah, I tend to be a realist on this.
• 1:06:18
  Damn, I should have led with that question because there's so
• 1:06:21
  many things that you just rattled off right there that I
• 1:06:24
  would love to dive into. But for the sake of time, I want
• 1:06:27
  to at least hear what your deep thoughts are on the debug
• 1:06:31
  ability because I know that is something that people it's
• 1:06:35
  almost like a freight train. You get on the train or as
• 1:06:38
  others put it by the ticket, take the ride.
• 1:06:40
  You can't quite hop off the roller coaster when it's mid
• 1:06:44
  flight per se, and so how are you trying to help that not be
• 1:06:51
  different? I have been trying to help that
• 1:06:53
  by talking to people smarter than me who are specialists in
• 1:06:56
  this, because I'm not. I have a friend, Julius Adebayo,
• 1:07:00
  who's I think currently a postdoc right now at NYU.
• 1:07:04
  He was in office made of mind during our PHD's and he's the
• 1:07:07
  person I turned to because he's actually written papers on it
• 1:07:10
  and studied the hell out of it. That is not an area where I have
• 1:07:12
  tonnes and tonnes of expertise and my job is not to pontificate
• 1:07:16
  on things I don't know, but to go and talk to smarter people
• 1:07:18
  than me. So happy to send you his info,
• 1:07:20
  you go talk to him. He's a cool guy to have on and
• 1:07:22
  you know, he's he's also known as the guy who ruined
• 1:07:25
  interpretability. That is to say, he wrote about
• 1:07:28
  papers during his PhD taking interpretability methods and
• 1:07:31
  falsifying them. He has a great experiment he ran
• 1:07:34
  where he took some vision works, ran interpretability methods on
• 1:07:38
  them to look for what they said, and then randomised various
• 1:07:41
  layers of the networks and ran the same interpretability method
• 1:07:43
  and got the same results even though the network was garbage.
• 1:07:47
  He applied science to a field that I think had diverged away
• 1:07:50
  from science and brought it to heel, and now he's done some
• 1:07:55
  cool work on debug ability and other neat stuff.
• 1:07:57
  That's someone who I would go and talk to.
• 1:08:01
  Incredible. Well dude, this has been
• 1:08:04
  fascinating. I appreciate you a tonne coming
• 1:08:07
  on here, Danny. I also appreciate you being here
• 1:08:11
  and helping me out with the with the support and questions now I
• 1:08:17
  had a blast and I look forward to getting to read the Mosaic
• 1:08:21
  blog again. I think by the time this podcast
• 1:08:25
  is out you'll probably have something up and ready for us,
• 1:08:29
  and then you'll have lots of other fun updates for us, and so
• 1:08:34
  I'll be waiting at the edge of my chair as time goes on to see
• 1:08:39
  what you all are doing. I know you're cooking up a lot
• 1:08:41
  of cool stuff that have special project names that I can't
• 1:08:44
  pronounce, and none of them surprisingly have anything to do
• 1:08:49
  with Q *, which is a bit of a let down, but I'll wait for when
• 1:08:54
  it comes out to see what actually happens.
• 1:08:58
  Thank you so much for having me. This has been a lot of fun as
• 1:09:01
  always. And, you know, maybe you'll
• 1:09:03
  invite me to a panel someday. You are not bad in my book
• 1:09:08
  exactly. You are definitely not bad and
• 1:09:11
  that is you keep it spicy. You've got other hot takes that
• 1:09:14
  I love and any last hot take before we go that you want to
• 1:09:19
  sign off with? My last hot take is, you know,
• 1:09:24
  do science and only say what you can back up with data.
• 1:09:30
  Boom. That's it.
• 1:09:31
  That's the podcast right there. Oops, you have to immerse
• 1:09:36
  yourself in the ML OPS content. The best way to do it is to
• 1:09:42
  subscribe to the E-mail OPS Community Podcast, so good luck
• 1:09:47
  and keep learning.