Latest / Agentic Conversations (formally mlops.community) / Rethinking Notebooks Powered by AI

Transcript

Rethinking Notebooks Powered by AI · Agentic Conversations (formally mlops.community)

• 0:00
  Oh, but how do you add the context?
• 0:01
  I do describe the data frame. Is the LM supposed to look at
• 0:05
  all of your code and infer what the column names are and what
• 0:07
  the types are? The way I thought about
• 0:09
  notebooks five years ago meets the change.
• 0:11
  If you want to use collab, that's fine, but you can rethink
• 0:13
  that and that shit. Like I care a lot about
• 0:15
  intellectual freedom. If you're a Python person,
• 0:17
  widgets are probably the most fun way to learn JavaScript at
• 0:19
  this point. OK.
• 0:25
  So you told me something that is very interesting.
• 0:29
  You guys just got acquired? Yeah, a few weeks ago, yeah.
• 0:32
  True. So in the span of me seeing you
• 0:35
  a month ago and now, what's changed?
• 0:40
  So I thought, OK, I saw you at π data the week after I was in San
• 0:43
  Francisco. And as one does, you get into a
• 0:45
  few weimos and do that fun. Take some video.
• 0:49
  Yeah, like up a hill, down a hill.
• 0:52
  But yeah, like during that week I also got word that we were
• 0:56
  getting acquired. The CEO and CTO were really good
• 0:59
  at keeping that sort of quiet within the company, but sort of
• 1:02
  a fun pleasant surprise. But the the short story of it is
• 1:05
  that we are still one group. Our road map hasn't really
• 1:08
  changed. We are part of wait some bite
• 1:12
  while we, if you look at the org chart, you've got Corweef, which
• 1:15
  is the company officially that acquired us.
• 1:17
  Then we have weights and biases and then we are part of the
• 1:21
  under the weights and biases umbrella.
• 1:22
  I guess you could say if you look at the org chart, but I
• 1:26
  mean again, the team is still the same.
• 1:29
  Our road map is definitely still the same.
• 1:30
  It's just that we are now part of a bigger group that's the
• 1:34
  longest short. Story of it makes a ton of sense
• 1:35
  why weights and biases would want you after you're breaking
• 1:39
  down all this notebook stuff and knowing like how they think
• 1:42
  about helping data scientists and what are some things that
• 1:47
  you're thinking about now? Has anything changed in that
• 1:49
  regard? Like your what?
• 1:52
  You want to build where you're going, I mean.
• 1:54
  Honestly, on a day-to-day basis, the main thing that's changed is
• 1:57
  I'm part of a larger Slack group now.
• 1:59
  So, so you learn to sort of ignore more channels and to
• 2:02
  really like do a little bit of that because before with eight
• 2:04
  people, like almost every channel is relevant.
• 2:06
  And now, you know, with a couple of 1000 people working there,
• 2:08
  it's that's not the case. So that's the, in a lot of ways,
• 2:10
  that's the main thing. And like one thing that's a
• 2:13
  little bit funny is I actually knew a bunch of people that
• 2:16
  already worked at weights and biases.
• 2:17
  So one thing you do see is that, you know, obviously the, our CEO
• 2:21
  and our CTO, they talk to some people over the weights and
• 2:23
  biases side, kind of like very much up the umbrella.
• 2:26
  And I know some people that aren't exactly up the umbrella,
• 2:28
  but we're like Twitter buddies. So then, OK, we are now each
• 2:31
  other. I'm I'm my gateway to this guy
• 2:33
  into Marimo and he's my gateway into weights and biases stuff.
• 2:36
  And you know that a lot of that is just very organically and
• 2:39
  natural. But like that, that, that is the
• 2:40
  main new thing because you know, we still do the YouTube stuff
• 2:45
  that do the education thing. We still do integrations with
• 2:48
  other open source packages and like all of that is basically
• 2:52
  still the same. The one thing is that we did
• 2:57
  kind of a funny story. So, you know, collab from
• 2:59
  Google, the number one ticket on the collab get up issue list was
• 3:04
  the people wanted it to have Marimo support in Google collab.
• 3:09
  Like so we built a thing called Moab and our intern wrote on the
• 3:14
  issue tracker like, yeah, we built it instead.
• 3:15
  It's kind of a funny little internal joke there.
• 3:18
  But we have we we made a thing called Moab, which is basically
• 3:21
  just hosting Marimo in the cloud.
• 3:23
  And one thing you can imagine is that we, you know, with the
• 3:25
  whole core we've backing us, we can do more elaborate things
• 3:29
  with cloud compute for MO Lab. Like that's definitely a thing
• 3:31
  that, you know, we can start dreaming a little bit there.
• 3:35
  So that's. Like how so?
• 3:36
  I didn't quite follow that, just basically hosted notebooks now.
• 3:41
  Yeah. So we've, we definitely, we've
• 3:43
  got free credits with our notebooks that was always CPU
• 3:47
  based. But you can imagine we might do
• 3:49
  like bigger machines and maybe a GPU or two because core weave
• 3:53
  has a bunch of them, right? That's like stuff that we, you
• 3:57
  know, we have in the pipeline. But the, the thing about the
• 4:00
  setup that I, I guess I want to emphasize that's really neat is
• 4:04
  one way to make MOLAP better is to make Marimo better.
• 4:09
  So you can definitely imagine that there's this incentive
• 4:12
  structure where if we want to, if we want MOLAP to be better,
• 4:14
  we definitely have to keep on investing in the open source
• 4:16
  thing. And that is the main thing.
• 4:20
  Like that was also in the back of our minds during the
• 4:21
  acquisition. Like we can do a lot of cool
• 4:23
  stuff, but only if we can find a nice way to keep that open
• 4:26
  source thing very much alive because we don't want to change
• 4:27
  that license. We just want that thing to just
• 4:29
  keep on existing. And by having this thing where
• 4:31
  Molap is, you know, a thing that we can work on, but the way you
• 4:35
  make that thing better is by investing in the open source
• 4:37
  thing. That is a very healthy incentive
• 4:39
  to have as well. So that's that's, that's the
• 4:40
  main thing I would like to emphasize.
• 4:42
  OK, real quick, if you happen to find yourself in the South Bay
• 4:46
  on March 3rd, we're going to be taking over the Computer History
• 4:50
  Museum for our Coding Agents conference.
• 4:54
  That's right, we're organizing another conference despite me
• 4:57
  telling myself so many times after we did our AI quality
• 5:01
  conference that I would never do another conference because of
• 5:03
  the stress that it caused. Somehow I seem to have broken my
• 5:07
  vows because the pole was too strong.
• 5:09
  Some of the notable speakers that we have already announced
• 5:12
  are Sid friend of the pod and Co creator of Claude Code, Harrison
• 5:17
  Chase, the founder of Lang Chain, good old Dex, the founder
• 5:21
  of Human Layer, and the dude who popularized the term context
• 5:25
  engineering and harness engineering.
• 5:28
  And last but not least, someone I consider a very close friend,
• 5:32
  Mikhail Eric is going to be doing a workshop.
• 5:35
  He's a Stanford lecturer and actually he was a past Co host
• 5:39
  of this here podcast that you're listening to, you know, before
• 5:43
  he got all famous and stuff. Come join us.
• 5:46
  It's intimate by design. There's only 450 people that
• 5:50
  we're going to let in to the room to try and keep the signal
• 5:53
  as high as possible. So I'll see you there real
• 5:56
  quick. Let me talk to you about
• 5:56
  hyperbolic GPU Cloud. It delivers NVIDIA H1 hundreds
• 6:01
  at $1.69 per hour and H2 hundreds at $1.99 per hour.
• 6:09
  And this is with no sales calls, no long term commitments or
• 6:14
  hidden fees. You can spin up one GPU or scale
• 6:17
  to thousands in minutes with VMS, bare metal clusters and
• 6:22
  high speed networking. You've also got attachable
• 6:25
  storage and you only pay for what you use.
• 6:29
  Save up to 75% less than legacy providers.
• 6:33
  And Oh yeah, by the way, you need steady production grade
• 6:36
  inference. Well, you can choose dedicated
• 6:39
  model hosting with single tenant GPU's and predictable
• 6:43
  performance without running your own hardware.
• 6:48
  Try now at app dot hyperbolic dot AI.
• 6:52
  Let's get back into this show. And what are the synergies with
• 6:56
  weights and biases the product? I mean, we're, that's like stuff
• 7:00
  that we are exploring right now, right?
• 7:03
  But it's definitely Weights and Biases has historically always
• 7:06
  kind of been, you know, they've been fairly friendly to like the
• 7:09
  academic community as well. I guess it's a little bit less
• 7:14
  in Europe, but my impression is definitely it's like a lot of
• 7:17
  companies in California just use Weights and biases.
• 7:19
  It's kind of a somewhat normal thing.
• 7:20
  So you can definitely imagine like some integrations, one of
• 7:23
  the first things that we've added.
• 7:27
  So Weights and Biases also has like an LM inference engine.
• 7:30
  The day after the acquirer, the acquisition was public, we just
• 7:33
  added support for that inside of REMA notebooks.
• 7:35
  We also have like very elaborate LM support.
• 7:37
  So that's like a click away now and you know.
• 7:39
  LLM training support or within the notebook.
• 7:43
  So weights and biases has a service where you can call Llama
• 7:47
  and a couple of those opens, I think they're all open source
• 7:49
  models if I'm not mistaken. So like that's, that was like an
• 7:53
  easy integration. We could just add on day one.
• 7:55
  Basically. I cannot talk about the main
• 7:59
  specifics just yet, right, because people are working on
• 8:01
  stuff, but you can definitely imagine integrations happen.
• 8:03
  Yeah, yeah, OK, awesome. You're like a treasure trove of
• 8:09
  cool stuff. I guess like one thing that that
• 8:12
  does make Marimo unique is the reactivity, right?
• 8:15
  But then came along these LLMS, so people were sort of banging
• 8:19
  on door like, oh, can you do LLM stuff too?
• 8:21
  And there is interesting stuff happening there right now.
• 8:23
  So we for the longest time, we have agent support.
• 8:29
  So you can open up a little window and you can give your API
• 8:31
  key of Claude and then you can, you know, do your Claude stuff.
• 8:35
  What you could do is you could say, oh, there's a data frame
• 8:37
  there. Hey, I want you to write SQL for
• 8:40
  that data frame. Oh, but how do you add the
• 8:42
  context? I do describe the data frame.
• 8:45
  Is the LM supposed to look at all of your code and infer what
• 8:48
  the column names are and what the types are?
• 8:50
  So that was always a bit of a conundrum.
• 8:52
  So what we do is you're able to point, do at and then some sort
• 8:55
  of variable. And then we do our best to add
• 8:57
  any context we can add to the prompts.
• 8:59
  So if you point to a data frame, we can automatically say things
• 9:02
  like, oh, that data frame has these columns, it has this many
• 9:04
  rows. These are the types of each
• 9:07
  column and what do you know that makes it a whole lot simpler for
• 9:09
  the LLM to actually write the sequel or the data frame code
• 9:12
  you want. So that's definitely interesting
• 9:15
  stuff that we can do. But then I think Zed, the editor
• 9:19
  started this thing with ACP. So can we have an agent protocol
• 9:24
  as well such that you can just say, oh, I want my editor just
• 9:26
  have clawed, have that be a protocol that's just added
• 9:28
  clawed code or Codex or Gemini. I think they've also added
• 9:31
  support for it, so there's an experimental feature in Marimo
• 9:34
  now that also lets you just run clod code within Marimo itself.
• 9:37
  So you can just use native clod code with your keys and your
• 9:40
  subscription within Marimo 2 via ACP protocol.
• 9:43
  But it will wait. ACP or MCP?
• 9:46
  So we also support MCPACP is a relatively new thing.
• 9:49
  OK, it's a new one. Great, all right.
• 9:52
  But that just means that now clod code is generating code
• 9:56
  inside of your notebook and you. Don't have to do the API key
• 10:00
  thing where we every time you make a request you have to pay
• 10:02
  for it. You can user Claude description.
• 10:05
  We also have a linter that's specific to like Claude because
• 10:09
  there's like some errors that are really specific to Marimo.
• 10:12
  You can have a linter trigger every time that Claude does
• 10:15
  something and then we do linting.
• 10:16
  And one thing you can't do Marimo, for example, is you
• 10:18
  cannot define variables twice. So if there's one cell that has
• 10:21
  a = 1, the other cell that says a = 2, we have to make
• 10:24
  assumptions about what the actual value of a is.
• 10:27
  So we just don't allow for that and there's like errors like
• 10:29
  that that we can then again send off to Claude, it can fix
• 10:32
  itself. So there's a lot of like LLM
• 10:34
  innovation that's happening too in the notebook is the main
• 10:37
  thing I would also like to stress.
• 10:39
  Yeah, tell me more. I mean, so my dream would be
• 10:41
  that eventually we might be able to.
• 10:44
  We're definitely not there yet. That's the one thing I want to
• 10:46
  mention first. But the one thing I would love
• 10:48
  is that you could say things like, oh, actually I want to
• 10:51
  have a copy to clipboard button and that button doesn't exist
• 10:54
  yet, but we can just generate one that's in any widget that we
• 10:57
  just want to generate on the spot.
• 10:59
  So that dynamic UI type thing. That's the thing I would love to
• 11:03
  do more of. And you can also imagine I need
• 11:06
  to find time for this. Like one thing I would also love
• 11:08
  is that we get these LLMS in such a place where, oh boy, this
• 11:11
  is a super interesting article. Look at this really cool diagram
• 11:14
  in it. Hey, LLM in the notebook.
• 11:16
  Can you reproduce this in some way?
• 11:19
  But at sliders so I can play around and we're definitely not
• 11:23
  there yet, right? But that will be a really cool
• 11:24
  learning experience. Like, OK, this is the latest
• 11:26
  paper. Oh, that mechanism looks kind of
• 11:28
  cool. I won't have a better intuition.
• 11:30
  Can I have sliders and interactive matrices and just
• 11:32
  the plot updates based on the thing that's in the paper and
• 11:34
  like what's the quickest path to maybe get there and another?
• 11:39
  On your own data or on the data from the paper, it would have to
• 11:42
  go out there and get that data or.
• 11:44
  Well, I mean the data has to be public and if the data set is
• 11:47
  huge like, you know, petabytes, then.
• 11:50
  It's going to be a little bit more it's.
• 11:51
  Going to be a little bit harder, but if you're dealing with the
• 11:53
  intuition, then you probably want to.
• 11:54
  Then it's more about generating the Pytorch code and can we get
• 11:57
  the phenomenon at least visualized and things like that.
• 12:00
  Another thing about Marimo notebooks that's good to mention
• 12:02
  is that these days a lot of them do work in web assembly, so you
• 12:07
  don't need the Python back end anymore.
• 12:08
  So if you want to have a blog with the sliders and all that
• 12:11
  jazz is generated with Python, you can actually do that.
• 12:14
  The downside is not every Python library supports WASM via
• 12:17
  Piadite just yet. But that's like also something
• 12:19
  like. I would love for people to maybe
• 12:21
  also have like a good caching mechanism, such as you write
• 12:23
  your notebook fully in Marimo and then you export it
• 12:27
  statically to HTML but you maintain.
• 12:29
  The dynamics, yeah, that's. We're not quite there yet, but
• 12:32
  like via Piadite, we do see a path towards it.
• 12:35
  Going back to that dynamic UI generation, what do you feel
• 12:38
  like would be behind the scenes happening in order for that to
• 12:44
  be like a reality? In my mind, the biggest hurdle
• 12:47
  right now is that some Python developers are a little bit
• 12:50
  arrogant and refuse to learn JavaScript.
• 12:53
  And if you refuse to learn JavaScript, it's going to be
• 12:56
  hard for you to debug JavaScript.
• 12:58
  We just heard last night in one of the talks from our agents in
• 13:03
  production conference, they were saying, OK, what do you think
• 13:06
  the hardest code to generate is for cloud code or cursor?
• 13:12
  And it had JavaScript, Python, Rust.
• 13:16
  And people are saying like, Oh well, the best is Python.
• 13:19
  By far the best is going to be Python.
• 13:22
  The worst? I don't know.
• 13:23
  But I think OK. My guess is the best is probably
• 13:26
  JavaScript, the worst is probably Rust.
• 13:27
  No, the worst is JavaScript, because there's so much
• 13:31
  JavaScript on the Internet and a lot of it is really bad.
• 13:35
  Interesting. So it just makes it much harder
• 13:40
  to generate quality JavaScript. Yeah.
• 13:43
  So I, so The thing is I, I can do JavaScript, but I know that
• 13:45
  I'm not an expert at it. So what the bias that I have is
• 13:48
  that I do spend more time like getting the prompts just right,
• 13:52
  planning it out really carefully.
• 13:53
  And with the Python, I kind of yellow it because I can always,
• 13:55
  it's easier to fix. So I've not really had that
• 13:57
  problem. But again, I'm, yeah, it depends
• 14:00
  on what you're building to. Like if you're building a web
• 14:01
  app, that's also a different story than the stuff that I'm
• 14:04
  doing. But getting back to this.
• 14:06
  So developers, Python developers reluctantly use JavaScript, yes?
• 14:11
  And you feel like that would be useful to dynamically create
• 14:17
  these UI elements? Or even not dynamically, just in
• 14:20
  general. Like one weird bit of feedback
• 14:22
  that I've gotten is like, Vince, did you make all these widgets?
• 14:24
  Why? And the honest answer is because
• 14:26
  for some reason, no one, no one else is making them.
• 14:29
  And it feels such a natural thing like if you, if you love
• 14:31
  if, I mean you don't even have to love it.
• 14:33
  If you work a lot in notebooks, you want to have widgets, they
• 14:34
  just make your experience so much better.
• 14:36
  So why am I the only one? I think it's because I, at a
• 14:41
  very early part of my career was just falling in love with D3.
• 14:45
  So I didn't care. I'll just learn JavaScript
• 14:47
  because I want to do D3 and I'll also learn Python because I want
• 14:49
  to do so I could learn. But I think a lot of people just
• 14:53
  are very reluctant to learn JavaScript just because of the
• 14:55
  Node stuff. That's my impression.
• 14:59
  Yeah. And but it does level you up to
• 15:01
  be able to go and execute so much more when you know that
• 15:04
  front end. I think so, yeah.
• 15:06
  Again, it's a mystery to me why someone would not teach
• 15:08
  themselves JavaScript in this day and age where everything is
• 15:10
  a web app. Yeah.
• 15:11
  And it's, it's like 90% generated for you.
• 15:15
  So a little goes a long way in this case.
• 15:18
  I would definitely agree, although I there's also that
• 15:24
  there's also that aspect of it of, OK, should people still
• 15:27
  still learn how to program And like, that's a, that's a theme
• 15:30
  that I hear people talk about. That's a.
• 15:32
  Little crazy statement in my mind.
• 15:34
  Yeah. So I'm actually kind of of the.
• 15:36
  So I like to do flash cards and things like that just to sort
• 15:39
  of. I think if you train your
• 15:40
  memory, it also makes it easy for you to remember and at least
• 15:44
  helps with the iteration speed and also makes it a little bit
• 15:47
  less intimidating. Like, OK, you know the steps to
• 15:49
  start a new project. You know the steps to like, OK,
• 15:51
  what, what can we expect from the document of the, of the
• 15:53
  browser and what are good methods?
• 15:56
  But I think also if you just really go deep a little bit,
• 15:59
  that is also less intimidating because otherwise you get into
• 16:02
  this self helplessness of, oh, I can only build stuff if I have
• 16:04
  an LLM around. Like yesterday Cloud Fire went
• 16:06
  down for a bit and like, I'm I'm hoping people were still
• 16:10
  productive. Productive day of the year.
• 16:12
  Yeah, but does the national intelligence go down whenever
• 16:16
  Cloudflare is now in because no one can resell what God knows
• 16:19
  where that sort of anti utopia we're living in?
• 16:22
  But you know that graph, right? That is like when you start a
• 16:25
  project, you think, oh, it's going to be so easy, and then
• 16:28
  you go into the. Yeah.
• 16:29
  Disillusionment. You basically got to get past
• 16:32
  that trove before or through I think is how it's pronounced.
• 16:36
  Or you have to train yourself to be good at project set up.
• 16:39
  Like if I remember very early on, I was sort of just playing
• 16:43
  around with these things. I was kind of thinking like, OK,
• 16:45
  let's see if I can I fully Yolo and I won't look at the code at
• 16:48
  all, but can I fully Yolo a web app and have it send emails and
• 16:52
  Postgres and set all that stuff up?
• 16:53
  And it did. And then I noticed that whenever
• 16:55
  I was running unit tests, I instructed it to start with an
• 16:59
  empty database. Apparently just.
• 17:02
  Delete the database. It would go to prod, delete the
• 17:04
  whole thing and run all the unit tests there.
• 17:07
  And if you just set up the project yourself manually or if
• 17:10
  you just have a good template or something like that, such a that
• 17:12
  that kind of crap just doesn't happen, I think you're already
• 17:14
  in a much better spot. And my impression is that all
• 17:17
  these, what's a polite term, teenage tech Bros.
• 17:24
  Vibe. Coders, yeah, if I whatever you
• 17:25
  want to call them, but if you missed that first bit, like how
• 17:28
  do you actually set up a proper project?
• 17:29
  You know that that's that's where the flaw is, not
• 17:31
  necessarily in the LLM. It's the first bit that
• 17:33
  foundation. I think so, yeah.
• 17:35
  Yeah, it is interesting to to see that now, since it is easier
• 17:42
  to put a front end on things with Vibe coding, the
• 17:47
  expectation has kind of been raised.
• 17:51
  So now, whereas before you would probably be doing a lot of stuff
• 17:55
  with Streamlit, now you can be like, well, let's see if we can
• 17:59
  do this in React. Yeah, but I would still be in,
• 18:03
  especially if it's experimental. I would still be in favor of
• 18:04
  first doing a widget then. If you're a Python person,
• 18:07
  widgets are probably the most fun way to learn JavaScript at
• 18:10
  this point because you might still not need the whole node
• 18:15
  pipeline and you might still be something you can use right away
• 18:17
  together with your Python code that you love to use, right?
• 18:21
  So that would be the way I would personally think about it.
• 18:24
  One other thing, I have a Flash card app that I'm making.
• 18:29
  And let's say you could do that with something like HTMX such
• 18:32
  that what we send over the wire is not Jason, but full HTML.
• 18:35
  And every, every time I see a new Flash card, we send HTML
• 18:38
  back, I see a new card. Next makes a request.
• 18:41
  I see a card that works, but it's not the ideal user
• 18:45
  experience. The ideal user experience would
• 18:46
  be, OK, I want to practice 20 cards, let's send one Jason
• 18:49
  payload back. I only make the request once
• 18:52
  that's in memory. Now I do all of my card stuff
• 18:54
  and then at the end I send the request back because that way
• 18:58
  it's subtle, but there's a noticeable time lag because it
• 19:00
  has to make every single request when it fetches a new card.
• 19:04
  In the end, if you care about user experience, you have to
• 19:06
  care about JavaScript on the front end.
• 19:07
  And it's fine if you say I don't care about it, but if you are
• 19:11
  product minded, you have to. And that's, that's what bother.
• 19:15
  That's what something that does bother me in the Python
• 19:18
  community sometimes, like there's a bit of an arrogance
• 19:19
  for JavaScript, which I, I plainly find unfounded.
• 19:24
  Yeah. Well, talk to me more about the
• 19:26
  LLMS in Miramo, because it feels like you've got other ideas.
• 19:32
  Some of this depends a lot on what standards become available.
• 19:36
  So a big reason that we were able to push for the LLMS the
• 19:38
  way that we have was that because at some point MCP was a
• 19:41
  thing and because at some point ACP was a thing.
• 19:43
  And so in a lot of ways we are dependent on that.
• 19:48
  So I can dream, but I, if we start planning, we have to sort
• 19:51
  of keep that in mind. A thing that we are
• 19:53
  experimenting with more now though is not so much, OK,
• 19:56
  people know how to use an LLM inside of a notebook, but can we
• 20:00
  bootstrap notebooks from the start?
• 20:02
  So we have this Remo new command and you can give it a prompt and
• 20:06
  what you are able to do is say like oh there's a CSV file right
• 20:09
  there. Make the first notebook version,
• 20:11
  it has to do these analysis and we'll just bootstrap on from the
• 20:14
  start. And the benefit of it being a
• 20:16
  notebook is that you can look at all the intermediate results and
• 20:19
  it's very visual you can add charts.
• 20:21
  So it's also a lot easier to spot the mistake.
• 20:24
  And I think that's kind of the the most interesting thing if
• 20:26
  you plump everything in an IDE and sure you can add some unit
• 20:29
  tests, but it's hard to see the intermediate results and it's a
• 20:32
  lot easier to poke around in a notebook.
• 20:35
  So that's a little bit more the thing that I'm I'm I'm
• 20:37
  personally very interested in so.
• 20:38
  You know what? I actually, yeah.
• 20:40
  What that just sparked in my mind is how notebooks are very
• 20:46
  made for humans. Yeah.
• 20:47
  And agents, I was trying to figure out like, well, is there
• 20:51
  a world where an agent would want to fire up a notebook and
• 20:54
  try and see things through the eyes of a notebook, Which I
• 20:58
  don't think really makes sense. But there is a world where
• 21:03
  you're interacting with the agent and when it's giving you
• 21:07
  results, you want to see it as a notebook or you want to be able
• 21:11
  to change different contexts that it has through a notebook.
• 21:15
  So one thing that I mentioned before, right?
• 21:17
  So you have the AT dataframe thing, you can point to a
• 21:19
  variable. You can also do AT errors and
• 21:22
  any errors that appear in the notebook go to the LM.
• 21:24
  You can also do that with a cell output.
• 21:27
  So if the cell output is some sort of weird widget that we
• 21:29
  don't necessarily have a good representation for, if the model
• 21:32
  in the back is multimodal, we just make a screenshot and send
• 21:36
  that. So when you say LMS can't see
• 21:38
  the notebook, well, they kind of can.
• 21:40
  It's just that they're not always as good as interpreting
• 21:43
  charts yet. Like that's more the thing, but.
• 21:45
  Data, just the raw data, right? It's.
• 21:47
  More. Well, we're not going to.
• 21:49
  Well, if you have a GB of data locally, you're not going to
• 21:51
  send that to the context anyway. But what you can do is say, OK,
• 21:54
  let's send the 1st 5 rows and the schema.
• 21:56
  That should already make it a whole lot simpler for the LM to
• 21:58
  write the right query, right? So there's definitely this
• 22:01
  interaction. It's just that we're only
• 22:02
  scratching the surface because we're just starting out with
• 22:04
  this. One thing I will say, under the
• 22:08
  hood, a Marimo notebook is a Python file.
• 22:10
  So we don't have that weird Jason structure thing that
• 22:12
  explodes and gets like, that's the thing we don't have.
• 22:14
  It's a Python file. So one thing that I've been
• 22:17
  experimenting with is you could build your Python package in
• 22:20
  Marimo because it's a Python file.
• 22:22
  So kind of what Jeremy Howard's doing with MP Dev, like there's
• 22:26
  no reason why you couldn't do that inside of Marimo.
• 22:28
  So a example, I like to write my command line apps in Marimo
• 22:32
  nowadays because with Typer, you can do like a decorator on a
• 22:35
  normal Python function. Marimo also supports π tests.
• 22:38
  So if you have a cell and there's functions in there, they
• 22:42
  all start with like test under score that we detect.
• 22:45
  Those are probably π tests. So you can add your unit test in
• 22:47
  notebook as well. And we can detect when the
• 22:50
  notebook is running in the browser or when it's running
• 22:53
  from the terminal. So I can define functions in
• 22:57
  Python that you can run from the command line.
• 23:00
  And with typer, I can reuse the same function inside of pytest.
• 23:02
  And it's all one Python file. Oh, and we support UV, so you
• 23:05
  can add your dependencies on top of the file and then UV will
• 23:08
  take care of all the dependencies and the right
• 23:10
  versions and all that for you too.
• 23:12
  Oh, and then we add LMS to the mix.
• 23:14
  So like, again, like I'm, I'm pretty comfortable saying that
• 23:18
  there's so much stuff still left to uncover.
• 23:21
  But the main thing that the notebook is really good at is
• 23:24
  you can stare at it and you can, you can fiddle around with it,
• 23:27
  you can play around with it. And something about that is very
• 23:30
  good for the human mind to also understand what's happening
• 23:32
  under the hood. Yeah, that exploration, yeah.
• 23:35
  And and that doesn't just mean data stuff, it also means
• 23:37
  command line apps. And there's like the flash card
• 23:42
  thing that I mentioned, I need charts to debug the like, timing
• 23:47
  between flash cards, learning aspects.
• 23:49
  There's a name for it, repetitive something, something.
• 23:52
  But for that aspect of the app I have, I need to have charts to
• 23:54
  debug. Yeah.
• 23:55
  And Marimo gives me a development environment that's
• 23:57
  really good, giving me charts as I'm developing.
• 24:00
  And it's a Python file, so I can export that one function that I
• 24:03
  really want in the notebook to my Flask app.
• 24:05
  And it just works. Oh, and we have LMS, Yeah.
• 24:09
  Actually, a common use case that I've seen with agents to debug
• 24:13
  them is creating heat maps to figure out if they called the
• 24:17
  right tools or if they did executed the right tasks.
• 24:20
  Now is the time to start dreaming about those things, I
• 24:22
  guess is the main thing I want to say.
• 24:23
  Because the way you think about the way I thought about
• 24:25
  notebooks five years ago needs to change.
• 24:29
  Because, you know, Python notebooks are better now.
• 24:32
  That's the. Word collab being restricted by
• 24:36
  if. You want to use collab, that's
• 24:37
  fine, but you can rethink that and that should like I care a
• 24:39
  lot about intellectual freedom. My number one fear is that
• 24:42
  people, I think if you want to succeed in this field, that the
• 24:45
  only thing you got to do is like watch YouTube videos and books
• 24:48
  of other people and that they stop thinking of new ideas on on
• 24:50
  their own. What I really like about Marimo
• 24:52
  is it does invite play, so you're able to explore your own
• 24:55
  ideas a bit more. And it's that intellectual
• 24:57
  freedom that I find. Yeah, we need more of that and
• 25:01
  we need more people making interesting stuff.
• 25:02
  You need to produce and not just consume content.
• 25:04
  I mean, we need more good ideas in the ecosystem.
• 25:07
  That's how we move the ball forward.
• 25:08
  And yeah, in my mind, when I first saw Marimo, the reactive
• 25:12
  nature of it actually corrected me a few times, which is neat.
• 25:15
  But I also just noticed like, oh, if I can mix UI with Python
• 25:19
  code, oh, that's super amazing for prototyping.
• 25:21
  And oh, I found a mistake in my data analysis there because a
• 25:24
  slider shouldn't work that way. OK, great.
• 25:26
  Like I'm aware of the fact that I'm like a guy who works in
• 25:30
  Marimo is like making a massive pitch from Marimo at this point.
• 25:33
  So like, dear listener, try to sympathize.
• 25:35
  But the main that I do want to emphasize, like the dreaming
• 25:38
  aspect of it is the cool bit. That's the one thing I want to
• 25:40
  like, double down on. I think that's a perfect place
• 25:42
  to end it, so that is awesome. Cool.
• 25:45
  Yeah, thanks for having me in this fancy studio.
• 25:49
  And it looks fancy from this side, trust me.
• 25:52
  On. This side not so fancy.
• 25:53
  This is all cables, yeah, and. Cameras.
• 25:55
  What the viewer cannot see is no problem.
• 26:01
  OK, final pitch then. We have a YouTube channel.
• 26:03
  There's videos that explain everything.
• 26:05
  Check it out. Check out Molab.
• 26:06
  We have a bunch of free compute as well.
• 26:09
  Feel free to use it and we have a cool discord.
• 26:12
  We have a great memes channel.