Latest / Agentic Conversations (formally mlops.community) / Trust at Scale: Security and Governance for Open Source Models // Hudson Buzby // #338

Transcript

Trust at Scale: Security and Governance for Open Source Models // Hudson Buzby // #338 · Agentic Conversations (formally mlops.community)

• 0:00
  There's a ton of value there that I think the industry is
• 0:02
  like just starting to explore and how they, yeah, diversify
• 0:07
  the requests more efficiently, use resources that starts with
• 0:13
  some form of a centralized gateway and some like semblance
• 0:16
  of control as to how these requests are being made.
• 0:21
  And if everything is just sporadic and every team is using
• 0:23
  a different model and you have no idea, those kind of an end
• 0:27
  goals won't be, you can't accomplish them.
• 0:35
  What's the obsession with point break?
• 0:39
  It's like one of the greatest movies ever.
• 0:41
  I don't think I need to. I don't think I need to defend
• 0:44
  Point Break by any means. You know, great pure surfing
• 0:50
  scenes, just peak Patrick Swayze, some great philosophical
• 0:54
  elements. Kathryn Bigelow's kind of debut
• 0:58
  movie leading into a fantastic career.
• 1:03
  All the above. Love that movie.
• 1:04
  And Keanu. Just anything with Keanu Reeves
• 1:08
  is well worth. Well worth.
• 1:11
  Very true. The reason I say that is because
• 1:14
  you mentioned that you've seen point break over 283 times, I
• 1:18
  think. But who's going on that?
• 1:20
  Yeah, a rough estimate. I think I would probably say I'm
• 1:26
  the same way for The Big Lebowski.
• 1:29
  Oh, OK. I can't tell you how many times
• 1:32
  I quote that movie like weekly or I send people memes when or
• 1:39
  gifts when I'm like, yeah, that's just your opinion man.
• 1:44
  So I feel you on that. That is great.
• 1:47
  So anyway, man, well, we're going to get into everything
• 1:52
  about the enterprise and their adoption of AII think it's
• 1:57
  timely because we've all seen the headline of like 95% of AI
• 2:04
  initiatives in the enterprise are doomed to fail.
• 2:08
  And it reminds me for anybody who's been in envelops for
• 2:12
  longer than a few years back in the day when we said that stat
• 2:18
  that either Gartner said or somebody said once and every
• 2:22
  single company quoted it on their website, like 80% of all
• 2:26
  data science initiatives are do not make it into production.
• 2:32
  And so everybody put like, but if you use our tool, you will
• 2:37
  make it into production. That was kind of the marketing
• 2:39
  speak. Now you you remember that one?
• 2:44
  I think we had it in, in one of our, our intro slides for Quack
• 2:50
  the, the previous product that I was working for.
• 2:52
  Yeah, absolutely. And it was true.
• 2:55
  I, I don't know that was in general.
• 2:57
  I talked to a lot of customers and prospects and the, when you
• 3:00
  ask them how is your ML environment, how is it going?
• 3:06
  Maybe 10% say it's in a good state, like we understand what
• 3:09
  we're doing. But more often than not, it's,
• 3:12
  it's generally stitched together, barely holding on for
• 3:16
  dear life. And but that's kind of the on
• 3:19
  the traditional side, I think with gender of AI, it's kind of
• 3:22
  switched into now everybody's in production with it or what they
• 3:25
  consider production, but it's a different kind of type of
• 3:30
  development that's taking place. Yeah, different kind of
• 3:33
  production, right? I was talking to a buddy who
• 3:35
  works at one of these big fintech companies and he said,
• 3:39
  yeah, it's super easy. We just, whenever we can
• 3:42
  outsource all of the infrastructure worries to open
• 3:47
  AI, Yeah. And you know, I think for
• 3:51
  certain implementation certain services that works.
• 3:56
  But yeah, I think it's kind of a a fundamental shift happening
• 4:01
  between kind of how data science and data science teams were
• 4:05
  originally defined as kind of their own little sector of an
• 4:09
  organization that like, you know, maybe C-Suite threw some
• 4:13
  money at them and they're kind of experimenting with projects
• 4:16
  for a few years, getting a few ideas.
• 4:18
  Maybe they get something that works, but it's not necessarily
• 4:20
  taken seriously as a part of the engineering org or strictly
• 4:25
  enforced from like dev OPS standards, infrastructure
• 4:28
  standards. And you know, now you're seeing
• 4:31
  that change with again, now they're getting things into
• 4:34
  production and now the rest of the organization, back end
• 4:37
  teams, front end teams are starting to adopt AI services
• 4:40
  where there's a bit of a disconnect in terms of, you
• 4:44
  know, they have not traditionally treated their
• 4:47
  services in a production like way with tests, multiple
• 4:51
  development environments, like all those things.
• 4:53
  You you ask data scientists that they're doing it and they kind
• 4:56
  of jump scare a little bit. But as all these services start
• 5:01
  getting incorporated into like more traditional engineering
• 5:04
  environments, they're going to be need to be treated with like
• 5:07
  a higher level scrutiny than what's kind of the norm
• 5:11
  currently. You know what I was thinking
• 5:13
  about this week and it was for the swamp up talk that I'm
• 5:18
  giving. I'm thinking like, oh, what is
• 5:20
  some stuff that I can mention has changed over the last year?
• 5:25
  And I was trying to figure out what the different development
• 5:29
  cycles are when it comes to predictive ML versus this new
• 5:34
  generative AI work flows. And I'm qualifying generative AI
• 5:40
  as almost like image video generation.
• 5:42
  And then the other stuff I would say is like a gentic development
• 5:48
  and the life cycle there, because I think you probably
• 5:52
  have RAG that fits in there somewhere.
• 5:54
  But most people are trying to do things agentic more than RAG, I
• 5:59
  think these days because RAG kind of hit, I don't know if it
• 6:04
  hit a plateau or if it just became something that was a
• 6:08
  whole lot of squeeze for the juice that you were getting.
• 6:12
  And so maybe you've seen different life cycles or you've
• 6:16
  seen different thoughts on on that.
• 6:20
  Yeah, I don't know. I think with a lot of the
• 6:23
  organizations I talked to, maybe it's just they're not throwing
• 6:26
  AI think a lot of them are still doing rag like services, not
• 6:28
  throwing as much of the rag label on it, where I almost
• 6:31
  think just additional context has become kind of the norm.
• 6:35
  Maybe that just kind of coincided with like the vector
• 6:38
  store kind of like boom and bust where everybody looking and
• 6:43
  building vector stores and then PG vector comes out and every
• 6:47
  database has some form of vector search now.
• 6:49
  And that kind of got like democratized and open source.
• 6:52
  But yeah, I think in terms of like the development life cycle,
• 6:57
  it it really varies. And I think that's kind of
• 7:00
  indicative of like the wide degree of skill, technical
• 7:06
  acumen implementation in the data science industry, you know,
• 7:11
  from kind of the low end of the spectrum of the kind of auto ML
• 7:16
  quick devs not writing a line of code or kind of like now looking
• 7:20
  into more like the vibe code world.
• 7:23
  To the opposite end of that, the the opening eyes and the
• 7:27
  anthropics were fine tuning massive foundational models.
• 7:31
  Those are two all those environments look drastically
• 7:35
  different even in the middle there in terms of how teams are
• 7:40
  thinking about development or thinking about the safeguard.
• 7:43
  So I mean, for a lot of organizations, you know, it's, I
• 7:47
  don't think it's deploying into production, but I don't know if
• 7:50
  there's that many steps before getting into production.
• 7:52
  I think a lot of it is a bit of trial and error and throwing up
• 7:57
  kind of agentic services, almost like a Lambda function or
• 8:01
  serverless function. And while it's working great and
• 8:05
  then we'll figure it out when it breaks, but not necessarily
• 8:07
  developing like this massive framework around fail safes and
• 8:11
  testing and what happens when this tweak is made to a model
• 8:14
  and it changes? So it's almost like in your
• 8:18
  eyes, the path to production has been shortcutted in a way.
• 8:25
  Yeah, or I just don't think that just because it's so easy to get
• 8:30
  things into production and it's still relatively new.
• 8:33
  I still, you know, like there aren't as so much in terms of
• 8:36
  standards that exist or are being developed that teams know
• 8:40
  to follow when it comes to development and deployment.
• 8:44
  I don't know if that's necessarily for, I mean, to a
• 8:46
  degree, I think there's a a sector of the developers who
• 8:51
  have never been exposed to a more sophisticated kind of style
• 8:55
  of not sophisticated, it's a bad word, but more yeah, legacy form
• 9:01
  of like software development. But I think it's something will
• 9:04
  emerge in terms of patterns and yeah, develop environments that
• 9:10
  will be, that will look a bit more, you know, what we were
• 9:13
  used to and have the SDLC of like the last 5-10 years.
• 9:17
  It's just, it's still new and it's still being developed.
• 9:19
  And teams need to kind of experience the pains of failing
• 9:25
  services and not having tests written until they, you know,
• 9:30
  decide that there's a, a better path.
• 9:34
  Yeah, wait, Gemini and Claude can't just write all my tests
• 9:38
  for me. What's going on here?
• 9:40
  Yeah. They can, they just like
• 9:42
  somebody has to run those tests and then they have to look at
• 9:45
  the results of the test and then fix them.
• 9:47
  That's the that's the part where, yeah, it takes, you know,
• 9:52
  just a little bit more effort that teams have to put in.
• 9:56
  But, you know, all of these, I think it's one thing that's
• 10:00
  consistent in computer science and software engineering, is
• 10:03
  that foundations and principles reemerge in the same way that
• 10:08
  event sourcing was hot five years ago.
• 10:12
  And those are patterns that were developed and written in the
• 10:15
  90s. And as more, you know, compute
• 10:20
  scales and new patterns become possible, then old ideas
• 10:25
  reemerge. So I'm sure, yeah, like I said,
• 10:28
  it's a temporary thing, but we're starting to see more
• 10:30
  organizations look at it like a regular software, I would say.
• 10:35
  Yeah. Because that begs the question,
• 10:36
  do you see these environments becoming a whole different thing
• 10:41
  or is it just bringing the old ways to what is new?
• 10:45
  Yeah. And I think that's, you know,
• 10:49
  the, the product that seemed that I'm working on, that's kind
• 10:52
  of a big thesis that we have is that they are at many
• 10:55
  organizations because of what I was mentioning earlier, data
• 10:58
  science teams kind of existed in a vacuum.
• 11:01
  They weren't necessarily following the same like dev OPS
• 11:04
  standards or practices. Now, as those AI services go
• 11:09
  from, you know, cool little feature that created a, a
• 11:12
  dashboard that somebody might look at to very core fundamental
• 11:16
  services, you're just not going to be able to have 1-2 different
• 11:21
  pipelines for managing the rest of your organization software
• 11:24
  and then your MO applications doesn't make sense functionally.
• 11:29
  And then also from a security perspective, from a governance
• 11:32
  perspective, from a legal perspective, those can't exist.
• 11:36
  As you know, weekend projects that are running in production.
• 11:42
  They need to have the same level of scrutiny or we'll start to
• 11:45
  see, you know, huge vulnerabilities and hacks in
• 11:49
  production, which I'm I'm sure is on the horizon either way.
• 11:52
  But I think, yeah, just kind of coming to a maturity point and
• 11:56
  realizing that they will need to be treated at the same level of
• 11:59
  scrutiny and we it's just easier to have them managed under one
• 12:02
  process. And there's another piece that I
• 12:06
  wanted to hit on with you, which is just how adopting
• 12:11
  specifically open source LMS is at the enterprise level and what
• 12:17
  that looks like, what you need, the kind of resources that you
• 12:20
  have to throw at it and where there's challenges that you've
• 12:23
  seen. Yeah, it's really interesting.
• 12:26
  And I would say kind of a complicated problem for a lot of
• 12:30
  enterprises, large companies, yeah, I would say on the whole
• 12:34
  of the Fortune 500 or big tech companies that I talked to,
• 12:37
  everybody is using Open a Iron Tropic to some degree that that
• 12:40
  wave has passed. I would say most organizations
• 12:43
  went kind of full force into adopting those managed LLM
• 12:47
  providers to some degree. And I think all of them kind of
• 12:51
  see on the horizon, you know either costs accelerating
• 12:58
  distrust of providers from like a competitive landscape or a
• 13:02
  data privacy landscape or just so you know desire to have an
• 13:05
  open source shop and we want to find an open source alternative.
• 13:08
  So I think everybody is kind of forward-looking to and also the
• 13:12
  convergence, I would say of like as open source models get better
• 13:17
  and better and kind of starting to reach that plateau, they,
• 13:21
  they, I think the long term want to have a large presence with
• 13:27
  open source LMS in their environment and they want to
• 13:30
  allow their developers and teams to freely build.
• 13:33
  But the problem is most organizations can't just, you
• 13:37
  know, allow their teams to build DeepSeek the day after it gets
• 13:42
  built. In fact, most of those
• 13:44
  organizations, their main concern when DeepSeek comes out
• 13:46
  is how do I block it from every single computer?
• 13:48
  And it's not just from, you know, the, the perspective that
• 13:53
  you would see on Twitter, the news of like, you know, the CCP
• 13:56
  stealing her data or something like that.
• 13:58
  It's more of like a from a licensing perspective, from a
• 14:01
  governance perspective, you know, even with licensing, it's
• 14:05
  really interesting. We speak to companies all the
• 14:08
  time that spend, you know, a great deal of effort deciding
• 14:11
  like which licenses they can adhere to or follow.
• 14:16
  But it's hard to keep track of those licenses on Hugging Face,
• 14:19
  especially as you start pulling and playing around with random
• 14:23
  fine tunings of any of the big models that might flip that
• 14:28
  license or switch it. Now it's kind of a a larger
• 14:31
  legal question as to do we follow the original models
• 14:35
  license? We follow Joe Schmo's new
• 14:37
  license, all this like, you know, very, very mundane stuff
• 14:42
  that is not like the the sexy vibe code stuff that you see on
• 14:46
  Twitter. But they're real problems that I
• 14:48
  think prohibit development right now.
• 14:52
  So organizations are, you know, and kind of leading into like
• 14:56
  the security element as well. Hugging Face has just gotten
• 15:00
  like any other open source package provider flooded with a
• 15:04
  lot of that models. I work for Jfrog.
• 15:07
  We have a our own kind of dedicated research team and
• 15:10
  we've done a lot of work. We scanned Hugging Face models
• 15:14
  several times a day and our results are often put into the
• 15:18
  Hugging Face console. And basically we're seeing like
• 15:22
  I think the rate of models growing on Hugging Faces like 5
• 15:25
  times month over month, three over year or something like
• 15:27
  that. And the rate of vulnerabilities
• 15:29
  is increasing by like 7 times. And just, yeah, I guess just
• 15:33
  like any other package provider going after misspellings or even
• 15:39
  going as far as just fine tuning models and including exploits,
• 15:42
  reverse shells, it's just like any other piece of software.
• 15:47
  But the delivery mechanism through, you know, it seems so
• 15:51
  friendly and so easy to pull into your environment and easy
• 15:54
  to deploy into production. And that's that's kind of the
• 15:57
  problem here is that, you know, these large enterprises
• 16:00
  recognize all of this, but they can't they need to have some
• 16:04
  degree of certainty with these open source models with the
• 16:06
  power and flexibility that they have that they are not, you
• 16:09
  know, exposing their teams and they can safely allow them to
• 16:13
  build. A little detour.
• 16:15
  I download a Hugging Face model and it is compromised.
• 16:20
  You're taught you're not talking just like it's data poisoned or
• 16:23
  there's something that is hiding in the latent space of the
• 16:27
  model. You're talking like the actual
• 16:28
  package that I pull from Hugging Face then is a virus that gets
• 16:33
  unleashed on my computer. Could just be opening a shell
• 16:37
  and then, you know, sending out your your keys or something like
• 16:41
  that, Sending out your GitHub key.
• 16:44
  That's one example. But yeah, that and it comes, it
• 16:46
  could be in the, the code itself of the model, or it could be in
• 16:51
  any of the packages that they're using.
• 16:54
  So it's important to, I think, scan for both because, yeah.
• 16:59
  But again, it's just like traditional software.
• 17:01
  You want to scan your Docker images, you want to scan your,
• 17:05
  your pit packages, your condo environment, whatever or poetry,
• 17:07
  whatever you're using or JDM, any of them.
• 17:11
  But yeah, both could exist. And we, we see instances and
• 17:13
  examples of both. So that's yeah, I think that's
• 17:17
  one aspect. It's just, you know, many
• 17:20
  organizations, teams, they see the long term benefit of these
• 17:23
  models, they see the cost benefits of these models, but
• 17:25
  they need to shrink the the marketplace a little bit, shrink
• 17:33
  what is available to maybe instead of, you know, thousands,
• 17:36
  hundreds of thousands of models on Hugging face to like maybe 5,
• 17:38
  maybe 10 that there are certain work, but still, because of just
• 17:42
  generative AI, they have the flexibility to be able to
• 17:44
  implement so many different work flows.
• 17:46
  But shrinking that down a bit and not letting developers
• 17:49
  basically go directly to a hugging face or O Llama, any of
• 17:55
  those, you know, large language model providers having a proxy
• 18:00
  or a middleman at least so that they're not going directly to
• 18:03
  the source just because it's it's that easy to pull in
• 18:06
  something malicious. That is fascinating you
• 18:09
  mentioned because I have seen a fair amount of questions pop up
• 18:13
  on the ML OPS community slack like hey what is an alternative
• 18:17
  to hugging face? My company doesn't let me use
• 18:20
  Hugging Face or, and the way that it's generally framed in
• 18:25
  the Slack questions is, oh, my company's so old school that
• 18:30
  they won't even add support to Hugging Face.
• 18:33
  But now seeing it from your side, it's like, oh, maybe the
• 18:35
  company is actually kind of forward thinking in a way, or
• 18:39
  they know what. That it's a it's a bit of both,
• 18:42
  yeah. I'm sure there's, I mean, one,
• 18:44
  there isn't really a place right now to get models out of, you
• 18:47
  know, besides kind of the more like traditional machine
• 18:51
  learning libraries, the for the tensor flows, the pie torches of
• 18:55
  the world. You get some models from GitHub.
• 18:58
  But yeah, hugging faces the marketplace right now for it.
• 19:01
  And I agree as it's both kind of antiquated and also, yeah, it's
• 19:07
  safe. It makes sense when you think
• 19:09
  about it. And also again, looking at like
• 19:11
  the the developers that are pulling those models.
• 19:14
  Yeah, it's not a wild statement to say data scientists don't
• 19:17
  really care about security. Data folk it's not either.
• 19:20
  There are definitely some that are that are concerned with
• 19:23
  that, but it is not top of mind by any means.
• 19:27
  So yeah, I think guardrails are really important, especially
• 19:31
  just as more and more models keep coming out.
• 19:35
  Yeah, it's like the double edged sword.
• 19:38
  What makes Hugging Face so amazing is the ability for
• 19:41
  anybody to just go and add their model to Hugging Face.
• 19:45
  But then what makes it absolutely crazy as far as
• 19:50
  security vulnerabilities is anybody can add anything on
• 19:54
  there. Yeah.
• 19:56
  And again, it doesn't even have to be, you know, I feel like one
• 20:02
  of the most common forms of like entry for malicious packages is
• 20:07
  just misspelling, you know, panic with AO or something like
• 20:11
  that or SP, you know, just swapping out a letter and
• 20:14
  that's, that's all it takes for you're not even.
• 20:18
  You're not even realizing that you're pulling in something bad
• 20:22
  happens all the time. So yeah, trying to add some kind
• 20:26
  of a proxy, some kind of a curated list of models that have
• 20:32
  been reviewed have been allowed do follow our organizations
• 20:36
  legal and governance. And then also having a somewhat
• 20:39
  of a, a metadata record of these, these events taking
• 20:43
  place, keeping track of the model cards that are being used
• 20:48
  from Hugging Face. We call it like an ML bomb, but
• 20:52
  it's really just, you know, a metadata stamp that has all the
• 20:55
  environments that were used, all the safe tensor files that were
• 20:59
  used. That's really, I would say
• 21:01
  purely for thinking. Is there a ton of benefit or
• 21:04
  values that today? No, but if a lawsuit comes out a
• 21:08
  year down the road and you need to figure out all of the
• 21:11
  relevant information around the training or a security leak
• 21:15
  happens, need to go back and kind of figure out what
• 21:18
  happened. It's helpful to have those
• 21:20
  things in some form of a, a registry, a central location
• 21:25
  that you can easily audit versus having to go, you know, dive
• 21:30
  through developer notebooks and do that hunt.
• 21:34
  Yeah, and you're saying something where it's like, oh,
• 21:38
  New York Times is suing open AI type of lawsuit.
• 21:41
  Sure. I.
• 21:42
  Mean that's the most extreme case, but yeah, it's I think,
• 21:46
  you know, just have being prepared for those events and
• 21:49
  having the relevant metadata just given like the flexibility
• 21:52
  and the power of these models, how extendable they are.
• 21:55
  I think that's that's one of the more kind of important
• 21:58
  governance perspectives you need to have is kind of like we don't
• 22:01
  know how this will be used. A new trend emerges in six
• 22:04
  months that completely throws out a gentic AI and has a new
• 22:07
  pattern for development. But we do want that metadata to
• 22:10
  exist somewhere. And yeah, or like I said, when
• 22:15
  the the licenses actually start getting enforced and then teams
• 22:19
  have to, you know, really care about that stuff or limit the
• 22:21
  types of models we're using. Similarly, like another problem
• 22:26
  that we've heard from a lot of our customers and something that
• 22:29
  we're releasing at Bay or in the air J Frog like stores, a lot of
• 22:34
  organizations, artifacts and code.
• 22:36
  We hear people basically saying for similar purposes, we need to
• 22:40
  keep track of governance, We need to keep track of licenses.
• 22:43
  We have no idea where generative AI is being used in our
• 22:46
  organization. They do.
• 22:49
  And because teams have just, you know, pip install open AI, start
• 22:52
  using the models. So that's one thing that we've
• 22:54
  been working on is basically just going through your whole
• 22:57
  organization's code base and detecting all right here, all
• 22:59
  the instances where this is being used.
• 23:01
  Let's generate some of that metadata for, you know,
• 23:04
  especially for regulated industries.
• 23:06
  If, you know, a law comes out next month that says, you know,
• 23:10
  any financial companies need to have like a specific list of
• 23:14
  auditing requirements related to any applications using
• 23:17
  generative AI or they ban generative AI with specific
• 23:20
  human use cases, whatever it might be.
• 23:23
  It's just important to know where those applications exist,
• 23:27
  where those packages are being used.
• 23:30
  Also for vulnerabilities purposes.
• 23:32
  Like you know more and more applications start using Open
• 23:37
  ISDK Cloud SDK and something happens.
• 23:40
  You would want to know where everything has been distributed.
• 23:43
  Well, I know I've talked to folks about this when it comes
• 23:47
  to auditing how AI is being used throughout the workforce.
• 23:53
  And one person that I spoke to said, yeah, I just got done.
• 23:59
  I think they were in a data governance role and they were
• 24:02
  expecting to find maybe like 12 to 20 uses of generative AI, and
• 24:10
  they found over 90. I think I was like, Oh my God, I
• 24:14
  would have never have guessed. But then you can also take it a
• 24:17
  step further and say, because you're, you're really talking
• 24:20
  about the stuff that you can see in the code base, right?
• 24:22
  Or the stuff that the developers are doing.
• 24:25
  But what about the marketers that are using that new shiny
• 24:28
  marketing tool that helps you write, copy and you send all of
• 24:31
  your like positioning documents to that SAS tool and that is
• 24:39
  using generative AI on the background.
• 24:41
  Like, does that count, You know, from a governance perspective?
• 24:45
  Now there's these third party SAS tools that are using AI and
• 24:49
  how do you qualify them too? Yeah, I mean, that's, that's a
• 24:54
  great point in a whole other Pandora's box, if that needs to
• 24:58
  be. Yeah, categorized and classified
• 25:01
  as well. Yeah, we're we're like focusing
• 25:04
  on purely just the code. But I mean, the example that you
• 25:06
  highlighted is exactly what we're hearing and seeing.
• 25:09
  It's like the it's already out-of-the-box, it's everywhere
• 25:13
  and it's only going to, you know, keep showing up
• 25:17
  everywhere. Not only, I mean from a
• 25:19
  development perspective, everybody is using some form of
• 25:22
  code Gen. tools at this point. And then the actual models
• 25:25
  themselves, I think, you know, we kind of anticipate like in
• 25:29
  three to five years still pretty much be a, any service will have
• 25:33
  some component of, of generative development.
• 25:38
  So yeah, just keeping track of everywhere it's being used.
• 25:44
  I think that's the, the important perspective, having
• 25:48
  some kind of a framework there. And yeah, treating, treating it
• 25:52
  like any other application or service, like that's the, it's a
• 25:58
  low bar, but it's important. And I think it's, you know, you
• 26:03
  see these like kind of very like diverging paths in, in
• 26:07
  development. Like, it almost feels like, in
• 26:10
  many ways, like the, the larger enterprise organizations I
• 26:13
  talked to are speaking like a different language of AI than
• 26:16
  the startups of the world or the, the, anything coming out of
• 26:19
  Y Combinator. Like it's truly like 2
• 26:21
  completely different pathways, two different focuses and two
• 26:23
  different like concerns that I feel like often get blended
• 26:29
  together, you know, when people are looking for AI tools, ways
• 26:35
  to improve their AI developing experience.
• 26:40
  But yeah, it's just there's, there's two very different kind
• 26:44
  of pathways emerging that do require kind of two different
• 26:47
  sets of solutions. And those solutions are
• 26:49
  ultimately serving like an AI purpose.
• 26:52
  But it's very different than I think the like the MO OPS
• 26:55
  landscape of five years ago where I think everybody was more
• 27:00
  or less kind of like on the same page to a degree, but.
• 27:04
  All suffering together in trying to get those damn Jupiter
• 27:07
  notebooks into production. Yeah, which is still a problem.
• 27:12
  Actually. I think teams still have a hard
• 27:14
  time doing that. But yeah, you just don't have
• 27:18
  the same, I mean, the startups should have the same security
• 27:22
  concerns. But I know we've all been in a
• 27:25
  startup where everybody has AWS console access and we'll figure
• 27:30
  that out next round or after we get the next customer to sign.
• 27:33
  So it's that's understandable. But it's, yeah, it is two
• 27:38
  different sets of concerns. Yeah, yeah.
• 27:40
  It's almost like PMF Trump's everything.
• 27:43
  And so you're building that organizational debt because you
• 27:47
  need to find PMF when you're at the start up.
• 27:51
  And if you're out of the enterprise, you really have to
• 27:53
  think about all of these key concerns since you have a lot to
• 27:58
  lose when you're at a start up, you don't have as much to lose.
• 28:01
  And so whatever you can kind of play fast and loose.
• 28:06
  Yeah. And the, the cost of an
• 28:09
  exploiter vulnerability is, you know, tremendous for enterprises
• 28:13
  when it comes to not only remediating the code, but like
• 28:17
  auditing, doing full investigations, autopsies of
• 28:21
  like every, you know, path that they could have been touched by
• 28:23
  it. It's A and that is how
• 28:25
  enterprises are thinking when they're looking at different
• 28:28
  tools to adopt, different platforms to adopt because,
• 28:31
  yeah, they have to be concerned. Yeah, it just sounds like a
• 28:35
  headache, even in that glimpse that you've given me.
• 28:39
  Yeah, I think one other aspect on that the open source that's
• 28:42
  interesting or a lot of a big motivation for a lot of
• 28:45
  organizations that want to adopt open source development is just
• 28:48
  on Prem development or air gap environments.
• 28:51
  That is still, and again, this isn't in like the the the most
• 28:57
  pressing technology or bleeding edge, but there's still so many
• 29:00
  organizations that have massive on Prem clusters that will have
• 29:04
  massive on Prem clusters that have moved stuff out of the
• 29:07
  cloud. It's not going in a way and it's
• 29:11
  kind of this neglected portion of the industry that doesn't get
• 29:16
  a ton of attention, but it's massive.
• 29:18
  So trying to find solutions for defense industry that yeah, they
• 29:23
  want to use open source models as well.
• 29:24
  They don't. They also can't in a lot of
• 29:26
  instances use an open AI use Anthropic because the maybe they
• 29:30
  can. But like those are large
• 29:33
  decisions that are made at the top level, like heavily
• 29:36
  negotiated. They would love to start
• 29:39
  implementing Jen AIA bit easier and they're trying to find ways
• 29:43
  to do that with open source. You did say something I wanted
• 29:48
  to go back to. It was on the.
• 29:52
  Yeah, trying to just sniff out all the places that AI is being
• 29:56
  used and if you're in that governance position, just how
• 30:02
  hard that is and how cumbersome it can be.
• 30:06
  And then. No, it it's, it's a, it is a
• 30:09
  problem. That's I think, you know,
• 30:10
  previously you're not necessarily concerned with where
• 30:16
  is pandas being used in my organization?
• 30:18
  Like that's an interesting question, but like what value
• 30:20
  does that really give or dictate?
• 30:23
  Whereas, you know, not only from like a commercial standpoint,
• 30:28
  you're negotiating a deal with open AI.
• 30:29
  Am sure they can give you the, you know, the top dollar amount,
• 30:32
  but you want to understand like where it's being used.
• 30:35
  It's just so easy to implement any of these services that it,
• 30:43
  there isn't always a clear auditing log.
• 30:45
  You could, you know, there's tools that can go through your
• 30:48
  entire code base and GitHub and detect these types of things.
• 30:51
  But I think that's one reason why Jfrog kind of has like a
• 30:56
  unique offering here and being able to, because we store a lot
• 30:59
  of organizations binary code. So we can in our scanning
• 31:03
  processes go through and kind of detect where those packages are
• 31:08
  being used, where those libraries are imported.
• 31:10
  And then provide a mechanism of creating guardrails or some kind
• 31:19
  of a administration system, also providing kind of like a gateway
• 31:24
  type feature where you can, yeah, limit access to models,
• 31:27
  limit who's calling models. Yeah.
• 31:31
  I think again going in just like the delivery mechanism of how
• 31:34
  most of these models are being pulled into products and
• 31:38
  services, it makes it just very easy to sneak in.
• 31:41
  Yeah, especially because you can almost like go around it too.
• 31:46
  And so I'm doing it in my CLI, but then I can just go to the
• 31:51
  website and maybe I'm using like 5 different ones.
• 31:55
  And so I have, I'm paying for quad code right now.
• 31:58
  I'm also paying for Gemini. And I'm, I've got cursor set up
• 32:02
  or I've got AMP and all of these different ones I'm using in some
• 32:06
  way shape or form. But maybe it's like not always
• 32:12
  the same. It's not just like a VS Code
• 32:15
  plug in that I'm using. I am going to Gemini directly on
• 32:20
  their website and talking to it or ChatGPT or whatever it may
• 32:24
  be. And maybe we have the enterprise
• 32:27
  version or maybe it's just me using the free version.
• 32:30
  Hopefully like people aren't doing that, but you know that
• 32:34
  there's going to be people that 100% are.
• 32:37
  Yeah, I mean either to test out a new model.
• 32:39
  I'm sure that's, you know, probably a challenging point for
• 32:44
  a lot of teams, like you're not getting something to work and
• 32:46
  then O 5 comes out and you're like, maybe it'll work 25 not O
• 32:51
  five. Yeah, but that yeah, it's a good
• 32:54
  point of like, how do you control those gateways?
• 32:56
  Again, it ultimate comes down to like you need to block off
• 33:00
  access to the chat TBT host in your organization.
• 33:05
  There has to be 1 gateway or one pathway.
• 33:09
  I think all of it does kind of come down to, yeah, some kind of
• 33:13
  a gateway that each organization manages, maintains on their own,
• 33:18
  has some form of a security, you know, user model where they're
• 33:21
  determining who can have access to certain models.
• 33:23
  Also from a cost perspective, like, you know, you don't want
• 33:26
  to be getting sequel analysts access to the PhD level models.
• 33:32
  We don't need that. Maybe some, but but yeah, I mean
• 33:38
  putting the yeah, the costs are that's a whole other thing.
• 33:40
  And I think there's a lot of really cool companies and
• 33:43
  products that are focusing on that.
• 33:47
  But yeah, I think just controlling the entry point is
• 33:51
  the kind of the one of the main kind of thesis that we have of
• 33:56
  what creates A framework that will be scalable, lasting and
• 34:01
  secure moving forward and can hopefully anticipate some of the
• 34:05
  different changes that might happen in the industry.
• 34:07
  But you're, you're almost like bottlenecking via the like if
• 34:14
  they're using the Wi-Fi of work or if it's the company computer.
• 34:18
  Because I can still see people inadvertently being like, oh,
• 34:22
  damn, I can't like reach ChatGPT on this computer.
• 34:25
  But if I bring my personal computer to work, I can still
• 34:28
  get to it when I need those questions answered that it won't
• 34:31
  let me answer. Yeah, it's, I mean, yeah, that
• 34:37
  that's just highlighting the differences between yeah,
• 34:39
  startups and enterprises, like enterprises.
• 34:40
  No, that's, that's a very standard thing that like
• 34:43
  specific hosts. I mean, even yeah, back in the
• 34:46
  day, like certain content websites are blocked from your
• 34:49
  work machine for obvious purposes.
• 34:52
  It's all, it's all pretty easy to do.
• 34:58
  But yeah, that's, that's how we see it.
• 35:00
  And I think a lot of other also some cool tools that I've seen
• 35:03
  coming out of like, you know, not even also like functionally
• 35:07
  creating like LM gateways that will allow you to really easily
• 35:13
  swap out models, direct requests to cheaper models.
• 35:19
  If there's like a intermediate model, they can determine this
• 35:21
  is a, you know, low complexity use case.
• 35:26
  Let's filter this to the small language model.
• 35:29
  There's a ton of value there that I think the industry is
• 35:32
  like just starting to explore and how they, yeah, diversify
• 35:36
  the requests more efficiently, use resources that starts with
• 35:42
  some form of a centralized gateway and some like semblance
• 35:45
  of control as to how these requests are being made.
• 35:50
  And if everything is just sporadic and every team is using
• 35:52
  a different model and you have no idea, those kind of an end
• 35:56
  goals won't be, you can't accomplish them.
• 35:59
  Fascinating. And building that gateway
• 36:01
  because I've heard this idea of a gateway before a few times.
• 36:07
  I know there's a few companies that are doing it, but it does
• 36:12
  feel like that's going to give the the company the most control
• 36:19
  of what is going out and coming in.
• 36:23
  But yeah, you have to think on various levels.
• 36:26
  You're not just thinking on the level of everything that's going
• 36:29
  to the research labs, but what's coming off of Hugging Face, like
• 36:34
  you were saying? And then what are we running
• 36:36
  locally? And are we setting up these
• 36:39
  smaller models and setting up like a platform team that's
• 36:43
  going to babysit the smaller models?
• 36:46
  And that's us. Maybe it's on Prem, maybe it's
• 36:48
  not. And how that looks is, is
• 36:52
  completely different. Also now I know that we
• 36:56
  mentioned before we hit record the whole evolution of Quack and
• 37:03
  how quack you've been there for a while.
• 37:06
  It was very ML OPS focused. And now it's kind of grown.
• 37:11
  It joined Jfrog, it now is security focus.
• 37:16
  But you're also seeing this type of stuff like on the platform
• 37:20
  engineering level, the DevOps level.
• 37:23
  These folks are tasked with the AI platforms.
• 37:29
  In a way, you have almost like the conglomerate of MLAI
• 37:35
  engineers, DevOps engineers, data engineers, and that's how
• 37:39
  you get the platform stood up. So what have you been seeing
• 37:43
  since this evolution took place of like ML OPS to platform maybe
• 37:51
  is the way that we could call it?
• 37:54
  Yeah. And you know, kind of the Segway
• 37:57
  that you mentioned of those teams managing those small
• 38:00
  language models, that is it's still ML OPS.
• 38:03
  Like it's a different type of model, it's a different type of
• 38:07
  latency and throughput. But at its core, it's, you know,
• 38:11
  it, it is ML OPS that can fit into what organizations have
• 38:14
  already built for ML OPS. Maybe they're not spending as
• 38:18
  much time on like the experimentation portion of that
• 38:21
  because unless they're really fine tuning, you know, they
• 38:24
  might just be, you know, constructing an image or
• 38:27
  artifact that they're then going to turn into a service.
• 38:30
  But you still want all those kind of core components of ML
• 38:34
  OPS. You like sophisticated AB
• 38:38
  testing. That's something that we see a
• 38:40
  lot with machine learning, the persistence of all of the result
• 38:44
  data into some easy, you know, easy to access, easy format
• 38:49
  where developers, data scientists can like query
• 38:51
  results, do drift analysis on those results.
• 38:55
  So all of that still relevant and I would say needed with
• 39:00
  generative AI was with large language models, it can fit into
• 39:03
  that pattern. I think it's just in like this
• 39:06
  kind of trailing way, everybody kind of forgot about it for like
• 39:09
  two or three years with like the excitement of what they could do
• 39:12
  from a development perspective. And now as they get into, you
• 39:16
  know, maturing these models and services over time, auditing
• 39:20
  them, needing of a system of record, they're kind of
• 39:23
  realizing that we do need, you know, our LLM results persisted
• 39:28
  somewhere or, or even like regulated industries.
• 39:31
  We have a healthcare customer that, you know, every single
• 39:34
  response with Jared of AI needs to be persisted somewhere.
• 39:39
  So I think yeah, in terms of the ML OPS industry, I think, you
• 39:44
  know, it's a you're deep in it. There's an insane number of like
• 39:49
  point solutions that I can point solutions platforms that I could
• 39:53
  cannot keep track of. And they're, you know, the ones
• 39:57
  that kind of die or fade like they're getting replaced by a
• 40:01
  similar solutions that are more generati focus.
• 40:03
  But you know, I think at the core.
• 40:07
  All of these organizations are looking for like a platform, a
• 40:10
  standard that will allow them to easily define workflows in a
• 40:14
  safe, reliable, repeatable way. I think that's why you see so
• 40:18
  much, you know, excitement around MCP, you know, MC PS not
• 40:25
  like revolutionary for a code perspective, but it's, it's
• 40:27
  somebody like, you know, defining and saying as a
• 40:30
  standard that we're going to follow when it comes to the
• 40:33
  access of external data sources or day that we can feed to a
• 40:38
  language model functionality, that type of standardization.
• 40:42
  I think like the whole market is very hungry for and there's a
• 40:47
  lot of opportunity for products platforms to kind of define that
• 40:52
  and to, you know, increase its use.
• 40:56
  That's such a good point that it is like it didn't change.
• 41:01
  It just changed in name in a way.
• 41:05
  Like the actual practice of ML OPS didn't change too much.
• 41:09
  The models got maybe a little bit bigger for a minute and so
• 41:13
  you had to do certain things. But if you were at a certain
• 41:17
  scale, you probably were already messing around with GP US and
• 41:21
  you were messing around with like trying to make your
• 41:25
  whatever feature creation feature extraction faster and
• 41:28
  you needed GP us and you needed to get BP GP us and then work
• 41:32
  with networking. But then that became very
• 41:35
  popular because of the training of LLMS. And so we lost our way,
• 41:43
  or not necessarily lost our way. We just got a little bit
• 41:45
  distracted for a minute and then came back.
• 41:48
  And we were like, Oh yeah. So we still got to deploy these
• 41:52
  models somewhere, and we still got to make sure that they don't
• 41:56
  drift and that we can monitor them.
• 41:59
  And so it's 100% the same problem, just almost like in
• 42:06
  different clothing. Yeah, yeah.
• 42:08
  Completely, I mean, yeah, and just, you know, expanded
• 42:11
  functionality, like the models aren't as as focused even with,
• 42:16
  you know, agentic work flows like there's still way more
• 42:19
  functionality and possibility than existed with your simple
• 42:22
  linear regression models. But yeah, it's, yeah, maybe not
• 42:25
  the exact same use case of how they're monitoring the data, how
• 42:29
  they're monitoring or configuring these services for
• 42:32
  testing, but a lot of it is the same or it's just a different
• 42:35
  spin on it. And yeah, organizations I think
• 42:39
  that were that did have mature sophisticated kind of ML OP
• 42:45
  systems in place, I think are definitely kind of leading the
• 42:49
  way in terms of what they can do with sharing of AI.
• 42:51
  And that at least it's more of like a lasting pattern versus,
• 42:57
  yeah, throwing up the ad hoc services without a lot of force
• 43:01
  at your thought. Yeah.
• 43:03
  And it's also certain things got like the time that you take on
• 43:10
  them disappeared in a way. Like you were saying before, you
• 43:14
  don't necessarily do all this experimentation unless
• 43:17
  potentially you're fine tuning. You're not doing data curation
• 43:22
  unless you're maybe fine tuning. But what you are spending more
• 43:26
  time on is figuring out the prompting and figuring out how
• 43:30
  that that fits into things. And so you're spending more time
• 43:34
  in different areas. But they're they're like they
• 43:38
  rhyme in that regard. Yeah.
• 43:41
  And that's, that's again, going back to kind of like the concept
• 43:44
  of a gateway, like, you know, this whole conversation is kind
• 43:47
  of focused on like everything that you can do, like pre
• 43:50
  deployments, pre prompt interaction to make sure your
• 43:55
  code is, you know, safely securely deployed.
• 43:59
  And then you have the whole, you know, everything you just
• 44:00
  mentioned creating guardrails around content, prompt
• 44:04
  injections, make sure the service is doing what you want.
• 44:07
  That's a whole other kind of separate topic that a lot of the
• 44:10
  industry is focused on. I think that also kind of kind
• 44:14
  of converges with we should have a centralized gateway where we
• 44:17
  can administer these different types of guardrails.
• 44:20
  We can administer these security policies that we want to
• 44:23
  enforce. Again, all of it just comes down
• 44:26
  to like organization, standardization, some methods,
• 44:31
  some system that needs to be in place versus, you know, Wild
• 44:37
  West ad hoc 5 code just throw services up.
• 44:41
  It works quickly in the short term, but to have the more
• 44:46
  mature systems there needs to be, yeah, some form of a, some
• 44:50
  form of a gate in place. But you're not seeing
• 44:53
  enterprises on the services side like the infrastructure side of
• 44:59
  oh hey, we're going to spin up GP us or we're going to go in
• 45:04
  contract a new GPU cloud because we can't get the GPUs for the
• 45:08
  price that we want. Your as Jfrog the gateways that
• 45:12
  you're talking about, you don't go into that side of the house
• 45:15
  do. You we are going to, yeah, our
• 45:19
  kind of the intent of our kind of rebuilds a bit of our MO OPS
• 45:23
  platform from the quack days to Jfrog.
• 45:26
  Mel is the new release term, but we're available on all three
• 45:29
  clouds now. We'll be available on Prem in a
• 45:31
  few months from now. But the idea would be that you
• 45:34
  could train on AGPU cluster locally and then deploy into the
• 45:38
  cloud, which is a common pattern that I hear from a lot of
• 45:40
  customers. You know, there's been enough
• 45:43
  boom cycles with GPU's where many organizations have bought
• 45:46
  GPU's and they are kind of are sitting there and it is cost
• 45:50
  effective or cheaper for them to to go that way rather than with
• 45:54
  on demand GPU's in the cloud. But I think most organizations
• 46:00
  at like an enterprise scale very it, it's rare to be limited to a
• 46:06
  single cloud at this point. I think the yeah, the default
• 46:10
  now is usually a combination of clouds, the combination of on
• 46:14
  Prem and cloud. And there is a desire to be able
• 46:17
  to seamlessly go through each of those environments without
• 46:20
  having to completely re engineer your machine learning
• 46:26
  philosophies of your machine learning practices.
• 46:29
  Being able to, yeah, switch seamlessly is huge for a lot of
• 46:32
  teams. What other features and ways
• 46:37
  have you looked at expanding or evolving the what was once the
• 46:43
  Quack platform into the now Jfrog platform or the ML
• 46:48
  capabilities inside of Jfrog or ML/AI if we're going to get
• 46:53
  technical? Yeah.
• 46:56
  So I mean, a lot of the the core ML OPS platform still exists.
• 47:01
  We still do. And then machine learning OPS,
• 47:03
  it's a really nice easy way to kind of deploy models and
• 47:07
  services. And then everything that we've
• 47:09
  kind of, you know what, what a lot of this conversation has
• 47:13
  been about focusing on the security aspects.
• 47:15
  When you trigger a machine learning training or a build,
• 47:19
  automatically scanning it, applying, you know,
• 47:22
  organizational policies around these packages are blocked or
• 47:26
  these packages are out of date. We only want to use open source
• 47:29
  packages that have 2000 stars in GitHub or have been updated in
• 47:33
  the last six weeks or 4 weeks, whatever that might be.
• 47:37
  That's kind of what the the larger Jfrog platform and
• 47:41
  ecosystem is about is basically managing open source development
• 47:46
  at scale for enterprise and organizations, taking that and
• 47:49
  applying it's machine learning kind of automatically.
• 47:51
  And that's the idea because we know the data sciences, machine
• 47:53
  learning engineers, they don't really want to worry about
• 47:55
  security. Slap it on that or, you know,
• 47:58
  put the the scans right in their face, make it easy for them to
• 48:02
  fix. And then we've kind of combining
• 48:06
  products with Jfrog has a cool product called Curation that
• 48:10
  basically kind of does what I just mentioned, like creates you
• 48:13
  can do like rule based systems as to what types of packages
• 48:18
  developers can pull in from open source languages, from
• 48:21
  repositories, combining that with like a really easy way to
• 48:24
  deploy LMS. You know, we've preconfigured the model.
• 48:29
  It has BLM already built into it where the, you know, the IT
• 48:33
  manager, security manager, DevOps manager at an
• 48:38
  organization can go and click, I want LAMA four, I want DeepSeek,
• 48:42
  I want Quen. And those models then become
• 48:45
  available to developers to either pull the artifact or
• 48:49
  deployed a service as well as they can control who's actually
• 48:52
  able to, to make requests to that model.
• 48:57
  And so we're going that AI catalog and that's kind of it
• 49:00
  also includes a gateway and some of the detection pieces that I
• 49:05
  was mentioning earlier about like, OK, here are all the
• 49:07
  services that are using open AI, like do you want to block them?
• 49:10
  Is this OK or here's everything, everything using Llama 4?
• 49:15
  Like we've determined that we can't use Llama 4 at an
• 49:18
  organization level. Now you can block those services
• 49:21
  from being redeployed and alert teams that they need to fix it.
• 49:25
  That's I would say, yeah, where the the product is kind of
• 49:29
  landed. But we still, yeah, do all the
• 49:31
  same MO OPS stuff that we've done before and have plenty of,
• 49:35
  you know, some of our older customers still using the
• 49:37
  platform the same way. Yeah.
• 49:39
  So it's like you thought about all of these different models
• 49:44
  that are the popular models and how can you just put your stamp
• 49:48
  of approval on them so that folks have the trust if like I'm
• 49:53
  going to use these? I know that it is not the
• 49:56
  Facebook Llama model, which I remember back in the day that
• 50:00
  people were downloading and they were like, no, it was actually
• 50:02
  the meta llama model that you had to get because the Facebook
• 50:06
  one was that was a virus. Yeah, we we do some suggestions
• 50:12
  like that within the platform, like we offer kind of like some
• 50:14
  statistics that show like the reliability or the viability of
• 50:20
  the model. And then we also provide like
• 50:22
  again, the kind of rules based framework where you can sort out
• 50:26
  if that model, you know, based on a security scan or based on
• 50:30
  the number of downloads. Like we can filter it so that it
• 50:33
  only includes like the the models that you want your team
• 50:37
  or you can just go and whitelist, select them.
• 50:39
  And that works too. Yeah, creating a yeah,
• 50:44
  management management system. And this is also, you know, we
• 50:49
  spent the first year kind of after getting acquired into
• 50:53
  Jfrog kind of just talking to a lot of customers and here's what
• 50:57
  we built. What do you want?
• 50:58
  And this was the kind of the biggest request that we got from
• 51:03
  so many customers that were basically saying we need a
• 51:05
  management framework. Just like a basically, we need a
• 51:08
  governance layer, yeah. I mean both governance as well
• 51:13
  as just we need to have something in place.
• 51:17
  We cannot let it let this be the Wild West and teams just using
• 51:21
  whatever packages and models they want.
• 51:23
  Well, I imagine most folks didn't even have any visibility.
• 51:27
  None, yeah. And that's, that's their, their
• 51:29
  initial reaction is when they don't have visibility like the
• 51:32
  the friend that you mentioned, they shut off hugging face
• 51:35
  access. And that's, you know, a
• 51:37
  temporary fix to or, you know, kind of delaying progress,
• 51:41
  right. Like I think you're ultimately
• 51:44
  going to be doing a lot of generative development in your
• 51:47
  organization. You need to find a solution for
• 51:50
  it. You can't just block it off.
• 51:53
  So that's, yeah, the and that's many organizations that we
• 51:55
  talked to that were like, we'd love to get into this type of
• 51:58
  development, but until we have something nobody's building.
• 52:04
  Wow, yeah. And then some more on this more
• 52:09
  like obviously a lot of them are using open eye still, but I
• 52:14
  think in more limited scope then like what they're kind of end
• 52:18
  vision would be. Yeah, Yeah.
• 52:20
  It's the using it right now again, trying to find like the
• 52:26
  PMF for the products that they're putting out with AI.
• 52:30
  And even if they are at that enterprise level, I think most
• 52:33
  folks are trying to experiment with how they're going to fully
• 52:39
  leverage AI. And I know that I have some
• 52:42
  friends at an enterprise and they are constantly saying like,
• 52:47
  I just want to know if it can work before I do anything.
• 52:52
  So the fastest way to know if it works is to use the best models
• 52:57
  out there. Because if I can't get it to
• 52:58
  work with the best models, then it doesn't matter if I can get
• 53:02
  like the 7B model running internally fully like secure and
• 53:08
  everything because it's not going to work.
• 53:12
  I pretty much guarantee that if it doesn't work on the best
• 53:15
  model, is not going to work on the small model.
• 53:18
  Yeah. I mean, yeah, it just highlights
• 53:21
  that you're not going to stop developers from using the best,
• 53:25
  the newest, best model, even, even when as that improvement
• 53:29
  gets more and more marginal, if you're stuck on a problem and
• 53:34
  the new model might solve it, Yeah, why not?
• 53:36
  Yeah. Yeah.
• 53:38
  And, and it is that though, like to what you were saying, I try
• 53:42
  and get it working, but then once it's working and I
• 53:45
  understand that it can be done with AI, that's what I'm going
• 53:49
  to try and optimize. And I'm going to make it fast
• 53:51
  and I'm going to try and make it cheap.
• 53:52
  And I'm going to figure out all the ways to make this
• 53:55
  bulletproof. But let's like almost try to
• 54:00
  figure out how fast we can get it up and running and does it
• 54:03
  work? All right, cool.
• 54:05
  It looks like I'm kind of there. And then you have that last mile
• 54:08
  problem of it kind of works. I think it could work maybe.
• 54:12
  And then you spend the last 10% of the time, which is way longer
• 54:18
  than that initial 90% of the time, trying to figure out the
• 54:20
  evals or figure out what if you need to do all these different
• 54:25
  prompt tricks or what you need to do.
• 54:28
  Arguing over the tone of one word.
• 54:30
  Just exactly, I don't know if. It's 9010% at that point, that
• 54:34
  might be like 80% of the time. Yeah, it's, it's all that
• 54:38
  development work. And again, that's, I think
• 54:42
  there's so many tools to help with that aspect of development.
• 54:45
  There's so many like go to Twitter and it's just flooded
• 54:48
  with examples of people and tricks.
• 54:50
  That stuff is has been discussed, is being discussed
• 54:57
  and will continue to be. But it is kind of like after
• 55:02
  that point, what do you do after you get something that's
• 55:05
  working? How do we turn this into a safe
• 55:08
  production service that actually is going to last?
• 55:11
  But I do think what you mentioned in terms of like
• 55:12
  portability, going back to that gateway concept, I think that is
• 55:18
  a should be a very significant focus for any organization.
• 55:22
  Just how quickly, you know, functionally can we sub out this
• 55:26
  model from a cost perspective, from a latency perspective, all
• 55:30
  those concerns because, you know, we're just going to see
• 55:34
  more and more models that are suitable for these use cases
• 55:37
  that have lower resources, that have lower parameter counts and
• 55:41
  are cheaper. So, you know, might and people,
• 55:44
  everybody spending right now pretty wildly on AI, but that
• 55:48
  won't always be the case. And there will be a time when
• 55:51
  shrinking occurs and people do have to get more optimized.
• 55:55
  So I'm fascinated by the whole product discovery that you did
• 56:01
  and you mentioned when you had that first year almost like of
• 56:07
  incorporating or acclimating into Jfrog.
• 56:10
  And so you got to go and talk with all their customers and
• 56:14
  they have a ton of enterprise customers and figure out like,
• 56:17
  what's the story? What's the biggest pain point
• 56:19
  that folks are looking for in this cross section of security,
• 56:24
  reliability, machine learning, AI?
• 56:29
  Can you walk me through a few of the stories of like, what were
• 56:34
  some ideas that you didn't go with?
• 56:37
  What were some things or how did you eventually land on what you
• 56:41
  decided to build? Yeah, it is.
• 56:44
  It is really interesting kind of place to be in.
• 56:51
  I think Jfrog is like almost as ubiquitous as like GitHub.
• 56:55
  So we work with over 80% of the Fortune 500 or something, so
• 57:01
  like 10,000 customers, there's access to way more accounts.
• 57:06
  So you go from the startup perspective where like you get a
• 57:09
  Fortune 500 company on the phone, you're stoked, that is
• 57:12
  your whole week and what your focus is on.
• 57:14
  And now you're able to have those like more meaningful
• 57:16
  conversations. And I think it's helpful when
• 57:18
  you have a really, our product was great to begin with.
• 57:22
  And like you're saying, it's, you know, the MO OPS solution
• 57:26
  still isn't really solved and people are still struggling with
• 57:28
  it. So those pain points still
• 57:30
  resonated with a lot of these enterprises.
• 57:33
  A lot of them had kind of solved MO OPS to a certain degree.
• 57:36
  They had built something because they had the resources over the
• 57:40
  course of a few years to spend on it, to make it functional.
• 57:45
  But yeah, again, it was just kind of just asking
• 57:49
  organizations like what you know, what do you need?
• 57:53
  What are your help? What would be helpful?
• 57:56
  What would allow you to build more and that this was the IT
• 58:01
  wasn't too difficult, I would say or like that much of AA menu
• 58:05
  of options that we had. It was really get this platform
• 58:08
  out on all the cloud providers and is and in having a flexible
• 58:12
  on Prem solution. That's definitely major, one of
• 58:16
  the major things we heard and then yeah, provided making open
• 58:19
  source models easier to play and ensuring more trust in them.
• 58:25
  One thing that we we did play around more with was I think on
• 58:28
  kind of the concepts a little bit before the acquisition we
• 58:31
  were focused on kind of creating like more like rag tracing
• 58:38
  elements are kind of like complex DAG style agentic
• 58:42
  pipelines. That was one thing that we
• 58:45
  started building. I think it all did kind of.
• 58:47
  We've had this idea of a gateway though for quite a while.
• 58:51
  We just weren't looking at it as a startup from a security
• 58:54
  perspective. It was more around prop
• 58:57
  management, content guardrails, those types of that type of
• 59:02
  functionality which we still want to implement down the road.
• 59:05
  Now we're just starting from kind of the the security layer
• 59:08
  in the enterprise layer.