Latest / Agentic Conversations (formally mlops.community) / What Happens When Every Developer Has 20 AI Agents?

Transcript

What Happens When Every Developer Has 20 AI Agents? · Agentic Conversations (formally mlops.community)

• 0:00
  If I have 5 or 10 or 20 agents that are working for me full
• 0:04
  time and I'm one person now you extrapolate that out by the
• 0:10
  millions of developers that are out there, where are we
• 0:13
  bottlenecking everywhere. Stephen O'Grady is the Co
• 0:17
  founder and principal analyst at Redmonk, where he spent over 20
• 0:20
  years tracking the technology shaping software development.
• 0:24
  His research spans open source, cloud computing, AI and
• 0:29
  developer tooling, making him one of the industry's leading
• 0:31
  voices on where software is heading next.
• 0:34
  Could he tell us which trends are real, which are just hype,
• 0:38
  and what developers should actually be paying attention to?
• 0:42
  I get the feeling you don't subscribe to this.
• 0:44
  Overly simplistic idea of SAS is dead.
• 0:47
  I do not. I find probably the notion kind
• 0:51
  of laughable. In fact, there was a phrase for
• 0:53
  many years it was you don't adapt SAP to your business, you
• 0:56
  adapt your business to SAP. But what you figure out is that
• 0:59
  customizing the software your business comes with some pretty
• 1:01
  significant costs. I think we're still figuring out
• 1:03
  the paradigms here, right? OK, let's talk about some of the
• 1:18
  research you've been doing. Well, what have you been up to?
• 1:21
  We have been, I think like a lot of people trying to sort of
• 1:25
  understand as best we can the implications of AI on
• 1:32
  developers, on other practitioners, on organizations,
• 1:35
  on open source. And, you know, I think the, the
• 1:40
  sort of clearest take away is not going to come as a surprise
• 1:42
  to anybody who's tracked the field or used the tools is that
• 1:45
  things are moving very, very quickly.
• 1:47
  So that makes a lot of work for us because it's, you know,
• 1:49
  you're, you're, you know, conclusions you come to one day
• 1:51
  or maybe different, you know, sort of the next day.
• 1:54
  So a lot going on. What are some key signals you've
• 1:56
  been seeing? So key signals, I think the
• 2:01
  couple things I think to, to think about.
• 2:04
  So we're here at the MCP Dev Summit, right?
• 2:06
  So I, I ran some numbers before I came down looking at downloads
• 2:11
  of MTP related software, the SDK primarily.
• 2:16
  The growth of that was pretty extraordinary.
• 2:18
  You know, it was a kind of a hockey stick going up and to the
• 2:20
  right, you know, with a caveat that's just downloads, right?
• 2:24
  That doesn't necessarily imply for actual usage and deployments
• 2:28
  and and so on. The sort of flip side of that
• 2:32
  coin is, is that if you, you know, with technologies, that's
• 2:35
  sort of an experience, an initial surge of interest.
• 2:39
  You tend to see sort of a lack of deployments in that the
• 2:43
  downloads and sort of ultimately rise and then greater as people
• 2:48
  figure out like, oh, hey, there actually isn't that much use
• 2:50
  here for me. That hasn't happened in the data
• 2:52
  that we've seen yet. So we can't prove that it's sort
• 2:55
  of being in use everywhere. But is like I said, at least the
• 2:59
  consumption of of the MCP related software is is sort of
• 3:04
  up and to the right. Wait, what are some past
• 3:06
  experiences that have? Probably one that comes to mind.
• 3:10
  We didn't sort of track it much because we were sort of
• 3:14
  convinced that there wasn't a there.
• 3:15
  There was the Web 3 stuff. So Web three saw a ton of sort
• 3:19
  of initial interest and then people figured out, I don't
• 3:22
  actually have much use for this. And so the that initial surge of
• 3:26
  interest in many cases just sort of Peter back out after sort of
• 3:32
  a lack of just a lack of sort of immediate applicability or use
• 3:36
  cases. So, so that's one thing that
• 3:39
  we've been looking at trying to measure, you know, sort of the
• 3:45
  as best we can, you know, sort of via a variety of sources as
• 3:48
  soon As for the uptake of of various other open or sort of AI
• 3:54
  sort of related packaging sort of across a wide variety of of
• 3:58
  software ecosystems. Again, it is just things are
• 4:02
  moving so quickly. There's tons of it, you know,
• 4:06
  and, you know, obviously as, as you know, folks who study what
• 4:10
  practitioners do, we use these tools ourselves, you know, quite
• 4:12
  a bit. So I was joking the other day
• 4:15
  that the, you know, your, your tokens are exhausted, you know,
• 4:20
  come back at 2:00 PM is kind of the, you know, sort of the new
• 4:23
  way I have to wait for my compiler to finish.
• 4:26
  So there's many cases where we're sort of waiting on on our
• 4:29
  our tokens to refresh and sort of as it were, as we use these
• 4:32
  tools. But yeah, I think that the
• 4:35
  biggest, I think the biggest takeaways are things that we've
• 4:38
  known sort of been in ways for a couple of years now, which is
• 4:41
  that the tooling has the ability to generate software at high
• 4:45
  velocity sort of on an individual basis.
• 4:48
  And the implications for that are vast, right?
• 4:51
  So the implications for that on an individual basis, right?
• 4:54
  If I can output, you know, some multiple of the software that I
• 4:57
  used to, what does that mean for me as an individual?
• 5:01
  What does it mean for the infrastructure that I rely on?
• 5:04
  What does it mean for the businesses in a sort of that,
• 5:06
  you know, developers work for? And what does it mean for sort
• 5:10
  of other ecosystems like open source, right.
• 5:13
  So yeah, it's a, it's a, it's a brave new world with a lot going
• 5:19
  on. I get the feeling you don't
• 5:20
  subscribe to this overly simplistic idea of Sassa's dead.
• 5:24
  I do not. I find probably the notion kind
• 5:29
  of laughable, in fact. So before becoming an analyst, I
• 5:33
  used to be a systems integrator. And part of my job was to take
• 5:36
  various pieces of packaged software, CRM portals, and sort
• 5:41
  of all these sorts of things and implement them.
• 5:43
  And one of the things you figure out pretty quickly is that if
• 5:46
  you sit and ask a business to describe how they want an
• 5:51
  application to function, it takes forever because everyone
• 5:55
  has an opinion, everyone wants their sort of little, you know,
• 5:59
  sort of a workflow sort of included.
• 6:01
  And so, you know, for many of the SAS offerings, you know,
• 6:07
  kind of the SAS is kind of like the, I don't know, there was a
• 6:11
  phrase for many years it was you don't adapt SAP to your
• 6:15
  business, you adapt your business to SAP.
• 6:18
  And you know, on the one hand, that seems, you know, kind of
• 6:23
  backwards, right? You think you, hey, I want to
• 6:25
  customize the software my business.
• 6:27
  But what you figure out is that customizing the software your
• 6:29
  business is it comes with some pretty significant costs.
• 6:32
  Right. And for what value?
• 6:35
  For what value, right? But it makes hiring more
• 6:39
  difficult, right? So in other words, I can go find
• 6:41
  somebody off the street that knows Work Day or Salesforce or
• 6:44
  SAP or whatever. I can't hire somebody off the
• 6:48
  street who knows my bespoke like a CRMERP, you know, HR system.
• 6:54
  It also to me places too much emphasis on the software as the
• 6:59
  only asset, right. So if your work day as an
• 7:01
  example and you are responsible for things like HR compliance
• 7:05
  and geographies all over the world, do you really want to
• 7:08
  have to recreate that from scratch?
• 7:10
  Probably not. So like, I think, you know, AI
• 7:14
  will have certainly an impact on, you know, various SAS
• 7:19
  businesses, but the whole SAS is dead.
• 7:21
  You know, to me is a, to me that suggests that the people saying
• 7:26
  that are people who think that the only thing that goes into
• 7:29
  these equations are software. And that is, in my experience,
• 7:32
  not the case. For me, it feels much more
• 7:34
  nuanced. It feels like you have certain
• 7:39
  chunks of a workflow that don't need to be part of a greater SAS
• 7:44
  solution, and those you can claim as your own, but that's
• 7:49
  not the whole story. Well.
• 7:51
  And also, I think a lot of it depends on, you know, look, if,
• 7:54
  if a SAS can't give you what you want, all right, if you are sort
• 7:59
  of a, a outlier, right, for whatever reason, OK, you know,
• 8:03
  AI can be a, a good way to, you know, present alternatives.
• 8:08
  But in many cases, do you really need to reinvent the CRM wheel
• 8:12
  or the ERP wheel? Probably not.
• 8:14
  Now, all that being said, and what I, I am certain you will
• 8:17
  see, and you know, we're seeing it a little bit already, is
• 8:21
  that, you know, a is uses the proverbial Amdahl mug, right?
• 8:25
  So the Amdahl mug was back in the day when IBM sold
• 8:31
  mainframes, right? They used to be able to come in
• 8:33
  and dictate terms to their customers because it's like,
• 8:36
  we're IBM, we're the only place you can get these mainframes.
• 8:39
  And Amdahl comes along and makes a, a competitive product.
• 8:43
  And so even if they had no intention of switching, what
• 8:46
  they would do is when the IBM salespeople came in, they would
• 8:49
  prominently have an Amdahl mug sitting on their desk, right?
• 8:53
  You know, to remind them like, hey, I have leverage from a
• 8:55
  negotiating standpoint. So I'm certain it will be used
• 8:58
  that way. But I don't buy the wholesale
• 9:01
  that's a replacement. It just the cost to me, you
• 9:04
  know, significantly outweigh any potential benefit.
• 9:06
  Well, for me, there are a lot of use cases that I find
• 9:11
  fascinating when. Folks.
• 9:15
  Can stand something up faster than they can get a tool through
• 9:21
  procurement. Yeah.
• 9:23
  So that's like the opposite end of the spectrum where you say,
• 9:25
  OK, these work days and Salesforce, those have been in
• 9:28
  my company since before the beginning of time.
• 9:32
  We're not going to really worry about that, but for a new tool,
• 9:36
  you really have to have a strong value prop as to why I'm going
• 9:40
  to use that and go through all the headache of taking six
• 9:43
  months to get it through my. Procurement team, yeah, Yeah.
• 9:46
  And as the person at our company who deals with procurement, I'm
• 9:49
  sympathetic to that stance. I think the flip side of that is
• 9:52
  that, you know, when I've had these conversations with folks,
• 9:56
  they're like, hey, what if I could take my licensing costs,
• 9:58
  you know, to sort of make them zero?
• 10:00
  And what if I could do XY and Z? And it's like, right, but then
• 10:03
  you have to operate it, right. And so there's a whole lot of a
• 10:09
  number of sort of very large banks, as one example, you know,
• 10:13
  sort of decade or more ago figured, OK, you know what, I
• 10:16
  don't want to pay, you know, sort of a, a platform tax.
• 10:20
  So I'm going to write my own essentially developer portal
• 10:24
  software. And what they found then is, is
• 10:28
  that now you you own that, right?
• 10:30
  You are responsible for keeping it up to date, you're
• 10:32
  responsible for securing it, you're responsible for operating
• 10:34
  it. And number of these banks have
• 10:37
  subsequently come back and said, you know what, this was not the
• 10:40
  best use of our time because this is not differentiating for
• 10:43
  us, right? Us being able to host
• 10:45
  applications doesn't give us a competitive advantage over other
• 10:48
  organizations. And so, you know, I think you'll
• 10:51
  see, you know, sort of in some cases, you know, companies that
• 10:54
  that sort of get frustrated with the procurement are like, OK,
• 10:57
  you know what, I'm just going to go build this, then realize
• 11:00
  6/12/24 months down the line, like I don't really want to be
• 11:04
  in the business of operating this.
• 11:05
  And it's not really giving me sort of anything, you know, I
• 11:08
  couldn't get through easily and more more easily and cheaper
• 11:11
  elsewhere. What I saw, which was very eye
• 11:15
  opening for me is a company that I feel is very advanced in the
• 11:22
  way that they're doing agents. Yep.
• 11:24
  They have had the talk to my data use case since basically
• 11:29
  when Chachi BT first launched. That was one of their first use
• 11:33
  cases that they set up. They've been tuning it for the
• 11:36
  past three years and it is very well done.
• 11:40
  They just hit me up the other day and said, can you tell me
• 11:42
  all of the talk to your data vendors because I think we're
• 11:45
  ready to outsource it, yeah. And I, I think you'll see a
• 11:48
  fair, a fair amount of that because it's look it, it's easy
• 11:52
  to, you know, I, I tell this story all the time now, right?
• 11:57
  So you know, up until November when you know, the first of the
• 12:03
  new models arrived right in Opus 4-5 and obviously 46 right 46
• 12:09
  now 4-5, I think was the one that was released in in late
• 12:13
  November. So we had this inflection point,
• 12:15
  right? There's a before and there's an
• 12:16
  after. And the tools are very truly
• 12:18
  useful before they would do interesting things in the wake
• 12:21
  of that. Oh my God, it's just a night and
• 12:24
  day difference, right? So the canonical example I point
• 12:27
  to there is Ala Dodgers fan who happens to do the website
• 12:31
  designed for CNN. So he built this essentially
• 12:35
  data analytics portal for the Dodgers and released it as open
• 12:37
  source software. So I've always thought like, oh,
• 12:40
  hey, I should pour this over for the Red Sox and so on, and just
• 12:44
  never found the time to do it. And I would always throw, you
• 12:47
  know, all of the tools like hero and cursor, windsurf, anti
• 12:52
  gravity, all these things at it. And they did get most of the way
• 12:55
  there, part of the way there, and then joke and sort of stall
• 12:59
  out and then over the holiday. And like a lot of folks did when
• 13:03
  Anthropic was giving people sort of more credits, I threw a clot
• 13:07
  at it and it worked. And it did all the things it
• 13:09
  needed to do, which was great. The flip side is, is that after
• 13:13
  realizing this, it's like, oh, OK, you know what?
• 13:16
  I want a dashboard for this. And, you know, I want an
• 13:20
  application that's going to go do that.
• 13:22
  So I ended up building, I don't know, a dozen, you know,
• 13:25
  different things like little small toy apps.
• 13:29
  And you quickly, this is just on an individual level, right?
• 13:32
  You get to the point it's like, right?
• 13:34
  Wait, which one was that and where did I put it?
• 13:36
  And what is that actually running on?
• 13:39
  I don't remember where it's hosted.
• 13:40
  What conversation was I having with Claude?
• 13:43
  Where's the context? I have no idea where we're at.
• 13:45
  And that's literally on an individual like, you know, sort
• 13:49
  of very amateur basis. And so, yeah, I think when a lot
• 13:56
  of enterprises figure out like, oh, I can start turning these
• 13:59
  things out, I think they'll be this burst of turning things
• 14:01
  out. And then after a little while,
• 14:03
  it's like, OK, this one is real and sort of new and
• 14:06
  differentiated. And I want to, I want to keep
• 14:08
  this. But in many cases it, you know,
• 14:10
  they're going to find, you know, the cost of those sort of
• 14:13
  operating, securing, protecting, maintaining, you know,
• 14:15
  continuing to develop, you know, these one offs is not going to
• 14:18
  be worth it. They're going to start throwing
• 14:20
  around the sexiest word in the English language.
• 14:23
  Yeah, right. Yeah.
• 14:25
  Exactly. Yeah, so, but I.
• 14:27
  You know what I noticed when playing with it over the
• 14:30
  holidays is I had this theory of oh wow, software is becoming
• 14:35
  local 1st as opposed to I'm going to create an app and then
• 14:40
  expose it and expect scale. Yep, it's like, no, I'm just
• 14:44
  going to have my own things and help my own workflow.
• 14:47
  But then it went from me going software is local first now to
• 14:54
  I've fully pivoted and now I'm like, but why do I even need an
• 14:58
  app if I can just build a skill that can do it?
• 15:01
  So I don't need the overhead of an app or even a local first app
• 15:05
  is not it's too much to worry about.
• 15:09
  Let's just do it with a skill because it's the lightest thing
• 15:12
  that I can think of. Yeah, yeah.
• 15:14
  And that's the thing. I think we're still figuring out
• 15:15
  the paradigms here, right, right.
• 15:17
  We're still figuring out, you know, to your point, like what
• 15:21
  do we need an app for? What does that app look like?
• 15:24
  Do we need an app at all? Do we need a skill?
• 15:26
  Do we need, you know, for some, you know, sort of stub or
• 15:29
  approximation like. So I think that the biggest
• 15:33
  thing is, is that the sort of, as we talked about at the top,
• 15:37
  you know, when you take the cost of producing code and drive it,
• 15:40
  you know, maybe not quite to 0, but near 0, we're going to have
• 15:45
  to rethink a lot of different things.
• 15:47
  We're going to have to rethink, you know, what we do
• 15:48
  individually. We're going to have to rethink
• 15:50
  the infrastructure we rely on. Businesses are going to have to
• 15:52
  rethink what they build. You know, because it's, I'm
• 15:56
  trying to remember the context in which I brought this up from
• 15:59
  research standpoint. But one of the things that we
• 16:01
  see, you know, we've been through this sort of, you know,
• 16:04
  multiple ways where we've taken something relatively monolithic
• 16:09
  and broken it up into pieces. So virtualization is probably
• 16:12
  the simplest example of that, right?
• 16:14
  And we had physical machine, you break that up into sort of some
• 16:17
  multiple of that virtual machines.
• 16:18
  And one of the things you figure out is, is that that comes with
• 16:21
  benefits, but it also comes with costs, right?
• 16:23
  And you know, I talked to one of the first large banks that move
• 16:27
  over to the VM Ware and said, you know, hey, how's it going?
• 16:30
  And they were kind of lukewarm when I thought, oh, is this, you
• 16:33
  know, a problem with the software?
• 16:34
  Is this an issue? And he said, no, no, like, look,
• 16:35
  we've had all the marks that we want.
• 16:37
  Our utilization rate is way up, you know, the, the deployment
• 16:40
  density is way up. And I was like, well, what's
• 16:42
  the, But he said, you know, like, you know, managing, you
• 16:45
  know, 500 physical machines is one thing, managing 5000 virtual
• 16:48
  ones is a different thing, right?
• 16:50
  And so I think what we're going to see for a lot of these
• 16:52
  organizations is, you know, if you now have 10X more
• 16:55
  applications to manage, that is going to be a special challenge
• 16:59
  of its own. Yeah.
• 17:00
  Or even I know folks are talking about that with agents.
• 17:03
  I've now got 5000 agents that our company uses, or 30,000
• 17:09
  agents it sounds like. They're.
• 17:12
  Testing that and they're running fast and they want to do that,
• 17:16
  but then later they'll recognize maybe it's not worth it.
• 17:20
  Yeah. Yeah, I, I think that's right,
• 17:22
  right. And it's, you know, we're in
• 17:24
  that initial burst of creativity where we're just creating things
• 17:27
  because we can and, and because it's easy and it's frictionless
• 17:30
  and so on. And, you know, we're going to
• 17:32
  get to a point where, you know, we, we have to reckon with that
• 17:36
  volume, right? And if we're not there yet, you
• 17:39
  know, I think many individuals are certainly I am.
• 17:41
  I'm like, Oh my God, where is, as I said, where is all this
• 17:44
  stuff and where do I put it? Organizations are going to get
• 17:47
  there pretty quick. Yeah.
• 17:48
  So then that's a very clear signal.
• 17:52
  What other stuff have you been seeing in the research?
• 17:55
  So. One of the things that is coming
• 17:57
  up in sort of various quarters is the strain on infrastructure,
• 18:03
  right? So you know, when you know, you
• 18:06
  think about individual developers, right?
• 18:09
  And you think about things like projects like Gastown, right
• 18:11
  from Steve Yagi, right, which is like, OK, you know, we have a
• 18:14
  team or swarm or pick your term of choice of, of individual
• 18:17
  agents, right? So if one developer suddenly
• 18:21
  becomes overnight a, a essentially a swarm of dozens of
• 18:26
  developers, like what does that do to our infrastructure?
• 18:29
  And is our infrastructure built for that?
• 18:31
  And the answer is no. And So what you're seeing from,
• 18:35
  you know, things like package sort of management repositories
• 18:38
  is that their traffic is through the roof and their cost for
• 18:43
  simply maintaining things like bandwidth, right, and scale and
• 18:47
  so on are dramatically up, you know, because there's much more
• 18:50
  usage. And that's without getting into
• 18:53
  all the other things that go into our infrastructure, right?
• 18:55
  Like, so, you know, triaging PRS, right?
• 19:00
  How do I sift through these, right?
• 19:03
  I just talked on somebody earlier that apparently the, the
• 19:06
  time to sort of resolve PRS is up significantly, you know, in
• 19:11
  part because there's just so many more of them.
• 19:14
  So, yeah, like I said at the, I think at the top, the the ripple
• 19:17
  effects from, you know, sort of taking this this code and making
• 19:21
  it so much easier our profound and we're seeing that sort of
• 19:25
  manifested in just various strains and and stresses on our
• 19:30
  infrastructure, you know, that we really have.
• 19:32
  Yeah. And which different parts of the
• 19:34
  infrastructure? Because I instantly think if I
• 19:36
  have 5 or 10 or 20 agents that are working for me full time and
• 19:44
  I'm one person now you extrapolate that out by the
• 19:49
  millions of developers that are out there.
• 19:50
  Yep. Where are we bottlenecking
• 19:54
  everywhere. Right.
• 19:56
  So everything from code review to build systems, production,
• 20:00
  deployment, security review, I mean Pickett, right?
• 20:04
  And any one of these things is getting swamped.
• 20:07
  And you see it's an open source, you know, where maintainers are
• 20:10
  just flooded with, you know, essentially AI slop PRS.
• 20:16
  And so it really isn't sort of an aspect of infrastructure
• 20:20
  that's not as we're being stressed, you know, by the
• 20:23
  volume, you know, that we're seeing.
• 20:25
  And, you know, sort of AI can help with some of this, right?
• 20:30
  So in other words, like, you know, there are companies doing
• 20:34
  AI security review, there are companies doing AI QA and
• 20:38
  software review and all these sorts of things.
• 20:40
  But it's going to be, you know, it's going to be a bit right
• 20:43
  before, you know, those technologies are sufficiently
• 20:45
  mature to to really help offset the stresses here.
• 20:48
  So honestly, at this point, you know, really the only thing
• 20:51
  that's gotten a lot easier is generating the code.
• 20:53
  Everything that's downstream of that is.
• 20:54
  Is more complicated. Yeah, the code reviews, I I have
• 20:59
  not been sold on these PR code review type of agents because it
• 21:04
  feels like just a different prompt and a different sub
• 21:07
  agent, yeah. Yeah, well, and it's, you know,
• 21:10
  you see people trying to get around it.
• 21:11
  It's like, let's talk about somebody last week who they just
• 21:16
  pit the agents against each other, right.
• 21:18
  Almost like a Ralph loop. You know, the Ralph Wiggum loop
• 21:22
  came out, I guess it was over the holidays, I think.
• 21:25
  But anyhow, they'll, they'll basically bounce, you know, sort
• 21:27
  of code between two different agents and, you know, from two
• 21:31
  different models from two different vendors to try to, you
• 21:34
  know, sort of get some, some, I don't know, outside
• 21:37
  objectiveness sort of as it were.
• 21:39
  But yeah, it's, it's, it's it's like I said, it's going to be a
• 21:44
  bit before we, you know, digest all this stuff is going to work
• 21:47
  together. Yeah.
• 21:48
  You did mention the package managers getting flooded and I
• 21:52
  find it fascinating that we also saw UV get bought and then BUN
• 21:59
  also got bought and you have any thoughts on that?
• 22:04
  Yeah. So, you know, I think the short
• 22:08
  version is, is that, you know, what you're seeing in terms of
• 22:10
  the acquisitions are people, you know, who want to control, you
• 22:15
  know, sort of critical pieces of infrastructure, right, and choke
• 22:18
  points. Yeah, right.
• 22:20
  And you know, I think, you know, that makes sense.
• 22:23
  You know, we'll we'll see sort of how that plays out and you
• 22:25
  know, whether that, you know, becomes a problem in terms of,
• 22:27
  you know, sort of how some of the projects are sort of
• 22:29
  developed, manifested, shared, etcetera.
• 22:32
  But on the package management front, yeah, really I think if
• 22:35
  you talk to any of this sort of major pack, you know, package
• 22:38
  repositories, you know, they're seeing significantly, you know,
• 22:43
  more traffic, right? And it's an inevitable
• 22:47
  consequence, right? If you, you know, as we were
• 22:49
  talking about a moment ago, if you take 1 developer and turn
• 22:52
  him or her into 20 developers virtually, right, the systems
• 22:57
  are simply not equipped. They were restraining, you know,
• 22:59
  sort of with the demand pre AI and now you have, you know, sort
• 23:03
  of this additional multiple, you know, sort of applied.
• 23:06
  It's yeah, like I said, it's inevitable in terms of how that
• 23:09
  plays out. I have some questions from a
• 23:12
  mutual friend, Alex, that I wanted to pull up.
• 23:15
  OK. And he was asking about we kind
• 23:19
  of answer this one. You know, Red Monk tracks
• 23:21
  developer adoption patterns better than almost anyone.
• 23:24
  Let me try. What signals are you seeing
• 23:27
  around MCP adoption? Is it following the trajectory
• 23:31
  of early integration standards like REST or Graph QL or is
• 23:36
  something different happening? Yeah.
• 23:38
  So the way I'll answer that is that I posted this is bad looked
• 23:44
  it up for coming down here. I posted on Blue Sky in May of
• 23:47
  last year that in, in our experience at Red Monk, we've
• 23:52
  been doing this for a very long time.
• 23:54
  MCP was the fastest project to become a de facto standard that
• 23:58
  we had seen. And Kelsey Hightower had
• 24:02
  replied, no, no, I think it's Docker.
• 24:04
  And obviously a lot of it depends on sort of what metrics
• 24:07
  you're looking at and and and so on.
• 24:08
  But I went back and for us, Docker had been our our sort of
• 24:12
  fastest growing de facto standard and that took about 13
• 24:16
  months. MCMCP took about 13 weeks.
• 24:20
  And so the sort of rise was meteoric.
• 24:24
  And I have, like I said, I've been doing this a while.
• 24:27
  I cannot remember another time when, you know, it was a period
• 24:31
  of like 2 to three weeks from our first hearing about it in
• 24:34
  what's it been in November, October, November, I guess at
• 24:37
  24, I think that's right. And then, you know, within less
• 24:41
  than a month talking to multiple vendors who were like, Yep, this
• 24:44
  is the standard. I'm like, I literally just heard
• 24:46
  about this last month. Like that doesn't happen and so.
• 24:50
  Why do you think that is? Why do you think everybody
• 24:52
  jumped on board so quickly? Well, I mean, I think for, for
• 24:55
  me, I think it's a simple, it's a fairly simple equation in that
• 25:00
  the models are incredibly capable, but we didn't have good
• 25:05
  sort of good standardized ways to connect them to different
• 25:08
  data sources, right. And they are that much more
• 25:11
  helpful, productive, useful, you know, capable, you know, when we
• 25:15
  can give them access to things that, you know, sort of aren't
• 25:18
  public in sort of one way or another.
• 25:20
  And so. So there's pent up demand.
• 25:23
  Pent up demand was like, hey, this is great, but I'd like to
• 25:25
  have this over here. So as an example, if you go and
• 25:27
  look at Google Trends data, all right, and you you look at data
• 25:32
  centers as a trend, there's a noticeable spike in December of
• 25:37
  2022, I think it is. And it occurs roughly a month
• 25:42
  after ChatGPT debuted. And so who knows what it is.
• 25:47
  But certainly one potential explanation for that is that
• 25:49
  people looked at this and said, wow, this is incredible
• 25:51
  capability and technology. What if I could have this on
• 25:55
  site and do it, you know, sort of feed it my data?
• 25:58
  And, you know, I think that was a lot of the sort of vision and
• 26:01
  promise for MCP upfront is OK, I can begin to connect this and
• 26:05
  sort of make it useful. And so, you know, my question
• 26:09
  coming in, you know, to this event was, is that still true?
• 26:13
  Right. And we we know that data was As
• 26:15
  for where it was, you know, about a year ago.
• 26:18
  But you know, you see in in what was it February, right?
• 26:21
  There's a piece that came out. It was like, oh, MCP is dead.
• 26:23
  Long live CLI on us. And so the question was, OK, you
• 26:27
  know, are we seeing anything in a in a way that would indicate,
• 26:30
  you know, sort of a drop off or as we talked about at the top
• 26:34
  right, we're not like the the charts, at least at this point
• 26:37
  are still sort of up and to the right.
• 26:39
  You know, they're growing tons of tons of downloads and so on.
• 26:42
  So yeah, I think the short version is, is that it is
• 26:45
  incredibly, incredibly me York in terms of its rise.
• 26:50
  And that's largely driven by the fact that it is.
• 26:55
  Yeah, just giving giving models additional data and context the
• 26:58
  sources they didn't. Have I want, I wanted to like
• 27:01
  meander down one of these streets, but I think I'm, I'm
• 27:04
  just leave it because I think that is, you said everything
• 27:07
  that needed to be said and it's like beating a dead horse a
• 27:09
  little bit. So I'm I'll go on to the next
• 27:14
  question. The next one that I had from
• 27:18
  Alex was asking about. In your research, have you seen
• 27:21
  what folks are actually building with MCP?
• 27:25
  Yes, so I don't, I'd have to look at that.
• 27:29
  I have to have the graphic in front of me.
• 27:30
  So folks that are listening to this, if you go to LinkedIn,
• 27:34
  look up Stephen O'Grady my, my LinkedIn page.
• 27:36
  I posted the graph that I'm talking about.
• 27:39
  I don't know, it's what is it today, Today's April 2nd.
• 27:42
  It would have been April 1st or March, the last day of March, I
• 27:45
  can't remember which they did. Anyhow in there, you know, sort
• 27:49
  of there's a sort of rough breakdown in terms of categories
• 27:51
  that we're seeing. You know, so most of it's
• 27:55
  predictably just sort of generic AI, there's CRM, there's
• 27:58
  commerce, there's all kinds of different things in there.
• 28:02
  But that's also sort of a rough assessment of categories from,
• 28:07
  you know, sort of base software descriptors.
• 28:09
  All right. So what we don't have insight
• 28:11
  into and what's hardest to get always is where people actually
• 28:15
  doing in deployment, right? Because not surprisingly, most,
• 28:18
  most folks don't let us come in and see their running in
• 28:21
  production infrastructure. So we have to sort of, you know,
• 28:23
  work backwards from all right, what are we seeing in terms of
• 28:27
  patterns of, you know, downloaded software and packages
• 28:30
  and, and things of that nature. But you know, the, the, I guess
• 28:33
  the gist of it from an MCP perspective is that there are
• 28:37
  lots of categories, right? It's not a, oh, hey, here two or
• 28:40
  three or four or five. And then I think we graphed, I'd
• 28:43
  say probably north of 20 sort of different sort of rough high
• 28:47
  level categories of of of usage. Well, it begs the question about
• 28:53
  the ecosystem around MCP also and what you've seen in that
• 28:56
  regard. Yeah.
• 28:57
  So the ecosystem, I think to me the critical step was, and this
• 29:03
  will sound self-serving because we're at a foundation event.
• 29:05
  But you know, the biggest issue to me was initially was OK, it
• 29:10
  sort of took office as de facto standard.
• 29:12
  But it's very difficult for for large vendors to contribute to
• 29:16
  projects that are wholly owned by another vendor, right?
• 29:19
  Because they are appropriately concerned about like, hey, am I,
• 29:23
  why am I sort of underwriting your software development costs?
• 29:26
  Like, you know, when you can on a whim do whatever you want with
• 29:28
  this project. So the, the key for me, and I
• 29:32
  frankly took longer than I thought it would, but the key
• 29:35
  for me was like seeing it through housed within a
• 29:38
  foundation. Because once you have a
• 29:40
  foundation, the ecosystem becomes much easier to form
• 29:43
  because the large vendors who have the ability to throw people
• 29:46
  at the project are comfortable then working on a, on a project
• 29:50
  in a, in a vendor neutral space, right.
• 29:53
  And like I said, if that project stays with Anthropic, you know,
• 29:57
  not that Anthropic can't do a good job of developing.
• 29:59
  They iterate very, very quickly. They're very good at what they
• 30:02
  do. But if you want industry
• 30:04
  support, you know, at some point it has to become a sort of
• 30:09
  neutral project because you just can't attract the same type of
• 30:12
  commercial community, at least in many cases the non commercial
• 30:15
  community as well. So that was the biggest step, I
• 30:17
  think in terms of allowing a a large ecosystem to form around
• 30:21
  it. And how about all of the
• 30:24
  different, so you have companies that are supporting MCP and the
• 30:30
  and the protocol, but then you have companies that are being
• 30:35
  built on top of it. Have you looked at that type of
• 30:39
  ecosystem also? Not as much, you know, we're
• 30:43
  we're sort of more focused on the lens at this point in terms
• 30:46
  of all right, you know, sort of what's the base level
• 30:47
  infrastructure. So in other words, like, you
• 30:50
  know, just on stage this morning, right, you have the
• 30:52
  Duolingos and you have this for all these companies here, like
• 30:54
  here's what we're building on, here's how we're using it and so
• 30:56
  on. Those are great.
• 30:59
  Those are interesting. But in other words, like our
• 31:01
  focus initially has been on, all right, what is this, you know,
• 31:05
  sort of what's been the growth of this project and what does
• 31:07
  this look like? Because without a certain level
• 31:10
  of critical mass, a certain sort of traction, you know, the kind
• 31:14
  of the, the end user use cases are academic because they're
• 31:18
  not, you're not going to get there without that kind of that,
• 31:20
  that traction. So, you know, we see some of
• 31:23
  those sort of common ones, you know, that you'd see in in, you
• 31:27
  know, things like the keynotes this morning, but don't have any
• 31:30
  quantitative insight into that. One thing that the last guest
• 31:33
  who is here, Jeremiah, just said is the majority of the use cases
• 31:37
  he's been seeing are internal use cases of NCPI.
• 31:42
  Wonder if you have any insights into that.
• 31:45
  I'd say that's very likely, you know, because like, you know,
• 31:49
  particularly in the early stage, I mean, you know, when, when MCP
• 31:52
  was really like there was no off initially.
• 31:54
  Right now it's, you know, there's a lot 22.1 I think.
• 31:58
  And you know, it's the, you know, you heard multiple times
• 32:02
  this morning, you know, the, the trifecta, right, of external
• 32:05
  access and you know, sort of internal date, all these sort of
• 32:08
  different sort of factors that can contribute.
• 32:10
  So in many cases, I think what we'll see, you know, the, the
• 32:13
  build out will begin internally, right?
• 32:15
  It'll say, OK, I want to build a system that is not exposing
• 32:20
  things to the outside world. And I want to limit the blast
• 32:24
  radius of any sort of learning experiences that I might have.
• 32:29
  So everything is going to be focused initially inwardly and
• 32:32
  then I'll get to the external piece, you know, eventually.
• 32:36
  You know, that being said, you know, I just had a conversation
• 32:38
  upstairs with you know who somebody was telling me they
• 32:41
  have three external facing MCP servers at their house, right?
• 32:45
  So. To turn on the lights.
• 32:48
  I have no idea. I didn't, I didn't get into what
• 32:50
  they were for. But but yeah, I think you know,
• 32:53
  the the sort of default pattern I would expect is for people to
• 32:56
  start internally because it's the safer thing to do.
• 32:57
  Yeah, in theory, but then when you delete some databases or you
• 33:03
  drop tables, then oh. Safer.
• 33:06
  Yeah, not safe. There's a difference, at least.
• 33:09
  Yeah, you're not ending up on the front page of the Internet.
• 33:11
  That's right. For something that wasn't the
• 33:13
  best. Yeah, no, not like I can't
• 33:16
  remember who the, you know, the meta researcher who is installed
• 33:19
  open claw. It was a yeah.
• 33:21
  It was like a I met a safety person there.
• 33:25
  It's so ironic. So one of their safety
• 33:27
  researchers talking about having to run over and unplug their
• 33:30
  mini, you know, it was like just using a bomb.
• 33:32
  Because it's like. Don't delete my e-mail.
• 33:34
  It's like OK I'll delete your e-mail so.
• 33:36
  So yes you will it. It's safer, not safe.
• 33:39
  Have you done any research into the open Claw ecosystem?
• 33:42
  Yeah, yeah, I had to stood up for a little bit and then it's
• 33:45
  like, yeah, you know what, I'm not, I'm not ready to dedicate
• 33:48
  the resources that I need to, to, to keep this safe right now.
• 33:52
  It's like, yeah, you know, so I, I was in the fortune of camp.
• 33:55
  I already had a Mac mini, so I didn't need to.
• 33:57
  I didn't need to wait on this. Premium.
• 33:59
  Yeah, it was one of these like, oh, hey, this is really cool.
• 34:01
  Then you dig into a little bit and you see the the breeches and
• 34:05
  I'm like, you know what, like there are people who are
• 34:09
  yellowing it and, you know, more power to you.
• 34:12
  But I don't have the time to, to worry about that right now.
• 34:14
  So I'll, I'll wait. And I like the vision.
• 34:17
  I like the promise. And you know, my, my hope is
• 34:19
  that ultimately I'll get to something that helps in sort of
• 34:21
  manage my day-to-day better, you know, much as Open Cloud does.
• 34:24
  But it's a little, it's a little early in that development
• 34:27
  process for me. Yeah.
• 34:28
  Well, this has been great man. I appreciate you coming on here.
• 34:31
  Not at all. My pleasure.