Latest / In The Loop / Top Three DevDay 2025 Announcements: ChatGPT Apps, AgentKit, Context Capture

Transcript

Top Three DevDay 2025 Announcements: ChatGPT Apps, AgentKit, Context Capture · In The Loop

• 0:00
  Open AI might have just made the biggest strategic bet yet, and
• 0:03
  it's not making AI smarter. They're shifting from raw
• 0:07
  intelligence with new language model updates to moving up the
• 0:11
  application layer to the top of the stock to try and make AI
• 0:14
  more useful. Just this week, Open AI Dev Day
• 0:17
  happened, and it's one of the biggest events of the year for
• 0:20
  companies like Open AI, where they tried to sell their vision
• 0:23
  and their new product features to developers all over the
• 0:25
  world. And for those that have been
• 0:27
  listening in every single week, you'll really understand and how
• 0:30
  important it is for companies like Open AI to retain the
• 0:33
  developer ecosystem. It always gives us a really
• 0:36
  interesting snapshot into the future of Open AI strategy.
• 0:40
  In today's episode, we're going to go through three of the
• 0:42
  biggest feature releases announced during Dev day.
• 0:45
  Whether you should even care about it, If so, what you should
• 0:48
  know about it. And finally, what this is going
• 0:51
  to tell us about the future of AI in the next few years.
• 0:54
  This is in the Loop with Jack Horton.
• 0:56
  I hope you enjoy the show. So let's start today's episode
• 1:11
  with the App Store announcement from ChatGPT.
• 1:16
  So what they actually launched was essentially this new feature
• 1:18
  of apps and ChatGPT that enables you to bring third party
• 1:22
  applications from major companies like Spotify, Zillow,
• 1:26
  Canva, booking.com, Expedia, Figma directly inside the
• 1:30
  ChatGPT interface. Now I think what's funny here is
• 1:33
  that obviously they tried with GPT plug insurance that failed.
• 1:37
  They definitely tried with the GPT store and that failed as
• 1:41
  well. So this is their third attempt
• 1:43
  at a platform strategy and ecosystem strategy around App
• 1:46
  Store ideas. And it basically lets you use
• 1:49
  the applications from ChatGPT and get the benefit of chat GBTS
• 1:53
  intelligence and it's context it has on you and other
• 1:56
  applications plus the third party applications context and
• 2:00
  intelligence. And so the key innovation here
• 2:03
  has been the deep context integration that they're
• 2:06
  enabling or they're saying they're going to enable.
• 2:09
  So these aren't just, you know, let's say Expedia or Zillow
• 2:15
  embedded into ChatGPT. They actually talk to ChatGPT
• 2:19
  and share information back and forth.
• 2:21
  So what they launched was an app's SDK that enables
• 2:24
  developers to build this into ChatGPT.
• 2:27
  And it's all built on top of model context protocols.
• 2:29
  So MCPS, which Open AI has aligned with, is an open
• 2:33
  standard just like pretty much everyone in the industry right
• 2:35
  now. They've launched an app
• 2:36
  directory and apparently going to be seen bringing out
• 2:39
  monetization features, including the Agentic Commerce protocol
• 2:43
  for direct in chat purchases. I might cover this in a future
• 2:46
  episode, but it's another one of those model context protocols or
• 2:50
  agent agent protocols that are emerging, but this time to
• 2:53
  standardise how AI systems can purchase goods via the likes of
• 2:58
  a ChatGPT. And this all makes sense.
• 3:00
  They've got huge distribution for 3rd party businesses and app
• 3:05
  builders and things like that. ChatGPT has 800 million weekly
• 3:08
  active users. The fact that that's weekly is
• 3:12
  ridiculous. And they've got 4 million
• 3:14
  developers building on Open AI's platform of some sorts.
• 3:18
  And just a caveat to this, although I say building on top
• 3:20
  of using their API, so they'll be using that API, many other
• 3:24
  providers, but just through the API, the processing about 6
• 3:29
  billion tokens through their APIs per minute.
• 3:32
  To put that in perspective, that's more weekly users and
• 3:35
  usage than TikTok ever had. And what I think is really
• 3:37
  important to understand about this is it really showcases the
• 3:40
  future of all software interactions because every
• 3:43
  application in the world, whether you're AB to C consumer
• 3:48
  product or your B to B SAS product or your fintech product,
• 3:52
  biotech product, whatever you do, it is going to be
• 3:55
  conversational. This is exactly what we do with
• 3:59
  our customers. We work with B to B SAS
• 4:01
  providers and product management and developers and designers to
• 4:05
  enable them to quickly turn all of their technology into a
• 4:09
  conversational interface that controls all the features, gets
• 4:13
  context about users, takes actions in other applications.
• 4:16
  And we see this happening in real time.
• 4:19
  And I think this is just another major proof point along that
• 4:22
  journey. And think about it, in many
• 4:24
  instances or use cases, you don't want to have to visit 10
• 4:28
  different application interfaces just to get something done.
• 4:30
  And the demos were really cool too.
• 4:32
  I have to say, in one, someone logged into Cosia through
• 4:35
  ChatGPT and was able to watch educational content and courses.
• 4:38
  And they did the same with Zillow, which for those in the
• 4:40
  UK is the platform for people to buy and rent and sell houses.
• 4:46
  And they also did it with Spotify.
• 4:47
  And this follows really a pattern we've seen many times
• 4:50
  before with big, big platform providers, you know, Facebook,
• 4:53
  Google, Apple, iOS. And really what we can see is
• 4:57
  that we're in step 0 right now, which is crazy competition
• 4:59
  between, you know, 5 to 7 to 8 providers in the world with no
• 5:03
  real winning yet. Opening eyes obviously
• 5:05
  identified them context to memory.
• 5:07
  And they're trying to obviously press that advantage by opening
• 5:11
  up a third party platform through the app SDK.
• 5:15
  But execute solution is going to be really important here.
• 5:18
  You know, back in the day of Apple building, you know, iTunes
• 5:21
  and platforms like that, there was many, many companies trying
• 5:24
  to do the same, Even those that should have done a really,
• 5:27
  really good job but failed because it felt just wrong.
• 5:31
  It didn't feel natural. They never nailed it.
• 5:33
  For example, Sony were trying to build the exact same all the
• 5:36
  music producer companies in the world were trying to produce
• 5:39
  their own version of iTunes, but they all did it just wrong.
• 5:43
  And I felt that sometimes with chat GB TS demos, it felt a bit
• 5:46
  gimmicky. In the demo, they asked you to
• 5:50
  design a company logo through a chat interface, speaking to
• 5:53
  Camber and then make a presentation deck.
• 5:56
  And I can see it with the deck maybe, but it just felt like, I
• 5:59
  guess weird, you know, I wouldn't go to ChatGPT to design
• 6:04
  A logo through Camber. I would just go to Camber for
• 6:08
  some of those tasks. I think that really speaks to
• 6:11
  how people in the world should see this, which is that for many
• 6:14
  use cases, it's perfect for buying and finding and
• 6:16
  researching houses. But for many use cases, you just
• 6:19
  want to be in deep work mode and really engage with the platform.
• 6:23
  Now that might change over time. Chat GB TS technology might
• 6:28
  evolve as to a point where you don't need to visit third party
• 6:31
  providers. But what I think this speaks to
• 6:33
  really is open a eyes bid to capture just consuming fact all
• 6:37
  user context possible. So once users, you know,
• 6:41
  experience Koscia with ChatGPT as their personal tutor, you
• 6:45
  don't want to go necessary back to this standard and experience.
• 6:47
  And ChatGPT has all the context of everything.
• 6:50
  So the conversation running up to that, the interactions that
• 6:53
  happened during that and after that, it's how Open AI could
• 6:56
  become the everyday assistant by owning all interaction context.
• 7:12
  I think this is a worthy thing to dig into and deep dive into a
• 7:14
  little bit more with this idea of context and opening eye
• 7:18
  trying to capture all that context.
• 7:20
  Because apps running in ChatGPT will access all the
• 7:23
  conversational context, they'll share information back and
• 7:25
  forth. They'll integrate with Chat GBTS
• 7:27
  memory feature for personalized experience across different
• 7:30
  applications. Open IO calls is talking to apps
• 7:32
  functionality from ChatGPT. The first time you'll see one of
• 7:35
  these applications, ChatGPT will show a privacy policy and what
• 7:38
  data is likely to be shared. Just like Google does.
• 7:41
  Apps will pull contacts from your chat history and provide
• 7:43
  personalized responses based on previous conversations.
• 7:46
  And what we're really seeing here is them trying to be that
• 7:49
  everyday system. But it raises really big
• 7:52
  questions because, you know, some of this makes perfect sense
• 7:54
  using Zillow to look at houses, then using GPT to, I don't know,
• 7:59
  look for data such as how close is this to a good park or
• 8:03
  school? Has there been any good
• 8:05
  restaurants opening soon? The IRS, all that context from
• 8:09
  your house browsing and can provide very relevant answers.
• 8:13
  And I think once users experience this level of
• 8:15
  contextual help, going back is going to be quite tough.
• 8:18
  This is how open AI probably will start to really stress out
• 8:21
  the likes of Google because although Google is way out in
• 8:23
  front, things could change quite quickly over time.
• 8:27
  And as I said, at Mindset, we see this all the time.
• 8:29
  We've just launched a memory feature in beta.
• 8:32
  Our customers are turning their applications into essentially
• 8:35
  lots of different agents that feel like a single conversation.
• 8:38
  And as a user, you can navigate their applications through
• 8:42
  simple conversations and then hop into taking actions across
• 8:44
  other applications. And so it's really interesting
• 8:46
  how the type of technology and work we're doing alongside some
• 8:51
  incredible companies will actually start to be further
• 8:55
  developed and improved by the likes of Chat Chi BT bringing
• 8:58
  out this contextual awareness feature.
• 9:01
  And I, I assume most providers will start to do the same.
• 9:04
  So when we combine these two worlds of our customers,
• 9:07
  applications becoming conversational with these new
• 9:11
  features by Chat Chi BT and soon all of the providers, no doubt
• 9:14
  technology will increasingly just feel quite human.
• 9:17
  Now, personally, I don't like the idea of opening and having
• 9:20
  all this context about me. I'm not a big fan about this
• 9:23
  idea. I'm not a big fan of my
• 9:25
  conversation history suddenly being used as part of web
• 9:30
  searching or the equivalent of modern day advertising as it
• 9:33
  probably will emerge in the next couple of years.
• 9:35
  Because the more details they have on me, the more they'll try
• 9:38
  to sell to me. For some, they live there.
• 9:40
  I, I always have my tin foil hat on with these things because
• 9:43
  really we never got social media right many years ago.
• 9:46
  And I think many of us regret the decisions or the lack of
• 9:50
  decisions, shall I say, that were made during that time on
• 9:53
  regulation, data protection and privacy.
• 10:07
  Anyway, going to the third part of what I think is really
• 10:10
  interesting here with the Dev Day release, which is agent
• 10:13
  care. So Agent Kit is essentially an
• 10:16
  agent and workflow builder for AI agents.
• 10:18
  It included a visual agent workflow builder with drag and
• 10:22
  drop workflows. I think like NAN Orchestration,
• 10:24
  there was a chat box interface that you could put your
• 10:28
  workflows into to test and play and access those workflows.
• 10:32
  It came with an evaluation platform with trace grading,
• 10:35
  which can essentially score multi step agent decisions and
• 10:39
  ensure that it's producing relevant and high quality
• 10:42
  outputs. They've got a connector registry
• 10:44
  which is their objective of trying to handle enterprise
• 10:48
  integrations, and they're also coming out the box with support
• 10:51
  for MCP, which is fantastic. And they're teasing the release
• 10:55
  of a reinforcement fine tuning feature for specialised agents,
• 10:59
  so really being able to make them hyper specialised at some
• 11:02
  use cases. I really like some of their
• 11:04
  approaches here. It's very similar to NAN, to
• 11:06
  Zapier, to Lindy AI. I'm yet to use the work for a
• 11:09
  builder from Open AI, but I've used all the other main
• 11:13
  providers that are basically the same as that Open AI set of
• 11:16
  features and it'll be really interesting.
• 11:19
  It reinforces something we've been saying for a long time at
• 11:21
  Mindset, which is that anything to do, building a genetic
• 11:23
  experiences must happen in configuration and not code.
• 11:28
  There are many cases that you want to build in code, that's
• 11:30
  fine. But really all an agent and a
• 11:33
  workflow is, is a series of different prompts using
• 11:36
  different tools and context of information at the right moment
• 11:40
  and prompts being sent to the model to give the model context
• 11:44
  or hyper specific context in that moment over a series of
• 11:47
  steps. That's all that's happening.
• 11:49
  All the magic happens in the language model layer, and we've
• 11:52
  got a workflow builder and an agent builder.
• 11:54
  So it's really good to see Open AI further prove our belief that
• 12:01
  building agents should happen in configuration and not code.
• 12:05
  But one thing that I would say here with this launch is that
• 12:07
  it's extremely myopic and it surprises me a little bit.
• 12:10
  To be honest, I thought that it'd been a little bit more
• 12:12
  expansive because everything is just Open AI ecosystem only.
• 12:17
  You can only use their models, their evaluation system, their
• 12:21
  deployment methods. It's all locked into their
• 12:23
  ecosystem. And although they're using some
• 12:25
  open standards such as MCP, their proposition really is once
• 12:30
  you're here in the Open AI ecosystem, you can't really
• 12:33
  leave. And that makes sense given their
• 12:35
  clear strategy of making ChatGPT into that next major
• 12:39
  distribution platform, just like Apple did.
• 12:41
  And it makes sense because Open AI is soon to launch their own
• 12:45
  product features with Johnny I've so the legendary designer
• 12:48
  behind most of the best Apple products ever released.
• 12:51
  All this further does is try and lock people in when people don't
• 12:54
  want that from a developer perspective, from a technology
• 12:58
  and vendor perspective. So I think it's going to be
• 13:00
  fantastic for internal use cases as a result, but not so much for
• 13:04
  those that want to be able to use many best of breed systems.
• 13:08
  It also creates this huge platform dependency problem
• 13:12
  because what happens then is these platforms open everybody
• 13:15
  up to cool exciting things to attract customers and developers
• 13:20
  with great distribution and new features and tools.
• 13:22
  But they just gradually close things down.
• 13:24
  Whether that's reducing organic reach on social media, whether
• 13:28
  it's increasing higher and higher revenue cuts like the App
• 13:30
  Store or just building competing first party products like
• 13:34
  Facebook did with the games. And it minds that almost all of
• 13:37
  our customers want to be able to use best of breed systems.
• 13:40
  They don't want to be locked in. Nobody knows what the next 10
• 13:43
  years will look like. So you want to be able to build
• 13:46
  and create value over time and you not have to RIP something
• 13:49
  out because you're locked into an approach.
• 14:02
  Anyway, let's let's conclude and summarize, I guess what we've
• 14:05
  been talking about. I think really there's been a
• 14:08
  bunch of nuclear features from the App Store, agent builder,
• 14:11
  workflow builder, new contacts being shared to companies that
• 14:14
  build on ChatGPT. And what we're really seeing
• 14:16
  here is an operating system war. Open AI is trying to own all
• 14:22
  interactions between humans and systems.
• 14:25
  And no doubt Google or Microsoft and Apple are going to respond,
• 14:28
  especially given that I've heard that Apple is looking to replace
• 14:31
  Tim Cook as CEO. So we're looking at this big
• 14:34
  platform battle that will shape how we interact with technology
• 14:37
  over the next few years. And I, for one, I'm really
• 14:41
  interested in how this all plays out.
• 14:44
  And I think what I've really been pleased with here is that
• 14:48
  this dev day by Open AI will only further open the world up
• 14:52
  to building agentic technology and systems and testing it out
• 14:56
  for themselves. Anyway, that's it for this week.
• 14:58
  I hope you enjoyed the show and I'll see you next week.