AI Engineer World's Fair 2025
A year of Gemini progress + what comes next
Read the talk
A year of Gemini progress, from model releases to a developer platform
A Gemini 2.5 Pro launch opens a roadmap spanning research transfer, proactive assistants, reasoning, context limits, and AI Studio’s direction as a developer platform.
From a talk by Logan Kilpatrick
A Gemini release before the announcement goes live
The announcement has not yet gone live when Logan Kilpatrick introduces a new Gemini 2.5 Pro update onstage. The immediate question is whether this release closes the gaps developers found in earlier versions well enough to become the final update. He calls it “hopefully the final update,” leaving room for further changes if feedback exposes problems.
Kilpatrick calls the update state of the art on Aider and HLE, without supplying scores or evaluation conditions. He connects the improvements to feedback on previous versions and describes 2.5 Pro as a turning point both inside DeepMind and in developers’ perception of Gemini. The complete launch slide also highlights LMArena and WebDev Arena gains, thinking budgets, a stable model candidate, and the Pareto frontier. The release is a candidate for stability, but the invitation to report failures remains open.
For trying the release, Kilpatrick points developers to ai.dev and users to the Gemini app. Those are the access routes he presents during the announcement; the rest of the talk explains how Google reached this point and where the model, app, and developer platform are heading.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Research breadth feeds the mainline model
The retrospective starts with the density of work packed into the preceding year. DeepMind’s research spans science, robotics, and general models, and Kilpatrick identifies that breadth as an advantage: specialized research can feed back into the mainline Gemini model rather than remaining isolated in separate projects.
He cites AlphaProof and AlphaGeometry as examples of custom-model work improving the mainline models in their respective domains. The important distinction is between developing a specialist system and transferring useful results into a general model. Kilpatrick describes that transfer as happening, but does not specify a training mechanism or quantify either system’s contribution to 2.5 Pro. He leaves a deeper discussion to Jack’s upcoming presentation.
Adoption has expanded alongside the research. Citing Sundar Pichai’s Google I/O keynote, Kilpatrick reports approximately 50× growth in monthly tokens processed across Google products and APIs over a year. The keynote’s totals are 9.7 trillion a year earlier and over 480 trillion at the time of its May 2025 report. This measures aggregate token volume, not inference speed or compute growth. Kilpatrick connects the increase to demand for Gemini within Google and from external developers.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Bringing research, models, and products together
Part of the explanation is organizational. Google previously had separate teams pursuing different strands of AI research. The April 2023 consolidation of Google’s Brain team and DeepMind created a shared organization with a remit extending beyond foundational research: build models and deliver them to the rest of Google and to external users.
Product teams subsequently moved closer to the research organization. Kilpatrick describes a further step earlier in the year of the talk; the Gemini app team’s own move had been announced in October 2024. The resulting structure puts research, model creation, and product delivery inside DeepMind, with the Gemini app serving consumers and the Gemini API serving developers. The slide pairs an organizational update with Kilpatrick’s post about moving AI Studio and the Gemini Developer API to DeepMind.
For Kilpatrick, that proximity makes it easier to collaborate with researchers and bring new capabilities to developers. His formula is direct: bring the best people together, find infrastructure advantages, and ship. The organizational change matters because it connects research progress to product delivery.
Veo provides an example of demand for the resulting products. Kilpatrick jokes that its reception is “burning all the TPUs down,” describing strong interest rather than a measured infrastructure failure. At this point in the talk, Veo is already available in the Gemini app; developer API access appears later in the roadmap.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
From shared sign-in to a universal assistant
The Gemini app’s intended destination is a universal assistant. Kilpatrick explains this through a familiar comparison: historically, a Google account connected someone to many Google products, but primarily by signing them into each one. It did not itself provide a rich, stateful experience spanning those products. Gemini is intended to become that connecting thread.
| Connection across products | Role |
|---|---|
| Google account | Shared identity and sign-in |
| Gemini’s intended role | A stateful assistant spanning products |
The next step is proactivity. In Kilpatrick’s description, most AI products still make the user initiate the interaction and do much of the work. A proactive assistant would take more initiative, changing not only the interface but also who starts a useful interaction. This remains a product direction rather than a demonstrated workflow in the talk.
For feedback on the app itself, Kilpatrick directs people to Josh and the Gemini app team rather than to him. That division becomes useful as the roadmap separates the consumer assistant from the underlying models and developer tools.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
More modalities, more reasoning, more context
Gemini’s original ambition was a single multimodal model handling audio, images, and video. Native audio and text-to-speech capabilities announced at that year’s I/O move it toward that goal: users can speak with the model, and conversational audio powers the Astra experience and Gemini Live. Kilpatrick describes the destination as an omnimodal model, bringing these capabilities together.
Kilpatrick characterizes Veo as state of the art across several areas, without naming evaluations or scores, and hopes to bring its video capabilities into mainline Gemini. He then points to early diffusion experiments as a route to much higher token generation rates. Diffusion is still a research exploration here, not a mainline capability, and he gives no numerical throughput result.
The next change concerns what a developer receives from a model. Historically, Kilpatrick thought of the interface as tokens in and tokens out, with surrounding software supplying the scaffolding needed to build a larger system. The direction he calls “agentic by default” would move more of that work into the model itself.
Reasoning may absorb some of the work now performed by external agent scaffolding. As models become more systematic, the reasoning step becomes a possible place for that behavior to live. Kilpatrick poses the amount of work that will move as an open question, with consequences for people building products around today’s models. He points again to Jack’s forthcoming discussion of scaling reasoning; he does not announce the replacement of developer orchestration.
The roadmap also includes more small models and more large models. Beyond size, Kilpatrick identifies a harder architectural frontier: infinite context. He argues that the current attention-based model paradigm cannot simply scale indefinitely, so expanding the amount of context models can use will require new innovations. Infinite context is a research ambition, not a promised context-window size. Tulsee is identified as the person leading Gemini model product work and receiving feedback about model capabilities.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Turning capabilities into developer APIs
The developer roadmap starts with embeddings. They remain important even as attention shifts toward agents because they support retrieval-augmented generation: retrieving relevant material to supply as context for generation. Kilpatrick calls the Gemini embeddings model state of the art, without specifying an evaluation, and expects a broader developer rollout in the next couple of weeks relative to the talk.
The proposed Deep Research API addresses a different layer. Rather than exposing only a model primitive, it would make research workflows—already popular in consumer products—available for developers to build into their own applications. Kilpatrick describes a bespoke API under development, not an endpoint or request format developers can use in this recording.
Veo 3 and Imagen 4 are also planned for API access. Kilpatrick hopes they will arrive very soon, tying availability to the work of scaling the developer platform. These rollout statements describe the roadmap at the time of the recording, not present-day API availability.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
AI Studio becomes a developer platform
The substantive roadmap ends with a positioning decision: AI Studio is being built as a developer platform. Kilpatrick expects it to move away from its consumer-like feel toward an experience designed more explicitly for developers. The accompanying roadmap slide places that focus alongside embeddings, media APIs, agents, a Computer Use API, evals tooling, a Deep Research API, and Gemini Code / CLI.
The intended developer experience includes agents built in. Kilpatrick hopes to integrate Jules and other coding agents natively into AI Studio, making them part of the environment where developers work with Gemini. That is the final product direction: not just a place to try a model, but a platform where model access and assistance with development come together. Native Jules integration remains an aspiration in this presentation.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
From the talk
Google's account of AI adoption, including the monthly token totals behind the 50× growth figure.
How formal proof search and geometry reasoning solved four problems from the 2024 International Mathematical Olympiad.
Further reading
An experimental approach to generating text and code by refining noise, announced in May 2025.
Updates since the talk
- Gemini 2.5 becomes stableArticle
The June 17 announcement explains stable Gemini 2.5 Pro and Flash, plus the Flash-Lite preview.
- Gemini API release historyDocumentation
Dated releases distinguish preview and stable availability for Gemini models, embeddings, audio, Imagen, and Veo.
Read the complete timestamped transcript
- 0:00
[on hold music] Awesome.
- 0:17
Thank you, Ben. Excited for the AI Education Summit. Should be fun. Um, my name's Logan. I do developer stuff at DeepMind, and I'm excited to talk about Gemini stuff.
- 0:26
Um, yeah, hopefully folks know what Gemini is, so no introduction needed. Um, I'll talk about three things really quickly. We'll do some fun announcement stuff. Um, we'll talk about sort of recapping a year of progress in Gemini, and then we'll talk about what's coming next across the model side, across the Gemini app side, and also across, uh,
- 0:45
of course, the developer platform. So the fun stuff, which is we announced a new Gemini model today. Um, so we haven't officially announced it, but we'll post live on the tweet. [cheers]
- 1:01
New Gemini model. Uh, this is hopefully the final update to 2.5 Pro. I think folks have given us tons of feedback, um, about the changes, and I think my slide has an animation which is hiding all the stuff.
- 1:14
But Gemini 2.5 Pro is awesome. Um, it's, it's super powerful. Uh, bunch of increases across, you know, benchmarks people care about. It's SOTA on Aider and, um, it's SOTA on HLE and some other benchmarks.
- 1:28
Um, I think it closes the gap on a bunch of the stuff that folks gave us feedback on from the previous versions of the model. Um, so hopefully it has great performance across the board.
- 1:38
It also, um, I, I think is like sort of setting the stage for the future of Gemini. I think 2.5 Pro for us internally and I think in the perception from the developer ecosystem was the turning point, which was super exciting.
- 1:51
Um, so it's awesome to see the momentum. We've got a bunch of other great models coming as well. Um, so 2.5 Pro, hopefully the final version. Send us feedback if things don't work, uh, and we'll, we'll continue to push the rock up the hill.
- 2:03
Um, you can go to ai.dev if you wanna try it out. It's also available in the Gemini app and all that other stuff. Um, and if you need anything, email us and we'll make it happen.
- 2:12
All right, new model launched. Let's talk about a year of Gemini progress. I think this has been the craziest thing. So I don't know if folks tuned in to, to Google I/O, but, um, Sundar showed this slide on stage, which I think was a, uh, was a great reminder for me of just how much...
- 2:27
Like it feels like ten years of, of Gemini stuff packed into the last, uh, twelve months, which has been awesome. Um, and it's, it's actually interesting to see as well, just to sort of opine on one of the points, like all of these different research bets across DeepMind coming together to like build this incredible mainline Gemini model.
- 2:44
And I think this is actually like I have a conversation with people all the time about like, what is-- what's the DeepMind strategy? What's the advantage for us building models, all that good stuff.
- 2:52
And I think the interesting thing to me is just this breadth of research happening across like science and Gemini and all these other areas like robotics and things like that.
- 3:02
Um, and all that actually ends up upstreaming into the mainline models, which is super exciting. Um, so you see like the, you know, AlphaProof and AlphaGeometry and a bunch of stuff that we did on, uh, in with custom models in those areas, actually improving the performance of our models, uh, for those domains.
- 3:19
And, uh, Jack will talk about that in a little bit, which I'm super excited about. Um, the other thing is just like not just the pace of innovation, but the pace of adoption.
- 3:28
Um, so I think, uh, Sundar also showed this slide, which was a fifty X increase in the amount of AI inference that's being processed, uh, through Google servers from one year ago to, um, last month.
- 3:41
And I think that is, [laughs] it is just remarkable to see the amount of increase in demand for, um, for Gemini models and for also from the external developer ecosystem.
- 3:52
So it's been, it's been wonderful to see that happen.
- 3:56
I think the other question, and I think this is like talked about a little bit, which is, uh, sort of what, what got us to this point. I think one of the critical pieces and like it's, you know, not super fun, uh, but is worth thinking about, uh, for folks who are building companies here, uh, i-is like
- 4:13
an organizational thing, truthfully. Like I think you bring together-- Google historically had lots of different teams doing lots of different AI research. Um, and in late 2023, early 2023, uh, Google brought a bunch of those teams together, um, and sort of charted this new direction for the DeepMind team to not only just like do theoretical foundational research,
- 4:30
but also to like build models and deliver them to the rest of Google and also the external world. Um, and then we took the second step of that journey later, uh, earlier this year, um, which was actually bringing the product teams into DeepMind.
- 4:42
So now DeepMind creates the models, does the research, um, but then also builds products and delivers, delivers those to the world. Then we have the Gemini app, which is our consumer product, and then we have the developer side of that with the Gemini API.
- 4:52
Um, and this has been like personally for me, super fun to get to collaborate with our research team and like help actually be on the frontier with them, um, and bring new models and capabilities to the world.
- 5:02
I think this is like the collaboration that works, uh, works incredibly well.
- 5:07
Yeah, and we ship lots of stuff. I think this is the, this is the most fun part, um, is there's so much stuff, so much innovation happening inside of Google.
- 5:14
It's, it's incredible to get to bring that to the world and bring that to developers, and I think we're actually very early in that journey and as we'll, we'll see in a couple of minutes.
- 5:24
Um.
- 5:26
So in summary, the formula is simple: bring the best people together, find infra advantages, and ship.
- 5:34
I don't know if folks have played around with Veo or not, but it's also been just incredible to see the reception to Veo. It's, uh, burning all the TPUs down, uh, which has been incredible to see.
- 5:45
Lots of demand, uh, lots of interest on the Veo front. Um, so hopefully folks get a chance to play around. It's available in the Gemini app right now. Um, all right, so let's talk about what next.
- 5:54
This is the fun stuff. So-
- 5:57
I think the, the sort of Gemini app piece is interesting just because people talk about it a lot, and it's, um, it's a fun product and it's cool to think about.
- 6:04
Um, and also sort of, I think, for folks building stuff, it's interesting to hear, like, what our strategy is from the app perspective. Um, but the Gemini app is trying to be this universal assistant.
- 6:13
I think what that means in practice is if you, um... I'm sure people don't think about this all the time, but I think a lot about, like, what Google's products do and, and sort of how we show up in the world.
- 6:23
And one of the interesting observations I had was that if you think about wh- what was the thing that, like, brought people, individuals through all of Google's products historically, like, the thing that comes to mind is, like, ki- like, your Google account, I guess, which, like, d- wasn't, like, super stateful.
- 6:38
You would sort of sign in to lots of different Google products with your Google account, but th- that didn't really do anything, um, other than just, like, get you signed into that individual product.
- 6:46
I think now we're seeing with Gemini that it's actually this thread that unifies all of Google, and I think the future for Google is gonna look a lot like Gemini as this sort of, you know, thread that brings all of our stuff together, um, which is really interesting.
- 6:58
And then hitting on all the trends, which I'm sure folks are also excited about building, I think the one that I'm most excited about is proactivity. I think most AI products today are still very, like, you have to go and do all the work as the user, and I think this proactive, uh, next step of, um, AI
- 7:15
systems and models coming into play is gonna be, is gonna be awesome to see.
- 7:20
Yeah, and the team is moving super fast. If you have complaints, please do not tag me on Twitter, please tag Josh. Um [laughs], he will make it happen. Josh is incredible.
- 7:28
The Gemini app team is amazing. Um, he's, he's pushing the team, uh, super hard, so it's incredible to see all the progress. Uh, but he is the person who can make stuff happen on the Gemini app side, not me, so please tag him.
- 7:40
Um, from a model perspective, like, again, there's, there's so much, uh... When Gemini was originally created, it was built to be a, a single multimodal, uh, model to do audio, image, video, et cetera.
- 7:53
We've made a lot of progress on that. At I/O this year, we announced, um, native audio capabilities in Gemini. There's TTS, there's audio. Uh, you can talk to the model.
- 8:01
It sounds, it sounds super natural, which is awesome. It's powering the Astra experience. It's powering Gemini Live. Um, so I think we're gonna get towards that omnimodal model, which is awesome.
- 8:12
We have Veo, which is SoTA across a bunch of stuff, so hopefully we'll get video into the mainline Gemini model. Um, if folks saw some of our early experiments with diffusion, which means you can get, like, crazy levels of tokens per second, um, really interesting.
- 8:26
That's, like, definitely a research exploration area, and it's not, uh, it's not mainline yet. Um, so it'll be, it'll be cool to see that come. The agentic by default thread I think is something that I've been thinking a lot about recently, which is, like, historically, for me as a developer, I've thought about models just as this thing
- 8:44
that gives me tokens in, in and out, and then there was lots of scaffolding in the ecosystem and to allow me to build with those models. I think this...
- 8:52
It's, it's becoming very clear to me that the models are becoming more systematic themselves, like, they're doing more and more, and I think the reasoning step is this, like, really interesting place in which a lot of that's going to happen.
- 9:02
And Jack's gonna talk about the scaling up of reasoning. Um, but I do think it'll be interesting to see, like, how much of the scaffolding work that's happened in the past ends up just, like, being a part of that reasoning step and, like, what that means for people who are building products and stuff like that.
- 9:17
So, um, it'll be interesting to see. We'll also have more small models soon, which I'm excited about, and big models. People want large models [laughs], which I know. Um, so I'm excited about that.
- 9:28
And then the last one is continuing to push the frontier on infinite context. I think the current model paradigm doesn't work for infinite context. I think it's just, like, impossible to scale up.
- 9:38
Attention doesn't work that way. Um, so I think there'll be some new innovations to hopefully help let people continue to scale up the amount of context that they're bringing in.
- 9:48
Um, and Tulsee is the person who drives all of our model stuff. So if you, if you have stuff, you wanna talk about Gemini models, uh, you have ideas for things that don't work well, uh, she is the person running the show on the Gemini model product side.
- 10:02
And then developer stuff. Um, so we have lots of things coming, which I'm excited about. Um, I think I'll highlight maybe three that I think people are super excited about.
- 10:13
Embeddings, I think we have, which is, uh, you know, feels like early AI stuff but I think is still super important. Embeddings power most people's, um, applications using RAG.
- 10:23
Uh, we have a Gemini embeddings model which is state-of-the-art, so excited to be rolling that out to developers more broadly in the next couple of weeks. Um, the Deep Research API I'm super interested in.
- 10:32
There's so many interesting products that are built around, um, this sort of research tasks, and people love the consumer products, so we're finding ways to bring a bunch of that together, um, into a, like, bespoke Deep Research API, uh, which will be awesome.
- 10:45
And then Veo 3 and Imagen 4 in the API as well, so hopefully we'll see that, uh, very, very, very soon, um, and as we work to scale and, and make that possible from a, from a developer platform side.
- 10:57
I'll make one other quick comment, which is the, um, AI Studio product positioning, which I also think is interesting. Like, AI Studio, just to, to be very clear, is being built as a developer platform.
- 11:09
Um, so we'll sort of move away from this, like, kind of consumery feel and move much more towards being a developer platform, which I'm personally very excited about because I think that's what developers want from us.
- 11:18
Um, so it'll be awesome to see that actually come to life with, like, many new iterations of our, of our developer experience with agents built in, and hopefully things like Jules and some of our developer coding agents, um, s- natively in that experience, which will be, which will be awesome to see.
- 11:33
Um, yeah, and that's, that's what I have. I appreciate all the people who send me [laughs] lots of great feedback about Gemini stuff. So we'll, we'll keep pushing the rock up the hill and, um, I'll be around, so if you have more feedback, come find me and we'll, we'll keep making Gemini great for everyone.
- 11:46
So thanks, and I appreciate it. [applause] [upbeat music]