Presentation by Raia Hadsell, VP of Research at Google DeepMind, on frontier AI research beyond standard large language models, covering three areas: advanced embedding models (including Gemini Embeddings 2, an omnimodal model producing unified semantic vectors for text, video, and audio), AI weather forecasting (GraphCast for 15-day forecasts, probabilistic GenCast, and FGN for cyclone prediction used by the US National Hurricane Center), and world models (the Genie series of interactive, real-time generative 3D environments). # Channel Identity Observations — Raia Hadsell Talk (Google DeepMind / AI Engineer Event) ## Source Basis - **Title**: "How Google DeepMind is researching the next Frontier of AI for Gemini — Raia Hadsell, VP of Research" - **Transcript**: Conference-style talk delivered to a live audience ("What a lovely full room") at an AI Engineer event in London. - **Curated metadata**: Not present; observations rest on title and transcript only. ## Explicit Self-Description (Creator/Channel Claims) - **Speaker affiliation**: "my name is Raya Hadsel. I've been a part of DeepMind for the last almost 13 years" and "now I am a VP within DeepMind." - **Organizational scope claim**: "I help run about a group of about 1200 scientists and engineers across 10 labs." - **Mission statement**: "The motto the mission of DeepMind is to build AI responsibly for the benefit of humanity. So I really take that seriously." - **Methodology claim**: "Our approach we look for root nodes... Find the deepest problems and solve those in order to then enable a lot of downstream impact." - **External role claim**: "I'm also very proud this year to be a UKAI ambassador. So I help the government academia industry sort of bridge those gaps." ## Demonstrated Presentation Patterns (Observable, Not Claimed) - **Non-language-model focus**: The talk explicitly frames itself around what isn't an LLM — "the theme of this talk overall is things that are not directly language models" — covering embedding models (Gemini Embeddings 2), weather prediction (GraphCast, GenCast, FGN), and world models (Genie series). - **Depth-first technical content**: Detailed explanations of Matryoshka representation learning, contrastive losses/Siamese networks, and spherical graph neural networks suggest an audience with technical ML background. - **Concrete results framing**: Specific metrics cited — "more accurate 97% of the time" against 1,300 benchmarks, "15-day forecast in 8 minutes on a single chip," Hurricane Lee landfall predicted "9 days out" vs. 6 for physics models. - **Interdisciplinary framing**: Opens with a neuroscience analogy ("Jennifer Aniston cell") and the speaker's own philosophy-then-CS background — presentation bridges fields rather than assuming uniform expertise. ## Intended Audience (Observed) - **AI engineers and practitioners**: Delivered at an AI Engineer event; audience responds with cheering/applause at "V03" mention and laughter at the Gemma 4 punchline. - **Assumed ML literacy but mixed specialties**: "I got a few neuroscientists in the room" — speaker checks for background knowledge rather than assuming it. ## Claimed Scope of the Channel/Content - **Per title**: "the next Frontier of AI for Gemini" — Gemini-adjacent research at DeepMind. - **Per speaker**: "we're thinking about what are the next architectures that we're going to use for Gemini" plus a broader ambition: "create the future of human intelligence as well and even robotics intelligence." ## Claims vs. Demonstration Distinction - **Claims**: DeepMind mission, research approach ("root nodes"), lab scale (1200 people, 10 labs). - **Demonstrated**: Actual shipped research artifacts described with evidence (Gemini Embeddings 2, GenCast, Genie 3 demos with interactive prompting), and partnerships with the UK Met Office and US National Hurricane Center ("already been used in the US by the National Hurricane Center and they are very very excited"). ## Notes & Limits - This is a **single talk by one speaker**, not direct channel self-description of the publishing channel itself; no uploader metadata was supplied. - No private motivations inferred; speaker's stated pride ("I really take that seriously") is quoted, not interpreted. # Delivery Observations — Raia Hadsell, "How Google DeepMind is researching the next Frontier of AI for Gemini" ## Tone & Register - **Conversational and warm**, opening with audience rapport: > "What a lovely full room and good smiles. I heard the dig on Google there at the end. I I I did catch that." - **Humor is light and self-deprecating**, especially in the biography section: > "Yes, I learned a lot and I'm glad that I did it and no it hasn't been very useful since." - **Enthusiastic about demos**, with exclamatory asides: > "Ah. I've just changed the world that I'm in." ## Recurring Phrases & Verbal Habits - **Repetition of words and phrases for emphasis** is the most consistent delivery feature: - > "very very hard problem" - > "extremely extremely good retrieval" - > "very very excited" - > "The weather is fundamentally chaotic" — followed by repeated "very" forms throughout - **Filler "you know"** appears repeatedly as a connective: > "You know, we're not going to waste time on the leaves" and > "and then you know, bring it to the to the leaf nodes." - **Stuttered restarts** occur throughout, e.g. > "the the ideal," > "is is is different," > "that that that we see more and more of." - The phrase **"A lot of fun"** recurs to describe work periods: > "with Jan LeCun a lot of fun" and, of the 2010s, "Atari video games go Starcraft and some robotics. A lot of fun." ## Structural & Teaching Habits - **Decade-based framing** for her own history: > "It's good to be as old as I am. You get to look at this by the decades." - **Signposting with ordinal transitions**: "So first," "Next I wanted to quickly talk about," "All right. Lastly, I wanted to use the last few minutes to talk about…" - **An explicit organizing theme**, stated twice: > "the theme of this talk overall is things that are not directly language models," repeated as > "This is not a language model at all. There was no language involved." - **Metaphor as a teaching device** — the "root nodes vs. leaves" metaphor: > "we're not going to waste time on the leaves," and the Jennifer Aniston neuron analogy to explain embeddings: > "I'll ask if anyone knows what a Jennifer Aniston cell is." - **Audience check-ins and interactive framing**: > "I got a few neuroscientists neuroscientists in the room," and rhetorical questions like > "Well, what happens if I prompt it at the same time? Then what happens?" - **Concrete numeric evidence** for claims: > "97% of the time," > "15-day forecast in 8 minutes on a single chip," > "8K tokens," > "128 seconds of video 80 seconds of audio." - **Narrative arc for research stories**, giving origin, iteration, and payoff: the weather story runs from a Met Office question ("Can you predict better rainfall…") through GraphCast → GenCast → FGN: > "this team is relentless. They're constantly coming up with new models." ## Personal Anecdotes as Anchors - **Ties technical topics to her own career**: embeddings to her PhD — > "One of the reasons why I like the space is because I did my PhD work in part looking at Siamese neural networks." - **Localizes examples geographically**: "work that we've done on the weather in London. It rains a lot," and > "I'm walking down a muddy lane in Kent. Looks not far from my house." ## Handling of Demo Issues & Time - **Recovers smoothly from a video playback problem**: > "oh, that's not playing. Well. Maybe this will play. I'm going to jump forward to Genie 1." - **Acknowledges running over time but continues**: > "now I am out of time, but I will still take another another minute or two." - **Ends with a forward reference and callback joke**: > "my uh colleague Omar is going to talk about Gemma 4, which is a language model" — met with audience laughter, tying back to her "not a language model" theme. ## Stability vs. Isolated Moments - **Stable patterns** (repeat evidence): word repetition for emphasis, "you know" fillers, signposted section transitions, numeric specificity, self-connection of topics to her biography. - **Isolated moments** (insufficient repetition): the playback glitch and recovery, and the closing joke about Gemma 4 — these are single occurrences, not habits. Audience reaction ("[cheering and applause]" at V03 mention) is a one-time observable event; her delivery response to it is minimal in the transcript. - **Uncertainty note**: snorts appear several times in the transcript markers ([snorts]) and may indicate a habitual breathing tic, but transcript annotation alone is insufficient to confirm a delivery pattern. # Position Record: "How Google DeepMind is researching the next Frontier of AI for Gemini" — Raia Hadsell ## Speaker Identification - **Speaker**: Raia Hadsell, VP of Research at Google DeepMind (introduced by an unnamed host; identified by name in the introduction and self-identification in the transcript). ## Background Claims (Self-Described) - At DeepMind for **almost 13 years**; also a **UK AI ambassador** this year, helping bridge government, academia, and industry. - Career trajectory as stated: - **1990s**: Undergraduate degree in philosophy of religion — she "really enjoyed it" and "learned a lot," but acknowledged it "hasn't been very useful since." - **2000s**: Pivoted to computer science; PhD in New York City on convolutional networks for robots with **Yann LeCun**. - **2010s**: Joined DeepMind when it was ~30–40 people; worked on Atari, Go, StarCraft, robotics. - **Present**: VP running ~1,200 scientists and engineers across 10 labs. ## Frontier AI: Approach and Principles - Focus on **"root nodes"** — the deepest problems in unsolved problem spaces — rather than "the leaves," to enable downstream impact. - Partners broadly with the world to find and solve root-node problems. - Cites DeepMind's mission: **"build AI responsibly for the benefit of humanity"** — she says she takes this seriously and wants to solve problems "worth solving." - Frames the goal as the future of **artificial, human, and robotic intelligence**, noting humans change along with the technology. ## Theme of the Talk - Explicit framing: the talk covers things that are **not directly language models** (embedding models, weather models, world models). She closes with the joke that a colleague will cover Gemma 4, "which is a language model." ## Claim Area 1: Embedding Models - **Jennifer Aniston cell** concept from neuroscience: a small number of neurons encode a specific person/thing; these activate robustly **across modalities** (name, image, video, voice) and support fast retrieval, recognition, and comparison. - **Claim**: We want the same properties — fast retrieval, recognition, comparison, and robustness to modality — in artificial neural networks, via **embedding models** trained with **contrastive losses**. - Personal note: her PhD included work on **Siamese neural networks**, an early contrastive-loss approach. - Embedding models are positioned as a **"critical companion to generative AI"**: "sometimes we want to generate sometimes we want to retrieve." - **Gemini Embeddings 2** (recently released, per Hadsell): - Fully **omnimodal**, derived from Gemini. - A single vector representing up to **8K tokens of text, 128 seconds of video, 80 seconds of audio, and a full PDF** — end-to-end without information loss when combining modalities. - Uses **Matryoshka Representation Learning (MRL)**: same network, variable embedding dimensions (e.g., start retrieval at 256 dimensions, expand for expressiveness). - Claimed result: **unified semantic space** and **state-of-the-art quality**. - Stated opinion: embeddings "don't get talked about quite as often as language models" but are "really important." ## Claim Area 2: Weather Prediction Models - **Origin**: A Met Office informatics scientist asked whether AI could predict rainfall better than physics-based models. Hadsell's response at the time was tentative: "I don't know. Interesting problem." - **Claim**: Weather prediction, though hard via physics simulation, is "quite tractable for neural network models" given 40 years of global data. - **GraphCast** (released a couple of years prior, per Hadsell): - Predicts atmospheric state up to **15 days out, globally**, for ~100 variables (wind speed, temperature, humidity) using a **spherical graph neural network** from Earth's surface to the lower stratosphere, autoregressive. - **Hurricane Lee (late 2024) example**: GraphCast's landfall prediction (Nova Scotia) was accurate **9 days out**, versus **6 days** for gold-standard physics-based models. She emphasizes "3 days is really important" for hurricane landfall. - **GenCast**: - Probabilistic, mesh-based; higher accuracy and efficiency. - Rationale: "The weather is fundamentally chaotic and we want to know what's happening on the tails" — probabilistic output enables operationalization. - **Claim**: More accurate than 1,300 gold-standard benchmark forecasts **97% of the time**; produces a 15-day forecast in **8 minutes on a single chip** instead of hours on a supercomputer. - **FGN (functional generative network)** — latest model: - Directly predicts **cyclones**, incorporating categorization, recognition, trajectory, wind speed, and eye formation into the network rather than as post-processing — "which means that it's much better." - **Claim**: Already used by the **US National Hurricane Center**, which is "very very excited" about the advantage; she expresses hope for worldwide use in coming years. ## Claim Area 3: World Models (Genie) - Origins in DeepMind's long-running games/simulation work (Atari, Go, StarCraft, MuJoCo robotics) — motivation: understanding **agency and the environment**, shifting from training agents to "creating an infinite environment." - **Genie 1**: ran only a few seconds; generated interactive 2D platformer worlds responsive to left/right input, with reasonable diversity. Interpretation offered: "we might have something here." - **Genie 2**: interactive but **not real time** ("you need to go awfully slowly"); 3D environments but not real-world quality/higher definition. - **Genie 3** (after Vo3 arrived, she notes): - **Embodiment**: prompted with walking down a muddy lane in Kent; "you look down at yourself and you realize that you actually have a body" — she notes it's "a little bit odd"; the model "really understood not just the appearance... but actually what it takes to engage with that." - Other examples: skiing; an artist's video fragment used as a prompt, which the artist was "so tickled" to see extended into a flyable, explorable world. - **Memory claim**: prompted "I'm an origami lizard in an origami world"; test described — run a minute in one direction, return, "everything is exactly as it was." - **In-world prompting**: while walking the Camden Canal, prompting mid-experience "immediately" changes the world. - **Speculative/visionary claims** (clearly framed as opinion — "I love the idea", "I think"): - A new form of gaming with **adversarially prompted worlds**. - Interactive worlds as "a whole new frontier" for **entertainment and education** — "the ability to be able to go into a world in order to learn about it... is incredibly powerful, and may well be something that we see more and more of." ## Preserved Qualifications and Tensions - Hadsell's initial uncertainty on weather AI ("I don't know") is retained alongside her later confident claims. - Genie 2's limitations (not real-time, lower fidelity) are stated explicitly, not smoothed over; Genie 3 limitations are not claimed to be absent, but real-time performance is not explicitly asserted for Genie 3 either. - Her assessment that her philosophy degree "hasn't been very useful" coexists with her statement that she is glad she did it. - Visionary claims about gaming/education are hedged ("may well be," "I think") and remain opinions, not settled claims. # Video Attribution & Provenance Record ## Video - **Title:** How Google DeepMind is researching the next Frontier of AI for Gemini — Raia Hadsell, VP of Research - **ID:** 01a15ca9-b34c-7496-b721-4c26e8888382 ## Speakers & Roles | Speaker | Role | Notes | |---|---|---| | Unnamed introducer | **Host/Introducer** | Introduces Hadsell as VP of Research at Google DeepMind. | | **Raia Hadsell** (transcribed as "Raya Hadsel") | **Creator/Keynote Speaker** — VP of Research, Google DeepMind | Self-identifies: ~13 years at DeepMind, PhD with Yann LeCun (transcribed as "Jan LeCun") on neural networks for robots, UK AI ambassador this year, oversees ~1,200 scientists/engineers across 10 labs. | ## Mentioned Third Parties - **Yann LeCun** — Hadsell's PhD advisor, named only as background. - **Unnamed Met Office informatics scientist** — originated the rainfall-prediction challenge (paraphrased quote: "Can you predict better rainfall than our physics-based models using AI?"). - **UK Met Office** — UK weather agency, source of the collaboration. - **US National Hurricane Center** — cited as user of the FGN cyclone model; Hadsell reports they "are very very excited." - **Omar (colleague, no surname given)** — announced as next morning's speaker on Gemma 4. - **An unnamed artist** — creator of the video fragment used to prompt Genie 3. ## Referenced Sources & Claims (per transcript) - **Gemini Embeddings 2** — omnimodal embedding model; up to 8K tokens, 128s video, 80s audio, full PDF per vector; uses Matryoshka Representation Learning (MRL). - **GraphCast** — spherical graph neural network, 15-day global forecasts, ~100 atmospheric variables; Hurricane Lee (late 2024) landfall prediction accurate 9 days out vs. 6 days for physics-based models (per Hadsell). - **GenCast** — probabilistic model; Hadsell claims more accurate than 1,300 gold-standard forecasts 97% of the time; 15-day forecast in 8 minutes on a single chip. - **FGN (functional generative network)** — direct cyclone prediction, in use by US National Hurricane Center. - **Genie 1, Genie 2, Genie 3** — world models; Genie 3 demos of interactive, persistent-memory environments and in-world prompting. - **Gemma 4** — mentioned as language model for next day's talk. ## Promotional Segments - **Flagged:** Product announcements/demos of Google DeepMind releases — Gemini Embeddings 2, GraphCast, GenCast, FGN, and Genie 3 — presented from the stage by a DeepMind executive; performance claims (e.g., 97% accuracy, 8-minute forecasts) are the speaker's own, not independently verified in the video. - **Flagged:** Audience applause/cheering at the Veo 3 ("V03") mention, per transcript stage directions. - **Flagged:** Promotional call-out for colleague Omar's Gemma 4 talk the following morning. ## Availability & Uncertainty - **No curated metadata was supplied** — speaker identity, roles, and provenance derived solely from the transcript. - Comment availability: **unknown/not stated**. - Transcript contains transcription noise (e.g., "Raya Hadsel," "Jan LeCun," "V03"), noted without correction beyond obvious identifications. \>\> Our next speaker is VP of research at Google DeepMind. Please join me in welcoming to the stage Raya Hadsel. \>\> \[music\] \>\> Hello everyone. Wonderful. What a lovely full room and good smiles. I heard the dig on Google there at the end. I I I did catch that. Um so my name is Raya Hadsel. I've been a part of DeepMind for the last almost 13 years and I'm very happy to have AI engineer come here to London. I'm also very proud this year to be a UKAI ambassador. So I help the government academia industry sort of bridge those those gaps and yes, I'm American by birth but I've been here for long enough that I can count myself among the proud Brits as well. So I'm going to talk a little bit about Frontier AI and the future of intelligence. To start a little bit of a introduction to who I am. It's good to be as old as I am. You get to look at this by the decades. So in the 90s I did my undergraduate degree in philosophy of religion. Definitely not not a computer scientist yet. But I really enjoyed it before you ask. Yes, I learned a lot and I'm glad that I did it and no it hasn't been very useful since. In the 2000s I did a bit of a pivot moved into computer science after some some good advice from those close to me and spent my PhD years in New York City working on convolutional networks neural networks for robots with with Jan LeCun a lot of fun. I then in the 2010s made the decision to join a small group of curious scrappy individuals working at DeepMind. It's a group of about 30 40 people at the time and we spent the rest of that decade working on things like Atari video games go Starcraft and some robotics. A lot of fun. And now I am a VP within DeepMind. I help run about a group of about 1200 scientists and engineers across 10 labs and we're working on a lot of different things. I'll tell you about three of those. So first Frontier AI is an area where we really are trying to make sure that we are staying in the front. So we're thinking about what are the next architectures that we're going to use for Gemini. What are the next problems that only AI can really address and how are we going to build the future of intelligence and that's thinking not just about artificial intelligence but it's create the future of human intelligence as well and even robotics intelligence as well. We are all on this journey together and I think that it's important to think about how how humans change as well as the technology. Our approach we look for root nodes. You know, we're not going to waste time on the leaves. We're going to really find for a big problem space that hasn't been solved. What are how deep can we go? Find the deepest problems and solve those in order to then enable a lot of downstream stream impact. We partner you know really with the world. I really think about it very broadly and think about who are the partners that can help us find those root nodes and solve those problems and also you know, bring it to the to the leaf nodes. And solving problems that are worth solving. The motto the mission of DeepMind is to build AI responsibly for the benefit of humanity. So I really take that seriously. We want to build build solve problems that are worth solving. All right. So we work in a lot of different areas within Frontier AI in DeepMind. These are sort of some of the different categories. I'm not going to tell you about all of them so you can just maybe keep those a mystery. But I'll just pick out a couple. So first in advanced models I actually wanted to bring up an embeddings model. So the theme of this talk overall is things that are not directly language models. And in the modeling space I wanted to talk about embedding models. To start that out I'll ask if anyone knows what a Jennifer Aniston cell is. I got a few neuroscientists neuroscientists in the room. So this is actually a concept from neuroscience where we've discovered that there are not just a single cell but a small number of neurons that will encode for a specific thing as in a specific person. And that those combinations of neurons that only activate for that one person or that one thing or that one place those cells are actually very robust. They activate regardless of modality. And this is used by the brain very very fast retrieval for recognition and for comparison functions. So that means that when I say the name Jennifer Aniston or if I showed you a picture or a video or if you even heard her voice if you knew her if you were enough of a fan then those all those different modalities lead to the same set of cells activating. So we want that in a artificial neural network for the same reason. We want fast retrieval recognition and comparison. And so we can trade what's called an embedding model in order to encode for those concepts in order to be more robust to different different ways the information can be presented and to be very very good at sort of understanding what is the comparison between these different activations. Use contrastive losses. One of the reasons why I like the space is because I did my PhD work in part looking at Siamese neural networks which was an early way of understanding what is a contrastive loss function. Um and so these these embedding functions are really critical companion to generative AI. Sometimes we want to generate sometimes we want to retrieve. So the group at Google has been working on this for a long time and just recently we've actually released Gemini embeddings 2. So this is exciting to me because it really is sort of the the ideal. It is fully omnimodal. It uses it's derived from Gemini. So it's got sort of that level of knowledge and understanding of the world and it is and it allows extremely extremely good retrieval. Um in a little bit more more detail than why is it good that it is unified and multimodal? It means that you don't have to have different steps to try to bring you can be truly end-to-end and not lose information by trying to combine audio information visual information text information together. So you can get a single vector that represents text up to 8000 8K tokens um \[snorts\] 128 seconds of video 80 seconds of audio and a full PDF and together that can give you a lot of information. You can then use that to be able to use it for retrieval for \[snorts\] for for querying for agentic logic and other things. We also use something called the Matryoshka representation learning MRL and that allows us to have be able to have the same network but represent different dimensions. So for instance you could start out doing a retrieval using only 256 dimensions for your embedding and then you can expand that to get to more expressiveness. Um So this also this gives us we can demonstrate that we have this allows us to have a unified semantic space and really state-of-the-art quality. So just something that's come out recently that I think should doesn't get talked about quite as often as language models but it's really important as that companion I think. All right. \[snorts\] Next I wanted to quickly talk about another thing that is not a language model. This is not a language model at all. There was no language involved. And this is work that we've done on the weather in London. It rains a lot and a few years ago there was a a informatics scientist at the Met Office, Meteorological Office for the UK. The UK's weather agency that said, "Can you predict better rainfall than our physics-based models using AI?" And I said, "I don't know. Interesting problem. Let me take this back to the team." Took this back to the team at DeepMind. We started working on this and we discovered, yes, you know what? Predicting the weather, even though it is a very very hard problem using physics simulation of the atmosphere, is actually quite tractable for neural network models given that we have 40 years of data, 40 years of global data on what the weather is. So a couple of years ago we came out with GraphCast. GraphCast predicts the predicts the state of the atmosphere up to 15 days out everywhere on Earth and for many different variables. And this uses a spherical graph neural network. Think about this encompassing the Earth and having nodes that go all the way from the surface of the Earth all the way up into the lower stratosphere. And we actually feed in and then predict in an autoregressively 100 different atmospheric variables. For instance, wind speed, temperature and humidity as shown here. Um And this worked very well. Here's a quick example. We were excited to see this in late 2024. This is Hurricane Lee. Sort of comes into the Atlantic, pauses for a moment and then takes a takes a turn to the north, speeds up and makes landfall in Nova Scotia. The total this total video is 9 days worth. So that's how far the the hurricane moves. And this is actually the output of the graph neural network. That's its prediction. And the prediction that it made is accurate 9 days out of where that landfall is going to be. In comparison, the best gold standard models that are that are physics-based were only accurate 6 days out as to where that landfall is going to be. When you're talking about a major hurricane hitting land, 3 days is really important. So with this we said, "Okay, this is important and we're going to keep on pushing the science." So the team developed the next model. We called this GenCast. And the difference here is that this model, while also based on a mesh, is probabilistic and it has a higher accuracy and a higher efficiency. The weather is fundamentally chaotic and we want to know what's happening on the tails. And so having a model that's probabilistic allows us to do that, allows this to be operationalized and used for actual weather prediction. Um GenCast also was more accurate. So when we compared it to 1,300 gold standard benchmarked weather forecast, then this was more accurate 97% of the time. And it was also could be we could produce that 15-day forecast in 8 minutes on a single chip instead of hours on a very large supercomputer. So much different sort of space of the solution that we were proposing. And just this last year, this team is relentless. They're constantly coming up with new models. And so the latest one is called FGN, functional generative network. This directly predicts cyclones rather than predicting the weather and then having to add on a cyclone detector as sort of post-processing. This actually incorporates the the categorization, the recognition of cyclones, their trajectory, their wind speed and the formation of the eye directly into the network. We train for that, which means that it's much better. So this has already been used in the US by the National Hurricane Center and they are very very excited by how much of an advantage this now gives. So this this will hopefully be used worldwide in the coming years. All right. Lastly, I wanted to use the last few minutes to talk about again something that is not language model-based. So this is world models. And this actually came out of work that DeepMind has done on games and simulation for a long time. We've been working on Atari and Go on StarCraft and then on you know Mujoco type environments for robotics because we wanted to understand agency and the environment. We started focusing more and more on not just the training the agent, but creating the an infinite environment. You know, when we when I did work on locomotion here um then oh, that's not playing. Well. Maybe this will play. I'm going to jump forward to Genie 1. So we wanted So this is Genie 1. It could only run for a few seconds, but you could say, "Hey, I want this type of a world." And then you could produce this little platformer 2D game environment where you could jump around for a few minutes and it would actually respond to whether you hit the the left or the right. And it could produce a reasonable diversity of different-looking platformer type of worlds. This was enough to say, "Hmm, we might have something here. Let's scale up. Let's scale up the data and train again. Improve the method and train again now on 3D games. Then we produced Genie 2. Genie 2 is is interactive, but it's not yet real time. So you need to go awfully slowly. And it can produce 3D environments, but it couldn't do anything that was really real-world type of quality and more higher definition. So we were working on that. And then along came V03. All right. Now \>\> \[cheering and applause\] \>\> All right. Well, now I am out of time, but I will still take another another minute or two to just show show you these. So this is saying that telling Genie, I want a world where I'm walking down a muddy lane in Kent. Looks not far from my house. The fun thing here is that you look down at yourself and you realize that you actually have a body. You're actually interacting with the world. It's a little bit odd to to know what's coming out of this model. It's really understood not just the appearance of a lane in Kent, but actually what it takes to engage with that, make the water move and to walk forward. Of course, it's not just scenes that are walking. We can very happily ski. And so you can create an environment where you can engage with the world in so many different different ways. Um Here's an example where it says original there. That's we started this. We prompted this with a fragment of video and now it's changed to Genie Genie 3. So this is an artist. He made those that those first few seconds. And then we used that to prompt Genie and bring this world to life. He was so tickled to see that we could take a little snippet of his world that he had laboriously created and bring that to life in a way that means that you can fly through it. You can bounce off of this thing and it remembers that oh, here's that here's that weird structure there and go back to that. And fly through there. So \[snorts\] these environments are not only diverse and interactive high quality. They also have memory. So the prompt here was, "I'm an origami lizard in an origami world." And this is what you get. And we use this as a nice little test that I can spend, you know, I can spend a minute running in one direction, run back to the start and everything is exactly as it was at the beginning because we have a really good memory. Working in these environments gives us consistency and control. And lastly, we have we're able to prompt this world as you're in it. So that means that while I'm in a world that might be a little bit boring. Here I am, you know, this is a world saying I'm walking down the Camden Canal in London here near the DeepMind office. Well, what happens if I prompt it at the same time? Then what happens? Ah. I've just changed the world that I'm in. Can change it again. There we go. Immediately, the world is is is different. And one more time just for fun. I love the idea of a new form of gaming where I could be adversarially prompting your experience of a world. It just creates a whole different sort of entertainment, a whole world, a whole new frontier um I think can be really amazing, not just for entertainment, but for education, as well. Um the ability to be able to go into a world in order to learn about it, I think is incredibly powerful, um and may well be something that that that we see more and more of. Um and uh with that, I will uh say thank you and just a quick call out that tomorrow morning, um my uh colleague Omar is going to talk about Gemma 4, which is a language model. \>\> \[laughter\] \>\> Uh thank you. \>\> \[music\]