How Google DeepMind is researching the next Frontier of AI for Gemini — Raia Hadsell, VP of Research

What comes next for intelligence when the research agenda extends beyond language models? Raia Hadsell approaches that question through a career that began outside computer science. At the time of the talk, she describes almost 13 years at DeepMind and a role as a UK AI ambassador, helping connect government, academia and industry.

Her undergraduate work in the 1990s was in philosophy of religion. In the 2000s, she moved into computer science and completed her PhD with Yann LeCun, studying convolutional networks and neural networks for robots. In the following decade, she joined a DeepMind of roughly 30–40 people, working on Atari, Go, StarCraft and robotics. She now helps lead approximately 1,200 scientists and engineers across 10 labs. That trajectory connects the three research areas that follow: learning representations, predicting physical systems and building environments in which agents can act.

The immediate questions include which architectures should come next for Gemini, which problems AI can uniquely address, and how advances will change human and robotic intelligence as well as artificial intelligence. The research target is broader than a better model in isolation: people and the systems around them change as the technology changes.

Hadsell describes the approach as looking for root nodes. A foundational unsolved problem can support many downstream applications once it is resolved. Partners help both identify those problems and carry the resulting capabilities into practical uses. The selection criterion is also a responsibility: building AI for humanity’s benefit means choosing problems worth solving. Embeddings, weather forecasting and world models are three examples of this agenda that are not directly about language models.

Suppose you hear the name Jennifer Aniston, see her picture, watch a video of her or recognize her voice. The sensory inputs differ, but the person you recognize is the same. Hadsell introduces the neuroscience analogy of a Jennifer Aniston cell: not necessarily one neuron, but a small set of neurons responding to a particular person, thing or place. In her account, that representation is robust across modalities and supports rapid recognition, retrieval and comparison.

An embedding model aims for a related property in an artificial neural network. It encodes inputs into representations whose relationships remain useful when the presentation changes. The important output is not a newly generated description; it is a representation that lets a system compare inputs and retrieve relevant material.

Contrastive losses train those relationships, encouraging related examples to have compatible representations and distinguishing them from unrelated examples. Hadsell connects this to her PhD work on Siamese neural networks, an early approach to learning comparisons between inputs. The resulting capability complements generative AI: sometimes the next operation should generate something new, and sometimes it should retrieve something that already exists.

Gemini Embedding 2 brings that idea to a model derived from Gemini and described by Hadsell as fully omnimodal. A unified semantic space removes the need to build separate representations for audio, vision and text and then map them together afterward. The intended benefit is an end-to-end representation that preserves information otherwise lost at those joining steps.

Hadsell describes a single vector representing up to 8K text tokens, 128 seconds of video, 80 seconds of audio or a full PDF, for retrieval, querying and agentic logic. For implementation, the March 2026 launch announcement specifies 8,192 text tokens, 120 seconds of video and PDFs of up to six pages; it does not establish the spoken 80-second audio limit. Those published limits should govern input preparation rather than the broader spoken account.

Matryoshka Representation Learning, or MRL, makes the representation’s size adjustable. The same network learns nested representations, so a shorter prefix can support an inexpensive retrieval pass and additional dimensions can support a more expressive comparison. Hadsell’s example starts with 256 dimensions and then expands. This means using more coordinates from the learned representation—not recovering coordinates that were discarded when only a short vector was stored.

A retrieval implementation can therefore keep full vectors available while using their first 256 coordinates to shortlist candidates. This Python example takes already-computed MRL embeddings, normalizes the chosen coordinates for cosine similarity, and reranks the shortlist using the full representation:

from math import sqrt


def cosine(a: list[float], b: list[float]) -> float:
    if len(a) != len(b):
        raise ValueError("Vector dimensions must match")
    norm_a = sqrt(sum(x * x for x in a))
    norm_b = sqrt(sum(x * x for x in b))
    if norm_a == 0 or norm_b == 0:
        raise ValueError("Cannot compare a zero vector")
    return sum(x * y for x, y in zip(a, b)) / (norm_a * norm_b)


def retrieve(
    query: list[float],
    documents: dict[str, list[float]],
    shortlist_size: int = 20,
    limit: int = 5,
) -> list[str]:
    if len(query) < 256:
        raise ValueError("Expected at least 256 dimensions")
    if any(len(vector) != len(query) for vector in documents.values()):
        raise ValueError("All embeddings must have matching dimensions")
    if not 0 < limit <= shortlist_size:
        raise ValueError("Require 0 < limit <= shortlist_size")

    shortlist = sorted(
        documents,
        key=lambda key: cosine(query[:256], documents[key][:256]),
        reverse=True,
    )[:shortlist_size]
    return sorted(
        shortlist,
        key=lambda key: cosine(query, documents[key]),
        reverse=True,
    )[:limit]

Only shortlisted documents receive the full-dimensional comparison. The richer pass cannot rescue a candidate excluded by the first pass, so shortlist size remains a retrieval-quality choice.

The broader aim is a shared semantic space that is both useful and economical to search. Hadsell describes the model’s quality as state of the art, without presenting a numerical embedding benchmark in this talk. Its role in the research agenda is clear: strong retrieval is a companion capability to strong generation.

The weather work began with a concrete question from a scientist at the UK Met Office: could AI predict rainfall better than physics-based models? Hadsell took the question back to DeepMind. Atmospheric simulation is difficult, but the team had a substantial learning substrate: four decades of global weather data. That made forecasting a tractable neural-network problem even though the underlying physical system remained complex.

GraphCast predicts atmospheric conditions around the globe. Hadsell describes a horizon of up to 15 days; the original GraphCast release documents a 10-day horizon. Its architecture uses a spherical graph neural network, which she illustrates as a structure surrounding Earth and representing the atmosphere from the surface into the lower stratosphere.

The model takes atmospheric state as input and predicts the next state autoregressively: predictions become inputs for subsequent forecast steps. Hadsell describes roughly 100 atmospheric variables, with wind speed, temperature and humidity as examples. The globe and three maps on the slide make the scope tangible: this is a prediction of a changing global field, not a single temperature or a text answer about tomorrow’s weather.

The Hurricane Lee demonstration follows a storm through the Atlantic: it pauses, turns north, accelerates and reaches Nova Scotia. Hadsell introduces it with a late-2024 reference, but the documented event was in September 2023. The animation shows nine days of the graph model’s predicted movement.

For Hurricane Lee, Hadsell reports accurate landfall guidance nine days ahead, compared with six days for the best physics-based models. The value of that particular result is operational: three additional days before a major hurricane’s landfall can materially change preparation time.

The next model, GenCast, retains a mesh-based approach but makes the forecast probabilistic. Weather is chaotic, so the goal is not only to identify a likely outcome. A useful forecast must also describe the tails of the distribution—the less likely outcomes that may carry severe consequences. Producing possible futures makes that uncertainty available for weather prediction and decisions about risk.

Hadsell summarizes GenCast’s accuracy as winning 97% of approximately 1,300 comparisons. The published evaluation is more precise: GenCast outperformed ECMWF ENS on 97.2% of 1,320 variable-and-lead-time targets, using 2019 test data after training through 2018. This measures performance across evaluation targets, not the percentage of individual forecasts that were correct.

The reported eight-minute runtime on one Cloud TPU v5 is for one 15-day ensemble member, not the full ensemble. Hadsell contrasts this with hours on a large supercomputer for conventional forecasting. The computational difference matters alongside accuracy: a probabilistic forecast requires multiple members to represent alternative futures, so the cost of producing each member affects what can be run operationally.

The next change concerns what the network is trained to predict. Hadsell introduces FGN, which she expands as Functional Generative Network. Rather than predicting the weather and then attaching a cyclone detector as post-processing, this model incorporates cyclone recognition, categorization, trajectory, wind speed and eye formation directly into training.

The distinction is where the application’s target enters the system. In the second approach, identifying and characterizing the cyclone is part of what the network learns to do, rather than only an interpretation of its output afterward.

Hadsell reports that the US National Hurricane Center has already used the model and is enthusiastic about the advantage it offers. Her account does not specify whether that use is experimental or part of routine operations. Worldwide adoption remains a hope for the coming years.

World models emerge from a different strand of DeepMind’s history: Atari, Go, StarCraft and MuJoCo-type environments for robotics. Those settings let researchers study an agent’s actions and their consequences. Over time, the question widened from how to train the agent to how to create an effectively unbounded supply of environments for it.

Genie 1 provided an early version of that capability. A requested world became a short 2D platformer environment that responded to left and right controls and allowed the player to jump around. Its worlds lasted only a few seconds, but their variety and responsiveness justified scaling the data, improving the method and training again on 3D games.

Genie 2 extended the idea into 3D. Its environments were interactive, but interaction was too slow to be real time, and their appearance did not yet reach realistic, high-definition quality.

The progression separates several requirements that can otherwise collapse into the word “generation”: diversity, responsiveness, dimensionality and visual quality each have to improve.

Hadsell then introduces Veo 3 as a turning point before moving to the newer world demonstrations. She does not give a Veo-to-Genie integration recipe; the emphasis shifts to what the generated environments now let a person do.

The first request is a world in which the user walks down a muddy lane in Kent. It resembles somewhere near Hadsell’s home. Looking down reveals a body, and moving forward affects the water. Hadsell interprets this as more than reproducing the appearance of a lane: the model also produces aspects of engaging with that environment.

The next example switches from walking to skiing. The first-person view includes gloved hands and ski tips beneath a snowy slope and distant peaks. The change illustrates the breadth of interaction the system is being asked to support: a generated world needs to respond to how a participant moves through it, not merely offer different scenery.

The next demonstration begins with an artist’s short video. Its on-screen label changes from Original to Genie as Genie 3 continues the scene into an environment the user can explore. The artist supplied only the first few seconds; the generated continuation allows flight through the world.

Interaction includes bouncing off a structure, moving away and returning to it. In the selected frame, an orange aircraft approaches a large blue ring over a green landscape. The significant behavior is the ability to revisit the structure: an explorable environment needs previously encountered places to remain available as the viewpoint changes.

The memory test becomes more explicit with the prompt, “I'm an origami lizard in an origami world.” Hadsell describes running in one direction for a minute, returning to the starting point and finding the environment as it was before. This example tests spatial consistency alongside visual quality and control. The Genie 3 announcement describes consistency over a few minutes, so the demonstration supports a bounded memory capability rather than indefinite persistence.

The final demonstration adds another kind of control: prompting the world during interaction. It begins with a walk in London, near the DeepMind office. Hadsell then supplies additional prompts and changes the environment repeatedly while remaining inside the experience. The world is no longer determined solely by the request that started it.

That suggests a different form of multiplayer interaction: one person could adversarially prompt another person’s world, altering the experience as they play. Hadsell presents this as a possible new gaming format. The educational possibility follows from the same capability—entering a world in order to learn about it, with the environment itself becoming part of the learning experience. Both are prospective applications of the demonstrated control.

Hadsell closes by pointing to a talk the following morning from her colleague Omar about Gemma 4, returning explicitly to language models after an agenda built around retrieving, forecasting and interacting.