Tag: DeepMind

Tech News : Google DeepMind Opens Project Genie for Real-Time AI World Creation

Posted on by Paul Stradling

Google DeepMind has opened access to Project Genie, an experimental world-building AI tool, as it looks to gather real-world feedback and accelerate progress on the world models it believes are central to the path towards artificial general intelligence.

What Project Genie Is and How It Was Built?

Project Genie is a web-based experimental research prototype developed by Google DeepMind that allows users to generate and explore interactive virtual worlds using text prompts or images. Technically, it is not a standalone model but a front-end experience built on top of several of DeepMind’s most advanced systems.

At its core is Genie 3, DeepMind’s latest general-purpose world model, which generates environments frame by frame in real time as users move through them. This is combined with Nano Banana Pro, an image generation model used to sketch and refine the initial appearance of a world, and Gemini, which handles higher-level reasoning and prompt interpretation. Together, these components allow Project Genie to turn a static description or image into a navigable environment that responds dynamically to user actions.

How Do You Use It?

Practically, users begin by creating what DeepMind calls a “world sketch”. This involves prompting the system with a description of an environment and a character, choosing a first- or third-person perspective, and optionally refining the generated image before entering the world. Once inside, the environment expands in real time as the user moves, with the model simulating basic physics, lighting, and object behaviour. Users can also remix existing worlds, explore curated examples, or download videos of their explorations.

Project Genie was built by DeepMind researchers including Jack Parker-Holder and Shlomi Fruchter, both of whom have been closely involved in the development of Genie 3 and earlier world model research.

DeepMind

Google DeepMind is the name for Google’s dedicated AI research lab, formed through the merger of DeepMind and Google Brain, and is focused on developing general-purpose AI systems. Its long-term stated ambition is to build AI that can reason, plan, and act across the full complexity of the real world, rather than being limited to narrow tasks.

Genie 3 Previewed Back In August 2025

DeepMind first previewed Genie 3 as a research model back in August 2025, positioning it as a major step forward in interactive world simulation. Five months later, the decision to open Project Genie to a wider audience appears to reflect a deliberate transition from closed research testing to broader, real-world experimentation.

In its own recent announcement, Google stated that “the next step is to broaden access through a dedicated, interactive prototype focused on immersive world creation.” Access is currently limited to Google AI Ultra subscribers in the United States aged 18 and over, reinforcing that this is still a controlled research rollout rather than a mass-market launch.

Why Now?

It should be noted here that the timing matters. For example, world models are moving from abstract research concepts into systems that can be directly experienced and evaluated by users. Therefore, by opening access now, DeepMind is hoping to be able to collect feedback, usage patterns, and behavioural data that are difficult to obtain through internal testing alone, while also demonstrating tangible progress in a competitive and fast-moving field.

What Genie Can Do and Who It’s Aimed At

Genie 3 enables real-time interaction at around 24 frames per second, with worlds that remain visually consistent for several minutes. Unlike traditional video generation models that produce a fixed sequence, Genie 3 generates each new frame based on what has already happened and how the user moves, allowing for exploration rather than playback.

Project Genie is actually aimed at several overlapping audiences. For example, in the near term, it is most accessible to creators, researchers, and technically curious users who want to experiment with AI-generated environments. The tool supports whimsical and stylised worlds particularly well, including animated, illustrative, or fantastical settings.

Beyond creative exploration, DeepMind also appears to see some real value in Genie 3 for education, simulation, and research. World models can be used to train and test embodied agents (AI systems designed to act within an environment), including robots or software agents that move and make decisions. Instead of learning in the real world, where training can be expensive, slow, or risky, these agents can practise inside simulated environments. For example, an AI-controlled robot can learn how to navigate difficult terrain or react to unexpected situations without any physical risk or real-world consequences.

DeepMind described world models as systems that “simulate the dynamics of an environment, predicting how they evolve and how actions affect them,” framing Genie 3 as part of a broader capability rather than a single product feature.

How Project Genie Fits Into DeepMind’s AGI Strategy

World models now appear to occupy a central position in DeepMind’s vision for AGI (artificial general intelligence), which are AI systems that can understand, learn, and reason across a wide range of tasks rather than being limited to a single narrow function. The lab has argued that this kind of intelligence requires an internal model of the world that supports planning, prediction, and counterfactual reasoning. In practical terms, this means being able to ask “what happens if” and simulate possible outcomes before acting.

Genie 3 builds on earlier models such as Genie 1 and Genie 2, but adds real-time interaction and longer-horizon consistency. This allows agents to execute longer sequences of actions and pursue more complex goals, which DeepMind sees as essential for general-purpose intelligence.

The company has already demonstrated Genie 3 being used to generate environments for SIMA, its generalist agent for 3D virtual settings. This reinforces that Project Genie is not the end goal, but a way to expose and test the underlying capabilities that future agents will rely on.

The Competitive Landscape and Why Timing Matters

The release of Project Genie comes as competition around world models is intensifying, with several AI labs and startups racing to build systems that go beyond static generation and towards interactive simulation.

For example, Runway has recently introduced its own world model concepts alongside its video tools. Also, World Labs, founded by Fei-Fei Li, has launched Marble as a commercial product aimed at interactive environments. Yann LeCun’s AMI Labs has also signalled a strong focus on world modelling as a foundation for intelligence.

By opening access now, DeepMind is hoping to position itself as a leader not just in theory, but in demonstrable, hands-on systems. This visibility matters for attracting talent, shaping industry standards, and influencing how developers and researchers think about the future of AI simulation.

Limitations, Guardrails, and Why This Is Still a Prototype

Despite its capabilities, Project Genie is explicitly framed as an experimental research prototype. Usage sessions are currently limited to 60 seconds of world generation and navigation, reflecting the heavy computational cost of auto-regressive real-time models.

With this in mind, Google has acknowledged several known limitations. For example, generated worlds may not closely match prompts or real-world physics, characters can be difficult to control, and latency can affect interaction. Some Genie 3 capabilities announced in August, such as promptable world events that change environments mid-exploration, are not yet available in Project Genie.

DeepMind has also been quick to emphasise responsible development. For example, safety guardrails restrict copyrighted content, realistic depictions of certain subjects, and other sensitive material. The company stated that “as with all our work towards general AI systems, our mission is to build AI responsibly to benefit humanity.”

These constraints help explain why Project Genie is not being positioned as a consumer product or game platform, but is currently a testbed designed to surface technical weaknesses and user expectations before wider deployment.

Entertainment Today, Embodied Agents Tomorrow

In the short term, Project Genie’s most obvious use is entertainment and creative experimentation. Its strengths in stylised, animated, and imaginative environments make it well suited to playful exploration and concept development.

However, longer term, DeepMind’s ambitions extend far beyond games. World models offer a scalable way to train embodied agents, including robots and autonomous systems, in simulated environments that mirror the complexity of the real world. This could reduce costs, improve safety, and enable faster iteration across industries such as logistics, manufacturing, and healthcare.

The same technology could also support training, education, and scenario planning, where exploring “what if” situations is valuable.

Business and Industry Implications

For Google, Project Genie is intended to reinforce its position at the frontier of advanced AI research and supports the premium value proposition of its AI Ultra subscription. It also strengthens Google’s influence over how world models are commercialised and evaluated.

For competitors, the move appears to raise the bar for what qualifies as a leading-edge AI system, increasing pressure to demonstrate interactive, real-time capabilities rather than static outputs.

For businesses and developers, Project Genie offers an early glimpse into tools that could reshape simulation, training, design, and creative workflows. At the same time, its limitations highlight that world models are still an emerging technology with unresolved challenges around realism, control, and cost.

For the wider AI market, the release highlights a broader transition from generative content towards generative environments, where interaction and agency matter as much as visual fidelity.

Challenges, and Criticisms

It should be noted that some key challenges remain for world models like Genie 3, particularly around scalability, realism, and controllability. For example, auto-regressive world generation is computationally expensive, which makes long-duration or large-scale simulations difficult to run. Critics have also questioned how quickly these systems can achieve reliable real-world accuracy, especially for safety-critical applications where errors or inconsistencies could have serious consequences.

There are also broader concerns around data use, intellectual property, and the environmental cost of large-scale compute. DeepMind’s cautious, limited rollout reflects an awareness of these issues, even as it pushes the technology forward.

Project Genie, as DeepMind presents it, is not yet a finished destination but a visible step in a much longer journey towards AI systems that can understand and navigate the world in ways that begin to resemble human reasoning.

What Does This Mean For Your Business?

Project Genie shows how world model research is now being tested outside the lab, with DeepMind deliberately exposing early capabilities to real users in order to gather feedback that research alone cannot provide. The limited access, short session lengths, and strict guardrails make it clear that this is about learning and validation rather than product launch.

For UK businesses, the immediate value is not in using Project Genie directly, but in what it signals. Interactive simulation has long-term relevance for training, design testing, robotics, and scenario planning, particularly in sectors where real-world experimentation is expensive or risky. As these models improve, they could become a practical tool for reducing uncertainty before decisions are made in physical environments.

For the wider AI market, the release raises expectations around what advanced AI systems should be able to do. The focus appears to be shifting from static content generation to interaction, consistency, and decision-making over time. Project Genie does not solve those challenges yet, but it does show more clearly how DeepMind is approaching them and increases pressure on competitors pursuing similar world model capabilities.

Tech News : GenCast : Ultra-Advanced Weather Forecasting

Posted on by Paul Stradling

Google DeepMind has introduced GenCast, an advanced AI model designed to revolutionise weather forecasting by delivering faster, more accurate predictions of weather uncertainties and risks up to 15 days ahead.

What is GenCast and How Does it Work?

GenCast is a diffusion-based generative AI model (one that transforms ‘noisy’ data into realistic outputs), a sophisticated approach typically used in creating high-quality images, videos, and music. In the realm of weather forecasting, it works by leveraging decades of historical meteorological data to simulate complex atmospheric dynamics. Trained on nearly 40 years of data from the European Centre for Medium-Range Weather Forecasts (ECMWF), GenCast can generate an ensemble of predictions, providing a probabilistic range of possible weather outcomes rather than a single deterministic forecast.

Google’s DeepMind explains why this ‘ensemble forecasting’ is essential, saying: “Because a perfect weather forecast is not possible, scientists and weather agencies use probabilistic ensemble forecasts, where the model predicts a range of likely weather scenarios. Such ensemble forecasts are more useful than relying on a single forecast, as they provide decision-makers with a fuller picture of possible weather conditions in the coming days and weeks and how likely each scenario is.”

Identifies Past Patterns to Predict Future Weather

Unlike traditional numerical weather prediction models, which rely heavily on computationally intensive physics-based equations, GenCast’s data-driven methodology identifies patterns in past weather events to forecast future scenarios. This unique approach allows it to outperform conventional systems, particularly in predicting extreme weather conditions like cyclones and storms, which are notoriously challenging for standard models.

Key Features and Capabilities

One of the standout features of GenCast is its ability to produce high-quality forecasts with remarkable speed. For example, as Google DeepMind points out: “It takes a single Google Cloud TPU v5 just 8 minutes to produce one 15-day forecast” (a Google Cloud TPU v5 is a specialised chip for accelerating AI computations). In contrast, traditional models often need many hours of processing time on supercomputers equipped with thousands of processors.

‘Ensemble’ Forecasting Method Means Better Forecasting

GenCast’s ensemble forecasting method, i.e. producing multiple plausible weather scenarios, may enable meteorologists and decision-makers to assess risks and uncertainties more comprehensively. This capability could prove to be crucial for industries and communities that rely on understanding the full spectrum of potential weather outcomes, particularly in today’s context of climate change and increasingly volatile weather patterns. As DeepMind says: “Better forecasts of extreme weather, such as heat waves or strong winds, enable timely and cost-effective preventative actions. GenCast offers greater value than ENS when making decisions about preparations for extreme weather, across a wide range of decision-making scenarios.”

Who Could Benefit from GenCast?

The versatility of GenCast makes it a real game-changer for a wide array of sectors, such as:

– Disaster management. Early and accurate predictions of extreme weather events allow governments and humanitarian organisations to plan evacuations, allocate resources, and mitigate damages more effectively.

– The energy sector. Renewable energy providers, particularly those in wind and solar power, can optimise energy generation and grid management based on precise weather forecasts.

– Agriculture and fisheries. Farmers and fishermen can better plan planting, harvesting, and fishing schedules, reducing losses due to unforeseen weather disruptions.

– Transportation and logistics. Airlines, shipping companies, and logistics providers can enhance operational efficiency and safety by anticipating weather conditions that may impact travel and delivery routes.

Impact on Futures Markets

GenCast may also hold promise for futures markets, e.g. in agriculture and commodities trading. For example, weather fluctuations heavily influence the supply and pricing of essential goods such as grains, pork bellies, and seafood. By providing early and accurate predictions, GenCast may enable traders to make more informed decisions, thereby stabilising markets and reducing volatility. Knowledge of an impending drought could, for example, prompt strategic planning, such as stockpiling or diversifying supply chains, to minimise financial losses.

A New Benchmark AI and Weather Forecasting?

While AI-driven weather prediction is not entirely new, GenCast’s performance appears to set a new benchmark. As DeepMind says: “GenCast showed better forecasting skill than ECMWF’s ENS, the top operational ensemble forecasting system that many national and local decisions depend upon every day. GenCast was more accurate than ENS on 97.2 per cent of these targets, and on 99.8 per cent at lead times greater than 36 hours.”

IBM’s Watson has previously ventured into this space with its weather-focused AI, but GenCast’s ability to forecast medium-range weather events and extreme conditions with superior accuracy looks like positioning it as today’s frontrunner.

Challenges and Criticisms

Despite its groundbreaking capabilities, GenCast is not without its challenges. For example:

Issues with resolution and local accuracy. GenCast operates at a lower resolution compared to some traditional numerical models, potentially limiting its precision for localised weather forecasts.

Integration with existing systems. Adoption of GenCast requires validation and acceptance by meteorological agencies, which must assess its reliability before integrating it into their systems.

Possible data limitations. GenCast depends on historical data, and its effectiveness may be constrained in regions with sparse datasets or when predicting unprecedented weather patterns driven by climate change.

Collaboration Between AI and Traditional Still Important

Although GenCast appears to bring a unique new and powerful method to weather forecasting, DeepMind is keen to point out that it’s not going to be a case of AI replacing all traditional methods, rather GenCast will be used as part of a collaboration between AI and traditional meteorology. DeepMind says: “We deeply value our partnerships with weather agencies, and will continue working with them to develop AI-based methods that enhance their forecasting. Meanwhile, traditional models remain essential for this work. This cooperation between AI and traditional meteorology highlights the power of a combined approach to improve forecasts and better serve society.”

What Could It Mean for the AI Sector?

GenCast’s promise highlights the transformative potential of AI in addressing complex global challenges. Its application in weather forecasting is another demonstration of the adaptability of generative AI, traditionally associated with creative industries, to scientific and practical domains. To advance its mission and hopes for the wide adoption of and collaborations with GenCast in the weather and climate community, DeepMind says it has “Made GenCast an open model and released its code and weights.”

At the same time, the introduction of GenCast raises the stakes for competitors. Companies aiming to replicate or surpass DeepMind’s achievements will need to tackle significant technical and computational hurdles. This competition could lead to even greater advancements in AI technology and its applications.

What Does This Mean for Your Business?

GenCast looks like being a remarkable leap forward in the integration of AI into weather forecasting. Its ability to provide accurate, probabilistic forecasts with unprecedented speed and efficiency appears to have set a new standard for the industry. By leveraging decades of historical data, and using the ‘ensemble’ forecasting method, GenCast can deliver insights that are not only scientifically impressive but may also be critically important in addressing real-world challenges. From disaster management and renewable energy planning to agriculture and futures trading, the potential benefits span a wide range of sectors.

However, as with any innovation, GenCast is not without its limitations. Its relatively lower resolution compared to some traditional models may restrict its utility in highly localised scenarios, and its reliance on historical data could pose challenges in areas with sparse records or in predicting unprecedented climate-driven phenomena. These constraints highlight the ongoing importance of collaboration between AI and traditional meteorological approaches, as DeepMind acknowledges. Their insistence on combining AI-based methods with existing systems demonstrates a pragmatic understanding that AI, while transformative, may not be a silver bullet.

The model’s openness, with its code and weights made publicly available, also signals a commitment to advancing the wider weather and climate community. This transparency may be a helpful way to both foster collaboration and ensure that GenCast can be scrutinised, validated, and improved upon by the global scientific community.

GenCast’s emergence is also likely to intensify competition within the AI sector. As other companies and research institutions strive to match or surpass its capabilities, the pace of innovation in AI-based weather prediction and other real-world applications could accelerate. This competition may benefit society at large by driving further advancements in technology and expanding the possibilities for AI integration across industries.

GenCast, therefore, is a vivid example of how AI can be harnessed to address some of the most pressing global challenges. While there is still room for refinement and integration, its launch signifies a future where advanced technologies like AI play a crucial role in safeguarding lives, improving efficiency, and fostering a deeper understanding of our planet’s dynamic systems.

Sustainability-in-Tech : Google’s AI Discovers 380,000 New Materials

Posted on by Paul Stradling

A new AI tool called GNoME from Google’s DeepMind artificial intelligence lab has reportedly discovered and contributed nearly 380,000 new compounds to the Materials Project, the open-access database founded at the Department of Energy’s Lawrence Berkeley National Laboratory (Berkeley Lab).

GNoME 

The Graph Networks for Materials Exploration (GNoME), is an AI-powered deep learning tool and a state-of-the-art graph neural network (GNN) model. Originally trained with data on crystal structures and their stability, it is particularly suited to discovering new crystalline materials.

Why Is Finding New Crystalline Materials So Important? 

As Google’s DeepMind says: “Modern technologies from computer chips and batteries to solar panels rely on inorganic crystals. To enable new technologies, crystals must be stable otherwise they can decompose, and behind each new, stable crystal can be months of painstaking experimentation.” 

380,000 New Stable Materials Discovered  

DeepMind reports that using its GNoME AI model, not only has it discovered 2.2 million new crystals (the equivalent to nearly 800 years’ worth of knowledge) but has identified 380,000 of these as being the most stable, making them promising candidates for experimental synthesis.

Faster And Cheaper Than Past Methods 

As DeepMind has highlighted, the traditional methods of scientists searching for novel crystal structures have been adjusting known crystals or experimenting with new combinations of elements. These methods have proven to be an expensive, trial-and-error processes that could take months to deliver limited results. Using the GNoME AI model, therefore, has dramatically speeded up and reduced the cost of this process.

Work Already Under Way On The New Materials 

Google says that researchers in labs around the world have already independently created 736 of the newly discovered structures as part of experimental work. Also, in partnership with Google DeepMind, researchers at the Lawrence Berkeley National Laboratory have published a paper showing how the AI discoveries can be leveraged for autonomous material synthesis.

What Does This Mean For Your Organisation? 

Many essential modern technologies rely on a supply of stable inorganic crystals, e.g. for computer chips, batteries, and solar panels. However, up until now, old methods of finding these crystals have involved time-consuming and expensive trial-and error process. Having an AI tool like GNoME has dramatically increased the speed and efficiency of discovery by predicting the stability of new materials. In doing so, it has demonstrated the potential of using AI to discover and develop new materials.

This could mean that AI models (such as GNoME) have the potential to develop a range of future transformative technologies which could include superconductors, powering supercomputers, and next-generation batteries to boost the efficiency of electric vehicles. Also, Google DeepMind releasing its database of newly discovered crystals to the research community could reduce development times for these new transformative technologies.

This could benefit society and businesses (new opportunities and new industries) as well as contributing to achieving environmental targets and improving sustainability by accelerating the development green technologies.