Google DeepMind's SIMA is a generalist AI agent for 3D games that follows natural-language instructions across various video game environments, marking a shift towards creating versatile, instructable AI systems.

ALOHA Unleashed by Google DeepMind is a new generation of AI-powered robots with impressive dexterity. DeepMind recently released a series of videos showing the robots hanging shirts, inserting precise gears, and even tying shoelaces. The robots can generalize to untrained objects. The videos are available in the link.

Following Anthropic, GDM has released some work from its alignment efforts. The main insightful piece here is using sparse autoencoders on Gemini Ultra. This is a substantial jump in interpretation size.

Google DeepMind's AlphaFold 3 can predict the structure of all of life's molecules. Previous versions of the model could only predict the structures of proteins. AlphaFold 3 can model DNA, RNA, and ligands. It shows a 50% improvement in prediction accuracy compared to its previous models. DeepMind is making the AlphaFold Server research platform, powered by AlphaFold 3, to some researchers for free for academic and noncommercial use.

Google DeepMind and Isomorphic Labs have developed the 3rd generation of AlphaFold, a powerful protein folding prediction model. It is launching AlphaFold Server, which is a free way to interact with the model. AlphaFold 3 is 50% more accurate than previous generations. It correctly predicted the folded structure of the spike protein on Coronavirus OC43.

Google DeepMind introduced the Frontier Safety Framework to address risks posed by future advanced AI models. This framework identifies critical capability levels (CCLs) for potentially harmful AI capabilities, evaluates models against these CCLs, and applies mitigation strategies when thresholds are reached.

A group of current and former OpenAI employees, along with two from Google DeepMind, have published an open letter criticizing OpenAI's prioritization of profits and growth over AI safety. The whistleblowers claim that OpenAI has used restrictive agreements to silence employees' concerns. While some see the letter as evidence that AI safety concerns are valid, others criticize it for lacking concrete details and focusing on hypothetical risks.

Google DeepMind has released Gemma 2, a family of high-performing and efficient open models for various AI tasks. Gemma 2 is designed to run efficiently on a single GPU or TPU host, and it is compatible with major AI frameworks like Hugging Face Transformers, JAX, PyTorch, and TensorFlow. It is available in 9 billion and 27 billion parameter sizes.

DeepMind and Harvard University developed a virtual rat with AI neural networks trained on actual rat movements and neural patterns to probe the brain circuits responsible for complex motor skills. This bio-inspired AI has the capacity to generalize learned movement skills to new environments, offering insights into brain function and advancing robotics. The research demonstrates that digital simulations can effectively mimic and decode neural activity related to different behaviors.

A post by Neel Nanda, a Research Engineer at Google DeepMind, about his favorite papers to read in Mechanistic Interpretability.

Google DeepMind's AlphaProof and AlphaGeometry 2 AI systems reportedly solved four out of the six problems from this year's International Mathematical Olympiad. AlphaProof uses reinforcement learning to train itself while AlphaGeometry 2 is powered by a Gemini-based language model. While the achievement is significant, the programs needed a lot more time to solve the problems than human contestants, and some of the problems needed to be manually translated by humans before the AI could understand them.

The computational cost of replicating a Google DeepMind paper that explores optimal hyperparameters for large language models is estimated to be around $12.9 million in H100 GPU hours. The paper involves extensive learning rate sweeps across different model sizes, parameterizations, and optimizers, leading to a large number of training runs. This post provides a breakdown of the compute cost for each type of experiment, including alignment experiments, learning rate variants, epsilon experiments, and weight decay.

Google DeepMind's AGI Safety & Alignment team shared a detailed update on their work focused on existential risk from AI. Key areas include amplified oversight, frontier safety, and mechanistic interpretability, with ongoing efforts to refine their approach to technical AGI safety. They highlighted recent achievements, collaborations, and plans to address emerging challenges.

Nearly 200 Google DeepMind employees signed a letter urging Google to terminate military contracts, claiming a violation of the company's own AI ethics principles. DeepMind technology has been bundled into Google Cloud and sold to militaries, sparking internal conflict with AI staff who value ethical standards. Google's response demonstrated an adherence to AI Principles, but workers remain unsatisfied, seeking stronger governance against military use of their AI.

Large language models sometimes fail at tasks like counting letters due to their tokenization methods. This highlights limitations in LLM architecture that affect their understanding of text. Nevertheless, advancements continue, such as OpenAI's Strawberry for improved reasoning and Google DeepMind's AlphaGeometry 2 for formal math.

The discussion surrounding artificial general intelligence (AGI) has gained significant traction, particularly among employees of major tech companies like OpenAI and Google DeepMind, who often assert that the development of AGI is inevitable. However, researchers from Radboud University and other institutions challenge this notion, presenting evidence that suggests the realization of AGI is not only unlikely but fundamentally impossible. Their findings, published in the journal *Computational Brain & Behavior*, highlight the complexities and limitations inherent in replicating human-level cognition through artificial means. Iris van Rooij, the lead author and a professor of Computational Cognitive Science at Radboud University, argues that while some theorists believe AGI could be achieved in principle, this belief does not translate into practical feasibility. The researchers emphasize that pursuing AGI is a misguided endeavor that squanders valuable resources. They propose a thought experiment in which AGI development occurs under optimal conditions, with access to perfect datasets and advanced machine learning techniques. Despite these ideal circumstances, the researchers conclude that there is no viable method to achieve the cognitive capabilities that tech companies promise. The paper elaborates on the inherent challenges of replicating human cognition, which involves complex processes such as memory recall and contextual understanding. Van Rooij illustrates this by explaining how humans can seamlessly integrate knowledge from various points in time during conversations, a feat that current AI systems cannot replicate. Olivia Guest, a co-author and assistant professor, adds that the computational power required to create AGI would deplete natural resources long before reaching the necessary capabilities. The collaboration among researchers from various universities underscores the importance of interdisciplinary approaches in understanding AI. The team advocates for critical AI literacy, emphasizing the need for the public to develop a nuanced understanding of both human cognition and AI capabilities. They warn against the tendency to overestimate what AI can achieve while underestimating the complexities of human thought processes. Van Rooij stresses the importance of skepticism towards the promises made by tech companies, urging individuals to apply critical thinking to claims about AI advancements. In summary, the researchers at Radboud University present a compelling argument against the inevitability of AGI, highlighting the significant cognitive challenges and resource limitations that make such advancements unlikely. They call for a more informed public discourse on AI, encouraging critical evaluation of the technology's capabilities and the motivations behind its promotion by the tech industry.

Concordia is a library developed by Google DeepMind designed for generative social simulation, enabling the construction and use of agent-based models that simulate interactions among agents in various environments, whether physical, social, or digital. The library employs a unique interaction pattern reminiscent of tabletop role-playing games, where a special agent known as the Game Master (GM) orchestrates the simulation. The GM acts as a narrator, interpreting the actions of player agents, which are expressed in natural language, and translating these actions into practical implementations within the simulated environment. In a physical simulation, the GM assesses the plausibility of agent actions and describes their consequences. In digital contexts, the GM may facilitate necessary API calls to integrate with external tools based on agent inputs. This flexibility allows Concordia to be applied across diverse fields, including social science research, AI ethics, cognitive neuroscience, and economics. It can also generate data for personalization applications and evaluate the performance of real services through simulated usage. To utilize Concordia, users need access to a standard Large Language Model (LLM) API, which is essential for generating responses and actions. The library can be installed via the Python Package Index (PyPI) using a simple pip command, or users can opt for a manual installation if they wish to work directly with the source code. The installation process includes cloning the repository and setting it up in an editable mode for development purposes. An illustrative example of Concordia's capabilities involves a social simulation scenario where four friends are trapped in a snowed-in pub, with two of them embroiled in a dispute over a crashed car. The agents in this simulation are designed based on a reasoning framework that prompts them to consider their situation, their identity, and appropriate actions in response to their circumstances. For those who wish to cite Concordia in their work, a specific article detailing its methodology and applications is provided for reference. It is important to note that Concordia is not an officially supported Google product, but it offers a robust framework for researchers and developers interested in exploring generative agent-based modeling.