Google's DeepMind, in collaboration with the Massachusetts Institute of Technology (MIT), recently announced a significant research breakthrough. The research team developed a new autoregressive model called 'Fluid', which has made groundbreaking progress in the field of text-to-image generation. After scaling up to 10.5 billion parameters, the model has demonstrated outstanding performance. This research overturns the industry's common perception; while autoregressive models have dominated the language processing domain, they have traditionally been seen as inferior to Stable Diffusion and Goog.

Recently, Google announced a series of important leadership changes and team restructuring in an official statement. These changes are primarily focused on the company's Knowledge & Information (K&I) team and the Gemini team. Google's CEO Sundar Pichai announced that Nick Fox will replace Prabhakar Raghavan as the new head of the K&I team. The K&I team is responsible for Google's search, advertising, geographic, and business products.

Recently, large language models (LLMs) with ultra-long context windows have become a hot topic of discussion. These models are capable of processing hundreds of thousands or even millions of tokens in a single prompt, opening up many new possibilities for developers. However, how well can these long context LLMs understand and utilize the vast information they receive? To address this question, researchers at Google DeepMind have launched a new benchmark called Michelangelo, aimed at evaluating long context reasoning capabilities. The research findings indicate that...

Recently, Tim Brooks, the research head of OpenAI's Sora video generation system, announced on social media that he will be joining Google DeepMind to continue his work on video generation and the world simulator. Brooks expressed his excitement about collaborating with a talented team like DeepMind and thanked OpenAI for the wonderful two years he spent there, especially for the achievements made in the Sora project.

Recently, researchers from the Mila Institute, Google DeepMind, and Microsoft Research conducted an in-depth investigation into the reasoning capabilities of AI language models. They found that small and inexpensive models exhibit significant shortcomings when addressing complex problems. The study utilized a test called 'Combinatorial GSM' to evaluate how these models perform in solving chain-based foundational math problems. Image source note: Image generated by AI, image licensed through Midjourney.

Recently, Google DeepMind announced its latest AI system - AlphaChip. This system is dedicated to accelerating and optimizing the development of computer chips, and the chip layouts designed by AlphaChip have already been implemented in Google's AI accelerators. AlphaChip operates similarly to what we have previously heard about AlphaGo and AlphaZero, utilizing reinforcement learning techniques to quickly generate optimized chip layouts. According to Google DeepMi

The Google DeepMind research team has recently achieved a significant breakthrough with the development of an innovative technology called SCoRe (Self-Correction through Reinforcement Learning). This technology aims to address the long-standing challenge of large language models (LLMs) struggling to self-correct, enabling them to identify and fix errors without relying on multiple models or external checks. The core of SCoRe's technology lies in its two-phase approach. The first phase optimizes the model.

Google DeepMind has released AlphaProteo, an artificial intelligence system focused on generating novel proteins tailored to specific target molecules, aimed at accelerating drug design, disease understanding, and health application research. AlphaProteo successfully designed binders for the protein VEGF-A, which is associated with tumor growth and various disease complications, outperforming traditional methods. Unlike AlphaFold, the proteins designed by this system can interact with biological processes, providing possibilities for developing targeted therapies against harmful proteins.

The Google DeepMind team, in collaboration with academic institutions, has developed an innovative method called Generative Reward Model (GenRM) aimed at improving the accuracy and reliability of generative AI in reasoning tasks. GenRM incorporates a validation process into text generation tasks, allowing the model to simultaneously generate and evaluate potential solutions while supporting Chain of Thought (CoT), enhancing the comprehensiveness of the validation process. Compared to traditional methods, GenRM has shown significant advantages in multiple tests, with accuracy improvements ranging from 16% to 64%, particularly in...

Google DeepMind has released GameNGen, a fully neural network-driven game engine that revolutionizes traditional game image generation methods. GameNGen can generate game visuals in real-time based on player actions, achieving high-quality images that are nearly indistinguishable from reality, with smoothness close to the minimum frame rate perceptible to the human eye. This technology showcases the immense potential of AI in the game development field, which could change game development efficiency, provide personalized experiences, unlock infinite possibilities, and impact game hardware demands and creative approaches.

At least 200 employees from DeepMind, Google's artificial intelligence research division, signed an internal letter on May 16, expressing concerns over Google's collaboration with military organizations, particularly in relation to reports of Google providing artificial intelligence and cloud computing services to the Israeli military. The employees believe this collaboration contradicts Google's mission statement and AI principles, undermining its image as a 'leader in ethical and responsible AI.' While these employees represent a small percentage of the total workforce at DeepMind, their actions reflect potential conflicts with Google's culture.

In a recent podcast, DeepMind co-founder Demis Hassabis shared his views on the current state of artificial intelligence development. He believes that the capabilities of AI are overly emphasized and exaggerated by the public, leading to some noise. The public's understanding of the significance of AI's future is an outlandish experience. Hassabis expressed surprise at the ability of AI models to infer real-world information based solely on language learning. He noted that many individuals and startups claim functionalities that current technology cannot deliver, referring to these as 'crazy ideas.' Hassabis believes that the long-term potential of AI...

DeepMind co-founder Demis Hassabis pointed out in a podcast that current AI capabilities are overstated, but its long-term potential, particularly in terms of General Artificial Intelligence (AGI), is still underestimated. The public's focus on AI has created a double-edged sword effect, showcasing its positive role in future everyday life while also generating a lot of noise. Hassabis appreciates AI models' ability to infer information about the real world through language learning, but he expressed concern over the public's overly high expectations of AI's short-term capabilities.

Andrej Karpathy suggested that Reinforcement Learning from Human Feedback (RLHF) may not be the ultimate solution for AI to achieve human-level problem-solving capabilities. Using AlphaGo as an example, he pointed out that true reinforcement learning techniques optimize neural networks through self-play, ultimately surpassing humans without human intervention. In contrast, RLHF resembles imitation of human preferences rather than solving problems, being effective in closed environments with clearly defined reward mechanisms, such as Go, but less so in open-ended tasks like article summarization and code rewriting.

Sequoia Capital has announced a $100 million investment in the startup Reflection AI, founded by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou. The company focuses on the development of artificial intelligence technologies, aiming to bring AI into a broader range of applications. This investment brings Reflection AI's valuation to $100 million, highlighting investor confidence in the company's future development. The team plans to increase R&D investment, expand their technical talent, and seek partnerships.

Google DeepMind's latest research, Gemma Scope, unveils the secrets of the 'black box' of language models by using Sparse Autoencoders (SAEs) to decompose and reconstruct the activations of language models, aiming to reveal meaningful features. Gemma Scope employs JumpReLU SAEs, which optimize reconstruction loss and regularize the number of active latent features by controlling activations to reveal the internal mechanisms of language models. The research found that the performance of residual flow SAEs is generally low, and sequence length affects the performance of SAEs.

Google DeepMind's AI systems AlphaProof and AlphaGeometry2 achieved remarkable results in the International Mathematical Olympiad, winning silver medals and demonstrating astonishing capabilities in solving mathematical problems. AlphaGeometry2 solved a geometric problem in 19 seconds, surpassing the speed and accuracy of top human mathematicians. This achievement has been highly praised by legendary figures in the mathematics community, who believe it expands the boundaries of AI applications in the field of mathematics. Terence Tao noted that advances in AI will aid mathematical research.