Chronological feed of everything captured from Demis Hassabis.
More details about the exciting partnership here: https://www.eveonline.com/news/view/a-new-era
Welcome @shimon8282 - excited for the multiagent work ahead!
Really enjoyed this conversation - thanks again @garrytan for hosting!
Fun to see Google DeepMind still making the paper headlines 10 years after AlphaGo!! 😀
Demis Hassabis expressed positive feedback after speaking at Y Combinator, highlighting the inspirational environment and enthusiasm for startups leveraging Google Gemma models. This endorsement from the DeepMind CEO signals growing adoption of Gemma in early-stage AI ventures. The post replies to YC President Garry Tan's announcement of the event.
Gemma 3 extends the Gemma family with 1B to 27B parameter multimodal models supporting vision, expanded languages, and 128K+ context lengths. It optimizes KV-cache memory via increased local-to-global attention ratio and short local spans. Trained with distillation and novel post-training, Gemma3-4B-IT matches Gemma2-27B-IT, while Gemma3-27B-IT rivals Gemini-1.5-Pro on math, chat, instruction, and multilingual benchmarks. All models are openly released.
Demis Hassabis discussed AGI pathways and AI's role in accelerating science and medicine with Harry Stebbings, highlighting UK deep tech talent. Stebbings equates Hassabis's impact to Turing, Newton, and Einstein for advancing humanity. The exchange underscores persistence in pursuing ambitious tech visions from humble beginnings.
Gemini Robotics-ER 1.6 upgrades robotic vision with superior object pinpointing in clutter, multi-view scene fusion for task completion detection, and precise analog gauge reading via spatial reasoning and self-generated distortion correction code. Integrated with Boston Dynamics' Spot, it enables autonomous industrial inspections while adhering to safety constraints like avoiding liquids, heavy objects over 20kg, and human injury risks. The model is deployable now via Google AI Studio and Gemini API.
Gemini Robotics-ER 1.6 upgrades robot vision models with superior visual and spatial understanding, enabling accurate object detection in cluttered environments and multi-view scene fusion for task completion verification. It excels at reading analog gauges with sub-tick precision by generating corrective code for distortions and integrates world knowledge with physical constraints for safer operations. Deployable now via Google AI Studio, it boosts industrial inspection on platforms like Boston Dynamics' Spot, improving human risk detection by 10%.
Google's Gemini 3.1 Flash now features integrated Text-to-Speech (TTS) capabilities, enabling sub-200ms latency for responsive voice interactions. This positions it as a frontrunner in efficient, multimodal AI models suitable for live conversational agents. Technical details are available in the linked Google blog post.
Demis Hassabis views Artificial General Intelligence (AGI) as the ultimate scientific tool, driven by a lifelong quest to understand the universe and apply this knowledge to solve critical global challenges in medicine, energy, and environment. He emphasizes a cautious optimistic approach to AGI development, acknowledging non-zero risks but believing human ingenuity can ensure safe deployment. He advocates for international cooperation and minimum standards among leading labs despite current competitive pressures.
The "Experience AI" class addresses common misconceptions and questions about artificial intelligence among students. The curriculum focuses on fundamental AI concepts like data importance, potential biases, and the volume of data required for model training. The initiative aims to equip future generations with AI literacy to foster problem-solving and innovation.
Demis Hassabis, co-founder of DeepMind, stands out in the AI landscape due to his early conviction in AI's potential, scientific rigor, and entrepreneurial drive. Unlike many peers, he envisioned and pursued artificial general intelligence decades before its mainstream recognition, combining deep learning with reinforcement learning inspired by neuroscience. He's a key figure in the dramatic and highly competitive race to develop advanced AI, navigating complex scientific and business challenges.
Demis Hassabis, co-founder of DeepMind and now leading Google's AI efforts, is characterized by an unwavering, missionary-like dedication to achieving Artificial General Intelligence (AGI) and leveraging it to solve humanity's most complex scientific problems. His career trajectory, from chess prodigy to game designer to neuroscientist and AI pioneer, reveals a consistent drive to understand intelligence and apply that understanding to create powerful, autonomous learning systems. Hassabis approaches challenges with intense competitiveness and pragmatism, learning from past failures and strategically navigating the corporate landscape to secure the resources necessary for his ambitious vision, often viewing fundraising and corporate maneuvers as necessary distractions from his core mission.
Demis Hassabis, CEO of Google DeepMind, engaged in a discussion with Cleo Abram, emphasizing AI's positive impact on scientific advancement and future potential. The conversation touched upon various facets of AI, including its optimal applications, drug discovery, and creative capabilities, while also addressing concerns regarding AI development and its implications for humanity and governance.
Demis Hassabis outlines a strategic transition from narrow, specialized AI (e.g., AlphaFold) to Artificial General Intelligence (AGI) to solve 'root node' problems in science. He emphasizes the synergy between deep reinforcement learning's creative search capabilities and foundation models' generalization to accelerate drug discovery, material science, and energy solutions.
Demis Hassabis posits that AGI is likely within five years, driven by a shift from pure compute scaling to new algorithmic breakthroughs in continual learning and long-term planning. He envisions AGI as a catalyst for a 'golden age' of scientific discovery, specifically in drug design and energy, while advocating for international, audit-based regulatory frameworks to manage existential risks.
Demis Hassabis, co-founder of DeepMind, has accepted an invitation to speak at Y Combinator. This event will likely focus on topics relevant to AI, startups, and technological innovation, given Hassabis's background and Y Combinator's focus.
The path to AGI necessitates a convergence of foundation models, world models for physical intuition, and a dual-track focus on scaling and innovation. While current models exhibit inconsistent 'jagged' capabilities, the integration of agentic autonomy and high-fidelity physics simulations is expected to bridge the gap toward general intelligence. This transition is anticipated to cause profound socio-economic dislocation at a velocity far exceeding previous industrial revolutions.
Demis Hassabis discusses Google DeepMind's roadmap towards Artificial General Intelligence (AGI), highlighting that while significant progress has been made, true AGI is still several years away. He emphasizes the current limitations of AI in reasoning, hierarchical planning, and long-term memory, and the need for systems to invent novel hypotheses rather than just proving existing ones. The conversation also delves into the transformative potential of AI in scientific discovery, particularly in biology and material science, and the societal implications of advanced AI systems.
DeepMind's Deep Think mode, integrating AI like Gemini, significantly accelerates the design process. This allows for rapid iteration and exploration of novel solutions, potentially reducing design cycles by a factor of ten. The AI acts as an accelerant, enabling designers to explore variations and materials more efficiently, ultimately bringing products to market faster.
A theoretical physicist leveraged Gemini for fact-checking, discovering a critical mathematical error in a peer-reviewed paper on infinite dimensional algebra and symmetry. This highlights AI's capacity for rigorous academic validation, even in fields with limited training data, and its potential to accelerate scientific discovery by identifying fundamental inconsistencies.
DeepMind's "Deep Think" AI is demonstrating significant advancements in materials science by optimizing the synthesis of 2D semiconductors. The AI, acting as a "Deep Tank" for design, effectively navigates complex parameter spaces (e.g., gas flow, thermal profiles) to achieve superior material growth, surpassing human expert capabilities in terms of efficiency and results. This accelerates the development of next-generation electronic materials as silicon approaches its theoretical limits.
The Music AI Sandbox, an experimental suite of tools, facilitates music creation by allowing artists to generate samples, extend clips, and edit sounds. This technology aims to augment human creativity rather than replace it, integrating AI-generated audio with human curation and artistic vision. The collaborative process demonstrates how AI functions as an instrument, extending creative possibilities through iterative refinement.