# Gemini Exponential, Demis Hassabis' ‘Proto-AGI’ coming, but … Source: https://www.youtube.com/watch?v=WHqaF4jbUYU Recap page: https://rapidrecap.app/video/WHqaF4jbUYU Generated: 2025-12-19T17:16:33.874+00:00 --- ## Quick Overview The exponential growth in AI compute spending, projected to consume half of OpenAI's revenue by 2028, is the single biggest blocker to achieving Artificial General Intelligence (AGI), according to Shane Legg, co-founder of Google DeepMind, who advocates for efficient algorithm development rather than just pure scaling to mitigate this trend. **Key Points:** - Shane Legg, Google DeepMind co-founder, stated that the exponential growth in compute costs is the biggest blocker to achieving AGI. - OpenAI's projected compute spending could reach half of its revenue by 2028, a trend he suggests is unsustainable without major breakthroughs. - Legg cited his 2019 prediction of AGI by 2029, noting that the current pace of scaling is outpacing his initial timeline. - He advocated for focusing on algorithm innovation (like finding better ways to simulate physics or math) rather than simply scaling models. - The release of Gemini 3 Flash demonstrates that smaller, faster models can achieve performance comparable to larger models, suggesting efficiency gains are possible. - The video referenced external reports, including one from The Information regarding specialized data sales and another from Reuters about compute crunch, illustrating the competitive landscape and cost pressures. - The discussion highlighted that while models like GPT-5.2 are improving, the underlying compute demands remain a critical constraint. ![Screenshot at 00:46: The host shows a performance benchmark chart comparing various LLMs across several metrics, setting the stage for a discussion on model capabilities and efficiency.](https://ss.rapidrecap.app/screens/WHqaF4jbUYU/00-00-46.png) **Context:** This content is derived from an interview on the Big Technology Podcast featuring Shane Legg, co-founder of Google DeepMind, discussing the state of AI development, particularly concerning the exponential growth in compute resources required for training large language models (LLMs) and the timeline for achieving Artificial General Intelligence (AGI). The discussion contrasts the pure scaling approach with the need for algorithmic efficiency, referencing recent model releases like Gemini 3 and the competitive dynamics between major AI labs. ## Detailed Analysis The discussion focuses on the primary bottleneck to achieving AGI: the exponential growth in compute costs. Shane Legg, co-founder of Google DeepMind, projects that OpenAI's compute spending will consume half of its revenue by 2028, a trajectory he deems unsustainable without fundamental shifts. He referenced his 2019 prediction of AGI by 2029, noting that recent model releases, such as Gemini 3, are accelerating progress faster than anticipated, but the underlying resource demands are the main constraint. Legg argued against relying solely on scaling, suggesting that advancements in algorithms—like those that allow models to better simulate physics or math—are crucial. He pointed to Gemini 3 Flash as an example of efficiency, achieving high performance at a lower cost than its larger counterparts. The conversation also touched upon external factors, like data scarcity (referencing a tweet about companies refusing to sell proprietary data to OpenAI/Anthropic) and the competitive landscape where Google is trying to keep pace with competitors like OpenAI and Anthropic, particularly in areas like image generation with models like Veo 3.1 and Nano Banana Pro. The overall theme suggests that the future of AI advancement hinges not just on throwing more compute at the problem, but on finding smarter, more efficient algorithmic solutions. ### Model Releases and Performance - Gemini 3 Flash offers comparable performance to larger models at significantly lower cost (0:15); GPT-5.2 Codex underperforms GPT-5.1 Codex-Max on MLS-Bench-30 (7:03); Companies like OpenAI and Google are fiercely competing for user data and training resources (17:54). ### AGI Timeline and Compute Constraint - Legg predicts AGI by 2029, but notes that the exponential rise in compute costs is the primary blocker (11:36, 11:53); OpenAI's compute spending projected to hit 50% of revenue by 2028 (15:34). ### DeepMind's Approach - DeepMind is training separate models for specific tasks (like gaming/simulation via SIMA 2) rather than just relying on a single, massive model for everything (8:55, 9:17); Focus should be on algorithmic breakthroughs, not just scale (13:19, 13:44). ### Data Scarcity and Competition - Specialized data is becoming scarce, with life science/accounting companies refusing to sell proprietary data to OpenAI/Anthropic (17:54); This forces companies to be more creative with available data (18:48). ### Future Outlook - The expectation is that the cost of compute will continue to rise exponentially, making efficiency breakthroughs critical (15:36, 17:46). ![Screenshot at 00:00: Title card: 'NEW GEMINI + THE 'PROTO-AGI'' featuring the speaker.](https://ss.rapidrecap.app/screens/WHqaF4jbUYU/00-00-00.png) ![Screenshot at 00:15: Benchmark chart comparing various LLMs across performance metrics, highlighting Gemini 3 Flash as a fast and capable model.](https://ss.rapidrecap.app/screens/WHqaF4jbUYU/00-00-15.png) ![Screenshot at 03:44: A chart showing the 'AA-Omniscience Hallucination Rate' where lower is better, illustrating that GPT-5.1 Codex-Max has the lowest hallucination rate among tested models.](https://ss.rapidrecap.app/screens/WHqaF4jbUYU/00-03-44.png) ![Screenshot at 07:03: A bar chart displaying performance on the MLS-Bench-30 coding task, showing GPT-5.2 Codex underperforming GPT-5.1 Codex-Max.](https://ss.rapidrecap.app/screens/WHqaF4jbUYU/00-07-03.png) ![Screenshot at 11:35: A graph tracking the 'Time-horizon of software engineering tasks different LLMs can complete 50% of the time' against LLM release date, showing GPT-5.1 Codex-Max as the current high performer.](https://ss.rapidrecap.app/screens/WHqaF4jbUYU/00-11-35.png)