2028 AGI, New Architectures, and Aligning Superhuman Models with Shane Legg, Deepmind Founder, on The Dwarkesh Podcast

Episode Summary

Executive Summary: In this episode of the Cognitive Revolution, Dwarkesh Patel interviews DeepMind co-founder and chief AGI scientist Shane Legg. Legg provides a candid assessment of AI progress, defining AGI as a machine that can perform the cognitive tasks humans typically do. He discusses measuring progress, the need for episodic memory and system-two thinking in LLMs, and his 50% probability of achieving human-level AGI by 2028. Legg emphasizes that alignment requires highly capable systems with deep world models, ethical understanding, and robust reasoning, rather than containment. He also reflects on DeepMind's impact on safety and capabilities.

Main Topics: Defining and Measuring AGI (Priority: 5/5): Legg defines AGI as a machine that can do the cognitive things people typically do, possibly more. He argues that measuring progress requires a comprehensive suite of tests spanning human cognitive abilities, and that AGI is achieved when no one can find a task where the machine falls below human performance. Shortcomings of Current LLMs (Priority: 5/5): Legg identifies key limitations: lack of episodic memory (rapid learning of specific information), absence of system-two thinking (deliberate reasoning and search), and issues with delusions and factuality. He sees clear paths to address these through architectural innovations and search mechanisms. Path to AGI: Scaling and Architecture (Priority: 4/5): Legg believes that current foundation models are powerful sequence predictors, but true creativity and AGI require search (system-two thinking). He expects multimodal integration (video, images) to be the next landmark, opening up grounded understanding and new applications. AI Alignment and Safety (Priority: 5/5): Legg argues that alignment requires highly capable systems with deep world models, ethical understanding, and robust reasoning. He advocates for a 'system-two' approach where AI reasons through ethical decisions, rather than relying solely on reinforcement learning. He supports concrete safety benchmarks and red teaming. DeepMind's Impact and Historical Context (Priority: 3/5): Legg reflects on DeepMind's role in accelerating both capabilities and safety research. He notes that the company was the first AGI company with a dedicated safety group, lending credibility to the field. He also discusses his 2009 prediction of human-level AI by 2028, based on exponential growth in compute and data.

Key Arguments: AGI is defined by generality, not specific task performance; a comprehensive test suite is needed to measure progress. Current LLMs lack episodic memory (rapid learning) and system-two thinking (deliberate reasoning and search), which are essential for human-level intelligence. These shortcomings are not fundamental; there are clear architectural paths to address them, such as adding separate memory systems and search mechanisms. Alignment requires highly capable systems with deep world models, ethical understanding, and robust reasoning; containment is not a winning strategy. The next landmark in AI will be full multimodality (video, images), leading to more grounded understanding and new applications. DeepMind has accelerated both capabilities and safety research, but counterfactuals are difficult to assess; the community is much larger than any single company.

Data Points: Probability of human-level AGI by 2028: 50% - Legg's modal estimate from 2009, based on exponential growth in compute and data. Year of first AGI safety hire at DeepMind: 2013 - Legg notes that the first hire only agreed to work part-time on safety. Training data scale relative to human lifetime: Beyond a human lifetime - Legg states that models can now be trained on more data than a human experiences in a lifetime, which he sees as a key unlocking step.

Pivotal Quotes: "I think there are kind of relatively clear paths forwards now to address most of the shortcomings we see in existing models, whether it's about delusions, factuality, the type of memory and learning that they have, or understanding video, or all sorts of things like this." — Shane Legg: Legg expresses optimism about overcoming current LLM limitations through further research. "To get real creativity, you need to search through spaces of possibilities and find these sort of hidden gems. That's what creativity is. I think current language models don't really do that kind of a thing." — Shane Legg: Legg distinguishes between LLMs' ability to blend existing data and true creativity requiring search. "I think there's a 50% chance that somewhere in 2028. Now, it's just a 50% chance. I mean, I'm sure what's going to happen is going to get to 2029, and someone's going to say, Oh, Shane, you were wrong. It's like, come on, it's 50% chance." — Shane Legg: Legg reiterates his 2009 prediction of human-level AGI by 2028, emphasizing it's a probabilistic estimate.

Implications: Legg's analysis suggests that AGI within the next few years is plausible, with clear technical paths to address current limitations. This underscores the urgency of alignment research, particularly system-two approaches, and the need for concrete safety benchmarks. The coming years will likely see transformative applications from multimodal models, but also increased risks from misuse.