What is AGI and where are AI bottlenecks today?

TL;DR

AGI is defined as a system that exhibits all cognitive capabilities of the human mind, and compute remains the biggest bottleneck for both scaling and experimentation. While progress in video models and world models is ahead of earlier expectations, continual learning and long term planning are still missing pieces that could unlock broader generalization.

Transcript

I would say about 90% of the breakthroughs that underpin the modern AI industry were done either by Google brain or Google research or deep mind. So one of our groups the returns are kind of still very substantial although they're a bit less than they were obviously at the start of all of this scaling. We have amazing guests on the show but very fe... Read More

Key Insights

• Compute is the main bottleneck for scaling and testing new ideas, acting as the workbench for research.
• Scaling laws provide substantial returns but are not guaranteed to stay exponential, requiring more efficient use of resources.
• Continual learning is a critical missing capability, as current systems do not easily learn after deployment.
• Brains use sleep and reinforcement learning for memory consolidation, a model researchers are exploring to improve integration of new information.
• Current systems struggle with long-term planning and consistent reasoning over years, revealing gaps in hierarchical planning.
• Open collaboration and pooling resources across teams helped DeepMind lead in several breakthroughs such as AlphaGo and transformer concepts.
• Leading labs with the ability to invent new algorithmic ideas are likely to gain bigger advantages as older ideas plateau.
• As models scale, interfaces and tools like coding and math aids shape the next wave of progress rather than merely scaling existing approaches.

Questions & Answers

Q: What is AGI according to the discussion?

AGI is defined as a system that exhibits all the cognitive capabilities the human mind has. This definition emphasizes that the brain is the existence proof we know of for general intelligence, and it sets the bar for what AGI should be. The concept guides research priorities toward broad, flexible intelligence rather than narrow task performance.

Q: How soon is AGI likely to appear according to the guest?

The guest suggests there is a very good chance of AGI arriving within the next five years. He notes that this assessment is based on a probability distribution around timelines and reflects ongoing progress in compute, model architectures, and scaling efforts, while acknowledging uncertainties in the field.

Q: What is the biggest bottleneck in AI today?

Compute is the biggest bottleneck, enabling not only model scaling but also the ability to run experiments at scale with many new ideas. The cloud serves as the workbench, and lacking sufficient compute can slow down testing and verification of new algorithms before they are deployed.

Q: Do scaling laws continue to produce exponential gains?

The guest believes that scaling gains are nuanced. While there have been enormous jumps with each generation of large language models, those gains are not strictly exponential forever. The field still sees substantial returns from scaling, but the rate of improvement may slow, especially as the frontier moves forward.

Q: What capabilities are still missing in AI systems?

Key missing capabilities include continual learning, advanced memory systems, and long-term planning. Current systems do not learn after deployment in a way similar to humans, and their planning over long horizons remains limited, leading to potential inconsistencies and errors in complex tasks.

Q: How did DeepMind accelerate its progress recently?

Progress was accelerated by organizational changes and by combining the company’s resources to build the biggest models. This approach mirrors a startup mindset, focusing on relentless pace and collaboration to push back toward the frontier and achieve breakthroughs in multiple domains.

Q: Why is continual learning challenging for AI?

Continual learning is challenging because researchers have not yet figured out how to integrate new learning into existing systems that have already been trained. Brain-inspired concepts like memory consolidation during sleep inspire approaches, but a robust, scalable solution for continuous adaptation is still under active development.

Q: What role does openness and collaboration play in AI progress?

Open sharing of research and models has long been a practice in AI, and the guest highlights that leading labs with access to broad resources and talent can push forward faster. Consolidating talent and tools across organizations helps accelerate breakthroughs and sustain momentum in a highly competitive field.

Summary & Key Takeaways

• AGI defined as systems with human-level cognitive capabilities, with compute as a primary bottleneck that limits scaling and experimentation.

• DeepMind and other leading labs have driven key breakthroughs, and organizational and resource consolidation helped accelerate progress.

• Ongoing missing capabilities include continual learning, memory architecture improvements, and long term planning that could enable more consistent general intelligence.