AlphaZero
A successor Go/chess-playing system that, like AlphaGo, did not rely solely on brute-force search.
As defined in et_648
Where it comes up · 16
- Demis Hassabis — Scaling, superhuman AIs, AlphaZero atop LLMs, AlphaFold4 claims
- 153 - Possible Minds3 claims
- The future of intelligence | Demis Hassabis (Co-founder and CEO of DeepMind)3 claims
- Ilya Sutskever: Deep Learning | Lex Fridman Podcast #942 claims
- Ray Kurzweil (USA) at Ci2019 - The Future of Intelligence, Artificial and Natural2 claims
- Demis Hassabis — Scaling, superhuman AIs, AlphaZero atop LLMs, AlphaFold2 claims
- What does the next training paradigm look like?1 claim
- Vishy Anand1 claim
and 8 more
Most important claims · 10 of 26Search all claims for “AlphaZero” →
- There is a trade-off between model quality and search: AlphaZero plays beyond world-champion level using only tens of thousands of evaluated positions versus millions for brute-force systems like Deep Blue, while human grandmasters use only a few hundred—because richer, more accurate world models let you pick which nodes to expand, so improving models makes search more efficient and lets you go further.Demis Hassabis — Scaling, superhuman AIs, AlphaZero atop LLMs, AlphaFold
- Calling these systems 'artificial intelligence' is a misnomer because it wrongly assumes they resemble human thinking, when in fact we don't understand our own brains well enough; a better term is 'alien intelligence,' because as self-propagating systems they grow more alien to us and begin making higher-stakes decisions in ways we don't understand—with the original creating group functioning as a 'God in that system.'Amy Webb on Artificial Intelligence, Humanity, and the Big Nine
- Human children are the best example we have of a system that truly works as an intelligence, generalizing from comparatively small amounts of data using innate structure, whereas current AI relies on statistical inference from huge datasets, is fooled by adversarial and superficial features, and suffers catastrophic forgetting when rules change slightly, so nothing we have is even in the ballpark of a four-year-old.153 - Possible Minds
- The recent AI spring was not driven by a new idea about how the mind works but by old ideas from the 1980s (neural networks, associative networks) becoming practical once large internet-scale datasets and Moore's-law compute power made them able to scale beyond toy problems, surprising even their designers.153 - Possible Minds
- RL agents like AlphaZero and AlphaGo are active learners whose next actions determine which next data or experience they receive, creating a feedback loop; this active learning, including in realistic simulation environments, is essential for understanding how actions affect the world and for robotics.Demis Hassabis — Scaling, superhuman AIs, AlphaZero atop LLMs, AlphaFold
- AlphaZero's superiority over Stockfish stems partly from roughly 4000 times more raw operations per second, but also from a form of meta-reasoning: rather than Stockfish's fixed-depth alpha-beta search, AlphaZero figured out for itself which infinitesimal fraction of the astronomically large game tree to explore deeply, embodying a general principle of deciding which computations will aid long-run decisions that applies to all reasoning.153 - Possible Minds
- LLMs must keep improving as reliable world models—a necessary but insufficient component of AGI—and AlphaZero-like planning mechanisms should be built on top to make concrete plans, chain reasoning, and use search to explore massive possibility spaces, a capability currently missing from large models.Demis Hassabis — Scaling, superhuman AIs, AlphaZero atop LLMs, AlphaFold
- Reinforcement learning with synthetic data—through simulation, self-play between systems, and multimodal data generation—can overcome data bottlenecks and provide the scalable learning signal needed for advanced AI systems.Demis Hassabis — Scaling, superhuman AIs, AlphaZero atop LLMs, AlphaFold
- Self-play systems (OpenAI Five in Dota, hide-and-seek agents, AlphaZero) produce creative solutions that surprise researchers with truly novel strategies not anticipated during design, which is an important property for AGI that current supervised learning systems routinely lack.Ilya Sutskever: Deep Learning | Lex Fridman Podcast #94
- Large language models need planning and search mechanisms on top of improved world models—similar to AlphaZero—to explore solution spaces efficiently and achieve AGI capabilities, rather than relying on LLMs alone.Demis Hassabis — Scaling, superhuman AIs, AlphaZero atop LLMs, AlphaFold