Scaling Test Time Compute to Multi-Agent Civilizations — Noam Brown, OpenAI

Mentioned alongside AlphaGo as an example of AI achieving superhuman performance through pre-training, test-time compute, and self-play.

An IDE tool that Noam Brown uses daily, though he notes that his preferred model, O3, is not yet the default interface, requiring manual selection.

A principle emphasizing that scaling and simple approaches often outperform complex, hand-engineered solutions in AI research. It's cited as a guiding principle for OpenAI's approach to multi-agent systems.

Mentioned for a finding that MCTS (Monte Carlo Tree Search) was not very useful to them.

A book by Richard Hamming that influenced the discussion on technological shifts and adapting workflows in new technologies.

A key figure at OpenAI with whom Noam Brown discussed AGI timelines, RL, and the importance of reasoning paradigms. His early emails emphasized the value of large experiments.

Used as an example of a game where a 'harness' might be needed for AI to perform well, but Noam Brown argues that improving the base model's capabilities is the better approach than relying on external harnesses.

Mentioned as a variant of poker with more hidden cards than Texas Hold'em, increasing the complexity of state enumeration for AI.

AI models that use techniques like chain-of-thought to improve their problem-solving abilities, discussed as a leap forward that requires a base level of capability and is now being integrated into more AI systems.

Used as a parallel to language model development, highlighting the stages of pre-training, large-scale test-time compute, and self-play in achieving superhuman performance.

An AI tool for coding that Noam Brown uses extensively for tasks ranging from research to generating pull requests, finding it effective and a valuable way to understand model limitations.

A stage of AI model development between pre-training and post-training, which OpenAI is explicitly hiring for, though its definition remains fuzzy.

The shape and form of robots, discussed in the context of whether humanoids are the optimal embodiment for AI, especially in domestic environments.

A company whose CEO's pitches influenced Noam Brown's view on the benefits of non-humanoid robotics.

Another social deduction game used as a comparison point for Blood on the Clock Tower, emphasizing its similar core concepts.

The model is mentioned in the context of the 'all you can eat' scaling paradigm and as a benchmark for reasoning capabilities. Its successor, GPT-4.5, is also discussed.

Author of 'The Art of Doing Science and Engineering,' cited for his ideas on technological shifts and the need to adapt workflows rather than simply replicating old tasks in new technology.

Referred to in the context of Diplomacy benchmarking and their potential to improve with scale and new paradigms.

Mentioned as a country that might have its own AI research or development landscape, in contrast to Western hubs.

Mentioned in the context of playing games like tic-tac-toe, where it performs reasonably well but can make mistakes, suggesting a need for system two thinking for perfect play.

Noam Brown's preferred model for daily use, even replacing Google Search, highlighting its utility for web browsing and research. It's also suggested as a primary tool for coding.

The podcast host is mentioned as someone Noam Brown has appeared on previously, indicating Brown's frequent appearances on podcasts.

The organization that the host implies is working on a specific type of model router that aims to balance fast and slow thinking in AI.

Mentioned as an example of another AI model that could be benchmarked against Diplomacy.

A benchmark for machine intelligence, discussed in the context of current language models like GPT-4 and O3 passing it, and how this affects human perception and the search for AI.

An AI system developed by Noam Brown and OpenAI that performed in the top 10% of human players in Diplomacy. Its development influenced Brown's own playstyle and understanding of the game.

A generative AI model for creating videos from text, whose initial announcement was seen as magical and indicative of AGI, but now shows flaws upon closer inspection.

Mentioned in the context of text diffusion, alongside auto-regressive imaging, as an example of different directions in generative media research.

Former president of OpenAI, whom Noam Brown suggests asking about the long-term future of AI and how to steer it towards positive outcomes.

A complex card game with imperfect information and a vast number of possible states, discussed as a challenging domain for AI that current poker AI techniques would not directly solve.

A variant of poker with a limited amount of hidden information, making it amenable to explicit state enumeration and probability assignment for AI.

A model that is discussed as passing the Turing test, with its capabilities improving since 2022. It's also mentioned in the context of agentic systems and conversational AI.

The organization where Noam Brown works, responsible for developing advanced AI models like Cicero and the GPT series.

Mentioned for his work on Voyager skill library and a talk on physical Turing tests, highlighting his role as an educator in AI and robotics.

A benchmark that features difficult but easily gradable problems, which Noam Brown suggests limits the scope of AI evaluation to more common, measurable tasks rather than fuzzier, more complex ones.

A game frequently used as a reference point for AI development, particularly in areas like game theory, sample efficiency, and the trade-off between exploitative and optimal strategies.

The embodiment of AI in human-like robots, discussed in terms of its advantages (familiarity, existing infrastructure) and disadvantages (limitations of human form, potential creepiness) compared to non-humanoid forms.

A game with a very large number of possible states (approaching 40 factorial), which breaks traditional AI approaches used for poker and requires different methods.

Mentioned as a model that likely would not have benefited from reasoning paradigms, highlighting the need for a certain baseline capability in models before applying techniques like chain-of-thought.

Mentioned by reference to a "hit talk" at ICLR about open-endedness and multi-agent systems.

Monte Carlo Tree Search, mentioned as not being very useful to DeepSeek but still a method engineers try for search in AI.

A social deduction game similar to Mafia or Werewolf, which has become popular for socializing and even recruiting in the tech industry, replacing poker as a favored activity.

Likely a misremembered or misheard name for 'Stratego', a board game with a vast number of possible states, making it challenging for traditional AI approaches.

A social deduction game that Blood on the Clock Tower is compared to, highlighting its similar gameplay mechanics.

A prompting technique that improves reasoning in LLMs, discussed as a paradigm that requires a certain baseline model capability to be effective, as seen when applied to GPT-2 versus larger models.

Mentioned as a prior iteration of language models, serving as a precursor to current capabilities and benchmarks.