The Constraints Are the Point

🌈 Abstract

The article discusses the role of constraints in artificial intelligence (AI) systems and argues that embracing the right kinds of constraints is crucial for the advancement of AI research. It examines the distinction between capability and action in human intelligence, and how AI systems have so far only achieved the former. The article then delves into the Neuro-Symbolic AI approach, using the examples of AlphaGeometry 1 and 2, as well as Cicero, to illustrate the value of constraint-based AI systems.

🙋 Q&A

[01] The components of human intelligence and Neuro-Symbolic AI

1. What is the key point made about the components of human intelligence and Neuro-Symbolic AI?

• The article argues that the components of human intelligence are constrained, and this is a good thing. Similarly, Neuro-Symbolic AI will solve some problems at the expense of others, and this will be a mark of progress.
• The article suggests that the constraints are the point, and that general/generally intelligent systems are not sound research objectives.

2. How does the article distinguish between capability and action in human intelligence?

• The article discusses how AI research only ever produces systems with capabilities (competencies), not systems that perform stimulus-free, unbounded, yet appropriate actions (performances).
• It argues that computational models do not act in the same way humans use language, which is stimulus-free, unbounded, and appropriate to the situation.

3. What is the significance of the capability-action distinction in the context of AI research?

• The article suggests that the stronger sense in which humans have "general" intelligence is that they use their intellectual resources effectively at will, whereas AI systems have their intellectual resources directed and steered by humans.
• In humans, the competencies enable the performance, but in machines, the competencies are all there is.

[02] Neuro-Symbolic AI and the examples of AlphaGeometry and Cicero

1. How does the article describe the Neuro-Symbolic approach used in AlphaGeometry?

• AlphaGeometry combines a generative language model (LM) with a symbolic engine (SE) to generate proofs for geometry problems.
• The LM is trained on synthetic data to focus on auxiliary constructions during proof search, while the SE handles the symbolic deduction proof steps.
• The two components work in a loop, with the LM generating auxiliary constructions and the SE expanding the proof state.

2. What are the key strengths and limitations of AlphaGeometry 1 and 2 as described in the article?

• AlphaGeometry 1 was able to solve 25 out of 30 problems on the IMO-AG-30, a state-of-the-art result.
• AlphaGeometry 2, along with AlphaProof, solved 4 out of 6 problems in the 2024 International Mathematical Olympiad, attaining a record-setting silver medal equivalency.
• However, the article notes that AlphaGeometry's limitations are real, and it cannot generalize to anything the symbolic deduction engine cannot verify.

3. How does the article describe the architecture and approach of Cicero, Meta's Diplomacy agent?

• Cicero's architecture can be thought of as a series of moving parts, with a planning engine that handles strategic reasoning and a dialogue agent responsible for natural language communication.
• The "dialogue-conditional action model" unifies these components, with the planning engine asserting control over the dialogue agent.
• Cicero makes use of a Transformer-based LLM for the dialogue agent, but the dialogue agent is subordinate to the planning engine, which is underpinned by variations of the "piKL" algorithm.
• Cicero is constrained to the game of Diplomacy and cannot defeat any human player, but it scores in the top 10% of human players.