Adam Marblestone – AI is missing something fundamental about the brain | Dwarkesh Patel

In this interview podcast, Dwarkesh Patel and Adam Marblestone delve into the complexities of the brain and its relation to AI. They discuss the differences between the brain and current AI models, focusing on the brain's efficiency and ability to generalize. Marblestone introduces Steve Burns' theory of a "steering subsystem" within the brain, responsible for innate responses and reward functions, and how it interacts with the cortex. The conversation explores the potential of omnidirectional inference in AI, amortized inference, and the role of reward functions in learning. They also touch on the importance of co-designing algorithms with hardware, the potential of connectomics, and the application of formal methods in mathematics and software verification.

Outlines

Part 1: Brain Architecture vs. AI Models

Part 2: Inference, Evolution, and Biological Constraints

Part 3: Reinforcement Learning and Hardware Trade-offs

Part 4: Neuroscience Research and AI Timelines

Part 5: Formal Logic, Math, and Future Capabilities

Part 6: World Models and Scientific Infrastructure

Sign in to continue reading, translating and more.

Continue

Adam Marblestone – AI is missing something fundamental about the brain

Dwarkesh Patel

Part 1: Brain Architecture vs. AI Models

00:00The Brain vs. LLMs: Unlocking the Secrets of Intelligence

The Brain vs. LLMs: Unlocking the Secrets of Intelligence

02:31Cortex as a Prediction Engine: Omnidirectional Inference and Lizard Brain Integration

Cortex as a Prediction Engine: Omnidirectional Inference and Lizard Brain Integration

06:06Evolution's Encoding of High-Level Desires and the Role of Loss Functions

Evolution's Encoding of High-Level Desires and the Role of Loss Functions

09:02The Steering Subsystem: Innate Responses and Generalization in the Cortex

The Steering Subsystem: Innate Responses and Generalization in the Cortex

14:32Academic Incentives, Loss Functions, and the Encoding of Desires

Academic Incentives, Loss Functions, and the Encoding of Desires

17:26Commingled Questions: Learning Algorithms, Connections, and Inductive Biases

Commingled Questions: Learning Algorithms, Connections, and Inductive Biases

Part 2: Inference, Evolution, and Biological Constraints

22:20Amortized Inference: Neural Networks vs. Bayesian Inference

Amortized Inference: Neural Networks vs. Bayesian Inference

27:03Evolution, the Genome, and Bootstrapping Cost Functions

Evolution, the Genome, and Bootstrapping Cost Functions

30:12Cell Types and the Steering Subsystem: Evidence from Single-Cell Atlases

Cell Types and the Steering Subsystem: Evidence from Single-Cell Atlases

32:41Genetic Wiring and the Complexity of Reward Functions

Genetic Wiring and the Complexity of Reward Functions

35:30Nucleotides, Megabases, and the Hominid Brain Explosion

Nucleotides, Megabases, and the Hominid Brain Explosion

37:37The Cortex and Omnidirectional Inference: Reward Functions and Social Learning

The Cortex and Omnidirectional Inference: Reward Functions and Social Learning

Part 3: Reinforcement Learning and Hardware Trade-offs

42:05Model-Based vs. Model-Free RL and the Dumbest Form of RL

Model-Based vs. Model-Free RL and the Dumbest Form of RL

44:35Dopamine, Reward Prediction Error, and Model-Based Systems

Dopamine, Reward Prediction Error, and Model-Based Systems

47:52Model-Free RL at a Civilizational Level

Model-Free RL at a Civilizational Level

50:31Biological Hardware vs. Computers: Advantages and Disadvantages

Biological Hardware vs. Computers: Advantages and Disadvantages

53:16Cellular Stuff and Synaptic Connections: The Algorithmic Thing

Cellular Stuff and Synaptic Connections: The Algorithmic Thing

57:22Different Kinds of Intelligences: The Steering Subsystem and Alignment

Different Kinds of Intelligences: The Steering Subsystem and Alignment

Part 4: Neuroscience Research and AI Timelines

1:00:02Algorithmic Conceptual Vocabulary: AI and Neuroscience

Algorithmic Conceptual Vocabulary: AI and Neuroscience

1:03:54Perfect Representation and the Learning Algorithm

Perfect Representation and the Learning Algorithm

1:06:11The Connectome as a Constraint: Refining Ideas from AI

The Connectome as a Constraint: Refining Ideas from AI

1:08:10Timelines and Relevance: The AI 2027 Scenario

Timelines and Relevance: The AI 2027 Scenario

1:11:32Funding and Practicality: Connectomes and Steering Subsystems

Funding and Practicality: Connectomes and Steering Subsystems

1:14:42Lessons from the Human Genome Project: Technology Development

Lessons from the Human Genome Project: Technology Development

1:16:09Funding and AI Labs: The Energy Efficiency of the Brain

Funding and AI Labs: The Energy Efficiency of the Brain

1:17:32Speculation and Neuroscience: The Grand Scheme of Things

Speculation and Neuroscience: The Grand Scheme of Things

Part 5: Formal Logic, Math, and Future Capabilities

1:19:12Behavior Cloning: Brain Data Augmented Training

Behavior Cloning: Brain Data Augmented Training

1:23:27Lean and Formal Math: RLVR and Proofs

Lean and Formal Math: RLVR and Proofs

1:25:58Conjectures, Creativity, and Provably Secure Software

Conjectures, Creativity, and Provably Secure Software

1:29:33Provable Programming Languages and the Specification Problem

Provable Programming Languages and the Specification Problem

1:31:21Usefulness and Satisfaction: New Ways of Thinking

Usefulness and Satisfaction: New Ways of Thinking

1:34:03Automated Cleverness and Provable Statements

Automated Cleverness and Provable Statements

Part 6: World Models and Scientific Infrastructure

1:37:02Symbolic Representations and World Models

Symbolic Representations and World Models

1:39:56Binding and Continual Learning

Binding and Continual Learning

1:42:46Fast Weights and Data Collection

Fast Weights and Data Collection

1:44:54The GapMap and Scalable Infrastructure

The GapMap and Scalable Infrastructure