Interview
Dileep George: Brain-Inspired AI | Lex Fridman Podcast #115
- Emerging technologies may defy century-scale predictions of infeasibility, with a non-zero probability that AI systems a hundred years from now will view current human brain science as flawed.
- Convolutional neural networks and static camera architectures are expected to become obsolete in favor of mobile robotic systems, as biological brains do not implement convolution or strict translation invariance.
- Future models like GPT-4, GPT-5, and GPT-10 are predicted to appear more impressive through scaling but will fail to achieve AGI or natural language understanding without architectural shifts enabling dynamic inference and counterfactual reasoning.
- Current deep learning approaches are viewed as insufficient for general intelligence because they slavishly mimic biology without constructing world models or running internal simulations.
- Brain-inspired AI is anticipated to be more valuable for engineering intelligence than theories from pure mathematics or physics, as the human brain serves as an existence proof of intelligence, though strict biological plausibility in learning algorithms should be relaxed to ensure functional convergence.
- Language is expected to evolve into a mechanism for controlling grounded simulations within a perceptual and motor system, a capability distinct from current text-corpus-based compression models.
- Epistemic memory is planned to be implemented as an indexing system over statistical or causal structural models in the hippocampus, allowing the cortex to reinstantiate past experiences.
- Vicarious operations involving the internalization of external actions are theorized as a pathway to consciousness when the modeling apparatus is applied to itself.
- Graph neural networks will increasingly merge neural network and graphical model properties to support more structure and dynamic inference.
- Reinforcement learning and systems like AlphaZero are expected to continue surprising humans by discovering reasonable local minima within nonlinear spaces.
- Brain-Computer Interfaces (BCIs) are predicted to succeed through bidirectional adaptation between brain and machine rather than perfect pre-existing protocols, though surgery and long-term health repercussions remain significant barriers.
- Directly connecting the brain to the internet or social media is forecasted to cause intense hallucinations due to neuroplasticity adjusting to new, overwhelming inputs.
- Ongoing research proposes a fully functional model where cortical columns encode concepts as binary random variables and connections encode relationships between these variables.
- Mortality and existential finiteness are posited as monumental properties of intelligence; therefore, copyable AI systems do not yet constitute AGI because they lack the urgency created by the finiteness of existence.
- The speaker plans to develop a system where perception, cognition, and language operate iteratively, noting that perception will only be fully resolved when integrated with higher-level cognition.