Interview
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
- Time travel forward is considered possible via cryogenic freezing or relativistic spaceflight, while active backward time travel is deemed very unlikely, and immediate time travel is considered improbable despite being desirable.
- Artificial General Intelligence (AGI) is predicted to arrive in approximately 200 years, contrasting with a 20-year estimate from Demis Hassabis; machines indistinguishable from human secretaries may appear soon or later, and existential AI risks are viewed as less immediate than nuclear threats.
- AI capabilities are expected to advance by helping humans think better in the short and mid-term, building on current tools, though true understanding of scenes, language, and actions remains far beyond current low-level vision and speech recognition success.
- Future AI breakthroughs are anticipated to be heavily inspired by neuroscience, particularly regarding reinforcement learning, deep learning, and the plasticity of brain circuits, though specific mechanisms like stochastic gradient descent are unlikely to be biological.
- The structural limitations of biological connectivity, such as the preference for short-range over long-range connections, likely explain the hierarchical and local connectivity found in deep neural networks, which allow efficient approximation of compositional functions and avoid the curse of dimensionality.
- Neural networks often utilize over-parameterization, having 10 to 100 times more parameters than data, which creates a solution space rich with good minima, contrasting with the traditional statistical requirement of more data than parameters.
- Learning in the brain relies on innate machinery, such as built-in motion detection for early object segmentation and a plastic cortical area that imprints the most frequent faces within the first two weeks of life, differing from the need for massive labeled datasets in current deep learning.
- Deep convolutional networks are powerful for compositional structures due to biological constraints, whereas shallow networks face exponential unit growth requirements, and Generative Adversarial Networks (GANs) are seen as useful for graphics but less critical for achieving general intelligence.
- Ethics in machines is expected to be a learnable problem grounded in the biological basis of ethical judgment, supported by fMRI and magnetic stimulation evidence, though the role of consciousness in creating an intelligent system or passing an extended Turing test remains a subject of debate.
- The Center for Brains, Minds, and Machines focuses on visual intelligence, self-awareness, and placing oneself in the world, with progress requiring collaboration between diverse expertise due to the intertwined nature of algorithms and circuits in the brain.
- The speaker emphasizes that curiosity, fun, and a collaborative, friendly, and ambitious research environment are essential for scientific breakthroughs, noting that heated arguments about revolutionary ideas are beneficial if they remain impersonal.
- Intelligence is defined as a learnable capability rather than a hardwired set of instructions, with evolution encoding general learning machinery rather than specific circuitry, a process that could eventually expand human capabilities to create intelligence significantly surpassing historical figures like Einstein.
- The speaker expects that understanding the future design of an AGI system will be difficult, though understanding the learning process itself does not require knowing every specific discovery, similar to how parents understand child development without predicting every preference.
- The speaker anticipates that machines will eventually reach a point where they can be created without necessarily fully understanding the human brain, although the difficulty of this task remains a significant open question in the field.
- The speaker highlights that the "no free lunch" principle applies to AI, meaning machines cannot learn more than the information provided, and that the "curse of dimensionality" necessitates deep hierarchical architectures for complex tasks like image processing.
- Biological differences between brain regions, such as the cerebellum and hippocampus versus the cortex, are noted, with the cortex acting as a universal hardware substrate for different modalities like vision, language, and motor control.
- The speaker expects that happiness may be invariant to intelligence levels, referencing literary inspirations like "Flowers for Algernon" and the views of Steve Jobs on mortality, while also noting that the stock market principle of avoiding consensus applies to scientific pursuits.
- Historical and institutional context is provided, including the speaker's role at MIT, the funding of the Center for Brains, Minds, and Machines by the NSF, and the application of these principles in companies like Mobileye and systems like AlphaGo.
- The speaker notes that while rational thought does not strictly require mortality, consciousness and mortality appear linked in biological systems, and that solving the problem of human intelligence is crucial for understanding the tool used to do science.