newsfilter.io
Interview, Fireside Chat, Roundtable

Why Scale Will Not Solve AGI | Vishal Misra - The a16z Show

  • Future AGI may be identified if a language model trained solely on pre-1916 or 1911 physics data can derive the theory of relativity.
  • Current scaling strategies are expected to hit limits, necessitating architectural shifts toward plasticity, continual learning, and causal modeling to achieve human-like simulation.
  • Within six months, cloud-based or Gemini-class models are predicted to handle well-defined coding tasks without human intervention.
  • While LLMs demonstrate high efficiency in Shannon information tasks, they are anticipated to stall when encountering new manifolds requiring fresh representations, likely demanding human intervention to generate these structures.
  • Experimental validation using 150,000 training steps on TokenProbe infrastructure reportedly achieved accuracy of $10^{-3}$ bits, with results reproduced by independent parties.
  • Mamba architectures are expected to outperform LSTMs in Bayesian updating tasks, whereas MLPs are forecast to fail completely in these specific environments.
  • Future systems must address data gravity by learning to ignore significant portions of previous data to form new representations, a capability current frozen models lack.
  • Integration of Judea Pearl's causal hierarchy—covering association, intervention, and counterfactuals—is projected to be essential for evolving from correlation-based to simulation-based capabilities.
  • Real-time weight updates carry a risk of catastrophic forgetting and the creation of random, chaotic models if proper plasticity mechanisms are not successfully implemented.
  • Bridging the gap between the lifelong plasticity of human brains and the static nature of post-training LLMs is viewed as a fundamental requirement for AGI development.
  • Progress involves creating mechanisms for generating new universal representations rather than merely mapping within existing bounded training manifolds.
  • Ongoing research trends, including recent Google papers, indicate a shift toward teaching LLMs vision learning through reinforcement learning from human feedback.
  • The TokenProbe tool remains active and is utilized in educational settings to facilitate student understanding of probability distributions.