Fireside Chat, Interview
Will LLMs Get Us To AGI?
- Current Large Language Models (LLMs) are predicted to plateau with incremental improvements similar to the smartphone industry's last seven to nine years, lacking the capacity to generate new science, new results, or new math such as theories of relativity, quantum mechanics, or Gödel's incompleteness theorem because they cannot reject existing axioms or training data paradigms.
- Scaling data and compute is expected to fail in evolving a new manifold, instead producing only smoother versions of the existing data distribution, meaning capabilities across major companies and open-source models will not fundamentally cross into a different realm despite becoming more efficient at existing tasks.
- A new architectural leap is identified as necessary to transition from current LLMs to Artificial General Intelligence (AGI), as simply solving problems like those in the International Math Olympiad involves connecting known results rather than inventing new axioms.
- Future development plans include building and training models that perform inference by following low or minimum entropy paths, with the "Token Probe" software expected to be useful for testing how confidence rises during in-context learning.
- Risks and limitations include the expectation that LLMs will likely never create new science even if they achieve high levels of autonomy, and that generating output falling off the existing distribution to learn new things would only occur if the system breaks from the current data distribution entirely.
- While current LLMs are not considered the answer to AGI, they are projected to remain fantastic tools that significantly increase productivity by becoming better at executing existing tasks without fundamentally changing the nature of their capabilities.