Limits of LLMs: Pattern vs. Reasoning

Eve argues LLMs mimic language patterns and lack true extrapolative reasoning, motivating a new architecture.

Play episode from 30:59

chevron_right

Transcript

chevron_right

Transcript

Episode notes

Eve Bodnia is the co-founder and CEO of Logical Intelligence, which is developing energy-based reasoning models (EBMs) as an alternative to large language models. She argues that LLMs, which operate by recognizing and recombining patterns within language space, are structurally incapable of genuine reasoning. Eve's alternative: Kona — an EBM that reasons in abstract latent space, learns rules about the world rather than surface patterns, and can interface with language models as one output channel among many. Eve traces the core ideas behind her architecture to decades of work in symmetry groups, condensed matter physics, and brain science — fields that share, as she explains, the same underlying mathematics. In a public demo, Kona solved a complex reasoning task for roughly $4 in compute, compared to an estimated $15,000 using frontier LLMs. With Yann LeCun serving as founding chair of its technical board, Logical Intelligence sits at the center of a small but growing effort to rethink AI beyond language-based models.

In our conversation, we explore:

Why Eve believes LLMs can’t truly extrapolate knowledge, even at larger scale
What energy-based reasoning models are—and where the “energy” concept comes from
The $4 vs. $15,000 benchmark, and what it tells us about the cost of guessing vs. knowing
How Logical Intelligence showed spontaneous knowledge transfer at just 16M parameters
Why systems like chip design, surgical robotics, and power grids need more than probabilistic AI
What formally verified code generation means for the future of programming
Why the math behind particle physics also explains how the brain filters signal from noise
How meeting Grigori Perelman as a teenager shaped Eve’s views on ego and ownership in science
Why Eve believes humans must remain the constraint-setters in advanced AI
How meditation, piano, and Eastern philosophy support her creative process

—

Thank you to the partners who make this possible

Granola: The app that might actually make you love meetings.

Persona: Trusted identity verification for any use case.

—

Transcript: https://www.generalist.com/p/everyone-is-betting-on-bigger-llms

—

Timestamps

(00:00) Introduction

(03:03) Eve’s encounter with Grigori Perelman

(05:38) Why bizarre people are Eve’s favorite people

(06:56) Her early obsession with math and physics