Logical Intelligence: A New Path to AGI Beyond Large Language Models

In the race toward artificial general intelligence (AGI), the dominant narrative has centered on large language models (LLMs) like those powering ChatGPT and other generative AI systems. Major technology companies have poured hundreds of billions of dollars into developing increasingly sophisticated LLMs, creating what some experts call a "groupthink problem" in Silicon Valley. However, San Francisco-based startup Logical Intelligence is charting a different course, developing what it claims is a more efficient and reliable approach to AI that could represent a significant step toward true artificial general intelligence.

The Limitations of Large Language Models

Yann LeCun, the prominent AI researcher who recently left Meta, has been vocal about what he sees as fundamental limitations in the LLM approach to AGI. In a recent interview, LeCun declared that the AI community has become "LLM-pilled," suggesting an over-reliance on language models that may not represent the most promising path to human-level intelligence. According to LeCun, LLMs essentially function as "big guessing games" that predict the most likely next word in a sequence based on patterns learned from vast amounts of internet data.

This approach, while impressive in generating human-like text, has significant drawbacks. LLMs require enormous computational resources, are prone to "hallucinations" (generating plausible but incorrect information), and struggle with tasks requiring genuine reasoning rather than pattern recognition. As Eve Bodnia, founder and CEO of Logical Intelligence, explains in an interview with WIRED, "LLMs are not allowed to deviate until they complete a task," making them inflexible in dynamic problem-solving scenarios.

Energy-Based Reasoning Models: A Different Approach

Logical Intelligence has developed what's known as an energy-based reasoning model (EBM), building on a theory conceived by LeCun two decades ago. Unlike LLMs that predict sequences, EBMs absorb a set of parameters—such as the rules to a puzzle or the constraints of a system—and complete tasks within those confines. This method is designed to eliminate mistakes and require far less computational power because there's less trial and error involved in the problem-solving process.

The startup's debut model, Kona 1.0, demonstrates the potential of this approach. According to Bodnia, Kona can solve sudoku puzzles many times faster than the world's leading LLMs, despite running on just a single Nvidia H100 GPU. In these tests, the LLMs are blocked from using coding capabilities that would allow them to "brute force" the puzzle, creating a more meaningful comparison of reasoning capabilities. Logical Intelligence claims to be the first company to have built a working EBM, transforming what was previously "just a flight of academic fancy" into practical technology.

Practical Applications and Business Potential

Logical Intelligence envisions Kona addressing thorny problems in industries with no tolerance for error. Bodnia expresses particular interest in the energy sector, where real-time processing of numerous variables is essential for efficient energy distribution. "Right now, it just dumps a chunk of energy, some of which is used and some of which is not," she explains. "People manage it, but we can automate it."

The company is also exploring applications in pharmacology for drug discovery and cancer research, as well as partnerships with major chip manufacturers and data center operators. Unlike the one-size-fits-all approach of massive LLMs, Logical Intelligence creates smaller, specialized models for each business client, with Kona 1.0 containing fewer than 200 million parameters—a tiny fraction of the parameter counts in leading LLMs.

The Road to Artificial General Intelligence

Bodnia contends that the path to AGI begins with layering different types of AI systems. In this vision, LLMs would handle natural language interfaces with humans, EBMs would take up reasoning tasks, and "world models"—like those being developed by LeCun's Paris-based startup AMI Labs—would help robots navigate and take action in three-dimensional space. Logical Intelligence expects to work closely with AMI Labs, creating complementary technologies that together might advance the field toward true artificial general intelligence.

For now, Logical Intelligence has opted not to open-source its model, with Bodnia expressing a desire to be a "responsible parent" to technology she sees as a "real step toward AGI." The company is currently seeking funding to scale its model, explore different use cases, and educate the industry about alternatives to text-based AI. As Bodnia notes, "People say, 'Oh, we're in an AI bubble.' But we're not. We're in an LLM bubble."

The emergence of Logical Intelligence represents an important counter-narrative in the AI development landscape. While LLMs have captured public imagination and corporate investment, alternative approaches like energy-based reasoning models may offer more efficient, reliable, and ultimately more intelligent systems for specific applications. As the company continues to develop its technology and seek partnerships, it could help diversify the technological pathways being explored in the quest for artificial general intelligence.