The AI Model Built for What LLMs Can't Do

Logical Intelligence aims to ensure correctness in mission-critical AI applications.

Eve Bodnia, founder and CEO of Logical Intelligence, introduces her company as a foundational AI enterprise. They are developing AI for mission-critical systems, specifically focusing on software and hardware correctness, including areas like code generation and chip design.

Bodnia highlights a critical flaw with current large language models (LLMs) in these high-stakes applications: their lack of inherent correctness and verifiability. She argues against deploying non-deterministic LLMs in systems where accuracy and reliability are paramount, emphasizing the market's need for deterministic and verifiable AI.

To illustrate the risks, Bodnia uses an analogy of an AI-driven car or plane that might hallucinate 20% of the time, leading to potentially dangerous outcomes. She underscores that while humans have managed without AI in many critical systems, the ongoing evolution necessitates integrating reliable AI everywhere, from banking to transportation, requiring systems that save time and prevent errors rather than introduce them.

people use LLMs today. but very few actually questioning of like how this results, actually, like, correct. Does it make sense what it produce? And it seems like there's a big gap on market today, having deterministic AI, verifiable AI. So we're trying to fill that gap.

LLMs Lack Internal Verification, While EBMs Offer Transparent Double Checking

Large Language Models (LLMs) operate as black boxes, lacking internal verification mechanisms. Their language-based architecture fundamentally involves a 'guessing game,' which means there's no inherent way to check their output internally. This absence of transparency makes achieving certainty with LLMs challenging.

To compensate, LLMs often rely on expensive external verifiers, such as proof languages like Lean 4 for mathematical checks. However, even with fine-tuning and external checks, the underlying token-based architecture remains costly. The expense comes from the compute required for the LLM's initial 'guessing game,' which external verification doesn't solve.

Energy-Based Models (EBMs), in contrast, are token-free models that do not engage in a guessing game. Their architecture is designed for self-alignment, allowing real-time inspectability. This means users can observe the model's internal processing at any point, making it transparent rather than a black box.

This transparency provides a powerful internal verification capability. When combined with external verifiers, EBMs offer a double-sided verification system—both inside and outside the model. This makes EBMs inherently more reliable and cost-effective for tasks that demand high certainty, directly addressing the limitations found in LLMs.

For verification tasks, EBM architecture allows self-alignment and absence of tokens makes it cheap, plus you have external verifiers, so you have verification on both sides, inside and outside.

Energy-Based Models Leverage Physics' Energy Minimization Principle

Energy-Based Models (EBMs) are an AI architecture fundamentally rooted in the physics principle of energy minimization. The term itself, 'energy-based model,' simply refers to systems that apply this core concept.

In theoretical physics, minimizing energy is a foundational approach. Scientists construct Lagrangians to represent a system's kinetic and potential energies, then minimize these terms to derive equations of motion and discover conservation laws. This principle posits that all systems inherently seek their lowest possible energy state.

EBMs apply this same universal principle to artificial intelligence. Instead of being an AI-specific concept, energy minimization is a broader idea where a system processes information by finding states that minimize an associated 'energy' function. This allows the AI to discover underlying laws or optimal solutions within its operational framework.

For example, just as humans sit down to minimize energy, EBMs utilize this natural tendency to process information efficiently. The core idea is to define an energy function for the AI's system and then derive its operational 'laws' by minimizing that energy.

So we're just using this minimization energy principle as AI is processing information in high level terms.

Energy-based models use energy landscapes to identify probable states.

Energy-based models (EBMs) interpret observed data by conceptualizing it within an "energy landscape." This landscape functions like a topographical map, where different configurations or states of data are assigned varying energy levels. Less probable or "uncomfortable" scenarios correspond to high points of energy, while more probable and "comfortable" states reside in the low-energy valleys.

For example, if you're tired and come home, your most probable state is relaxing on the couch, representing a low-energy configuration. Conversely, dancing around the house would be a high-energy, less probable state. The model naturally seeks to minimize this energy, guiding it towards the most likely and stable outcomes.

This energy landscape is not predefined but is learned through observation. By repeatedly seeing real-world behaviors—such as a person's activities under different levels of tiredness or workload—the EBM refines its landscape. Over time, the model trains itself so that the consistently observed, most probable states, like relaxing on a couch when tired, become the lowest energy points on its map.

Essentially, all of this picture can be mapped into something we call energy landscape, when... The highest points we can associate less probable scenarios.

The Limitation of Language-Dependent Intelligence in LLMs

Energy-based models (EBMs) operate by mapping raw data directly onto an energy landscape, which is then navigated using algorithms. A crucial distinction is that EBMs do not rely on tokens or predict sequences of words, setting them apart from large language models (LLMs).

A fundamental critique of LLMs is their reliance on language for intelligence. Human thought processes, by contrast, are often abstract and can occur independently of any specific language. The speaker argues that linking intelligence to the prediction of next tokens based on language can lead to different reasoning processes across various languages, which feels inherently flawed.

Consider tasks like navigating a house, which are inherently visual and spatial, not linguistic. LLMs attempt to process such non-linguistic information by first mapping it into a language space, translating visual data into words and embeddings, and then associating tokens with probabilities. This approach is seen as problematic because it forces non-linguistic realities into a linguistic framework.

The widespread popularity of LLMs has led to a perception that language is the universal tool for interacting with the world. However, many real-world actions, such as driving a car or moving around a house, rely on visual data and body states rather than predicting words. Over-relying on language-dependent AI overlooks these crucial non-linguistic forms of intelligence.

what usually bothers me about LLMs, it's intelligence which is language dependent.

Energy-Based Models Offer an Efficient Alternative to LLMs for Non-Language Tasks

Large Language Models (LLMs) operate by modeling intelligence as sequences of tokens, relying on weak correlations to predict the next element. While LLMs can be creatively adapted to model non-language tasks, such as tracking human movement or performing image recognition by treating these as token streams, this approach is not inherently the most efficient.

The token-based "guessing game" employed by LLMs for non-language tasks incurs significant computational expense and leads to slow processing. For applications demanding rapid responses, such as real-time control of an electrical circuit where reactions are needed in milliseconds or microseconds, LLMs are fundamentally unsuitable due to their inherent latency.

Energy-Based Models (EBMs) provide a more resource-efficient architecture specifically designed for non-language related tasks like spatial reasoning and applied engineering. Unlike LLMs, EBMs do not rely on token prediction, making them faster and less costly. Furthermore, EBMs excel at working with sparse data, making them particularly advantageous in scenarios where comprehensive training datasets are limited.

If you want to minimize your resources and you don't have opportunity to wait, like for example, if your AI controls the circuit, you probably cannot wait even a second, it's all milliseconds, microseconds, so this form of AI is not suitable for those tasks.

Energy-Based Models Use Latent Variables for Deep Understanding

Energy-Based Models (EBMs) go beyond simply recognizing patterns in data; they aim to 'understand' the data by learning fundamental rules about the world. Unlike Large Language Models (LLMs) that primarily predict the most probable scenarios based on vast amounts of input, EBMs strive to grasp the underlying reasons for data's appearance.

This deep understanding translates into knowledge about basic rules, such as inferring that a couch behind someone is likely for sitting or aesthetic purposes. EBMs are designed to identify and store these kinds of relational rules from various data points, building a conceptual model similar to how a brain constructs a mental world model to comprehend its environment.

Latent variables within EBMs are the storage mechanism for this acquired knowledge and these rules. Instead of explicit key-value pairs, this knowledge is stored in the form of an energy landscape, which the model can navigate. This allows the EBM to infer what to do with new, unseen data based on its established understanding of the rules.

This capability makes EBMs particularly powerful for data analysis, especially with numerical and relational data. By directly working with the data and extracting its inherent rules, EBMs overcome the limitations of language-based models, which can lose significant information when trying to map complex data relationships to natural language.

It's not equivalent to a role, but it's equivalent to something which holds the knowledge about the roles of your data. It's like a knowledge storage.

Energy-Based Models Offer a Bird's-Eye View Compared to Large Language Models' Tunnel Vision

Energy-Based Models (EBMs) avoid the issues of tokenization common in Large Language Models (LLMs) by mapping data directly into a different data structure. This makes EBMs non-autoregressive, meaning they don't process information in sequences of tokens like LLMs do, which is a fundamental difference in how they operate.

Imagine trying to navigate a map, like one of San Francisco. An LLM's approach is likened to having "tunnel vision"; it can only choose one direction at a time, without seeing the broader map or other potential routes. If it chooses to walk in a certain direction, it's committed, and turning back or reassessing is not part of its operational model.

This linear, step-by-step decision-making makes LLMs prone to taking wrong turns or even hallucinating paths, leading them into computational 'holes' from which they cannot easily recover or backtrack. An LLM might keep searching and spending compute without reaching a correct destination because it lacks a holistic view of the problem space.

In contrast, EBMs operate with a "bird's-eye view," allowing them to see all possible routes and assess the entire landscape at once. This capability means EBMs can proactively identify and avoid pitfalls like 'holes' or dead ends, making them inherently more robust and efficient in problem-solving compared to the constrained, step-by-step nature of LLMs.

You cannot turn back because you're auto-regressive LLM. You have to go into the hole.

Solving 'Vibe Coding' with Formally Verified AI and Natural Language Programming

Vibe Coding describes software solutions that are locally correct but form a patchwork of hotfixes, lacking a unified design. This happens when systems struggle to integrate context into a cohesive solution, often getting distracted by immediate problems rather than maintaining a consistent concept throughout.

Engineered Bayesian Machines (EBMs) address Vibe Coding by aiming to automate the entire coding process through formally verified code generation. This shifts programming from specific languages like C++ or Python to natural language, allowing developers to "code" in plain English.

EBMs incorporate both internal and external verifiers to ensure logical compatibility. These verifiers check if new code logic is consistent with existing codebase and 'vibe code specifications,' providing mathematical proofs and certificates. This machine-verifiable process identifies incompatible code and suggests fixes in natural language, reducing the burden on human engineers.

For mission-critical applications like self-driving cars or aircraft, EBMs offer a crucial advantage over large language models (LLMs). While LLMs can hallucinate and cannot be constrained, EBMs can be forced to follow human-defined rules and constraints, ensuring the AI behaves predictably and safely, thereby preventing misbehavior where human lives are at stake.

LLM can misbehave because you cannot constrain it, it just hallucinates. And EBM can be constrained. You can come up with a set of constraints and EBM just forced to follow it.

LLM Progress Plateauing and Economic Inertia in AI Investment

The AI industry experienced a significant "aha" moment with Large Language Models (LLMs) between 2021 and 2023, leading to substantial investment. Investors believed that LLMs, initially impressive for conversation, would eventually master complex tasks like data analysis and taxes.

However, progress in LLMs is now plateauing, delivering incremental improvements rather than major breakthroughs, despite increased compute and architectural modifications. Billions of dollars have been invested in the LLM ecosystem, creating immense economic inertia. It is challenging for investors to abandon these substantial investments and pivot to radically new AI architectures.

This inertia leads investors to favor LLM-based solutions with minor changes, enabling them to leverage existing portfolio companies and past commitments. The current LLM ecosystem is a complex web of dependencies involving LLM creators, data center providers, and hardware manufacturers, forming a "giant thing which is impossible to break."

Logical Intelligence's EBM (Entity-Based Model) architecture offers a compatible solution that does not require abandoning LLMs. EBMs can function as a valuable layer, handling tasks like spatial reasoning or data analysis, which can make existing LLM operations more efficient and cost-effective without discarding previous investments. This approach aims to create a new ecosystem while integrating with the existing one.

People start believing that, okay, if it's really good at talking to me, eventually it's gonna be good at doing data analysis, my taxes, and other stuff. So all the investment communities start pouring money into LLMs.

Mission-Critical Industries Shun LLMs for Core Tasks Due to Privacy and B2B Gaps

Mission-critical sectors such as digital assets, drug discovery, and energy grid management have not integrated LLMs for their core operations. These industries rely heavily on human-driven data analysis and precise decision-making, like interpreting blood markers for drug discovery or calculating energy distribution for power grids. The current state of LLMs does not meet the complex, high-stakes requirements of these fields, which continue to use human expertise or human-controlled programs.

A major barrier to LLM adoption in businesses is the concern over data privacy. Existing big tech LLMs are primarily B2C, designed for general consumer use. Companies are unwilling to share proprietary and sensitive data with these models. There is a clear demand for custom B2B AI solutions tailored to specific tasks that can operate within private environments, a capability that current mainstream LLMs lack.

This creates a substantial void, especially in applied engineering and data analysis where verifiability is paramount. LLMs are not yet equipped to provide the level of accuracy, control, and trustworthiness required for applications where errors could have severe consequences. This gap underscores the need for AI systems that can offer reliable and auditable results.

Despite the limitations of current LLMs, there's a notable shift in the industry. Some large technology companies that initially focused on LLMs are now reportedly developing Energy-Based Models (EBMs) in-house. This internal development of EBMs by established leaders suggests an acknowledgment of the current limitations and a move towards more suitable, verifiable AI paradigms for demanding business applications.

The big tech LLMs, they're mainly like B2C... for businesses, they don't wanna share their data with them.