From Magic Boxes to Real Solutions: Why Not All AI is Created Equal

Aron Laszka, PhD (Co-Founder)

Grayson McClain, PhD (CRO)

Overview

Artificial intelligence startups seem to be everywhere, promising to revolutionize everything from writing e-mails and planning vacations to operating fleets of vehicles. You have likely wondered how there can suddenly be so many AI experts at so many startups.

As advisors for Mobius AI, we are often asked a more specific question, “How is your company different?” This article offers answers to these questions to help you separate hype from reality.

The AI Boom: A New Foundation

Today’s AI boom is overwhelmingly driven by a specific type of artificial neural network known as Large Language Models (LLMs). These LLMs are massive, general-purpose models trained on vast amounts of data from diverse sources, which can be applied to a wide range of tasks.

While AI has been a field of study for decades, the current boom was ignited by a few breakthroughs from the past 10 years. A pivotal 2017 paper from Google, “Attention Is All You Need,” introduced the transformer architecture, which is the engine behind nearly all modern LLMs. In 2018, OpenAI's GPT-1 demonstrated the power of generative pre-training. But it was the 2022 release of ChatGPT, based on GPT-3.5, that shifted public interest into high gear. It amazed users because of its breadth of knowledge, its conversational tone, and its self-confidence. 

Today, we are in the midst of an intense innovation cycle, with technology companies competing to build ever-larger and more capable models—from OpenAI's GPT series to Google's Gemini and Anthropic's Claude—at enormous development costs.

The breadth of knowledge spanned by these LLMs requires continuous spending to build, train, host, and regularly retrain the foundational models.

Proliferation of AI Applications

So, how does this lead to a sudden “expert” in every startup? The answer is that these powerful, revolutionary LLMs are increasingly becoming commodities. To drive commercial usage, the LLM owners provide Application Programming Interfaces (APIs) to outside developers. These APIs allow almost any developer to integrate LLMs without the investment in understanding of how the underlying neural networks actually work.

For many new startups, the LLM is treated as a magic black box. The “AI” part of their products consists of engineering clever prompts, which are the instructions given to the LLM, or connecting the LLM to external data through a process called Retrieval-Augmented Generation. While useful, these are often thin veneers on top of a commoditized, general-purpose LLM technology.

The growth of AI applications and agents that we see today is due to the coupling of veneers and LLMs to extract a small portion of the LLM’s value. This black box approach has serious limitations, especially for critical industries like transportation.

Three Limitations of One-Size-Fits-All AI

First, these applications are expensive. LLMs require massive hardware investments upfront. Accessing an LLM through a cloud provider incurs significant ongoing costs. One useful analogy: if a photographer only needed a digital camera, few would still choose to take photos with a mobile phone with all its battery-draining distractions and high monthly service charges. Instead, many professional photographers spend their money on a purpose-built high-quality camera.

Second, general-purpose LLMs are not explainable or interpretable. Users rarely know how the responses were generated, and the LLMs provide no guarantees for their outputs. As a result, these applications can behave erratically or unpredictably when given inputs that are different from their training data. This is a major risk for systems that manage real-world logistics. An additional risk for fleets in highly regulated industries is the need to audit why certain operational decisions were made.

Finally, the general-purpose LLMs can also be relatively slow, which may be a non-starter for real-time operational planning. Some consumers may accept that their browser’s search engine slows its response rate from 0.1 seconds to 2 seconds when it incorporates AI. But very few fleet operators would agree for their operational software to increase its processing time by that same 20x factor. 

Mobius AI: Purpose-Built AI Solutions

Some AI startups address these limitations by fine-tuning a generic LLM for a specific task or by “distilling” it into a smaller model. While this can improve the system’s speed and can partially address the cost, it does not robustly address all three of the limitations above.

This is where Mobius AI's approach is fundamentally different. We do not just use AI; we create purpose-built, neurosymbolic AI solutions that are designed from the ground up to optimize transportation. Neurosymbolic means that we combine the best of both AI worlds: the pattern-recognition and generalization power of neural networks (the “neuro” part) with the logical, structured reasoning of classical algorithms (the “symbolic” part).

Our team is not just jumping on the AI bandwagon. Mobius AI's researchers and advisors are leading experts with more than a decade of experience in both core AI research and transportation optimization, including deep learning and traditional Operations Research (OR) methods. This expertise is backed by peer-reviewed academic publications and successful research, development, and demonstration projects applying AI to optimize transportation services.

Instead of wrapping a generic black box, our solutions combine the advantages of traditional OR techniques (like search algorithms for planning) with highly specialized deep neural networks that we build to perform specific tasks (like estimating objective values). These approaches build directly on our team's extensive experience in these domains. This neurosymbolic method means that our solutions are made of components that we purpose-built to be the most efficient and reliable for their specific job. This allows Mobius AI to deliver solutions that have the trustworthiness of traditional, field-tested techniques combined with the flexibility and power of state-of-the-art AI.

Beyond the Hype

Not all AI is right for your business. While GPT-3.5 amazed consumers with its breadth of knowledge and conversational tone, most fleet owners are asking their transportation management system to do a smaller set of tasks with crispness and reliability.

While the sheer number of AI startups grows, Mobius AI focuses on building a different product: deep, explainable, and trustworthy solutions engineered for the specific, complex challenges of the transportation industry.