The rapid advances in LLMs produce new, mind-blowing demonstrations almost weekly. LLMs are the future of many software applications, from software development to writing to decision-making. It’s the closest to magic that humans have created!

The question is not where LLMs will be useful but where they will not. Where are the areas where LLMs will not be the best solution both now and in the future? The answer tells us where some of the biggest opportunities will be for startups in the coming decade, as the big companies dominate the LLM model market.

We use the following model to identify these areas, a simple system we call the LLM Opportunity Model:

Where accuracy is not essential, LLMs can do it fast and well enough. Examples include drafting documents, generating images, and entertainment. Large company (OpenAI, Google, et al.) LLMs will dominate all of these use cases.

Where accuracy is critical, but you only do the task once, a human is the best option. It will take a human longer to verify the LLM did not make any mistakes than to just do the task themselves. LLMs might still be helpful to draft or explore, but humans will have to complete the task.

An example is data analysis. LLMs can do data analysis (quite well), but they still make errors, and their work needs to be verified by a human to ensure correct results. In the case of a single analysis, the cost of finding errors is often more than the cost of just doing the work.

Where accuracy is critical, and you need to do the task frequently, it is better to have a human supervising the LLMs. The speed of LLMs outweighs the cost of errors at scale. The errors will be random and unpredictable enough to make structured checks and balances ineffective. The human is a supervisor, ensuring the LLMs don’t produce significant error rates.

A great example is psychotherapy, where the therapist must handle delicate situations correctly to ensure the patient's health. Unsupervised LLMs are too likely to produce errors that might harm patient health, but with human supervision, they can service many more patients than human therapists. Jimini Health (a Near Horizon company) is the leader in this approach and is already demonstrating how it can greatly improve mental health outcomes.

LLMS and Hallucinations

Underlying the LLM Focus Model is the realization that hallucinations (mistakes made by LLMs) will not be removed. They are a fundamental part of the auto-regressive design of LLMs, and new approaches will be required to avoid them. Those advances are not a linear extension of the work already done; they will require entirely new inventions and innovations.

As a result, LLMs are an amazing but fundamentally flawed technolgy. That flaw is where startups can challenge big tech companies because addressing flaws is often a nuanced and application-specific problem. This gap in LLM capabilities is where the next industry-defining software companies will be built.

Today, all of our companies at Near Horizon are using AI to take advantage of this opportunity. We use the LLM Focus Model to identify the types of problems to tackle, which gives us confidence that they will not be commoditized by some future LLM foundation models.