Not All Problems Are Great Fits for LLMs

by Dylan Huang on November 18, 2024

Many startups are racing to find product-market fit at the intersection of AI and various industries. Several successful use-cases have already emerged, including coding assistants (Cursor), marketing copy (Jasper), search (Perplexity), real estate (Elise), and RFPs (GovDash). While there are likely other successful LLM applications out there, these are the ones I'm familiar with off the top of my head. Through my experience building and selling LLM tools, I've discovered a new important criteria for evaluating an idea.

Are LLMs Especially Good at Solving This Problem?​

Traditional business advice emphasizes finding and solving urgent, critical problems. While this principle remains valid, not all pressing problems are well-suited for LLM solutions, given their current capabilities and limitations. As non-deterministic algorithms, LLMs cannot be tested with the same rigor as traditional software. During controlled product demos, LLMs may appear to handle use-cases flawlessly, creating an illusion of broader applicability. However, when deployed to production environments with diverse, unpredictable inputs, carefully crafted prompts often fail to maintain consistent performance.

Where LLMs Excel​

However, LLMs can excel when their non-deterministic nature doesn't matter or even provides benefits. Let's examine successful LLM use-cases where this is true.

Coding Copilots​

Think of coding assistants like Cursor that help you write code and complete your lines.

When you code, there's usually a "right way" to solve a problem. Even though there are many ways to write code, most good solutions look similar—this is what we call "low entropy", like how recipes for chocolate chip cookies tend to share common ingredients and steps. LLMs are really good at pattern matching, which is perfect for coding because writing code is all about recognizing and applying common patterns. Just like how you might see similar ways to write a login form or sort a list across different projects, LLMs have learned these patterns from seeing lots of code, making them great at suggesting the right solutions.

Copywriting Generator​

Marketing copy is more art than science, making non-deterministic LLM outputs acceptable. Since copywriting involves ideation and iteration rather than precision, it has a naturally high margin of error.

Search is unique because users don't expect perfect first results - they're used to scrolling and exploring multiple options on platforms like Google or Amazon. While search traditionally relies on complex algorithms, LLMs can enhance the experience by leveraging their ability to synthesize and summarize information within their context window. This enables a hybrid approach where traditional search algorithms surface results that LLMs can then summarize to guide users to what they're looking for.

Real Estate Assistant​

Leasing agents primarily answer questions about properties to help renters find suitable homes and sign leases. Since their core function involves retrieving and relaying property information, a real estate assistant effectively becomes a specialized search problem.

RFPs​

RFP responses combine two LLM strengths: extracting questions from lengthy, unstructured documents and searching internal knowledge bases for relevant answers. Since extracting questions from RFPs is time-consuming but straightforward, LLMs can work in parallel to identify all requirements that need addressing. This makes the RFP response process essentially a document extraction and search problem that's perfect for automation.

Conclusion​

When building an LLM startup, focus on problems with two key characteristics:

  • Low Entropy - solutions follow common patterns, as seen in coding
  • High margin of error - tasks like copywriting where art trumps science

Or can be solved in a simlar way to common well-suited problem types such as:

  • Search
  • Document extraction

Beyond traditional business evaluation, ask yourself: "Are LLMs particularly well-suited to solve this problem?" If not, reconsider unless you have unique insights into making it work.

← Back to all posts