chatgpt is so slow a tortoise walks past it

7 Principles for Reducing Latency in LLM Applications

When I first started developing with LLMs the logical way of providing your own data to an LLM was to add it to the prompt.  Imagine my suprise when the LLM got slower and slower.  Of course I blamed GPT4! 🙂

In the rapidly changing world of AI, large language models (LLMs) like GPT-4 is the default choice in various applications where chat is needed. However, as the power of models grow, so does the potential for latency issues, which can significantly impact the time requests take. Fortunately, there are several principles you can apply to optimize latency and ensure your LLM-powered applications run smoothly.

Ilan Bigio an engineer at OpenAI has written about 7 principles improving latency in your LLM applications:

  1. Process tokens faster: Explore techniques like model size optimization, prompt engineering, and fine-tuning to increase the rate at which the LLM processes tokens.
  2. Generate fewer tokens: Reduce output size by teaching the model to be more concise or minimizing output syntax for structured outputs.
  3. Use fewer input tokens: Techniques like fine-tuning, context filtering, and maximizing shared prompt prefixes can help reduce input tokens, albeit with a relatively minor impact on latency.
  4. Make fewer requests: Combine multiple steps into a single prompt to avoid the round-trip latency of multiple requests.
  5. Parallelize: Leverage parallel processing and speculative execution for non-sequential steps to reduce overall latency.
  6. Make your users wait less: Implement techniques like streaming, chunking, and progress indicators to create a better perceived experience for users.
  7. Don’t default to an LLM: Consider alternative approaches like hard-coding, pre-computing, leveraging UI components, and traditional optimization techniques when appropriate.

These principles are derived from real-world experiences working with a wide range of customers and developers on production LLM applications. By applying them thoughtfully, you can create efficient, responsive, and better AI-powered experiences.

If you build these kind of applications and want to apply these principles, please read the original article here:


Leave a Reply

Your email address will not be published. Required fields are marked *