3 Underappreciated LLM Techniques that every builder should Know
There are thousands of pages of research/hours of content published on LLMs- both surveying existing capabilities and testing new ones that can drive up the performance of LLM-based systems. However, not all of them are created equal.
After an ungodly amount of research (no cap, I dreamt about Random Forests the other day)- involving both reading a lot of papers/listening to talks and also talking to a lot of people in the industry- I’ve identified 3 powerful LLM techniques that I fit the following criteria:
Immediately Practical: Nothing super out there (like Toaist AI).
General: Nothing niche like Jailbreaking, which isn’t super practical to most people.
Underappreciated: Ideas that don’t get as much appreciation as they deserve. For example, here is a list of ideas that are not included: Agentic AI, Reasoning Models, Synthetic Data, Multi-Modality, Evolution Based Exploration, Sparsity, Mixure of Experts, or Reasoning in Latent Space. These are all ideas I talk about extensively (and will likely continue to discuss in the future).
In this article, you will learn about techniques that will-
Help you improve the robustness of your system.
Reduce both AI training and inference costs.
Improve the Quality of Your Generations.
Help you Detect Hallucinations.
and more.
Since this article is paywalled, I want to clarify the intended audience before we go deep. This article is specifically directed towards two personas-
The technical builders who build (or want to build) LLM-enabled systems- whether AI Engineers, Researchers, or people in other fields who are actively building on top of LLMs (and not just using LLMs directly). Applying these techniques will require a lower intermediate level of technical comfort with AI Engineering.
Leaders with builders working for them.
These are the two personas that will gain the maximum benefit from learning about these techniques.
Assuming this is well understood, let’s cover these techniques.
For access to this article and all future articles, get a premium subscription below. Your premium subscription allows me to spend more time improving the quality of both my research and my presentation. If you believe that high-quality work deserves compensation and if my work meets your standards, please support our chocolate milk cult-
Many companies have a learning budget that you can expense this newsletter to. You can use the following for an email template to request reimbursement for your subscription. You can also pay what you want here.
I provide various consulting and advisory services. If you‘d like to explore how we can work together, reach out to me through any of my socials over here or reply to this email.