Improving LLMs Isn't About Making Them More Powerful
Like the steam powered engines of the industrial age supercharging human muscle, LLMs are now doing the same with the human brain. They are powerful, and getting more so through self-learning via ML algorithms and human validation. However in these early days of open access LLMs are vulnerable to misuse, and that should be the primary focus of LLM development and governance, not making them run faster or even improving accuracy.
TECHNOLOGY
Security, interpretability, and ethical use are the next frontier of LLM improvement.
LLMs aren't that different than humans.
Like us, they learn through training.
Like us, they pick the right tools to execute a task.
Like us, they don't reproduce things exactly (lossy compression) but they can recall, using it to summarize and make conclusions.
Like us, they sometimes get things wrong before they get it right.
And that's where they are both like and unlike us.
We generally (not always!) know what we do and don't know. Humans (especially wiser ones) often express doubt, acknowledging the limits of their knowledge. Do LLMs know what they don't know, or when less than fully confident in their knowledge do they fairly represent it in their answers? Or do they simply make stuff up? And when they do, how would we know?
With humans, you can look at their credentials and weigh their opinions accordingly. If not satisfied, we go seek a 2nd opinion.
Can we do that with LLMs? Or do they share common enough training data where they may hallucinate about the same things?
This video is terrific intro to LLMs by Andrej Karpathy.
There are problems with LLMs some which he explains in the video:
- identifying what's true and what's made up (hallucinations)
- identifying what is human- and what is AI-generated (deepfakes)
- understanding how conclusions are made (interpretability)
- preventing jailbreaking and malicious use
So 'improving' a model is fundamentally about equipping it with the right self-governance and security mechanisms where it serves for good. Not about making it bigger or faster.