Working Up

by Dwayne Phillips

Large Language Models appear to have all sorts of problems. I wonder why companies that build such don’t hire adults to help build them.

I recently learned that large language models (LLMs) have “sleeper agents” in them. Given some inputs, the LLM starts doing crazy things that it shouldn’t do.

Gosh.

LLMs are pieces of simple software that trigger large matrix multiplications and pick words based on the results of multiply, add, subtract, divide, etc. LLMs don’t “act” strange.

What is happening? Here is the simple answer. The well-meaning folks that build the LLMs are not testing them sufficiently before releasing them as products. The input space, all those possible combinations of words, is huge. The builders don’t exhaust the input space and correct mistakes. “Well,” the logic goes, “We tested this thing using a million or so test cases. It’s pretty good. Let’s release it.”

Some business manager looks at the numbers, decides that enough is enough, and releases the product to recoup some cost $$$. Well, that makes business sense.

Engineering? Are you kidding? Systems engineering? Are you kidding? Testing? Are you kidding?

Do these companies hire any old engineers who understand how to build, test, and release products when they are ready? Where are the adults? Where are the people who are old enough to have seen bad things happen?

There is much value in having people who have experienced tragedy on your development team. Tragedy is no fun. It is, however, real life. It does, however, sear scars into the heart and mind. Have such people on your team. Pull them aside and talk with them. Learn what could possibly go wrong and the results. We can do better.