GPT, Use your Tools!

GPT, Use your Tools!

“The right tool makes the job easy, son.” -My dad, Shlomo Avni

I had been joining my dad for the odd job around the house (not as much as I’d like anymore), and this phrase comes up again and again.

“Use your tools!” -Emily Bache

These both stuck with me since I heard them.

I’ve been fortunate to work with Emily and to hear her tell this to us developers many a time. And she’s right. Using tools and becoming proficient at it offloads effort, freeing up resources for better things. And yes, many tools have been tuned for a narrow use case, and are just ‘perfect’ for the job.

Knowhow Loss

Some tools might worry you.

“If I rely on this tool, I will lose the knowhow and skill required to get the job done without it.”

A valid concern. We become dependent on our tools.

An examle that comes to mind is navigation.

Learning is Required

Using a tool is a skill in and of itself, we have to learn it.

GPT

Large Language Models (LLMs) are much like brains.

  • Context / focus
  • Short term memory
  • Knowledge with far from perfect retrieval capabilities
  • They’re slow, when compared to other software
  • Can learn

LLM’s primary tool, you might say, is word generation. It is the main mechanism by which they interact with the world.

Word generation on its own, goes a long way, but we get to experience its limits very often.

Example: ‘1283645 * 1763520931313 = ?’

A simple math problem. However, simply generating words is obviously not the way to go here. LLMs are not endowed with built-in calculators, they’re just neural nets.

See one experiment with ChatGPT here.

Good tools are reliable, predictable controllable tend to be specialized GPT is much like a brain it’s very shares same characteristics with the human brain it can learn unpredictable, stochastic generative makes things up perceptive I found (and many others) that some practices are useful for better cognitive performance.

How to Provide Tools to GPT?