Hammers and Nails

I think I would post more if I wrote shorter posts, so I’ll try writing a short post.

Hammers are nice tools. You can fuse two pieces of wood together with a hammer and some nails. You could build an entire house with a hammer and some nails. And I’m sure we could, if we tried, rebuild most of our infrastructure with a hammer and some nails.

When new technological advancements arise, there’s a tendency in business to treat the new technology as a hammer and every prior difficult problem as a nail. This is especially true in AI.

In the 80s, expert systems were all the rage due to their success in limited environments. Unfortunately, expert systems are costly to maintain and scale. Eventually, they lost that “new tech sheen”, and we stopped treating them like hammers.

The corollary is also true: expert systems still have their place in modern AI, but since they’re no longer “new”, their utility is overlooked in favor of the more “chrome-y” options. Statistical machines have largely taken the place of the expert systems of past.

• How can we get language models to perform perfect math?
• How can we get language models to perfectly translate code?
• How can we get language models to reason?

Like the old hammer, you’ll find no shortage of projects forcing a statistical solution to structured problems. Like the old hammer, these solutions are costly to maintain and scale.

Could you just use a hammer for every problem? Of course. Should you?

We’re still in the honeymoon phase with our new hammer. It still has that new tech sheen – that new hammer smell. It’s just, we could do so much more if we put down the hammer and picked up a chainsaw every once in a while.