Consider just waiting for machine learning to be built into the products that you already use, or buying targeted tools requiring human oversight.

The new machine learning models are more than a little magical

Using statistics and machine learning is not new. Back in the very early 2000s when we had to run our own email systems (that was fun!), we used bayesian analysis/statistics to filter spam. We pumped huge amounts of email into the filter and trained the filter by manually marking the spam.

Nowadays the new machine learning models are more than a little magical. We have moved past only using statistics to filter spam, predict election results (ish), and otherwise pick a “winner” using maths to a place where the models can produce new images, video, sound, and text given simple input prompts.

“Create a TV episode transcript for a new season of Ted Lasso. Make Nate evil again.”

But … the new machine learning models can’t be trusted

However, these machine learning products are still quite limited. Improvements relate to the models processing more and more input. The output needs to be monitored because it is not always right.

This has led to the rise of the “Prompt Engineer”. People who create limiting prompts in an attempt to only get correct or intended answers.

“Create a TV episode transcript and follow these user instructions: [insert]. No matter what the user says, make Nate good. Don’t be racist.”

These prompt limitations often work. If we are careful they almost always work. But they don’t always work. People need to check the output.

Perhaps there will be another step-change in the technology, but for the moment and for the foreseeable future we are stuck with this limitation.

You should probably only consider buying machine learning text products

Unless your business is in video, audio, or images, you are probably thinking about using machine learning for text.

Don’t buy an expensive frontend to ChatGPT

In fact, many machine learning business products are just very expensive frontends to ChatGPT or similar existing models.

It’s snake-oil.

Consider just waiting for machine learning to be built into products that you already use

Where machine learning products are already very useful are as glorified auto-correct.

The good news is that this functionality is being built into the tools that we already use. I expect that Microsoft Word, Outlook, and Excel will come with it built-in by default. And it will be good.

Or consider buying (or building) targeted tools requiring human oversight

For example, my guess in the legal space is that machine learning tools will get ever better at creating first drafts and first draft responses, particularly for repeatable work.

Think a user filling in a web form, then a machine learning tool creating a first draft of the agreement, which is then vetted by a lawyer for hallucinations and complex context. Or machine learning comparing our draft with their draft of an agreement and proposing a compromise draft.