The Well-read Parrot
Why AI can't run your shop (yet)
Regular readers will remember that a post earlier this year explored some of the ups and downs of working with AI.
I was reminded of that post this week when I was tagged in a LinkedIn post in which Andrew Thomas-Derrer of Artefact linked to a really very funny write-up of an attempt to get an AI to run a shop for a while.
The short version is that whilst Claudius (The AI persona involved) did some things well, most notably taking and adapting to requests and suggestions from his customers, he did pretty badly overall, selling things for below cost and generally not making much money.
More spectacularly, he also had a number of hallucinations during the experiment, including at one point apparently deciding he was an actual human being wearing a blue blazer and a red tie and getting a bit upset when it was pointed out that he wasn’t.
All of that is amusing enough and leads us to the usual trap with these kinds of articles - the current experience was terrible enough that I could frame this post as “AI can never replace real shop managers” only to find that a couple of technological evolutions into the future it can do exactly that.
In fact, the point I really want us to take out of this story is one we touched on at the end of the post in January - that we should be careful when deploying AI-based technologies to use the right one for the right job.
Ask most people in business leadership roles what they think of AI and pretty soon you will realise that they are only talking about one particular type - Large Language Models. In a sense, that isn’t surprising because using LLMs like ChatGPT is probably the AI tech experience most of us have had as curious consumers.
It is ironic that this is the case, though, because LLMs are a relatively new implementation of AI principles, having come about really only as the existence of massive computing power has been matched by the existence of huge amounts of written data (e.g. your Facebook posts) for it to learn from.
In fact, much older applications of AI have impacted your life as a consumer for decades longer - in particular, Machine Learning models deployed on large blocks of numerical data about customers and their purchasing habits, which have driven loyalty schemes, direct marketing campaigns and digital advertising for much longer than LLMs have even existed.
Notwithstanding that, the LLM is a powerful tool. In recent weeks I’ve used ChatGPT to:
Read and summarise a huge Powerpoint document, and pick out the key talking points.
Generate the Python code necessary to scrape and process some web data
Sort out a particularly nerdy Linux problem
Come up with a set of holiday itineraries
And more. And by and large, it has done those tasks really well.
But, in every case, I have checked its work manually before making any decisions that would be expensive to reverse.
Why? Because an LLM is simply a programme which has read as much content as it can, and synthesised it into an ability to construct sentences of its own.
Imagine a parrot that had read the entire internet.
You’d probably get some interesting anecdotes out of it.
You might even ask it some questions (tell me everything about the Cuban Missile crisis).
But I’d be very surprised if you asked it for investment advice, or to generate your business strategy for you, or to diagnose the odd pain in your elbow. Because in the end, it’s a parrot.
The same is true for your favourite LLM. Having read so much, it is brilliant at tasks which require reading or producing lots of text. That’s why it is good at summarising long documents, or producing essays for wayward students. But it does all of that without really understanding anything it is saying. That’s why it is prone to random hallucinations and to simply getting answers wrong.
AI tools will be useful for a huge range of the tasks we have to do in our businesses, whether that is talking to customers via chatbots or working out your optimal pricing and promotional strategies.
But surprisingly few of those successful applications will be simply asking an LLM to produce an answer for you. Maths based applications like Machine Learning will remain the best way to process large amounts of data, for example.
When you are seeking advice about how to use AI to improve your bottom line (and I’ll bet you are asking those questions already) the real secret is to figure out the right tool for the job you are asking for help on.
Get that wrong, and you might end up with a parrot making a mess all around your business.