This blog is part of my 38 before 38 series. I write a blog for every single day for the 38 days leading up to my 38th Birthday.
Note: I am aware that AI is a big umbrella term that comprises of many Machine Learning paradigms. Most of them are already deployed in various industries. They are successful not only commercially, but in practice as well. This essay is about LLMs. But nobody would click that link if write that in the title.
Large-language models(LLMs), or “AI” for normies, are a cool technology. Apps like ChatGPT. For the first week you use it. Then it becomes a cute toy. It is not very good at anything interesting. It is very good at things that nobody likes anyway. Can that be converted into sales?
What LLMs can’t do
If you ask an LLM, “Write a poem about the beauty of nature in the style of william wordsworth”, it would produce an out like this:
In valleys green where golden meadows lie,
The morning light spills softly from the sky;
The lark ascends,
her song a silver thread,
The winds a whisper in the oak tree’s head.Behold the river’s dance upon the stone,
A gleaming path where light and shadow roam;
The lilies bow in grace beside its way,
Their petals kissed by dawn’s first gentle ray.
This is terrible. It can reproduce the structure of a poem, and how Wordsworth might have written. It can’t actually produce anything worthwhile. You can ask it to write a short story, movie script, or even an essay such as this. It would read like advertising copy, or a memo from your department head.
Apologies to copywriters and managers everywhere, nobody wants to read that. It is very little value to the intended audience, let alone someone looking for a short story.
What LLMs are good at
Structured repetitive language with quantifiable measures. Which includes corporate/business writing, but also computer program code. Code has a rigid structure, and quantifiable outputs. There is also a lot of it available publicly, with clear instructions on how to implement it, or explanation of what it does.
All the LLM vendors will show you various benchmarks, that are meaningless. In practice, even the free-tier “lesser” models can produce code that is as good as newly hired junior engineer or intern.
How to make money off of this
The decision to implement any given technology in a business is made on the basis of productivity. At least in principle*. So how can any given firm enhance their productivity using these use cases?
An advertising firm, for example, can fire most of its copywriting team. And have the rest do the job. Lowering labor cost, increasing revenue. A smaller firm may try to go after more clients and use the productivity boost to increase revenue by increasing capacity.
A tech firm may fire their junior engineers and let senior engineers generate grunt code, fix it and implement it. A manager can save the time drafting memos, and employees can save time by the llm summarize this
Why isn’t anyone doing it
The prevailing theory for all the layoffs is that the big corporations are replacing those workers with LLM. Even if that is true, getting your employees to do 3 people’s work rarely does result in productivity gains. A senior engineer would end up spending more time cleaning up the ChatGPT code than reviewing code or applying their domain expertise.
A copywriter working 15 hour days would not last long. And if you are an executive at any given business, why would you spend money so a machine can write memos for a machine to read**?
The main reason for the layoffs seem to be the pandemic overhiring. Somehow the CEOs that made that decision and destroyed
Picks-and-shovels
This might read like another cranky rant. But I am genuinely looking for people making money off of this. Because I want to make money off of this. That is while VCs are subsidizing these services. This a gold rush, but without any gold. The picks and shovels are being sold at a loss.
*As companies like SAP and Salesforce show, in function it is based on how good your sales team is.
**Apple and Google are advertising this as features, by the way.