Recently, while waiting for a flight, I was sitting next to someone who was calling several companies he was doing business with – his bank, his health insurance provider, and some others.
In every case the call was answered by a chatbot.
But in every case he also did whatever he had to do to get past it and talk to an actual human being.
So much for the promise of AI, I thought!
As anyone who has ever phoned up companies, to hear the expression “In a few words tell me what you’re looking for” knows, the responses they get back are generally useless.
Any standard query such as looking up a document or bill, can be better done on the website (so it tells you), while asking a non-standard or unstructured question is a near total waste of time.
Which surely begs the question, what can AI actually do?
In the time since ChatGPT became available, the hype around the capabilities of AI has reached the stratosphere.
We’ve all heard the proclamations that AI will transform work completely, and (in doing so), will eliminate millions of jobs.
The master of hype – Elon Musk – recently predicted that AI will take all our jobs and robots will provide any goods and services that we want. (He may do better at predicting how long California can keep subsidizing Tesla given the size of the state’s budget hole, but that’s another story).
But here’s the thing. Use cases for AI are proliferating but few can actually demonstrate that the technology can be deployed at scale.
Companies are seeing some benefits from AI. For instance, employees at PwC who use a tool built on ChatGPT report a 20% to 40% increase in productivity.
But that’s hardly a game changer.
Chatbots have improved, but they seem to be better at getting customers to stop calling out of frustration or just drop the call, waiting for a human, than they are at solving real problems.
They are being considered for mental health counseling, but they are still no substitute for an actual human therapist.
Moreover, the advice they offer is not always appropriate or safe, and needs to be highly supervised.
Yes, AI tools do improve productivity, but not nearly as much as the hype would lead one to believe.
What we know about AI is that it’s best suited for summarizing text and generating output based on prompts.
And even that requires humans to review the results and write the prompts, i.e., human intelligence.
The one thing we can’t get away from is the fact most jobs require more from the people doing them than just summarizing text and responding to questions.
So AI cannot replace people because it can’t do all the work people do.
Even the simplest jobs require a person to do multiple tasks, many of which can’t be automated.
To have an impact on employment, jobs would have to be limited to just the tasks AI can’t do, which is rarely possible.
Plans of trucking firms to use self-driving trucks have floundered on the reality that drivers do much more than just drive. The driver has to check the truck is correctly loaded and unloaded, and often does that task, as well as fix a flat tire, clean the cab, and fill it up with gas.
Even for tasks such as coding, AI can help write the code given the correct prompts.
But writing the prompts does itself require understanding many requirements, which are rarely simple. And the results of the code usually have to be cleaned up too.
To deliver the limited functionality – summarizing and responding to prompts – AI tools require vast amounts of data. But the data must be organized first – by humans. No LLM can simply ingest raw, unstructured data.
The best of any technology is always yet to come.
AI will improve as all technology does.
So perhaps the question we really need to ask is by how much?
The iPhone 15 is a huge improvement over the original iPhone, but the improvements in each generation are increasingly small.
Just how much better is the camera in the 15 compared to the last version? AI is already facing this problem.
It’s actually difficult to tell the difference in the results produced by ChatGPT version 4 compared to ChatGPT version 3.5.
Two factors constrain the ability of AI products to get better: data and cost.
AI needs vast amounts of data to train the LLMs.
But they’ve already consumed virtually the entire internet. And there isn’t a second one.
The second limitation is the cost of training AI because of the energy requirements of AI data centers.
AI requires a lot of energy. Did you know that a simple ChatGPT request uses nearly 10 times more electricity than a regular Google search? Supporting that requires specialized chips, which are incredibly power hungry.
Every AI chip uses about as much electricity each year as three EVs. Over 10 million have been installed already and thousands more are being installed in data centers daily.
The cost of all this is massive. Estimates are that the tech industry spent more than $50 billion on chips and energy to train AI in 2023.
But they’ve only brought in about $3 billion in revenue. The difference will have to be made up by increased costs to users.
The impact of any new technology is almost always impossible to predict.
A quote attributed to many people, from the Nobel prize-winning physicist, Niels Bohr, to legendary baseball player Yogi Berra states: “It is difficult to make predictions, especially about the future.”
Much as we would like to, we cannot predict the future because there are too many factors and trends that influence what will occur.
How humans use a technology is ultimately what determines the impact. Anyone remember Google Glass?
Jeff Bezos predicted back in 2013 that soon delivery by drones would be commonplace. But who wants to see the sky filled with drones carrying boxes?
We’re still getting Amazon orders by Fedex and UPS trucks – even though drones are now a cheap and mature product.
Some things we just can’t always predict. When PCs first became available they were intended to be used just by secretaries and clerical staff.
But they quickly became popular because people wanted this cool new gadget on their desks – even though many of the people using a word processor would have considered it beneath them to use a typewriter.
So what can we predict about the impact of AI on jobs? There’s no model or methodology that can reliably predict what will happen as AI becomes widely used, but one study by researchers at MIT suggests the impact on jobs may be minimal.
The researchers focused on jobs that mainly require completing “vision-related” tasks. That is, tasks currently performed by workers which could be carried out by a sufficiently sophisticated computer vision system.
For example, checking products for quality at the end of a factory assembly line or scanning medical imagery for anomalies. These are tasks AI is extremely well-suited for in the current state of development.
But even here the researchers found that it would only be cost effective to automate about a quarter of these tasks – mainly due to the large upfront costs of AI systems.
Even with rapid decreases in cost of deploying AI it will still take decades for such tasks to become economically efficient for full-scale automation.
In other words, any expectation of widespread replacement of jobs with AI is a pipe dream.
One prediction we can make is that AI will create more jobs than it eliminates, because all technology ever developed always has.
