AI Has A Giant, Expensive Flaw
Turns out, your job is safe.
I’d be lying if I said I wasn’t worried about AI. For one, being a writer as a career is one of the few creative jobs that is already being threatened by it. But I have also taken a negative stance against many of its applications, most notably Tesla’s self-driving system and the claims around it. So, if it does take off and works well, I will have egg on my face. But a recent MIT study has calmed my fret and should cure any AI anxiety you might be feeling, whether that be of the job-taking or Skynet variety. You see, it turns out AI is too damn expensive and will be for a very long time!
These MIT researchers wanted to scythe through the marketing bullshit and over-exaggerated claims of AI CEOs to find out what the commercial viability of AI actually is. To achieve this, they calculated if AI was more cost-effective in 1,000 visually assisted tasks in 800 occupations. As such, the type of AI they are checking here is specifically computer vision combined with peripheral robotics and computer systems. Computer vision AI allows computers to understand the world around them through camera feeds. It is the technology behind Tesla’s self-driving and the world-famous ASIMO robot.
One example of a job/task the researchers looked at replacing with AI was in a bakery, using computer vision AI to assess the quality of the ingredients before they enter the kitchen. In this example, the AI would only replace 6% of the work of a single job; as such, the cost of developing, installing, operating and maintaining such a system is far more expensive than hiring a human for the same task.
In fact, the study found that only 23% of workers’ wages could be replaced by AI for less expenditure. Only the most mind-numbing, constrained and simple tasks could be replaced by AI, and even then, it would mostly only augment a job and only partially replace it. So, the vast majority of jobs are utterly safe from computer vision AI.
But the researchers also found that it will be a long time before they can. Even if the cost of AI declined by 20% each year (which it definitely isn’t), it would still be decades before it become cost-effective for businesses to automate themselves with AI. The big hurdle here is the vast upfront costs, as programming the AI, gathering the enormous amount of data needed to train it, training it and then setting up the computer systems to run it costs an extraordinary amount of money.
So, not only is your job almost undoubtedly safe from AI, but it will be for years to come.
But, this situation also limits any rouge AI Skynet situations. If AI is too cost-probative to replace straightforward jobs, it won’t be cost-effective or viable to replace any more crucial world-threatening jobs.
But what about the handful of jobs that AI can entirely replace? Well, multiple studies have found that AI will create a net gain in jobs. AI and computer vision are labour-intensive to develop, program, train, install, operate and maintain. So, one low-skill job will be replaced by a multitude of jobs with varying skill levels. This isn’t to imply that people losing their jobs to AI will not be painful, but that it won’t gut the job market.
Now, this MIT study did have one giant caveat. To quote the actual paper: “An important exception is that if a task requires prohibitively complex supplementary systems, e.g., “Piloting aircraft” or “Driving ground vehicles,” we did not consider it exposed [to the risk of being replaced by AI] even if it can be done with computer vision.” Why this obvious omission in their calculations? The idea that AI could replace taxi drivers and lorry/semi-truck drivers is one of the most discussed applications of AI, so surely, these jobs deserve to be investigated by the MIT team.
Well, taxi drivers and lorry/semi-truck drivers earn not that much more than bakery workers. Yet, the AI needed to replace their job/tasks is exponentially more complex (in a legal and program sense) and exponentially more expensive than the ingredient-checking baker AI that wasn’t commercially viable. In other words, there is no point in analysing this scenario, as we know it is explicitly commercially unviable, as these extra peripheral tasks and systems and a need for far more comprehensive training and programming make them too expensive.
But there is also a development issue with these types of AI. You see, it might not be possible for these systems to work as predictably as we require them to, as diminishing returns could render them developmental deal ends.
Musk said in a recent investor call that their self-driving system “barely works at 2 million [training examples]. At 3 million, it’s like, wow, OK, we’re seeing something. But then, you get to, like, 10 million training examples, it becomes incredible. So there’s just no substitute for massive amount of data. And obviously, Tesla has more vehicles on the road collecting this data than all the other companies combined. I think maybe even an order of magnitude.” But computer vision AIs (and deep learning algorithms in general) aren’t that simple. They don’t get predictably better the more you train them with more data. There is no way for Musk to know that 10 million training examples are enough to make his self-driving software work reliably enough to be legally used as fully autonomous vehicles. The returns get unpredictably smaller the more extensive the dataset. What’s more, the AI can pick up incorrect or non-existent trends in the data and learn to react to that, meaning they get worse the more you train them.
It is 100% within the realm of possibility that Musk will need such a vast amount of data and have to undertake such lengthy, vast and intricate AI training that the cost of creating a fully autonomous vehicle through this method is well beyond commercial viability, even with Tesla’s massive scale advantage (more cars will use it, spreading out development costs).
So, how can these researchers evaluate the commercial viability of self-driving cars or planes?
Not only does this research vindicate my opinion that Tesla’s claims that, shortly, all their cars will be driverless are effectively shilling snake oil, but it also indicates that the AI revolution is still far, far in the future. What’s more, it’s not like the costs associated with AI will drop any time soon, as the researchers theorised might happen. In fact, considering geopolitics, the crucial silicone chips that are needed to power AI could skyrocket in price soon! In short, you can sleep soundly knowing that robots aren’t going to render your job obsolete for a good while.
Thanks for reading! Content like this doesn’t happen without your support. So, if you want to see more like this, don’t forget to Subscribe and follow me on BlueSky or X and help get the word out by hitting the share button below.