In the early 2010s, it became increasingly common to suggest that robots were stealing our jobs. Dozens of headlines breathlessly warned (e.g. MIT, BBC, The Conversation, Fast Company) nearly 50% of 700 common occupations would be completely automated – filled by machines, rather than mankind.
Only it didn’t really happen.
Turns out it wasn’t even a new prediction. There was a similar wave of panicked predictions in the early 1980s. In fact, The Omni Future Almanac created a long list of jobs which were supposed to be eliminated by the year 2000, including secretaries, dry cleaners, farm workers, real estate brokers, bank clerks, and toll booth operators.
Well, at least they got the toll booth operators correct. Sort of. Even in the epicenter of technology, we are still discussing eliminating toll booth operators 40 years after the original prediction.
So forgive me if I pause a bit amidst the latest claim prediction hysteria that AI agents will replace most jobs “faster than you think.” Yes, robots and AI agents – technology in general – will streamline many processes and even eliminate others. Some relatively quickly but most over time.
As the saying goes, “we tend to overestimate the effect of a technology in the short run and underestimate the effect in the long run.”
How close are we now – the midpoint of the 2020s – to a fully robotized company? Can we eliminate most, if not all, roles in a company and replace them with AI agents?
To find out, researchers designed a simulation of a small software company with a variety of tasks that must be performed by its workers, including browsing the Web, writing code, running programs, and communicating with other coworkers. They staffed this company – dubbed TheAgentCompany – with large language model (LLM) AI workers, powered by both closed API-based and open-weights models.
How did the LLM employees perform?
All the LLMs struggled with administrative, finance, and data science tasks; many completed none of these successfully. On the other hand, software engineering tasks, arguably more difficult for most people, had a much higher success rate. As the authors wrote, “this suggests there exists a gap between the perceived difficulty of the tasks for humans versus the difficulty for LLM agents.”
Overall, the LLMs weren’t very good employees. The best-performing LLMs completed just 24% of their tasks and took an average of 30 steps to complete each one.
At least as of now, robots and AI agents are automating some tasks, but they aren’t stealing our jobs.
No comments yet.