A few years back, when I started covering Silicon Valley’s push to replace human labor with artificial intelligence, most tech leaders had the decency to at least pretend otherwise.
“We’re not automating workers, just enhancing them,” they’d say. “Our AI tools won’t eliminate jobs; they’ll serve as helpful assistants, liberating employees from tedious tasks.”
Yet, while such claims aimed to soothe anxious employees and obscure corporate automation agendas, they reveal more about the technology’s limits than about executives’ intentions. Back then, AI lacked the capability to automate a majority of roles, particularly for degree-holding professionals in sectors like technology, consulting, and finance.
Things are beginning to shift. Modern AI systems can now develop software, generate comprehensive research reports, and tackle intricate math and science challenges. The new AI “agent” enables users to carry out lengthy task sequences and verify their own output. While many regions still face labor shortages, some experts fear that the recent uptick in unemployment among college-educated individuals is already being attributed to AI taking over certain entry-level positions.
On Thursday, I witnessed a glimpse of the future at Laborg, thanks to an event hosted by Mechanize in San Francisco. They have an ambitious mission: to automate every job—including those held by miners, doctors, lawyers, software engineers, and designers responsible for our buildings and even our children’s education.
“Our aim is total work automation,” stated Tamay Besiroglu, one of Mechanize’s 29-year-old founders. “We aspire to achieve a fully automated economy as swiftly as possible.”
The aspiration for complete automation is not new. Economist John Maynard Keynes foresaw in the 1930s that machines would take over nearly all jobs, generating material wealth and freeing individuals to pursue their passions.
Naturally, that vision never materialized. However, recent strides in AI have rekindled hopes that technology capable of large-scale labor automation is on the horizon. Dario Amody, CEO of Humanity, recently cautioned that AI could displace half of entry-level white-collar positions within the next five years.
Mechanization represents one of many startups aiming to make this possible. Founded this year by Besiroglu, Ege Erdil, and Matthew Barnett, who has a background at Epoch AI, a firm researching AI system capabilities.
They have attracted investments from prominent tech figures, including Stripe’s Patrick Collison and Google’s chief AI scientist, Jeff Dean. Currently, they operate with a team of five and collaborate with major AI companies. (They declined to disclose specifics due to a non-disclosure agreement.)
Mechanize’s strategy for job automation via AI employs a method known as reinforcement learning, the same technique utilized to train computers to play board games nearly a decade ago.
Today, top AI firms enhance their language models’ performance using reinforcement learning for additional calculations prior to generating responses. Frequently termed “thinking” or “inference” models, these are stunningly proficient at narrow tasks, like coding and solving complex math problems.
However, most roles demand handling multiple tasks, and today’s leading AI models struggle with complex workloads or navigating intricate enterprise systems.
To address this, Mechanization is designing a novel training environment for these models. Essentially, they create intricate scenarios to teach the AI what actions to take in specific contexts and assess its performance.
For instance, to automate software engineering tasks, Mechanize simulates an environment reminiscent of the one software engineers use, complete with a virtual machine encompassing an email inbox, Slack account, coding tools, and a web browser. AI systems are tasked with utilizing these tools to accomplish given objectives. Success yields rewards, while failure incurs penalties, prompting the system to try again. With sufficient iterations and well-structured simulations, AI could ultimately replicate what human engineers accomplish.
“It’s akin to designing a rather mundane video game,” Besiroglu remarked.
Mechanization starts with the field of computer programming, where reinforcement learning has already shown some potential. The hope is to extend this strategy to automate roles across various other white-collar sectors.
“We’ll know we’ve truly succeeded when we develop an AI system capable of undertaking nearly all responsibilities that can be handled by a computer,” the company articulated in a recent blog post.
There are lingering questions about the efficacy of Mechanize’s methods, particularly for non-technical jobs where success isn’t as straightforward to measure. (For instance, what does it mean for an AI to succeed as a high school teacher? Even if students achieve high standardized test scores, what if they remain dissatisfied and disengaged? Could reward hacking result in merely giving students the correct answers to boost their scores?)
The founders of Mechanize are realistic about the challenges in automating such positions. Mr. Barnett estimates that achieving full automation may take between 10 to 20 years, while Erdil and Besiroglu anticipate a timeline closer to 20 to 30 years.
These timelines are modest by Silicon Valley standards. However, I commend Mechanize for its transparency regarding its objectives, unlike many AI enterprises developing labor supply technologies behind closed doors.
Nevertheless, I noticed that their proposition seemed to lack empathy for those whose jobs are at stake and bore no consideration for whether society is ready for such monumental change.
Besiroglu maintains that AI will ultimately generate wealth that can be redistributed to displaced workers through concepts like “radical abundance” and universal basic income to help sustain their quality of life.
However, similar to many AI firms delving into labor supply technology, Mechanization has yet to propose new policies for easing the transition into an AI-driven economy, nor do they advocate for enhancing the social safety net or retraining workers for new employment.
During the Q&A, I raised the ethical question of whether it’s morally right to automate all labor.
Barnett, identifying as a libertarian, replied that he believes AI fosters economic growth, drives life-saving advancements in medicine and science, and that a society fully automated is preferable to one that still relies on human labor in a low-growth environment.
“If society as a whole becomes significantly wealthier, I believe the benefits outweigh the downsides for those losing their jobs,” Barnett noted.
Well, at least they are honest.
Source: www.nytimes.com