AGI Won’t Change the World, AI Infrastructure Will

5 min readJan 31, 2025

1909, Henry Ford had a problem.

A year before, Ford Motors unveiled the Model T, the first vehicle affordable for everyday Americans. Demand skyrocketed. The problem? There weren’t enough roads in the United States for people to drive. Some car companies floated the idea of privately funded highways, but Ford had a bigger vision. He realized private funding alone couldn’t cover the nation’s highway needs. He funneled effort into public relations and lobbying, urging states, counties, and the federal government to support building roads. With his trademark persistence, he pushed this agenda, playing a major role in establishing the U.S. highway system. [1]

But it wasn’t all smooth sailing. Before automobiles hit the scene, roads were shared community spaces — people strolled, biked, shopped, and kids played without worry [3]. Suddenly, these roaring metal boxes caused widespread accidents. By 1923, roughly 18,000 people died in car crashes — half of them were children [2]. Public sentiment soured, with activists calling to ban cars outright.

US streets before automobiles were communal places. This looks surprisingly similar to the Bangladeshi village I grew up in, where the street markets still today are the heart of the community.

In retaliation, the automotive companies banded together and started a massive ad campaign, coining the term “Jaywalking”. They turned the blame for automotive accidents from manufacturers to pedestrians — the people were faulty, not the machines [2]. It worked. In 1925, Los Angeles formally criminalized jaywalking in its traffic code. By 1930 all over America, the roads now belonged to machines, and humans needed permission to cross them.

These major social changes weren’t caused by technology alone; they required cultural and infrastructural shifts that allowed the technology take hold.

Let’s leap forward.

After all the hype, why don’t we have self-driving cars yet?

For about a decade, people have claimed self-driving cars were just a year away. But we’re still waiting. The engineering is dazzling — spinning LiDAR that maps 3D environments dozens of times a second, an array of cameras capturing a 360-degree view, giant neural networks spotting everything from pedestrians to cats and traffic signals, plus onboard computers crunching it all at lightning speed. And yet, these wonders still can’t outperform the flawed human driver who’s sometimes tired, sometimes distracted, and sometimes hungover. Some argue we need to solve AGI to have full self-driving [8]. I disagree.

Self-driving car is an infrastructure problem.

Since Ford’s days, roads have been built around human sight. Highways, signs, traffic lights, parking spots, and vehicle designs have gone through countless iterations to make them optimal for human vision. Now, we force these brilliant supercomputers on wheels to awkwardly wade through this muddy road made for humans, too careful not to hit one of them, or worse, hurt their feelings. They are driving with their eyes blindfolded, and hands tied behind the seat. But imagine if the road infrastructure was built “for computers”…

Waymo selfdriving vehicle on a street (Photograph: Allen J. Schaben/Los Angeles Times/Getty Images)

No need for a computer vision system to recognize traffic lights, because you can do this faster from miles away by broadcasting the signal with radio waves. No need for ultra-complex obstacle detection systems, because every car on the road can tell you its precise location, speed, road condition, and possible obstacle, from miles away via the Internet or radio. Traffic jams would vanish because the grid could find optimal routes for the entire grid. No need for speed limits, because the car's reaction time is in microseconds. No bulky computers, because all cars on the road can process their data as a distributed system (some of that processing can be offloaded to roadside traffic computers). Car windows will be optional because cars will be for entertainment and work (sigh!). No need to be on the lookout for unpredictable human drivers, because all cars on the road are self-driving. Tickets — are you kidding?

Although I am massively oversimplifying. But all of this, we can do TODAY. No AGI required.

Do we really need AGI to make LLMs superhuman?

If you are reading from the future, I am writing this sometime after OpenAI made o3 and smashed multiple AGI benchmarks, DeepSeek R1 stirred up a lot of geopolitical drama, alongside making high-quality, lightweight AI models open to everyone. Some time ago, Google promised to Spend 100 Billion for AI infrastructure [9], so did Microsoft [10], Meta 65 Billion [11], and Softbank, OpenAI, and Oracle are spending 500 Billion on project Stargate [12] — an AI data center. The next wave of AI infrastructure looks gigantic, and its impact will match that scale.

But that’s only one kind of infrastructure. A while ago, Anthropic released “computer use” ability in Claude 3.5 Sonnet [13]. The AI takes recurrent screenshots of a computer screen, moves the mouse and presses the keys, and (poorly) attempts tasks humans do (and sometimes takes a break to look at Yellowstone National Park [14]). Soon after OpenAI released their Operator [15] — an agent framework that can browse the web and book Airbnb for you by clicking the mouse and typing text. Other than being some fancy demos, these tools are barely useful for anything practical.

Can anyone smell muddy roads again?

The operating system and the web were built “for humans”. The future of OS and the web will come with simple API calls that AI models can access. In fact, all original Operating Systems were built for text-based terminal commands, and the Graphical User Interface (GUI) we enjoy today was an (expensive) afterthought. With this small infrastructure shift, these AI agents would leap from clumsy interns to unstoppable staffers — even if AI advancement froze today. What sort of societal ripples would that cause? Let me know if you want me to elaborate. But no matter what, the disruption will be huge.

In short, AGI would be an incredible feat, but it’s not required for AI to reshape our daily lives right now. All that’s needed is an upgrade to our current infrastructure so it’s AI-ready. When that happens, we’ll see the true scale of AI’s capabilities — AGI or not.