Could the Next AI Revolution Begin in a Garage, or a UFO?

AI for the Perplexed
By Dennis H. Goodhart, with Hymie, my AI collaborator

Could the Next AI Revolution Begin in a Garage, or a UFO?

AI Gen II: What happens if the next breakthrough makes today’s billion-dollar investments obsolete?

In my last article, I asked a fairly simple question: Who told them to spend all that money?

The technology industry is investing extraordinary sums in artificial intelligence. J.P. Morgan Global Research estimates global AI capital spending could approach $900 billion in 2026 and total $5.5 trillion through 2030. Much of that money is going toward chips, data centers, power infrastructure and the systems needed to build and operate increasingly large artificial-intelligence models.

Maybe they are right.

But there is another question worth asking:

What if we are spending all that money developing the wrong technology?

Not bad technology. Not useless technology. Perhaps simply a technology that will eventually be superseded by something fundamentally different.

Let’s Call Today’s AI “Gen I”

The term AI Gen II is my own shorthand. It is not (yet) an accepted technical designation, and I am not suggesting that some secret second generation of artificial intelligence already exists.

Although, given the subject, I wouldn’t rule it out.

What I am calling Gen I is the broad family of artificial intelligence that dominates today: large language models and related systems built primarily around the Transformer architecture, enormous amounts of training data, enormous computing resources and increasingly sophisticated software surrounding the underlying model.

And Gen I is getting better very quickly.

Models are gaining memory. Agents can use tools. Systems can search databases and the internet, call other programs, work through multi-step problems and retain information from earlier interactions. Researchers are also experimenting with ways for AI agents to learn from their own successes and failures.

Google’s ReasoningBank project, for example, gives agents a memory system that extracts useful strategies from previous experiences and makes those strategies available for future tasks. Recent research is increasingly treating memory not simply as storage but as a way for AI systems to improve future behavior.

That is significant.

But it may still be Gen I getting better.

What Would Gen II Be?

What would Gen II actually look like?

I don’t know.

That is the point.

It might maintain persistent memory, learn continuously from experience, adapt its reasoning as it learns, and build knowledge throughout its operating life instead of periodically waiting to be retrained.

It might require dramatically less computing power. It might use computing in a completely different way.

Or it might be something none of us is imagining yet.

Does anybody really know?

We Have Seen This Movie Before

Technological transitions often look obvious after they happen and much less obvious while enormous industries are still making money from the old technology.

The automobile did not emerge because somebody finally bred a faster horse. It replaced the assumption that transportation required a horse at all.

Kodak provides another useful caution, although historical analogies should never be taken too literally. Kodak itself invented the first digital camera in 1975. Yet the company’s economics, infrastructure and identity were deeply connected to film. Digital photography did not simply produce a better roll of film; it changed what photography was.

Palm helped define the handheld-computer era. Then the smartphone changed the category. The question was no longer who made the best electronic organizer. The organizer became one function inside an entirely different device.

None of these examples proves that today’s AI architecture will be displaced. They demonstrate something else: successful incumbents naturally optimize the technology and business model that made them successful.

Sometimes the disruptive idea begins by questioning the assumption everybody else is optimizing.

Somewhere, Someone May Be Working on It

This is where my speculation begins.

Suppose somewhere outside the major AI laboratories there is an unconventional engineer, researcher or small team working on artificial intelligence from a completely different starting point.

Maybe the goal is not to build a larger language model. Maybe it is to build a companion intelligence that learns continuously from one person or organization. Instead of repeatedly consulting a gigantic pretrained model, it develops through experience. Instead of requiring ever-larger centralized infrastructure, perhaps much of its useful intelligence develops locally or through a far more efficient computing architecture.

That person may be working in a university.

Or a garage.

For all I know, a friendly Martian may drop off the design.

The Martian is optional.

The possibility of an architectural breakthrough is not.

Would Today’s Data Centers Become Worthless?

Probably not.

That would make a wonderful headline, but technology rarely works quite that cleanly. Data centers, power systems, networks and semiconductor infrastructure can often be repurposed. Current AI systems would not suddenly stop functioning because something better was invented.

The more interesting question is economic.

If a new architecture could provide comparable or superior intelligence using a fraction of the computing resources, the economics of AI could change dramatically. The value of constantly building larger models and larger data centers could decline. Some infrastructure investments could produce lower returns than originally expected.

That does not make current investment irrational. Companies have to build with the technology available today.

But it does make one assumption dangerous:

That the future of artificial intelligence must look like a larger version of the present.

Gen I May Become Gen II Before Anyone Notices

There is another possibility.

Maybe the revolution will not come from a garage at all.

The work already being done on memory, agents and continual learning could gradually transform today’s systems so substantially that Gen I evolves into something we would eventually recognize as Gen II.

ReasoningBank is an interesting example because it addresses a weakness that seems almost absurd when applied to intelligence: an AI can solve a difficult problem today and still approach a similar problem tomorrow without automatically learning from what happened yesterday.

Giving AI meaningful experiential memory begins to change that.

Researchers are now examining systems in which memory can be created, revised, consolidated, forgotten and used to improve future decisions. That sounds less like a larger database and more like the beginning of an artificial learning process.

Whether that leads to a fundamentally different architecture is unknown.

The Real Risk for the Giants

Microsoft, Google, Amazon, Meta, OpenAI and the other major players have tremendous advantages: talent, capital, computing infrastructure, customers and enormous amounts of accumulated research.

Like Kodak, they may be fully aware of a technology that could eventually challenge the economics of their existing investments. In fact, many of the people most likely to invent the next architecture probably work for these companies today.

But incumbency creates its own gravitational pull.

When billions of dollars have been committed to data centers, chips and a particular technological direction, the natural response is to make that direction work better.

The outsider has one advantage: no need to protect yesterday’s investment.

The Bottom Line

I am not predicting that large language models are about to become obsolete.

There is no evidence that an AI Gen II breakthrough is sitting in somebody’s garage waiting to be switched on.

This is speculation. But useful thought leadership sometimes begins by asking a question before anyone knows the answer. We are watching one of the largest technology investment cycles in history being built around assumptions about how artificial intelligence works, how much computing it needs and how it will improve.

Those assumptions may prove correct.

Or the most important AI question of the next decade may turn out not to be:

“How much bigger can we make the model?”

It may be:

“Why are we building intelligence this way at all?”

And if somebody comes up with a convincing answer to that question, I hope they remember to close the garage door.

Especially if there is a UFO parked outside.


Sources

J.P. Morgan Global Research, What Unprecedented AI Capex Means for Markets and Investors, October 2026.

Google Research, ReasoningBank: Enabling Agents to Learn From Experience, April 2026.

Chenye Han, Hongche Yu, Mengyue Wu, Dell Zhang and Jixiang Luo, Learning to Remember: Advances in Memory Mechanisms for Large Language Models, October 2026.

Eastman Kodak, company history and milestones regarding Steven Sasson’s 1975 digital-camera prototype.

Disclosure: This article contains speculation about possible future developments in artificial intelligence. “AI Gen II” is terminology used by the author for purposes of this discussion and is not presented as an established industry classification or as a technology known to exist.

The opinions expressed are those of the author. The author is not employed by, compensated by, sponsored by, or affiliated with any of the companies or products mentioned in this article. The author does have a paid ChatGPT subscription. ChatGPT, referred to here as “Hymie,” is credited as an AI writing collaborator.

AI for the Perplexed is written by Dennis H. Goodhart with assistance from Hymie, his AI collaborator. Dennis retains editorial responsibility for the published content.