You can be deeply familiar with AI and still be skeptical of “agentic AI.” I was.

For a while, I dismissed most “agentic” startups as VC-funded cron jobs. An LLM integration wrapped in a thin scheduler, rebranded as autonomy.

But there’s something real underneath the hype, and understanding it requires zooming out. LLMs haven’t evolved in one continuous line. They’ve moved in distinct acts, each driven by hitting a ceiling and finding a way around it.

We’re now entering their fifth act. The first four largely occurred in parallel, so forgive the linear framing that follows.

Dario Amodei, Anthropic’s CEO, recently suggested that their Claude AI models might be conscious.

This may seem jarring: large language models generate text by predicting the next likely token in a sequence, a process that appears purely statistical and mechanical.

But both the claim and the dismissal miss the deeper question.

The End of the Crawlable Web : For most of the modern internet, there was an implicit contract between websites and search engines. Websites made their content accessible. Search engines indexed it. And in return, search engines sent users back to the original site.

It wasn’t always perfect — news aggregators and social media links were persistent points of contention — but the incentives aligned well enough that the web became broadly searchable. You could discover obscure blogs, old research papers, forum discussions from fifteen years ago, or technical documentation buried deep in a site.

As software development shifts from requiring specialized skills—built on multiple layers of technical understanding—to describing intent in plain English (or your language of choice), the act of producing software appears to become accessible to a much wider audience. The people best positioned to excel may not be well versed in software at all, but rather those who are good at expressing ideas clearly, thinking iteratively, and breaking problems down.

At first glance, that sounds like perhaps the most egalitarian shift our industry has ever seen.

But there’s a quieter change brewing while we’ve been distracted by the expanding capabilities of AI tools. As AI reshapes how software is written, it reintroduces something we spent decades deliberately removing: cost.

And once cost becomes the bottleneck, software stops being democratic very quickly.

If you paint a picture of Sonic the Hedgehog in your living room, you are technically creating an unauthorized derivative work—but in practice, no one cares. Private, noncommercial creation has always lived in a space where copyright law exists on paper but is rarely enforced.

Gift it to a friend? Still functionally tolerated—a technical act of distribution that copyright law mostly ignores at human scale. Take a photo and post it on Instagram? Now you’ve crossed into public distribution of a derivative work without permission. Under the letter of the law, that’s infringement, although a fair-use defense might apply and Sega almost certainly won’t care. It’s fan engagement, free marketing, and good PR.

Sell that painting, though, and the tolerance disappears. You’re no longer a fan, you’re a competitor.

The rapid expansion of context windows has become one of the most visible metrics of progress in modern language models. In just a few years, we have moved from a few thousand tokens to systems advertising over a million. The value proposition seems straightforward: providing more context should lead to a better understanding for the task.

In practice, it’s a bit more complicated. Longer context windows unlock new use cases, but they also introduce costs, failure modes, and design tradeoffs that are easy to overlook. For teams building production systems, these tradeoffs matter as much as raw capability.

Growing up in the 90s and early 2000s, tech was a foundational part of my childhood.

I built more physical computers than I can remember. We went from paper maps to GPS (which itself evolved from DVDs with static maps to internet-connected real-time navigation). CD players became MP3 players, then streaming services. We had Palm Pilots and early attempts at “smart” phones, which were anything but. Our computers could search for extraterrestrial life through SETI. We emerged from the pager era to portable phones to the entire internet in our pocket (which evolved from charging per SMS or megabyte to unlimited data plans).

Although unintentional, the software patent system has evolved into a classic game-theoretic dynamic.

The core issue is that offensive and defensive patent stockpiles look identical from the outside. That ambiguity triggers a feedback loop that sustains the system.

Patents are meant to protect specific inventions, but in software where ideas spread quickly, they have taken the role of strategic assets in an arms race. Useful concepts are rediscovered independently or unintentionally overlapped. As a result, nearly every company is vulnerable to infringement claims at all times, often through no fault of their own.