Hello world. This is my first blog post. I decided to start writing mostly as a way to think out loud. I want to use this space to explore technical concepts I am curious about, connect some disparate ideas, or just share non-technical thoughts that inspire me. Hopefully other people will enjoy learning along with me.
As I start the spring semester here at UT Austin, it is almost impossible to ignore the noise around AI. But so much of the discourse online feels entirely disconnected from reality, swinging between hype and doom with little in between.
Then I read Dario Amodei's new essay "The Adolescence of Technology" that he shared this month. It helped me reframe my perspective on the topic. He frames the next few years not as a sci-fi movie but as a very concrete, imminent engineering challenge.
Bootleggers, Baptists, and Validating the Critics
Before diving into the tech, I want to touch on the criticism his essay received. A lot of people online immediately dismissed Amodei's essay because Anthropic obviously has a financial interest in heavy regulation. The argument is that calling for government oversight is just a classic moat-building strategy to lock out open-source competitors.
I think dismissing him on those grounds is a mistake. It reminds me of the economic theory of "Bootleggers and Baptists" introduced by Bruce Yandle in the 1980s. Yandle observed that regulations often pass because two entirely different groups advocate for them: moral advocates (the Baptists who want to ban alcohol on Sundays) and self-interested groups (the Bootleggers who want to sell illegal alcohol without competition).
Anthropic might be playing the role of the bootlegger here. But that does not mean the whiskey isn't actually poisoned. There was a Medium article I read that pointed this out really well; it basically says that Amodei is willing to write extensively about biological risk, autonomous weapons, and authoritarian capture. He does not use abstract language. He outlines specific scenarios and mechanisms. That takes a massive amount of courage when your company's valuation depends partly on market confidence. He genuinely believes someone is going to build this infrastructure. It might as well be someone who is actively worrying over the consequences.
The Duplicator and Abnormal Technology
Amodei talks about the arrival of a "country of geniuses in a datacenter." This phrasing immediately made me think of Holden Karnofsky's concept of "The Duplicator".
Karnofsky asks a simple question: what happens if you build a machine that can perfectly copy a human mind and run it on a server? If you can duplicate your smartest systems researcher a million times and run them at high speeds, economic growth stops acting normally.
If you look at the chart of human history in "This Can't Go On", economic output is basically a flat line for thousands of years. The industrial revolution bent it sharply upward. But algorithmic systems capable of autonomous engineering create a completely different, self-referencing feedback loop. We are dealing with a profoundly abnormal technology. The growth line goes completely vertical. That spike is the "adolescence" we have to somehow survive.
Mechanistic Interpretability as Xenolinguistics
This brings me to the technical rabbit hole I am currently curious about and I hope to fall into. Amodei highlights mechanistic interpretability as a core defense against AI risk. We have to look inside the black box and understand the internal math of these models before they act in the real world.
For some reason, reading about this made me think of Wittgenstein's famous quote: "If a lion could speak, we could not understand him."
Wittgenstein meant that a lion's conceptual universe is so different from ours that their language would be fundamentally incomprehensible to us. I feel like this perfectly describes neural networks. They are alien intelligences. When we look at their high-dimensional latent spaces, we are looking at concepts and features that do not map cleanly to human language.
Mechanistic interpretability is basically applied xenolinguistics. It is the science of building translation tools for an alien mind. How do we map these polysemantic neurons into human-legible intent? How do we build visual interfaces that let researchers watch a model in the act of deception? This space, to me, feels like one of the most fascinating puzzles in computer science right now.
Keynes, Vonnegut, and Purpose
I want to close with a thought from the essay that has been stuck in my head. In a section he titles "Player Piano" (after Vonnegut's novel), Amodei discusses how machines eventually displace human cognitive labor. He notes:
"I think human purpose does not depend on being the best in the world at something, and humans can find purpose even over very long periods of time through stories and projects that they love. We simply need to break the link between the generation of economic value and self-worth and meaning."
This is a beautiful sentiment. It echoes what John Maynard Keynes predicted in his 1930 essay Economic Possibilities for our Grandchildren. Keynes thought technology would solve the economic problem entirely, leaving humanity with 15-hour workweeks and the new challenge of figuring out how to live meaningfully. We obviously did not hit Keynes' timeline. But Amodei is arguing that we are finally approaching that exact threshold.
The idea is intimidating. It means my generation's primary value will no longer just be our economic output. Figuring out how to navigate that transition feels incredibly important. I am excited to dig deeper into the math of alignment this semester!
Things I referenced
- Dario Amodei — The Adolescence of Technology
- Holden Karnofsky — The Duplicator and This Can't Go On
- Bruce Yandle — Bootleggers and Baptists
- John Maynard Keynes — Economic Possibilities for our Grandchildren
- "Complications from the Ground" — a Medium response to Amodei's essays