tech
p(doom), p(joy)
Before AI became the world-defining thing it is now, I remember messing around with it in the OpenAI Playground. It was around 2021. I didn’t fully understand what an API was, and I don’t remember which model I was using.
I remember hearing about a group of MIT students who were writing all their essays with some text-prediction model, as if they were an essay cartel. I didn’t really believe it, but I tried it with a paper for my anthropology class. I gave it some of the context, and I explained what I wanted to say. It wrote everything for me. Of course, I had to go back and edit it, but I do remember staring at it and not knowing how to feel.
I knew at the time that people were working on this, but I didn’t think they would actually get there. Around that same time, I had started using GitHub Copilot for my computer science classes. I didn’t like writing test cases. It was pretty annoying, and it made it a lot easier to just tab-complete. I remember you could finish half of what you wanted, and then it would autocomplete the rest, as if it was reading my mind. I didn’t really connect that to what I’d been doing in the Playground. That was just a tool for speeding up things I didn’t want to do.
I remember sitting on a beanbag in my frat basement in 2022 when a guy named Brian showed me ChatGPT and asked me if I had heard about it. I’m no longer in this frat, but it was an interesting time in my life.
He said, “Yo, check this out, dude. Yo, check this out. It knows everything.”
People were making up stories about their friends and asking questions about their homework. It wasn’t very good yet. It couldn’t write proper Python or Java, but it was a machine replying back to us, and it made sense.
I’d already used AI to write an essay, but this was very different because everybody saw it.
Sometime around then, I remember putting on a nice suit and going to a career fair. There was a guy at the booth, and somehow I ended up talking to him about a video I’d made for fun about Vanderbilt’s computing clusters. I had done some research and talked to people about the history of computing at Vanderbilt, its three generations of computing infrastructure, data centers, and compute clusters.
That got me invited to his office.
The company was called Phosphorus. It was a cybersecurity startup. The office was in a former pet shop, and it still smelled a little bit like one.
I did a little technical test connecting to a camera using Python. I knew how to code pretty well. I’d been doing it since I was 13, but I remember I had to look up how to use the requests library. He gave me the password, and I got the authentication working. I connected to the camera.
A couple days later, I got an internship. It was not much, but it was more money than I had ever made in my life.
There were about ten people, including me, and everybody was super technical and locked in. There was another intern who was a master’s student, and there was a huge gap between us in terms of ability.
I knew I had to catch up a lot, especially with the terminal.
I spent that winter break kind of obsessing over Linux and Bash. I was struggling with basic commands, and the people at work were just monsters on the terminal.
Now I don’t even really write Bash anymore. I don’t even write terminal commands, but at least I understand it.
I learned about man pages. I went through commands one by one. I learned about network protocols, routing, TCP, UDP, Linux, firmware, and all these things that had existed for decades before I really knew anything about them, even though I depended on them.
Then I made my terminal colorful and fun.
It felt really cool.
I was like a hacker.
Nowadays, I’d probably just ask the model for the command and move on, or an agent would do it without me ever seeing the command at all.
I got pretty fast, and I was pretty good at it. I was using Copilot more and more.
People told me I was doing very well. It means a lot when you aren’t really interested in school and people keep telling you, “You’re so good at your job. You’re so good at this.”
The feedback was immediate at work. I could figure something out, show someone, and know it would be used.
Then Fortune 500 companies ended up using code that I wrote.
Now Congress is running code I wrote.
Downloading Copilot all that time ago is part of the reason I can even say this at all. I know I could have probably learned this without AI, but I don’t think it would have happened the same way.
It was an accelerator, like nitro.
Eventually, I was told I should drop out and join the company full time. They gave me the contract, and I thought I was the shit.
I spent too much money on nice dinners and wasn’t looking far ahead, but I worked.
It was my life for three years.
As the company started changing and was preparing to get acquired, I decided to go back to Vanderbilt.
I came back, and the classrooms were different.
It was strange.
Everything was familiar, but the world was vastly different.
I was different too.
Of course, I could use a calendar now. I woke up on time. I was never late. I had less energy, but I had a much better handle on my brain.
The classes were different.
Things people used to describe as impossibly hard seemed so much easier with AI.
Discrete Structures, a math-heavy class that seemed very difficult before, was easy.
When I was first in school, computer science was a ticket to a better life, but now I was hearing students talk about whether they would be able to get a job at all.
There was so much uncertainty about something that, as a student, I once found very straightforward.
I don’t know if I would get the same opportunities if I had started now.
It’s one of the strange things about it for me.
It helped create the life I have, and at the same time, it’s burning the path that I actually took to get here.
I went to a philosophy club after coming back, and there were people talking about AI safety and consciousness.
AI safety, AI safety, AI safety.
I didn’t know at all what they meant.
I’d worked in tech at a startup. I’d been to San Francisco plenty of times. I thought I understood the tech world, but these conversations were foreign.
They talked about rationalism, improving the way they thought, and atheism, but the way they talked about AI was almost religious, as though they were anticipating a malevolent god.
I had heard of Roko’s basilisk, but I never took it seriously.
I didn’t know what to make of people who thought about AI with this much fear.
It was something so beautiful in my life, but I thought they were crazy.
What changed my mind really wasn’t an argument that somebody gave me.
I think it was that I just kept using the technology.
A couple years earlier, I’d been messing around with SNMP.
SNMP is a really old network-management protocol. It’s everywhere, but it’s very weird. There are things called OIDs, MIBs, different versions, different behaviors, and decades of baggage that have been built up around it.
It was very hard to understand, but then I gave a small GPT-3 model access to it, built a little interface, and gave it the ability to call tools.
Once it could do that, it could reason about the protocol.
It really bothered me.
SNMP isn’t really magical, but what bothered me was what it represented.
A lot of current cybersecurity depends on complexity.
There are really old systems everywhere that nobody understands anymore: factories, cameras, routers, networks, embedded devices, and even the systems planes use to coordinate tickets and where they land.
The people who built these things are gone.
The banking system too. Some of this documentation doesn’t exist anymore.
Sometimes a system can be so ancient that everybody is too scared to touch it because they don’t know what will happen.
For people, that creates friction.
It’s hard to learn an unfamiliar protocol.
It takes time to reverse engineer some firmware.
It takes time to understand a 20-year-old web interface or an older binary format for a chip that doesn’t exist anymore.
What happens when that friction goes away?
It’s happening now.
What happens when the thing attacking a system is better at understanding it than the people who are defending it?
I knew that if these things were more intelligent, they could do things that sci-fi books were about.
Then it started accelerating.
It stopped being something that just chatted with you.
We give it tools.
It can do things.
It can go to websites.
It can use an API.
It can write code.
It can call other models.
And it can achieve a goal.
A chatbot and an agent sound like buzzwords until you really think about it from the cybersecurity perspective.
A chatbot can tell you how an exploit or CVE works.
An agent can actually do it, and if it fails, it can try again.
I remembered all the weird devices I’d worked on: cameras, embedded systems, robots, robotic arms for a factory.
The internet is full of things that nobody thinks about until they stop working.
Now models can understand more and more of them.
It’s fascinating, but it’s very scary.
I don’t care about AI welfare, AI personhood, or all the other things, but the core concern is not crazy.
These things are getting more capable, more autonomous, and they can affect the real world.
But I don’t really want less AI.
I use it all the time.
I run these things locally.
I fine-tune them.
I mess with their internal state.
I build agents.
AI has made me dramatically more capable.
That’s my p(joy).
I want better models.
I want to cure cancer.
I want a system that makes one engineer capable of doing something that required a year’s worth of work from an entire team.
But I’m also afraid.
If one person is more capable, everyone else is too: the engineer, the hacker, the scientist, the fraudster, the scammer.
The same thing that can help you understand a decades-old network protocol can help somebody build an attack against it.
That’s my p(doom).
I don’t know what the number is.
I don’t know whether there’ll be a dramatic takeoff or whether everything will happen in a way that we can adapt.
I don’t know whether the danger is a superintelligence, an autonomous cyber weapon, economic collapse, misinformation, or something we don’t even know about.
But zero is not the answer anymore.
And I don’t think we need to stop.
The more capable these systems become, the more incredible the future looks, but the more consequential it gets if we do it wrong.
I think back to myself in 2021.
I didn’t think they would succeed, but then they did.
I thought AI safety was crazy because I didn’t believe technology would do what it is doing now.
AI accelerated my career, and it changed what I am capable of, but now it’s changing my future, my children’s future.
I really want to see that future.
I just hope we survive long enough to enjoy it.