I consider myself fortunate enough to work with people on the forefront of computer science, data science, journalism, and the humanities. Sometimes my conversations with these people send me into a tailspin.
The other day, I had a conversation with a computer scientist. By all accounts, including mine, a true genius. He asked me if I had used Claude Code, I replied that I have actually been using OpenCode, but only because I prefer using open-source software. We continued to talk about AI’s coding abilities and such. He asked something that I cannot quite remember, but I remember my reply being along the lines of “I only really use it for research. I like to write my own code. […] I really really like writing code. It is something that engages my mind. I don’t like letting other people/things do my thinking for me.” I do remember his reply very clearly: “very interesting.”
It has, to a certain extent, sent me into a kind of quarter-life crisis as I approach my quarter-century birthday in two days. I now can no longer tell myself that AI is slower in the long-term than me at coding, especially for relatively simple tasks. Again, I really enjoy the challenge that coding brings me. I wrote a RAG CLI app as a challenge for me to learn Rust over the course of a few days. Could an AI have built that in less than 10 minutes? Certainly. Could it do all of my daily work much faster, with fewer mistakes, and (in the short run) cheaper? Certainly.
In this, I see two options: the first is locking the fuck in, start doing industrial quantities of amphetamines and nootropics, become better and faster than the AI at coding, and kill God; the second more feasible option is to lock the fuck in and read a bunch of Anthropic engineers’ substack blogs to learn how to better prompt AI. Perhaps I have a paranoiac perception of AI. In the long-run, humans will adapt to a new economic reality with AI, just as they did with computers, just as they did with the internet, just as they did with the introduction of currency. Human art will never really go away any time soon. What I am paranoid about is the occluding of human intellect behind a paywall. If the Anthropic API goes down, then who will respond to security incidents? Who knows our network topology? If and when there is a state-sponsored targeting of AI data centers, will that be considered a supply chain attach? Will a threat-actor use that opportunity to launch a massive cyber-attach against which AI cannot defend?
We are creating the most incredible single point of failure in the history of human kind. In the past, we have relied on humans doing incredible feats of simultaneous intellect and physicality to get us out of emergencies. I asked Claude to spend six hours and $800 (Sorry Northwestern) to build a system for digitizing books. It needed authentication, it needed model access, it needed database. It built a modern react application. I do not know how to use JavaScript and I am utterly uninterested in learning it. Claude had built a massive network of APIs, routes, database queries, etc. When it kept failing to implement a feature. I had no idea how to even use the system. I did not know how to use ECS and Fargate and Lambda and Step Functions and all that shit. Again, I am uninterested in learning how when I can just spend $1000 on a GPU and get a better price/performance ratio that I can just plug a keyboard into.
My solution was far simpler, but more importantly, I know how to turn it off in an emergency. I know how to rotate the keys. I know how to query the database. I know where the docker containers are running. All my solution does is just load a bunch of compute jobs into the Fireworks workflow management database, I manually spin up a bunch of EC2 instances, and I just check on it every hour or so. It works fine. It is like 200 lines of code. It does not require VPCs and ECS instances that I do not understand nor have the time to learn.
I think this is what AI will fail to do. It can just… turn off. Humans cannot just turn off. I can, have, and will get up at 2AM to have a call with some guy on the other side of the world because our databases just surpassed a latency threshold. When AI is asleep, it cannot be roused. When it asks for double the money, there is no negotiation and there is no competition.
When I mean AI is asleep. I mean that the data centers are offline. Another critical infrastructure target for sophisticated state-sponsored attacks. One will respond with the (true) assertion that there are always Open-Source models to run locally. These models are indeed competitive with frontier models. The problem is that who can afford the compute to run them on their own hardware? Individuals cannot. Most of them anyway. Organizations, meanwhile, are married to the Cloud! Consistently relying on other people! With no backups in the (inevitable) future where a sophisticated state actor has the bright idea to attack multiple availability zones at once! Hell, if the power goes out in Ohio, Virginia, and Oregon at the same time!
This has turned into a long rambling rant about my paranoid (possible) delusions about Cloud infrastructure and the folly of putting AI systems into it and letting AI systems configure it.
So what’s a girl to do? AI is clearly very good at some things. It already has revolutionized information retrieval on an academic and public level and can code decently well in Python and JavaScript. I admittedly do find it very helpful when trying to read and understand mathematics, as I have had relatively little formal training in the area.
I don’t think I really have an answer to this question. It is something I have been troubled by for many months now. My most recent attempt to solve the problem was to attempt to convince my business unit to purchase a DGX Spark for our “lab”. But funders and donors are less than willing to purchase such a thing nowadays. Something something AI use is a dialectical process something something.
Perhaps my issue is that I ascribe so much magic to the feeling of coding. There is something that just feels right in my brain when I am in a flow state and creating something very cool. Maybe I just need to find a way to get into the flow state while being assisted with AI.