State of Agentic Coding #7 with Armin and Ben
Every voice is named by two or more agreeing sources.
“Ruby actually is in many ways the worst language you can pick for AI.”
Armin argues that Ruby is poorly suited for AI because of its design choices.
“in July, we started writing the first code for Lephos”
Armin recounts that the first code for their agent Lephos was written in July.
“something like $1.2 million in raw token cost a month”
Peter Steinberger reportedly spends about $1.2 million on tokens each month.
“I don't think that scales down to the world of you and me”
Armin doubts that high‑token‑spend approaches can be applied to typical developers.
“There has never been a time when there's been this much asymmetry between the haves and the have-nots in programming.”
Ben believes the current gap between resource‑rich and resource‑poor developers is unprecedented.
“I'm paying for an account right now. $200 that they basically don't use, right? It's free money for them.”
Armin notes that Anthropic earns profit from unused subscription money.
“Well, I think there's another part, and this is my conspiracy theory, but there's a, there's a big difference if what the machine is doing is the input from a machine versus what the machine is doing is the input of a human.”
Armin predicts that Anthropic will favor human‑generated prompts because they provide more valuable data.
“Well, when Opus 4.7, was it 4.6 launched? I forgot. They changed the tokenizer. So everything became magically 20% more expensive.”
A tokenizer update in Opus made token usage about 20% more costly.
“But in some sense that, that number didn't actually really that you could change that number. But even if you change that number, it didn't automatically also change your revenue.”
Changing usage metrics like MAU doesn’t automatically boost revenue.
“Never before in our industry have we had a sort of metric, a usage metric that correlates so closely to dollars, right?”
He believes token usage is a uniquely dollar‑linked metric.
“I want to make a comment here that Enron, the famous, the famous, um, 2001 bankruptcy. I saw the documentary a long time ago, not the documentary, there's a fictionalized case. The Smartest People in the Room.”
Ben references the Enron story and a related documentary.
“The most obvious place right now where companies see the need/desire to spend tokens is to solve security problems.”
Token spend is primarily motivated by security research needs.
“It's an incredible product because it both creates problems. And it solves the problems you created and you pay both times.”
He sees AI security tools as a double‑edged sword that costs users twice.
“Your comment being like, look, if, if agents are just producing assembly, no one will be able to evaluate anything.”
Ben disagrees with the idea that LLMs will output raw assembly that humans can’t assess.
“Finally, COBOL can die.”
Armin declares that COBOL is effectively obsolete.
“I think open weight models should exist.”
Armin expresses belief that open weight models need to be available.
“you can download my extension for Py, which after you installed it gets you the model, the inference engine, the kernel, the everything that you need running on your computer without you having to do anything.”
Armin recommends using his Py extension to easily install and run a local model.
“One of the takeaways here is if you are, if you think you want to run a local model, Uh, you might want the 16‑inch.”
Ben advises using a larger 16‑inch Mac for running local models due to power and cooling constraints.
“I was using DeepSeek v4 Flash, but not the— not DwarfStar 4, not the quantized version, just the original. And I was working with it. It would, you know, aggressively refactor things in the ways that I didn't want it to.”
Ben describes his livestream where the model kept refactoring his code in unwanted ways.
“Microsoft released one. Microsoft, I think, even called it like ethically trained or something, which I found interesting.”
Armin notes that Microsoft has released an ethically trained open weight model.
“It's 218 billion total, 25 billion active parameters.”
Cohere's Command‑A+ model has 218 billion total parameters, of which 25 billion are active.
“my suspicion is that pretty much every single company next year is going to pay for tokens in one form or another, at the very least for security reasons.”
Armin predicts that essentially all companies will have to purchase AI tokens for security purposes within the next year.
“you will not be able to run a software company and not pay a ton of money to get security audits and security reviews and all of this going because someone else will do.”
Armin believes software companies will soon be forced to spend heavily on security audits or be disadvantaged.
“I think dead internet theory is real and I think that it's actually gonna sort of like, I actually think it will negatively impact our progress on things.”
He predicts that a dead internet will hinder technological progress.
“90% of issues are agent-generated.”
He says the majority of issue‑tracker entries are created by AI, not humans.
“This is why maybe the crash is going to come in the next 12 months. I don't know.”
He speculates a crash could happen within a year.
“I have a project called Slop Scan, which tries to identify like slop TypeScript coding patterns. And the way that, that the origin of it is that I pin a bunch of notable open source projects from a variety of different, like solving different problems, like a web app or a library. And I pin them. I call it the epoch date, like of the before times. I put January 1st, 2025.”
He explains his Slop Scan tool that compares pre‑AI open‑source code to current code using a fixed epoch date of Jan 1 2025.
“My advice to you is have some conviction and work on things a little bit longer.”
He advises developers to stick with projects longer instead of abandoning them quickly.
“I don’t believe in the Ralph loop, agentic loop.”
Ben expresses skepticism about the effectiveness of the so‑called Ralph/agentic loop for library creation.
“I’m pretty sure that a year from now, there will be a lot of AI-generated code in the kernel because it will be very hard to keep it away.”
Armin predicts AI‑generated code will become common in the Linux kernel within a year.
“Courts have already indicated that. A machine can generate.”
Armin notes that courts have ruled a machine can generate code, implying legal precedent.
“There’s actually only a handful of countries that even have, possess anybody making models, right? And as far as I know, that’s obviously the US and China.”
Ben points out that only a few nations, mainly the US and China, develop large AI models.
“I’ll make one prediction that I think one tremendous discovery will happen in a year.”
Ben predicts a major, yet‑unknown breakthrough will occur within the next year.