He also revealed that AI adoption has expanded beyond engineering to include design, planning and financial sectors, and that AI is even contributing to internal decision-making processes. “We're testing that limit. Can we actually start to become decision makers on a few things and do better than humans?” Armstrong said.
John Collison: What other ways does Coinbase encrypted AI pildo work differently than a company founded 10 or 20 years ago?
Brian Armstrong. Like many companies, we are leaning as hard as we can towards AI.
What does that mean, and what exactly does it mean?
We do a lot of best practices. We made a big push to take all the engineers to the cursor and the copilot. Then another question was, “Well, are they actually going to use it?” Because many of them were on my own –
Was it you or Tobi? [Lütke, member of Coinbase board] Who mandated it?
I obliged it. Yeah.
You have requested people to call you –
it's true. I did that.
If you were not using AI code, you had to justify your CEO.
it's true. Originally –
Sorry, but maybe I'm not going to talk about it?
No, I don't care. That's actually a good story. They were originally coming back and saying, “Well, the next quarter will be 50% hires over two quarters.” I said, “You're telling me – why can't all the engineers be on board by the end of the week?” So I went a bit cheating. Posted on the All-Inslack Channel.
Light dusting in founder mode.
I said, “Ai is important. We need all of you to learn it and at least have it onboard. You don't need to use it every day until you do your training, but at least by the end of the week. If not, we have a meeting with everyone who doesn't do it on Saturday. I want to meet you to understand why.”
A few people were on vacation now. There was a list of. Anyway, I jumped on this call on Saturday and there were a few people who didn't do it. They just came back from a trip etc so some of them had good reason, and some of them didn't, so they were fired.
oh.
Incidentally, some people really didn't like that strong approach, but I think it set the clarity that we should at least lean on this and learn about it.
What have you experienced with AI coding so far? It is clear that helping AI write code is extremely helpful. It's not clear how to run an AI-coded codebase and what is the best way to do it.
I agree. I think we still understand that.
One of the things we started is hosting each month. We call it AI Speed Run, where one of the engineers volunteers that month and trains them on how they use it. We try to cherry pick people, the team that's doing it best.
Currently, we currently carry out approximately 33% of the code written by AI. We have a goal of reaching 50% by the end of the quarter. Let's see if we get there. You can probably go too far with that. You don't want the atmosphere where people run and code these systems. We've really been encouraging people – you need to review it and get the right checks on it. However, some parts, such as the front end, can iterate faster.
We want to make sure that it is being used not only in the engineering team. It really should be which team. The design uses it frequently. Product Manager. I think FP&A uses this as “take all the data and tell me what the revenue is.”
Incidentally, I use it a lot as a CEO. Using a decision-making process called Rapids, everyone writes their input. AI has a line of AI that writes input as one of the people who help you make decisions. We are testing the limits of it – we can actually start to become decision makers on some things and do better than humans.
