Mick and SPEAKER A discuss how companies are burning through AI token budgets by over-using expensive coding agents, creating "zombie programmers" who manage dozens of tools instead of reducing work. Token costs now outpace salaries, raising fears of layoffs and lost institutional knowledge.
We've been hearing this, this, this term token maxing, right? There have been companies who have blown through their budget for the entire year within just a couple of months. That is referred to sometimes as token maxing. Mick, can you explain what that means?
What we're seeing is a lot of companies having this realization of, wow, these coding agents are tremendously powerful and we just need to get people using them as much as possible. And so one of the incentives is set up a leaderboard, right? Reward the people who are using them the most. Sounds great, but like you say, particularly when you're talking about the very frontier models, that comes with a significant cost. I heard a startup CEO interviewed yesterday talking about how the token budget's bigger than the salaries budget. There is this question of, does this mean I need to reduce my headcount, right? And so there is this fear among, among employees that AI is essentially here for their jobs. There's a real danger of people going so hard at AI at this maximum frontier model cost, and they'll lose all of the institutional kind of knowledge and understanding of their business that was residing in those people.
There's been this concept that, oh, we're going to need to work less, right? So because these tools are just going to do all the work for us. But what we're seeing is the inverse has happened. They're able to build so much that these programmers are seeing all these things that are capable, all these things that they dreamed of doing, all these bugs that they needed to all these new programs and abilities that they wanted to do. And so there's this FOMO. There's a new term called like zombie programmers, which are, you know, programmers that are essentially they cannot go to sleep because they have 20 different agents building all these different things for them and they need to manage them. So they're afraid if they're going to go to sleep, they're going to be less productive.
Early in the kind of post-ChatGPT moment where people were starting to imagine new possibilities, I saw the argument being made that as a result we would have less software engineers, right, going forward was wrong because actually we have a huge deficit of software in the world. We need way more software.
Yes, it's the cost is going down. It's much easier. It's much faster to create applications. But at the same time, if you're just going to use Opus 4.8 and just have it like work 24/7 and build these things and you're not managing it correctly, choosing between which models you're using for which tasks, It can really rack up your bill very, very quickly.
