I rant for 9 minutes about company stupidity

MMaximilian Schwarzmüller
컴퓨터/소프트웨어경제 뉴스경영/리더십

Transcript

00:00:00Let's talk about one of the most stupid trends we saw over the last couple of weeks and months,
00:00:05which, as it seems, is coming to an end already. Rightfully so, because it makes no sense.
00:00:12Token maxing. Token maxing, in case you don't know, is simply about using, or one could say
00:00:18burning, as many AI tokens as you possibly can per month, per year, whatever time period you
00:00:24were measuring. So, the idea, from a company perspective, because this is a term coming from
00:00:30the enterprise world, the idea really is that you want to incentivize your employees to use as many
00:00:37AI tokens, for example, through tools like Cloud Code. And just as a side note, that is a useful tool,
00:00:44just like Codex and these other tools. You can get work done through them. I got courses on Cloud Code
00:00:50and Codex if you want to learn more. They're really in-depth and show you some tips and tricks. But
00:00:54the idea is that you use these tools to burn or use as many tokens as possible, because
00:01:00that will give you great outputs, right? No. As mentioned, these tools are valuable. As a developer,
00:01:09I believe you need to be able to work with these tools, but use them as assistance. The idea behind
00:01:16token maxing, or the incentive behind token maxing, clearly, of course, is that you just waste tokens in
00:01:23the end, that you mindlessly spend them, that you prompt after prompt after prompt, that you look at the
00:01:29output as little as possible, or not at all, of course, because that will just keep you from prompting
00:01:36more. We've heard about companies having internal leaderboards, where the people that spend the most
00:01:42tokens would, well, be on top and potentially get some rewards. And of course, that makes no sense. And of course,
00:01:50I'm mostly talking about AI being used for development here, because that's where I'm coming from. But I would
00:01:57say it doesn't make sense in any context. But especially if we're talking about using AI for writing
00:02:03code or for generating code, you want to understand and review that code. It's not about spitting out as
00:02:11much code as possible. It never was. Even before AI, it was not a good idea to measure the productivity of a
00:02:20developer by the lines of code they can write on a given day. And it's not different with AI. The quality
00:02:27matters. And I know that this seems to be something not all companies would agree with these days. But yes,
00:02:35it does. If you go down the, in the end, wipe coding rabbit hole, and you have AI generate all that
00:02:43spaghetti code, and you totally lose track of what's going on, and you don't understand what the code is
00:02:50about, and you were not able to dive into the code manually at some point, because it's just too much,
00:02:55then you lost. Then you truly lost. Because AI is far from perfect, as we probably all know. So of course,
00:03:03we need that human touch, that human control, to use AI efficiently and to get good results. And
00:03:11that's why I've been saying for all my videos, and I still strongly believe in that, AI can be a useful
00:03:17tool. But it's a tool. It's not the replacement of developers, no matter how much certain companies are
00:03:23wishing for that. And of course, the entire idea behind token maxing in the end is that, hopefully,
00:03:30from the company's perspective, you can get to a point where your employees are using AI so much that
00:03:37they are producing lots of amazing output. And as a company, you can at some point say, great, now we
00:03:42don't need all those employees anymore, or at least we need less of them. Now, turns out this is not
00:03:49working out too well. There is this report about Uber, which got quite popular on X, for example, over the
00:03:57last weeks, where the Uber COO, and I think also their CTO, in the end mentioned that they burned through
00:04:05their entire 2026 AI budget in four months. So they had a budget of tokens they wanted to pay for or use,
00:04:14and they burned through it within a few months. Now, of course, and that's important, I think,
00:04:20one reason is that the budget was likely set in 2025, one would assume, or at the end of 2025 or early
00:04:292026. And then we had agentic coding take off in early 2026. So that happened. And that happened,
00:04:40of course, because certain models like Opus 4.5, but also GPT 5.4, or Codex before that, got really
00:04:50good or got better, especially at following instructions at the end of last year. And then
00:04:55those tools, Cloud Code, Codex, on which I have those courses I mentioned, which are great, the tools and
00:05:00the courses, those also got better and used those models really efficiently. And of course, also other
00:05:05tools like Pi, which is an amazing coding agent, and so on. Now, the combination of that led to more
00:05:12usage of these tools. But since we're talking about agentic coding here, where these tools,
00:05:18or where the models in these tools think and use tools, call tools, do searches, analyze the search
00:05:25results. That all costs way more tokens than the way we or these companies used AI last year, where it was
00:05:35shorter sessions, not so many long running agentic sessions. And of course, the longer a session runs,
00:05:40the more tokens it burns. So the calculation that happened in 2025 has nothing to do with the reality
00:05:47of how AI is being used in 2026 with those enhanced agentic coding models and the tools around them.
00:05:55But nonetheless, Uber burned through its entire budget. Now, if they were getting amazing results,
00:06:02they would surely increase their budget, but doesn't look like that is what happened. An NVIDIA executive
00:06:10also said the cost of compute is far beyond the cost of employees. So it's more expensive right now to
00:06:18use AI than to use humans. Now, of course, you could say doesn't matter if AI is 10 times as productive
00:06:25as a human employee. It's fine if it's 10 times or eight times as expensive, right? Maybe it would even
00:06:31be fine if it were 15 times as expensive because it can get even better, whereas for the human
00:06:39productivity, it can also increase, but probably not as sharply as that of AI.
00:06:45But we're also not near to these numbers, 10x, 15x, because again, the number of lines of code generated is
00:06:54not a good measure. And we need human employees with their experience, with their empathy, with their
00:07:01understanding of a code base, with their connection to other departments and a company, with all those
00:07:08nuances that make up a job. Of course, with all the trust that is assigned to a human. And of course,
00:07:15also with their deep understanding of what makes a good code base, what will likely come next in a
00:07:21code base, which future capabilities may be needed. All things AI models are missing, of course. So it's so
00:07:29stupid from so many different angles to compare productivity from AI models with human productivity.
00:07:36And the first companies are seeing that, I think. Which is why all that token maxing here is coming to
00:07:43an end. You can read about more and more companies like Amazon, Meta and many, many others that are
00:07:48cutting back on their token leaderboards, that are cutting back on their AI budgets or on the token
00:07:54maxing approach here. And I truly hope, I don't know though, I hope that we'll soon enter an era where
00:08:02things will settle down a bit more. AI is here to stay and AI is useful. It's a useful tool.
00:08:09It can make you more productive. It's great for doing extra research. It's great for producing that
00:08:15boilerplate code or also the non boilerplate code. But based on clearly defined specifications with human
00:08:22review, ideally based on some code base that was at least shaped and fine tuned by a human, AI can be
00:08:30really useful there. And it can even be useful for wipe coding if you need a little tool that just does
00:08:38something you need to get done right now, which you don't plan on publishing to the world, where you don't
00:08:43care about all the bugs and where you will not add a lot of features, which you don't have to maintain.
00:08:48It can be great for that too, for these one-off tools. There are many great use cases for AI and
00:08:55it's a technology that's here that will stay and that will become better, of course. And nobody knows
00:09:00what will be the case in 10 years or so. But right now, I really hope things will settle down a bit more
00:09:07and we'll use AI for what it is, a useful tool, but not that magic thing right now that changes
00:09:15everything and we'll get rid of all the jobs and we'll replace all employees and all humans within
00:09:20the next 12 months or so. And it looks like, probably for publicity reasons though, that even our favorite
00:09:28tech CEOs, Sam Altman and especially also Dario Amodei, are kind of retreating regarding those pretty
00:09:36strong statements of how soon AI will replace pretty much all white collar work, right? Sam Altman said
00:09:45in an interview that he was pretty wrong about AI's economic impact. And Anthropic CEO Dario Amodei,
00:09:52who not so long ago mentioned that most or pretty much all of white collar work will be replaced by AI
00:09:59relatively soon, now says automation may actually expand the work people do. Probably though,
00:10:06because their PR department told them that whilst it's amazing for selling their tools to companies
00:10:13when they say how many employees they can replace, it's not so amazing if the entire world turns against
00:10:21them. So I didn't care too much about their statements before and I still don't now that they reverse
00:10:28them, I always was pretty convinced that nowhere near in the future will AI replace all white collar work.
00:10:37I'm sure it will actually lead more to more work. That has been the case with all those technological
00:10:43breakthroughs. And like with all of them, we just don't see how future roles will look like. But when we
00:10:48take a look at coding, we're not even close to the point where you would want to let AI write all the
00:10:56code and not care about it at all for any serious product. At least I definitely wouldn't and I think
00:11:03any company that would, would make grave mistakes. But as it seems, companies also are hopefully starting to
00:11:11realize that AI is better used as a great tool instead of an all-in do-everything solution.

Key Takeaway

Companies are curbing aggressive 'token maxing' strategies because the high compute costs of agentic coding do not yield productivity gains superior to human developers.

Highlights

  • Companies are abandoning 'token maxing' programs that incentivized employees to maximize AI token usage through internal leaderboards.

  • Uber exhausted its entire 2026 AI token budget in just four months due to the rapid adoption of token-intensive agentic coding workflows.

  • NVIDIA executives reported that the cost of compute required for AI-driven development currently exceeds the cost of employing human developers.

  • Measuring developer productivity by lines of code generated remains ineffective, even when using AI tools for code production.

  • Tech leaders like Sam Altman and Dario Amodei have shifted their public stance, moving away from claims that AI will soon replace most white-collar jobs.

Timeline

The Failure of Token Maxing

  • Token maxing programs incentivize employees to mindlessly burn AI tokens to climb internal leaderboards.
  • High token usage does not correlate with better software development output or improved code quality.
  • AI is effective only when used as an assisted tool requiring human review and understanding of the code base.

Corporations implemented incentives to maximize token consumption, mistakenly believing high volume output equals high productivity. This approach ignores that AI-generated code, especially spaghetti code, requires human oversight to be useful. Relying solely on AI to produce code without human review leads to a loss of control over the codebase.

Budget Realities and Agentic Coding

  • Uber depleted its 2026 AI budget in four months, driven by the shift to agentic coding workflows.
  • Agentic coding models incur significantly higher token costs than standard AI interactions due to tool usage and recursive reasoning.
  • The current cost of AI compute for development tasks exceeds the cost of human employee compensation.

Agentic coding, which became prominent in early 2026, involves long-running, recursive sessions that consume far more tokens than previous manual AI prompting methods. Uber's experience demonstrates that budgets calculated in 2025 failed to account for this shift. Current economics show that replacing humans with AI is not yet cost-effective.

The Future of AI in Development

  • Amazon and Meta are among the companies scaling back on AI token leaderboards and aggressive budget allocations.
  • AI remains a useful tool for specific tasks like boilerplate generation and one-off scripts, not as a complete replacement for human developers.
  • Prominent tech CEOs have retracted prior assertions that AI would rapidly replace the majority of white-collar work.

The industry is beginning to recognize the limits of AI-driven automation. Instead of replacing roles, technology historically expands the scope of work. Successful development requires human empathy, nuance, and an understanding of future product requirements, which current AI models cannot replicate.

Community Posts

No posts yet. Be the first to write about this video!

Write about this video