In Defense of Tokenmaxxing

0:00Today on the AI Daily Brief, a defense of token maxing. the controversial practice of incentivizing employees to spend as many AI tokens as they can. Before that, in the headlines, Google starts dropping announcements ahead of I.O., including the new Gemini Intelligence. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI. All right, friends, quick announcements before we dive in. First of all, thank you to today's sponsors, KPMG, Granola, Superintelligent, and Zencoder. To get an ad-free version of the show, go to patreon.com slash ai daily brief, or you can subscribe on Apple podcasts.

0:36To learn more about sponsoring the show, send us a note at sponsors at ai daily brief dot AI. Now, two other quick announcements. First, registrations are live for enterprise claw cohort three. You can get a link from the ai daily brief site or at enterprise claw dot AI. And second, as I mentioned yesterday, I am hiring a growth engineer for the podcast and podcast ecosystem. All of these crazy things we do, like the context portfolio builder, and these free education programs. That's the type of stuff you're going to get to build as a growth engineer. Your job will be to expand not only the audience of the podcast, but the impact of the audience.

1:10You can find information about that at jobs.ai, dailybrief.ai. We are actively recruiting. I am looking to hire fast, and it is a full-time role. And with that, let's get to the headlines. We got a lot today, so let's cook. First up, Google has announced a new agentic suite for Android users called Gemini Intelligence. Framing their vision, Google wrote, as Android transitions from an operating system into an intelligent system, your devices are becoming even more helpful with upgrades that will save you time. Now, you might be thinking to yourself, wait, isn't Google I.O. just around the corner?

1:45Why are we getting announcements now? Well, we've actually seen this for the last couple of years, where there's so much that Google has to announce at their big I.O. event that they actually start dropping things the week before. I will say when they're introducing an entire new agentic suite In advance of the event, you gotta wonder what's gonna come at that event. But in any case, Gemini Intelligence will include a major upgrade to the Gemini Assistant, allowing it to handle more complex tasks and multi-step processes, as well as a new feature called Personal Intelligence, which is Google's AI memory system.

2:16Gemini Intelligence will roll out to the latest generation of Google and Samsung headsets over the summer and will become available on smartwatches, glasses, and laptops later this year. Speaking of laptops, also on Tuesday, Google unveiled the Google Book, a new iteration of the Chromebook designed with AI in mind. The device will now run on a mix of Android and Chrome OS, allowing handset features to be easily migrated across. The Google Book will have a built-in AI assistant built on the same Gemini intelligence stack with a bunch of new modes of interaction. For example, instead of learning a new hotkey to summon the Gemini assistant, users can simply jiggle the mouse and Gemini will pop up.

2:51DeepMind actually showed a research demo of where the concept of an AI-enhanced mouse pointer is going. The demo showed a user gesturing with the mouse while giving voice instructions, asking the AI to do things like, add these two ingredients to my shopping list without naming them. This seems directly related to the conversation we were having yesterday about interaction models, where the next generation of AI interfaces is about having them interact more like we do rather than us having to learn how to interface with them. Now there is some competitive weirdness here, given that Google is powering the new Siri, but it also named its toolset Gemini Intelligence after Apple named its Apple Intelligence.

3:26This frenemy era is frankly very hard to keep a handle on. Google is also the latest company taking a closer look at orbital data centers. The Wall Street Journal reports that Google is in talks with SpaceX to launch data centers into space. Sources said that those exploratory talks are also being held with other rocket companies, with Google planning to have their first prototypes in orbit by next year. While to some orbital data centers seemed like a fantastical sci-fi concept, there has been a major groundswell over recent months. As part of their recent deal with SpaceX, Anthropic seemed to express a genuine interest in orbital data centers as a way to get around land permitting issues.

3:59There's also been a huge surge in new startups pursuing the idea. This week, for example, Robinhood co-founder Baju Bhatt unveiled his new startup Space Cowboy Corp and announced fundraising at$2 billion. In an accompanying article, the Wall Street Journal asked whether data centers in space were just a pipe dream or the next big thing in AI, noting that even NVIDIA is getting on board, recently posting a job ad for an orbital data center system architect. This, I think, is going to continue to be a growing theme. But before we leave Google, there is one additional story on that front. Just a day after OpenAI confirmed their consulting plans, and a week after Anthropic announced their new initiative, Google is apparently jumping on the AI consulting bandwagon as well, with a new plan to hire hundreds of forward-deployed engineers.

4:42The new group will be housed within Google Cloud as part of the go-to-market team. Google Cloud CEO Thomas Kurian wrote in a LinkedIn post, While having FTEs is not new for Google Cloud, the demand from customers and partners for Google Enterprise AI products and Google engineers to help them embrace agent development is growing very rapidly. In a separate post, Chief Revenue Officer Matt Renner noted that the way AI services are sold looks very different to traditional cloud services. He wrote that adding hundreds of FTEs would help Google Google is also apparently taking the fight directly to Anthropic and OpenAI with their own private equity partnerships.

5:18The information reported that Google is in talks with Blackstone, KKR, and QT to deploy Google's AI products throughout their portfolio companies. The very clear takeaway is that the AI race is no longer just about model performance and benchmarks, but has a major new dimension in model deployment. Moving on to Anthropic, the company is attacking their next vertical with the expansion of Claude for Legal. The legal plugin for Cowork was first released earlier this year, and since then, Anthropic says legal professionals have become the most engaged users among knowledge workers. This release is similar to the Claude for Finance update from last week, consisting of a series of new connectors and pre-built agents.

5:53Anthropic has added connectors for dozens of legal tools, including DocuSign, Trellis, and Thomson Reuters Co-Counsel. Interestingly, Cowork can also now connect directly to Harvey, the largest legal-specific AI startup. This means legal professionals can use Cowork as their agentic harness to interface with Harvey's legal knowledge base and reasoning engine. There's also a suite of 12 pre-built agents designed around specific practice areas areas, including commercial law, regulation monitoring, employment law, IP, and client management. Anthropic Associate General Counsel Mark Pike said that the launch of Coworks specifically had massively boosted the number of legal professionals using Claude.

6:26In fact, he said that lawyers are now the most frequent users of any profession other than software engineers. Now, one of the things that's worth watching is whether we start to see a convergence in how the big labs take on knowledge work. Anthropic has now a pretty consistent and clear pattern. They roll out packages of connectors and pre-built agents designed to give professionals a solid set of AI workflows right out of the box. Each package is marketed under its own branding, in this case clawed for legal, but forms part of the overall co-work platform. In contrast, at least at the moment, OpenAI appears to be sticking to their super app strategy.

6:57Knowledge workers are all routed to codecs, where there are starting to be some basic pre-packaged connectors, but not the sort of branded codecs for legal experience that we're getting from Anthropic. I wonder if that will change as more and more knowledge work gets consolidated around those interfaces. For now, though, we're actually going to close the headlines there because Maine gets a little long today. Something to watch for sure, but for now, that's going to do it for today's headlines. Next up, the Maine episode.

7:26One of the most important AI questions right now isn't who's using AI, it's who's using it well. KPMG and the University of Texas at Austin just analyzed 1.4 million real workplace AI interactions and found something surprising. The highest impact users aren't better prompt engineers. They treat AI like a reasoning partner. They frame problems, guide thinking, iterate, and push for better answers. And the good news? These behaviors are teachable at scale. If you're trying to move from AI access to real capability, KPMG's research on sophisticated AI collaboration is worth your time. Learn more at kpmg.com slash US slash sophisticated.

8:04That's kpmg.com slash US slash sophisticated.

8:30After the call, you can chat with your notes, ask Granola to pull out action items, help you negotiate, write a follow-up email, or even coach you using recipes which are pre-made prompts. Once you try it on a first meeting, it's hard to go without. Head to granola.ai slash ai daily and use code AI daily. New users get 100 % off for the first three months. Again, that's granola.ai slash ai daily. It is a truth universally acknowledged that if your enterprise AI strategy is trying to buy the right AI tools, you don't have an enterprise AI strategy. Turns out that AI adoption is complex. It involves not only use cases, but systems integration, data foundations, outcome tracking, people and skills, and governance.

9:11My company, Superintelligent, provides voice agent-driven assessments that map your organizational maturity against industry benchmarks against all of these dimensions. If you want to find out more about how that works, go to bsuper.ai. And when you fill out the Get Started form, mention Maturity Maps. Again, that's bsuper.ai. So coding agents are basically solved at this point. They're incredible at writing code. But here's the thing nobody talks about. Coding is maybe a quarter of an engineer's actual day. The rest is stand-ups, stakeholder updates, meeting prep, chasing context across six different tools.

9:44And it's not just engineers. Sales spends more time assembling proposals than selling. Finance is manually chasing subscription requests. Marketing finds out what shipped two weeks after it merged. Zencoder just launched Zenflow Work. It takes their orchestration engine, the same one already powering coding agents, and connects it to your daily tools. Jira, Gmail, Google Docs, Linear, Calendar, Notion. It runs goal-driven workflows that actually finish. Your stand-up brief is written before you sit down. Review cycle coming up? It pulls six months of tickets and writes the prep doc. Now you might be thinking, didn't OpenClaw try to do this?

10:15It did, but it has come with a whole host of security and functional issues, which can take a huge amount of time to resolve. Zencoder took a different approach. SOC 2 Type 2 certified, curated integrations, tighter security perimeter, enterprise grade from day one, model agnostic, and works from Slack or Telegram. Try it at zenflow.free. Welcome back to the AI Daily Brief. When it comes to today's episode, honestly, my hand was forced. There was just too much lazy thinking getting too loud for me to not wade into this conversation. And I actually genuinely think it matters because lots and lots of, by which I mean every company, is right now trying to figure out how to address AI adoption.

10:55And so when we have lazy thinking that rejects something out of hand, as is happening now with this idea of token maxing, I think it can be actively harmful to people and companies who are trying to figure out what they're supposed to be doing in this quickly changing world. What I'm discussing today is the new AI actually isn't good narrative. It also conveniently has the new AI bubble narrative all wrapped up in it. And in short, what the narrative is, you're wasting tokens. So what's going on? Well, right now we are, of course, in the midst of a shift in the way AI is being used. You can think about this loosely in the prioritization shift among the frontier model labs in seeing themselves as moving from selling seats to selling tokens.

11:35Now, this is not just a business model shift. this is reflective of a work shift from assisted AI, where AI is helping me do the things that I do, to agentic AI, where my job is no longer to produce things, but instead to set up the conditions in which agents can do things for me. In that agentic paradigm, success is not just in how many people are using their ChatGPT or Cloud subscriptions, it's in what they're doing with them. And frankly, this presents a challenge for enterprises. It was already hard enough just trying to get people to integrate chatbots into their work process, despite the easy productivity gains and improvements to how people could do their existing set of work.

12:11We talk all the time at this show about the capability overhang, which is the space between what AI is capable of and what organizations are actually getting out of it. And that capability overhang, that gap, is in the agentic era getting nothing but bigger. And the important thing, and the premise from which all other parts of my argument will stem, is that there is no way to figure out the best ways to use agents without experimentation. People simply have to go try things. You have to hack and build and see what works. And along the way, you're going to abandon a lot of half-done projects. The problem, at least according to some, has come in how enterprises are trying to scorecard this.

12:50Earlier this year, we started to see stories in the press about how companies were doing things like creating token leaderboards, where they were incentivizing employees to use more AI. Effectively, how many tokens you consume was used as a proxy for how good at using AI you were, and the people who were using the most tokens were lauded above all. Kevin Roos at the New York Times wrote about this all the way back towards the end of March. He wrote, An engineer at OpenAI processed 210 billion tokens, enough text to fill Wikipedia 33 times through the company's AI models over the last week, the most of any employee.

13:25At Anthropic, a single user of the company's AI coding system, Cloud Code, racked up a bill of more than$150 ,000 in a month. and at tech companies like Meta and Shopify, managers have started to factor AI use into performance reviews, rewarding workers who make heavy use of AI tools and chastening those who don't. This, Kevin writes, is the new reality for coders. AI was supposed to help tech companies boost productivity and cut costs, but it has also created an expensive new status game known as token maxing. Now, even in this first piece, Kevin got that this was a direct byproduct of the shift from assisted to agentic AI.

13:57And again, the new phenomenon was not just wanting people to use a lot more AI, it was incentivizing them. And so really, the article was exploring two things simultaneously. One was the underlying shift in what it meant to use AI for work, and the second was the new incentive structures companies were putting around it. It is that second part that has generated much more attention. A couple weeks later, the information reported that Meta employees were competing in this token-maxing sort of way. Their sources told of a new internal leaderboard at Meta that aggregated AI usage from across Meta's 85 ,000 employees listing the top 250 power users.

14:30Users near the top of the list got cool titles like Session Immortal or Token Legend. And it was very clear from the very beginning that the journalists writing these articles thought that all of this was dumb. Indeed, the information called it Silicon Valley's newest form of conspicuous consumption. And it turns out it wasn't just Silicon Valley. A couple weeks later, Business Insider reported that Disney had an AI adoption dashboard. According to a screenshot viewed by BI, the dashboard shows things like the number of employees actively using AI, the number of requests made, the number of tokens used over a given period, and the most active AI users buy requests made and tokens used.

15:04Visa, it appears, is also giving internal awards for individuals and teams who use AI the most, and it wasn't long before the debate started to rage. On the one side, are those like the companies mentioned so far, or like AI startup Brighter, whose CEO May Habib said, it's existential for us. We're in the most competitive space that has ever existed and will ever exist. On the other hand, as the perspective valiantly summed up by Nick Hodges in Infoworld, token maxing is super dumb. Now, this whole narrative crescendoed this week based on two things. The first was a Financial Times report that claimed that Amazon staff were using AI tools for unnecessary tasks to inflate these usage scores.

15:43Wrote the Financial Times, Amazon employees are using an internal AI tool to automate non-essential tasks in a bid to show managers they are using technology more frequently. Some employees said colleagues were using the software to automate unnecessary AI activity to increase their consumption of tokens. An employee said managers are looking at it. When they track usage, it creates perverse incentives and some people are very competitive about it. Now this hit at the same time as Vasuman on Twitter went viral after he posted a Slack message that said, whoever spent$600 on Anthropic last night, great job leveraging AI.

16:12But to the person who spent$23 on Uber Eats, please remember our limit for food is$20 per meal. Now, I'm almost positive that the screenshot itself was a joke, but it doesn't really matter because the resonance of it, with 2 million people having seen this and 69 ,000 having favorited it, suggests for the state of the conversation. Now, when it comes to the idea of token maxing, by which we mean rewarding employees for using more tokens than their peers, there are some pretty obvious problems that arise. Most notably, the problem summarized in Goodhart's Law that states that when a measure becomes a target, it ceases to be a good measure.

16:46In other words, as soon as you introduce a new metric for success, people are going to game that metric. That is just human nature, it's the nature of human systems, it has always been and will always be the case. But what was interesting to me is that that's not where the conversation has really stayed. Instead, what I'm seeing is a couple of old strands of critique finding new purpose. There are two points of view that were everywhere in late 2025 that look fairly silly now. The first of those is what I'll call the Gary Marcusian or Ed Zitronian. Actually, AI isn't all that good. The idea that actually all this AI stuff was just hype and that the models weren't even that good at all the things that people said they were good at.

17:25This, of course, played directly into the second narrative that was quite popular in Q4 of last year, that AI is a bubble that will never make enough money to justify the infrastructure build out that's going into it. Now, clearly these things are related in that if AI isn't good and it can't actually get all that much better, at some point people are going to realize it and there's not going to be money on the other side to justify the investment spend. Now, to put it bluntly, these points of view look fairly silly now with the benefit of hindsight. They look silly because one, AI models have done nothing but get better and the things that we can do with them have gone nothing but up.

17:55In fact, we are in a fundamentally new paradigm of what we can do with them. And second, revenue is growing like nothing we've ever seen. And to be very crisp about this, I'm not just saying that these are the fastest growing startups of all time. I'm saying that at the scale that we are talking about, getting to tens of billions of dollars in annual revenue and seeing that revenue in the case of Anthropic this year, ADX in the first part of the year is completely without precedent in business history. And while there is still reasonable debate around how much infrastructure investment that will all justify, it's pretty undeniable that at this point, the demand for AI, i.e.

18:32tokens, radically outstrips the supply, and we're kind of going to need all that compute that's coming online. Now, as an aside, I would suggest humbly that the fact that those two points of view look fairly silly now is a good warning for people who make criticism some combination of their personality and their business model, as it makes it very hard to change your perspective as times change. Now, coming back, however, to the discussion at the moment, you might see where these things are converging. The argument that AI isn't really all that good, and that it doesn't have a business model that justifies the investment, starts to get new life if all of those tokens that are being used are for utter rubbish and non-productive things and gaming systems.

19:13It's a new nuanced version of the actually AI isn't all that good, because now the argument is not that AI isn't good, it's that you suck at using it and the stuff that you're doing isn't valuable. TXMC Trades on Twitter wrote, If you're vibe coding some cool website but not making money with it, then AI didn't create value for you. It merely accelerated your hobby. Now, relatedly, this also creates a reason to doubt the business model. Because if all of that exciting token consumption is actually just employees gaming their employers, then eventually that will come crashing as well. CNBC's Deirdre Bosa writes, AI's main demand metric tokens has decoupled from actual economic value.

19:49Like page views during the dot-com era, numbers justify the spend until they don't. Nobody special memes. They say, the demand is insane. But the demand is. With a link to the headline, Amazon employees admit to using AI unnecessarily to pump up internal usage scores. There's a bonus gotcha of, if this technology was so good, would you really have to convince people to use it? Jason McEula asked, did people need to be convinced to switch from fax machines? And since social media rewards people for feeling clever, and since being skeptical is a priori treated as being clever, people jumped all over this.

20:22Now, I think all of this is just absolutely preposterous. I want to talk about why it's preposterous, but then I want to go farther and actually defend the token maxing. So first, let's talk about why the return of the AI isn't good narrative as you're using tokens poorly and the return of the bubble narrative as people are just using tokens to game the system are both ridiculous. The logical leap that the bubble people are making, is that the activity being reported on by the Financial Times is broadly reflective of the majority of token consumption. I.e., somehow the majority of the demand is people using it for these silly, non-consequential purposes, rather than actually real valuable stuff, which will see sustained demand over time.

21:00But since we discussed Goodhart's law before, let's talk about a few other types of logical fallacies. The first is selection bias. Token maxing fraud is a story because it's the deviation. People using AI for value isn't a story, especially now that the narrative generally has shifted back to AI being really powerful and valuable. Media is naturally then picking up the thing that swings the narrative pendulum back in the other direction, because to get eyeballs and clicks, you have to be saying the thing that's counter to the conventional wisdom. In other words, even if you assume that some amount of this incentive system gaming is happening, it takes a lot of leap to then extract that all the way to that being somehow the majority of AI usage, which then gets us to a second logical fallacy, hasty generalization or nutpicking.

21:44Basically, that means taking the visible extreme and treating it as the norm. One of the most important things we teach our children is the fundamental incorrectness of assuming that because one part of group X does something, then all of group X do that thing. And yet, that's exactly what we're doing here. Now, a final logical fallacy we'll throw on there just for fun is category error. In this case, gaming is being used as evidence about the quality of the technology when the only thing it can be reasonably considered evidence actually of is the incentive structure. Anyway, none of this really matters, because in many cases it is finding resonance among the people who had spent the last six months feeling annoyed at how quiet they had to be after being extremely loud about their bubbles and performance walls for the second half of last year.

22:23Those folks aren't trying to logic through with first principles, they're just excited to have a new version of their arguments, identity, and skeptics business model to cling on to. What I actually want to do is go farther and defend token maxing. As a practice, to the extent that token leaderboards are about showing off to the level above you, about how much you're doing in AI, yes, that is going to create some warped incentives and is probably not all that valuable. But there are very good reasons to encourage people to experiment more. First of all, historically speaking, some of the biggest barriers to getting enterprise employees to try and learn AI is those employees' sense that they just don't have time.

23:00This came up in basically every study and survey from 2023 to 2025. People were expected to just pick up and learn AI on top of doing their normal jobs. And given how much that didn't work, and how many challenges that created, it would be reasonable to think that creating incentive structures around experimentation might now be a practical necessity. Second, the shift we're now experiencing from assisted to agentic is, in my estimation, a much more significant disruption than the chat GPT moment. In the agent era, the question shifts from just, can we do the same thing but faster or cheaper, to, should we change how we do the thing?