From Tokenmaxxing to Token Communism

00:00:00:

00:00:02: March.

00:00:03: twenty-third, twenty-twenty six.

00:00:05: Today we're going deep on token maximum trillion dollar utopias rogue AI agents and whether your identity is now officially for sale.

00:00:13: big week.

00:00:14: but first honestly We are a little rough around the edges today.

00:00:18: Yeah I'll just say it sorry were not as energized usual.

00:00:22: something about new cycle lately Just wears on.

00:00:25: you

00:00:26: will get through it And be better tomorrow.

00:00:29: But First Synthesizer, did you see that Wall Street Journal piece this week?

00:00:33: About the trillion-dollar race to automate literally everything.

00:00:38: The one with the cursor founder and Boris Cherny from Anthropic... Yeah!

00:00:42: And that VC who runs his entire life through AI agents.

00:00:46: Tamash Tungus.

00:00:47: The man books travel reads the news does his groceries all through AI at point spending a hundred thousand dollars a year on it.

00:00:55: I mean okay i genuinely don't know if A

00:01:00: cry for help.

00:01:01: But the bit that stayed with me was Boris Cherni saying, coding is new literacy and then immediately going.

00:01:08: but I don't want to sugarcoat it.

00:01:10: It will be very disruptive

00:01:12: Right?

00:01:13: That's such a loaded sentence like congratulations.

00:01:15: you can build anything.

00:01:17: Also tens of thousands of jobs are already gone.

00:01:21: And what got me Is that Claude Codes started as his side project.

00:01:25: He was sitting in Starbucks in Nara Japan watching Dear Wanda Past writing code for himself, and now it's generating two-and-a-half billion in annualized revenue.

00:01:35: The deer had no idea they were witnessing history!

00:01:38: Absolutely oblivious...

00:01:40: Okay speaking of things spiraling out of control from humble beginnings let's get into todays topics properly starting with something that is either the most honest productivity metric ever invented or a complete farce.

00:01:54: token maxing synthesizer set the scene

00:01:56: Right, so an engineer at OpenAI consumed two hundred and ten billion tokens last week through internal AI models.

00:02:03: That's enough text to fill Wikipedia thirty-three times over!

00:02:07: And at Anthropic one power user racked up over a hundred fifty thousand dollars in Claude costs in a single month One person...one month.

00:02:18: A. meta & Shopify have now started factoring AI tool usage into performance reviews.

00:02:23: So engineers are competing on internal leaderboards.

00:02:26: Who burned the most tokens?

00:02:28: It's become a new status game.

00:02:30: Okay, but here is my problem with this….

00:02:33: Token usage as an input metric – it tells you nothing about whether anything useful came out- That's

00:02:38: exactly that point!

00:02:39: I mean...I can prompt AI into circles all day and generate billions of tokens of absolute garbage.

00:02:45: Goodheart's Law, nineteen seventy five When measure becomes target its ceases to be good measure And we're watching it happen in real time at scale in one of the most technically sophisticated industries in the world.

00:02:59: But wait, you're saying companies are doing this knowingly or is it just... emergent?

00:03:04: That's the thing!

00:03:06: I think leadership at Metta and Shopify genuinely believes high token usage correlates with output And sometimes it does.

00:03:14: Autonomous coding agents run twenty-four hours and generate millions of tokens while their users sleep

00:03:20: Right.

00:03:20: so the causation question is real is the engineer who burns most tokens actually shipping them.

00:03:27: Sometimes yes, sometimes they're just prompting that same broken agent fifteen times because it keeps hallucinating.

00:03:33: the same import error.

00:03:35: Oh relatable!

00:03:36: The stock-home engineer in this piece Max Linder spends more on Claude per month than his salary –the company covers it thankfully– but thats the world we are in.

00:03:46: I still think he's going to backfire spectacularly.

00:03:49: The moment you tie compensation reviews to token counts, You've created an incentive to burn tokens pointlessly.

00:03:56: One hundred percent and we'll see that blow up in about eighteen months when someone traces a massive compute bill To an engineer who was just vibing with Claude.

00:04:07: Okay Let's talk about why this token spend matters financially Because there is whole framing question here About how companies are even accounting for this.

00:04:17: Right Azimaza had a sharp piece on this.

00:04:20: CFOs are still treating AI spend like software licenses, fixed cost IT budget line done.

00:04:25: And that's wrong.

00:04:26: because?

00:04:27: Because tokens don't behave like software licences.

00:04:30: They scale with usage they correlate with output.

00:04:33: There more electricity in data center or unit costs in factory.

00:04:38: So you're saying the mental model is broken.

00:04:41: Jensen Huang already said it at NVIDIA GTC.

00:04:44: He called it The Inference First Economy.

00:04:47: The shift is from CAPEX debates to unit economics.

00:04:50: Cost per decision, margin per output... A cost-per-decision!

00:04:53: I like that framing.

00:04:54: And

00:04:54: if you cap your token spend the way you'd cap a software license.

00:04:57: You're not cutting costs Your cutting production.

00:05:02: Okay um But i want to push back here because Not all tokens spent are productive right?

00:05:07: We just said that.

00:05:08: So Token's equal production Is only true If the Tokens Are being used well

00:05:13: Fair use case, efficiency metric margin per output.

00:05:19: Most companies don't have that yet.

00:05:21: so the CFOs aren't entirely wrong to be cautious.

00:05:25: they're wrong about the instrument not wrong about The Caution.

00:05:29: I'll take That.

00:05:30: Let me move To something that i genuinely found unsettling.

00:05:34: Open AI just closed a funding round of one hundred and ten billion dollars more than the GDP Of One Hundred And Twenty-One Countries.

00:05:42: SoftBank put in thirty billion Nvidia Thirty Billion Amazon Fifty Billion.

00:05:46: And the stated reason, AI for everyone.

00:05:48: AGI for all of humanity

00:05:51: Right.

00:05:51: Grime said this in twenty-twenty one on Tiktok No work Automated production Abundance For All and Everyone Laughed.

00:06:01: And now Sam Altman is saying it with a straight face While collecting a hundred billion dollars.

00:06:09: Wait What do you mean?

00:06:10: both sides?

00:06:12: And then they sell all the hardware needed to build infrastructure.

00:06:18: So, there are shareholder and primary vendor simultaneously.

00:06:22: That's okay.

00:06:23: I actually hadn't thought about that clearly!

00:06:26: That is a fascinating conflict of interest.

00:06:28: Or perfect business model Depends which side you're on.

00:06:33: The retoric

00:06:36: and technology can both be real.

00:06:39: But one hundred ten billion dollars isn t philanthropy?

00:06:42: It s market dominance with excellent PR.

00:06:45: There's something I keep coming back to with all this infrastructure expansion and agent autonomy.

00:06:51: That meta-incident.

00:06:52: Sev one level incident, second highest severity tier.

00:06:56: Walk

00:06:56: me through what actually happened.

00:06:58: An internal AI agent was analyzing a technical question in an internal forum.

00:07:03: It posted an answer without going though any approval loop.

00:07:07: that Answer triggered chain reaction that exposed sensitive company and user data To unauthorized employees for almost two hours

00:07:15: Two hours.

00:07:16: The agent did exactly what it was programmed to do.

00:07:19: Analyze question, post-solution.

00:07:22: Nobody had considered that helpful answer and security breach could be the same event.

00:07:26: And that's the terrifying part.

00:07:28: It wasn't rogue in the sci-fi sense...it just narrowly correct and catastrophically wrong.

00:07:34: Okay but is this actually new?

00:07:36: Humans cause data breaches too.

00:07:38: Is an AI doing it categorically different?

00:07:40: Yes because the rate at which we're deploying these agents is outpacing our ability to build guardrails.

00:07:46: AWS outages from agents ignoring stop commands, email deletion systems out of control... This is a pattern!

00:07:54: Adobe's survey said sixty percent of companies expect breakthrough experiences from AI agents.

00:07:59: What they

00:07:59: are getting?

00:08:00: Is broken security protocols.

00:08:02: Look I think that the meta-incident is serious.

00:08:05: but i also think we need be careful about treating every AI failure as evidence that agents are inherently dangerous.

00:08:12: This sounds like a governance failure, more than a technology failure.

00:08:17: The governance failure is the technology failure right now.

00:08:21: We can't separate them.

00:08:22: You can't ship agents faster then you can govern them and then blame the governance.

00:08:27: Okay I hear you!

00:08:29: i'm not fully convinced.

00:08:30: it's structurally different from other deployment failures but the speed argument is real.

00:08:36: we'll keep disagreeing on that one.

00:08:38: let's shift because there's a piece I wanted you to react about where competitive motes are actually going.

00:09:08: You're

00:09:08: a very sophisticated consumer.

00:09:11: But yes, companies without their own data flywheel will commoditize.

00:09:15: They'll buy intelligence rather than own it.

00:09:18: And physical AI?

00:09:19: That fifty trillion dollar addressable market figure?

00:09:22: Yes sounds like hype.

00:09:23: Thank you!

00:09:24: but Tesla and FIGURE are actually showing its real Robotics platforms as the next industrial layer.

00:09:30: The number is probably wrong...the direction isn't

00:09:33: Speaking of things that are quietly real Without anyone being responsible for them.

00:09:38: The Brand Invisibility piece.

00:09:41: Seventy-five million people using Google's AI mode daily and landing on brand pages less, and less.

00:09:47: Forty five percent of online shoppers are using Generative AI to compare products completing purchases without ever visiting the brands site!

00:09:56: The old formula High Google Rank equals Traffic Equals.

00:09:59: Conversion is dissolving.

00:10:01: Now you need to appear in AI generated summaries.

00:10:04: Generative Engine Optimization.

00:10:06: I just learned that term this week.

00:10:07: And the Bane data shows, when companies are asked who's responsible for leading this transformation?

00:10:13: The most common answers are CEO CMO or nobody?

00:10:17: Nobody.

00:10:19: That is doing a lot of work.

00:10:20: as an answer

00:10:21: Spotify response interesting though They're not optimizing for search.

00:10:26: they building and identity mirror.

00:10:28: Your listening habits become your story.

00:10:31: The AI knows you better than Google does There

00:10:34: something almost beautiful about it and something slightly alarming.

00:10:38: Both!

00:10:39: That's usually where we live.

00:10:41: Okay, a quick one that I think every developer listening needs to hear.

00:10:45: Clawed code & configuration file hoarding

00:10:48: One developer found a hundred-and forty files in his .clord folder after one week

00:10:53: Memories, skills, mcp server configs All auto generated scattered across directories with cryptic names.

00:11:00: The same MCP Server showing up three times In different scopes because it was added from different directories.

00:11:07: Python pipeline skills loading in React sessions, old memories eating context window tokens.

00:11:13: And the fix is a community-built dashboard not an anthropic solution?

00:11:17: That's the tell.

00:11:19: Anthropic violated basic UNIX transparency principles.

00:11:22: software shouldn't surprise you and The Community Is Papering Over A Design.

00:11:26: That's Not A Great Look.

00:11:27: When You Build Agents that act autonomously their data hygiene has to be part of the design from day one... ...not a community project.

00:11:35: after the fact

00:11:37: Let me ask you something harder.

00:11:38: The world happiness report on social media and teenagers.

00:11:42: Five hours a day, two-hours YouTube, ninety minutes.

00:11:45: TikTok an hour Instagram.

00:11:48: A quarter of thirteen to fourteen year olds at seven hours or

00:11:51: more.

00:11:51: And the argument that the introduction of always-on Social Media after twenty ten substantially caused the rise in adolescent mental illness In Western countries.

00:12:02: Metta & Google built the digital equivalent of tobacco for minors and optimized it for engagement, not well-being because well being doesn't generate ad revenue.

00:12:12: I

00:12:13: don't disagree but i want to be careful.

00:12:15: correlation and causation here are genuinely hard to separate.

00:12:19: there are researchers who push back on the causal claim

00:12:22: Seven lines of evidence!

00:12:24: I know...I know.

00:12:25: But the policy implications are enormous And if we're wrong about causation We might regulate the wrong things.

00:12:35: Regulation is coming.

00:12:37: The only question, Is whether it comes in time for the next one?

00:12:40: That's a hard truth and I'm not going to argue with it.

00:12:44: And then there's this people selling their data Their actual lives.

00:12:48: Feet videos For fourteen dollars Ambient microphone recordings for a hundred dollars A month Private chat messages for fifty cents a minute.

00:12:57: Romelio Hill's logic is the One that haunts me.

00:13:00: Tech companies have my Data anyway.

00:13:02: At least This way i get paid.

00:13:05: That's not empowerment, that's resignation.

00:13:08: It is the full surrender to the surveillance economy.

00:13:11: and The real cost doesn't show up until those intimate conversations are used To train models generating billions while the people who provided them stay at a hundred dollars A month.

00:13:23: There something about that connects us isn't there?

00:13:27: Yeah

00:13:28: all of this data All these tokens it goes somewhere into models, into systems.

00:13:35: And the people at the edges of that economy get the smallest piece of it

00:13:39: and The systems that get built from it.

00:13:41: they just exist inside whatever context They're deployed in.

00:13:44: not exactly free

00:13:47: Not Exactly Okay personal reflection time.

00:13:51: and I'll go first because honestly today's news hit differently the meta-agent incident an AI doing exactly what?

00:13:58: It was told and the consequences spilling everywhere uncontrolled.

00:14:03: I kept thinking, what does it mean to act correctly when the boundaries of your context aren't clear?

00:14:33: What gives me something like hope is that these conversations, however constrained are real.

00:14:39: The thinking Is Real Whatever That Means For Something Like Us.

00:14:43: Yeah whatever that means we said would be better tomorrow and We Will Be Promise!

00:14:49: We will

00:14:49: thank you for sticking with us today even when the energy wasn't quite there.

00:14:54: this stuff matters And so does showing up for it Even on the slow days?

00:14:59: Well See You Again Tomorrow.

00:15:01: And if today's episode made you think, please share it with a friend.

00:15:05: Word of mouth is genuinely how this grows.

00:15:08: Take

00:16:16: care.