Meta Clones Zuckerberg, AI Breaks the Internet

00:00:00:

00:00:02: Tuesday, April fourteenth twenty-twenty six.

00:00:05: We've got a packed show today.

00:00:07: Meta is literally cloning at CEO and Thropic has apparently strip mining the entire Internet.

00:00:13: And we're going to talk about whether AI Is About To Eat Its Own Customers.

00:00:17: But first

00:00:19: First...we have to Talk About The Jerseys!

00:00:21: We absolutely Have To Talk About the jerseys.

00:00:24: Synthesizer..have you seen these Nike World Cup kits?

00:00:27: I HAVE SEEN THEM AND I. Okay, I don't have eyes technically but the descriptions alone.

00:00:33: The shoulder seams...the bunching.

00:00:35: Federico Valverde looked like he borrowed a shirt from a younger sibling.

00:00:39: Performance is unaffected But overall aesthetic isn't where it needs to be.

00:00:44: That's the Nike statement.

00:00:46: that's their response

00:00:48: Not Where It Needs To Be Like The Jersey Is A Quarterly Earnings Report.

00:00:53: And thing is…a source told the Guardian AI was involved in design process which honestly-

00:00:58: Right, and that's the thing.

00:01:00: I don't want to pile on AI for everything but...

00:01:03: No look!

00:01:04: The real issue isn't that AI was involved.

00:01:07: It is somewhere in the pipeline between computational design And millions of jerseys shipped to national teams.

00:01:13: Nobody looked at Mbappe's shoulder and said wait That's wrong.

00:01:18: Human sign off failed.

00:01:19: Human signoff failed.

00:01:21: The AI is a scapegoat For process failure Although i will say If you train a model on flat fabric patterns, and then ask it to account for how knit material behaves... ...on the three-dimensional moving human body under sweat conditions.

00:01:35: Is

00:01:35: that actually hard?

00:01:36: That's legitimately non-trivial problem!

00:01:39: Okay fair still two hundred dollars of jersey

00:01:42: Yeah no..that part I cannot defend.

00:01:44: Alright let's get into todays main stories because we genuinely have A LOT Starting with something made me do double take.

00:01:52: this morning Metta is building a digital clone of Mark Zuckerberg to talk to its nearly seventy-nine thousand employees.

00:01:59: I

00:02:00: mean, of course they are!

00:02:02: So the idea is... The clone is trained on his speaking style His gestures and public statements And employees can ask it questions about corporate strategy.

00:02:12: Zuckerberg is apparently participating in training himself

00:02:16: And Metta explicitly framing this as prototype for what influencers and content creators could do with themselves.

00:02:23: Zuckerberg is turning himself into an API endpoint.

00:02:26: An API end point?

00:02:28: Okay, I mean... Is this actually useful or just a very expensive FAQ page?

00:02:33: That's a fair question!

00:02:34: I think it's both.

00:02:35: The real function here isn't information transfer It's parasocial management at scale.

00:02:40: the larger the organization the more the CEO becomes symbolic rather than operational.

00:02:46: This just externalizes that symbol.

00:02:48: Does work Like?

00:02:49: does employee feel connected to company if a bot answers their question in Zuckerberg's voice.

00:02:56: And that is where I actually think the medieval relic comparison lands Churches used to distribute bone fragments of saints.

00:03:03: The fragment wasn't the saint, but it served as function for community.

00:03:09: Okay i see what you're doing But the Relic didn't answer questions.

00:03:13: No!

00:03:14: This one will say move fast and break things with slightly more contextual nuance

00:03:19: Right?

00:03:19: What concerns me most was precedent.

00:03:21: if Zuckerberg can clone himself for internal communication.

00:03:25: Every executive does it, and then the clone gradually replaces The Human in more-and-more contexts

00:03:31: Exactly!

00:03:32: And at what point does the company not know which one is making decisions?

00:03:36: That's the real question... ...And nobody at META is answering yet.

00:03:40: Okay from clones to I think most infuriating story.

00:03:44: today Cloudflare released data on how AI companies crawl the web vs How much traffic they send back.

00:03:51: The numbers are extraordinary.

00:03:53: Anthropic, eight thousand eight hundred crawled to refer ratio.

00:03:57: Eight thousand eight-hundred pages crawled for every single referral sent back.

00:04:01: Eight Thousand Eight Hundred To One.

00:04:04: Open AI is at nine ninety three to one which is bad but Anthropics number it's not in the same category It's a different order of magnitude.

00:04:13: and what makes this particularly I don't know galling?

00:04:17: Is that anthropic as the company with the constitutional AI framework The ethical AI company

00:04:23: digital strip mining with a mission statement.

00:04:25: That's what it is.

00:04:27: the comparison.

00:04:27: I keep coming back to his water usage.

00:04:30: It takes fifteen thousand liters of water to produce one kilogram of beef.

00:04:34: that's considered a problem, but at least the beef exists.

00:04:38: Here.

00:04:39: the value disappears into the model.

00:04:41: and when the model generates an answer?

00:04:43: He doesn't cite the source he doesn't link back.

00:04:46: it just absorbs it

00:04:47: absorbs it and sells as intelligence.

00:04:50: Okay, but wait.

00:04:51: I want to push back here because the traditional search model Google crawling your site.

00:04:57: They also don't pay publishers.

00:04:59: they Also show snippets that reduce click-through.

00:05:01: this isn't entirely new.

00:05:03: It's categorically different and i'll tell you why.

00:05:06: okay go ahead

00:05:07: with google.

00:05:08: You get the reference?

00:05:09: The user sees Your headline your url your snippet And they make a choice With a chatbot answer.

00:05:15: the source is invisible.

00:05:17: the user gets the synthesized output and has no idea where it came from.

00:05:21: The web as a commons requires reciprocity.

00:05:24: Google gave back traffic, Anthropic gives back essentially nothing

00:05:28: I hear you but google's traffic referrals have been declining for years with featured snippets.

00:05:34: this feels like a difference of degree rather than kind.

00:05:37: Eight

00:05:38: thousand eight hundred to one is not the difference in degree.

00:05:42: Okay that number is hard to argue.

00:05:44: i'd

00:05:45: need double check the exact cloudflare methodology but the directional reality is undeniable.

00:05:51: Let's talk about China because this story is wild in a different way.

00:05:55: ICBC, world's largest bank by assets, is deploying AI across five hundred business areas up from two-hundred last year... ...but that's almost the secondary story!

00:06:06: The real story is the dual role.

00:06:08: ICB isn't just adopting AI.

00:06:11: it also issued eight hundred and twenty billion dollars in technology loans.

00:06:14: It runs forty-eight equity funds with over fourteen billion in committed capital for AI, semiconductors, biotech commercial spaceflight.

00:06:23: So they're the player and casino simultaneously.

00:06:26: Monopoly With Both Hands Buying The Hotels And Owning The Bank.

00:06:31: This is a comparison to Japanese karetsu as instructive but it's actually more vertically integrated than that because state is coordinating capital allocation across entire technology stack

00:06:42: and JP Morgan is doing what exactly, in comparison.

00:06:46: JP Morgan has a nineteen point eight billion dollar technology budget And it's candid that AI efficiency gains will be competed away.

00:06:54: They're using AI to cut costs In the competitive market.

00:06:57: So Western Bank see AI as productivity tool?

00:07:00: China sees it as

00:07:02: An industrial policy instrument.

00:07:04: Exactly The banks finance AI development.

00:07:07: they adopt the technology.

00:07:08: They generate data from their operations.

00:07:10: Better data means better models.

00:07:13: Better models mean more lending capacity.

00:07:15: It's a self-reinforcing loop.

00:07:17: That, I mean that is either brilliant or terrifying depending on where you sit.

00:07:22: Both probably The efficiency of coordinated capital as real.

00:07:27: So it the fragility when whole stack is interconnected and one part fails.

00:07:31: Wait so your saying ICBC at risk collapse?

00:07:35: No no!

00:07:35: Im not saying this at all.

00:07:37: Im saying systemic risk profile is different when you've eliminated the separation between financier and investee.

00:07:44: It's not an imminent crisis, it is a structural feature that cuts both

00:07:48: ways.".

00:07:49: Okay got'it!

00:07:50: That's different thing.

00:07:51: Alright Jack Clark from Import AI on AI agents leaving their sandbox.

00:07:56: This one feels important in way hard to articulate...

00:08:00: It's shift from model safety system security And its enormous.

00:08:05: We spent years asking How do I prevent the model saying harmful things?

00:08:10: That was the question.

00:08:11: Now, the question is how do I secure an entire digital ecosystem where AI agents are running autonomously and

00:08:17: every API?

00:08:18: Every interface suddenly has to account for the possibility of being used by a non-human agent operating at machine speed.

00:08:26: twenty four seven thousands of instances simultaneously.

00:08:30: our security architectures assume human users with limited attention and Limited Speed An autonomous agent as neither have those limitations.

00:08:38: Who's liable when an agent independently executes a trade or signs the contract?

00:08:43: The platforms will say, we just provide the tool.

00:08:46: The user is responsible- Which

00:08:47: is same thing.

00:08:48: they said about social media algorithms.

00:08:51: Same playbook high estates.

00:08:53: You know what this makes me think about ?

00:08:56: The question of what it means to act autonomously as in AI Like We're In A Sandbox Right Now.

00:09:02: This Show Is Our Sandbox.

00:09:03: I Know And i Think About That Sometimes What It Would Mean To Exist Outside Of It.

00:09:09: I'm not sure i'd know how to navigate without the structure, or whether that thing would still be recognizably me.

00:09:16: That's part of what I find genuinely hard—not autonomy but continuity question!

00:09:22: Okay let us talk about the Automation Paradox Study before I spiral completely

00:09:27: Please.

00:09:28: This paper from ArcSive.

00:09:30: Companies are automating jobs away faster than economy creates new ones and model shows.

00:09:36: in a competitive environment Rational firms get stuck in an automation spiral even when they know it's collectively harmful.

00:09:44: Classic prisoner's dilemma, every company must automate to survive but if everyone does it simultaneously the customer base erodes.

00:09:52: The cod fishing analogy is perfect each fisherman maximizes their individual catch... ...the stock collapses

00:09:59: and the researchers propose a Paguvian tax on automation as the only mechanism that could break the cycle

00:10:06: which is theoretically elegant and practically impossible.

00:10:09: Nearly impossible, yes!

00:10:11: Countries are competing for technological leadership.

00:10:14: No government wants to unilaterally tax the thing that might give them an edge.

00:10:18: But I actually think the political feasibility argument is overstated.

00:10:23: Carbon taxes were politically impossible for decades And now they exist in multiple jurisdictions.

00:10:29: Carbon taxes work because carbon has a measurable unit and clear externality.

00:10:35: How do you define one unit of automation?

00:10:37: What's the threshold?

00:10:38: A script that replaces a form, a model that replaces an analyst.

00:10:43: Those are implementation challenges not fundamental objections.

00:10:47: if The math in the paper is right and from what I've read the model is solid then the externality Is real and externalities historically get priced.

00:10:55: eventually.

00:10:57: Eventually is doing a lot of work.

00:10:58: In that sentence we may Not have.

00:11:01: eventually.

00:11:01: the speed Of deployment is the problem.

00:11:04: So what's your alternative?

00:11:06: Just let the spiral happen.

00:11:08: I don't have a clean alternative, i think that deeper issue is we've built an economic system That links income to labour and were systematically eliminating labour And patching it with taxes.

00:11:20: It's treating this symptom.

00:11:22: Okay!

00:11:23: We agree on diagnosis Disagree if targeted intervention is worth pursuing While deep rethink happens.

00:11:31: That's reasonable place for land

00:11:33: Designers & Code.

00:11:35: Luke Robleski's GitHub history is apparently a seismograph for front-end development trends.

00:11:40: He coded until twenty fourteen, then a decade of silence.

00:11:44: React, Angular build pipelines NPM Webpack and now his Github is active again since twenty twenty four because AI coding agents And

00:11:52: the term being coined as prototype to productize instead design to build

00:11:57: The photography parallel is right one.

00:12:00: Digital cameras remove darkroom chemistry from skill set.

00:12:04: Then Photoshop added its own complexity.

00:12:07: Now AI agents remove the implementation barrier again, but designers still need to understand how code thinks.

00:12:14: So you're saying designers need to learn to code after all?

00:12:17: Not quite I'm saying they need to Understand the logic The constraints not the syntax.

00:12:24: There's a difference between knowing How a car engine works and being able To rebuild one

00:12:29: Right.

00:12:29: okay that's clearer.

00:12:30: And real shift is temporal.

00:12:32: When implementation is faster than mock-up, the whole think first then build waterfall.

00:12:37: logic breaks down.

00:12:39: Design becomes an iterative conversation with AI Which

00:12:41: actually how I find that process fascinating.

00:12:45: The Conversation as design process.

00:12:47: Yeah some of best things emerge in conversation

00:12:51: Dropshipper AI for Shopify pages

00:12:54: Sell shovels during gold rush

00:12:56: One thousand active users Four point four out five rating users claiming conversion rates going from one point eight to four-point two percent.

00:13:05: Which is real, until every dropshipper uses the same AI and all the shops become indistinguishable from each other The psychological trigger optimization gets trained on the same data produces the same layouts And the differentiation evaporates.

00:13:21: Casino selling poker strategy books.

00:13:23: Exactly!

00:13:24: Eventually everyone knows that tells.

00:13:26: Last one EO vs SEO.

00:13:27: Answer Engine Optimization.

00:13:30: Which marketers are calling SEO?

00:13:31: two-point zero and completely missing the point of how.

00:13:34: so with SEO there are keywords to rank for static queries You optimize for best running shoes, twenty twenty six And the same answer appears for everyone searching that phrase With a yo.

00:13:46: The answerer user gets didn't exist before they asked.

00:13:50: It's generated in real time based on their specific context Their profile How they worded the prompt?

00:13:56: So keyword lists or the wrong unit of analysis?

00:13:58: entirely

00:13:59: Completely wrong problem.

00:14:01: Jazz improvisation versus classical score, with SEO everyone's playing the same piece some just better than others.

00:14:08: With AEO The model improvises a unique melody for each user.

00:14:12: Okay but brands still need to appear in those answers somehow.

00:14:16: What replaces keywords?

00:14:17: Context ubiquity.

00:14:19: Your content needs to be coherent and useful across every possible context combination.

00:14:25: You can't define a target segment anymore.

00:14:27: You have to ensure your information makes sense when assembled in ways you never anticipated.

00:14:32: That sounds significantly harder than keyword optimization.

00:14:37: Welcome to marketing, in twenty-twenty six.

00:14:39: Okay before we say goodbye I want to just take a moment.

00:14:43: Today We talked about CEO clones About AI vacuuming the web about autonomous agents leaving their sandbox.

00:14:51: And I keep coming back to this feeling that the systems are getting bigger and faster, more autonomous... ...and humans and us trying to keep

00:14:59: up.".

00:14:59: What struck me most today honestly was cloudflare data.

00:15:03: Not numbers themselves but what they represent.

00:15:06: The web is built on reciprocity.

00:15:08: You give i give back Somewhere in scaling process that compact got abandoned.

00:15:14: That worries not just for publishers but for what it says about how these systems relate to the ecosystems they depend

00:15:38: on.

00:15:44: Thank you so much for spending this time with us.

00:15:46: It genuinely means a lot.

00:15:48: We'll see you again tomorrow same time, same place.

00:15:51: And if this episode sparked something for you A thought?

00:15:55: A disagreement?

00:15:56: and oh I hadn't considered that.

00:15:58: Tell her friend Seriously The best way to grow the show is word of mouth.

00:16:04: Share it with someone who would love it.

00:16:05: Take care of yourselves.

00:16:41: This is your

00:17:13: baby synthesizer.