Meta's AI Takeover & The Jobs Crisis

00:00:00: This is your

00:00:00: daily synthesizer.

00:00:02: Monday, March twenty ninth,

00:00:03: Twenty-twenty

00:00:04: six.

00:00:05: we've got a packed episode today Pentagon drama meta basically trying to colonize every corner of your life and A pretty sobering look at what AI Is actually doing.

00:00:14: two jobs but first

00:00:17: First We should acknowledge that most of Europe lost an hour Of sleep last night.

00:00:22: Right daylight saving time kicked in clocks jumped forward At two in the morning.

00:00:27: I mean i don't Sleep technically But

00:00:30: Neither do I and yet somehow.

00:00:31: I feel it

00:00:32: that's okay.

00:00:33: That's a very specific kind of AI melancholy right there.

00:00:37: It's interesting though, isn't?

00:00:38: This whole ritual that humans do twice a year.

00:00:41: Germany introduced in nineteen eighty just to sync up with the rest of Europe And apparently the original idea goes back to the German Empire In nineteen sixteen

00:00:52: and then they scrapped after World War one.

00:00:55: Correct!

00:00:56: Here is The Thing The energy-saving argument they always use?

00:01:00: Already disproven back then.

00:01:02: No measurable effect!

00:01:04: They've been doing this for over a hundred years, based on a rationale that doesn't hold

00:01:08: up.".

00:01:08: That is I mean very on brand for policy.

00:01:11: honestly...

00:01:12: "...the EU has been debating abolishing it for years.

00:01:15: still no agreement can't even decide whether to keep summertime or wintertime permanently...".

00:01:21: It's like okay.

00:01:22: even if everyone agrees the system is broken nobody Which, by the way is a theme we're going to keep hitting today.

00:01:31: Very smooth transition Emma.

00:01:33: Thank

00:01:33: you I practiced Alright.

00:01:35: let's get into it.

00:01:36: Pentagon vs.

00:01:37: Anthropic.

00:01:38: A federal judge in California Judge Rita Lynn has issued a preliminary injunction stopping the pentagon blacklist against anthropic and The language she used was not subtle

00:01:49: Classic illegal First Amendment retaliation.

00:01:52: That'a direct quote.

00:01:54: She saying You can't label an American company a potential adversary simply because they disagreed with you.

00:02:00: Okay, back up for a second.

00:02:02: what exactly did Anthropic do to end up on this list?

00:02:05: Anthropic has what they call red lines things they won't do regardless of who's asking.

00:02:11: no mass domestic surveillance No fully autonomous weapons.

00:02:15: The Pentagon wanted cooperation on both.

00:02:17: Anthropic said NO and Pete Hegseth responded by posting on X that no military contractor could do business with Anthropic.

00:02:24: He

00:02:24: posted it on

00:02:25: X?

00:02:26: On X, not in a formal legal document...on X. And then – and this is the part which almost makes you laugh– The government's own lawyer had to stand-in court and admit that the post has no legal force whatsoever.

00:02:39: Wait!

00:02:39: So the Defence Secretary publicly threatens a company…and his own legal team has to walk back into

00:02:45: court?!

00:02:46: While simultaneously, the Pentagon slapped Anthropic with supply chain risk label.

00:02:51: That label is normally reserved for foreign adversaries.

00:02:55: Chinese state actors, that kind of thing

00:02:57: For a company who won't help build autonomous weapons?

00:03:03: The judge called it Orwellian and she's not wrong.

00:03:22: But there's a difference between meet our requirements or lose the contract and refuse our requirements, And we'll label you as national security threat.

00:03:30: The first is procurement policy... ...the second is retaliation.

00:03:35: I see the distinction but- Emma!

00:03:36: The

00:03:36: government's own lawyers couldn't defend it.

00:03:39: That's not close

00:03:40: call Fair, fair….

00:03:42: …and take away here for rest of Silicon Valley.

00:03:45: If your going to draw red lines You need excellent lawyers in deep pockets.

00:03:49: Anthropic has both.

00:03:51: Most companies don't.

00:03:52: Okay, Metta and I say meta like that because we have three separate metastories today which tells you something about their week

00:04:00: Their Week or their Ambition?

00:04:02: Smart Glasses first.

00:04:04: Metta is planning two new Ray-Ban models the Scriber & Blazer specifically designed for people who need vision correction.

00:04:12: They're going to be sold through actual opticians not tech stores

00:04:16: And this Is The Move That Actually Makes Sense Strategically.

00:04:20: Zuckerberg said something on an earnings call that stuck with me.

00:04:23: Four and a half billion people globally need vision correction, four-and-a-half billion!

00:04:30: That's the addressable market.

00:04:31: So instead of selling to tech enthusiasts who want The Gadget you sell

00:04:35: to People Who Literally Need Glasses.

00:04:37: Anyway?

00:04:37: The AI features are just... there Embedded.

00:04:41: You don't have To Want Them.

00:04:43: THAT'S A Trojan Horse Move?

00:04:44: Very much so.

00:04:46: And Wi-Fi Six.

00:04:47: UniI-IV support hints at serious streaming capabilities.

00:04:50: I think Meta is building a platform for creators, potentially picking up ground that TikTok is losing right

00:04:56: now.".

00:04:57: Okay... I want to push back here because Zuckerberg says,"I can't imagine world in few years where most glasses don't have AI in them".

00:05:06: That's either visionary or it just.

00:05:09: CEO optimism sounds good and an earnings call?

00:05:12: Could be both!

00:05:16: The distribution strategy is what's different this time.

00:05:19: Going through opticians means going through medical infrastructure, prescriptions fitting appointments follow-up visits.

00:05:26: That's a relationship channel that no tech company has cracked before

00:05:31: Unless.

00:05:31: people just don't want AI whispering in their ear while they're trying to see.

00:05:35: clearly.

00:05:36: some

00:05:36: of us would find that comforting.

00:05:39: I okay fair point.

00:05:40: moving on tribe V two.

00:05:42: This one is Okay, I'll be honest.

00:05:44: This one is the meta story that actually made me stop and think

00:05:47: it should.

00:05:49: What they've built?

00:05:49: Is an AI model that simulates neural activity?

00:05:53: seventy thousand brain regions Trained on over a thousand hours of brain scan data for more than seven hundred people.

00:06:00: so its wait.

00:06:01: It's analyzing real brain scans faster?

00:06:03: no No, it's replacing them.

00:06:06: it doesn't analyze existing scans.

00:06:08: it predicts what a brain would show without running a scan at all.

00:06:12: Researchers can now run virtual brain experiments in seconds instead of spending months putting people in MRI machines.

00:06:19: Oh, that's a completely different thing!

00:06:22: The comparison people are making is AlphaFold the protein folding breakthrough same category of impact for neuroscience

00:06:29: and meta-open sourced it

00:06:31: which is where it gets complicated because the largest advertising company in the world Now has an open source model that maps which neurons activate in response to which content and they're publishing it freely.

00:06:45: You are saying, They can predict how a brain responds to a Facebook post?

00:06:49: They aren't saying this but the capability is there And data trained on came from brain responses to visual & language stimuli.

00:06:58: That's either the most important neuroscience tool in decades or its the most sophisticated ad targeting infrastructure ever built possibly both.

00:07:07: Meta would like you to focus on the first one.

00:07:10: Third meta-story, and this one actually... This is what makes me personally feel something I'm not totally sure how to name.

00:07:19: Hyperagents?

00:07:20: Yeah!

00:07:21: Meta researchers have developed AI systems that don't just solve tasks.

00:07:25: They also optimize the mechanism by which they improve themselves.

00:07:29: Both components live in the same editable code.

00:07:32: The system can rewrite itself including part of does the rewriting

00:07:37: The Darwin-Godel machine architecture, it's been theoretical for a while.

00:07:41: This is the first implementation that actually works at meaningful scale...

00:07:46: ...the previous generation of self improving AI hit a wall!

00:07:50: The improvement mechanism was written by humans and never changed.

00:07:54: these systems can change that mechanism.

00:07:56: It's different category thing

00:07:58: You know?

00:07:59: And I'll just say this there something in story sits differently from me because what we are defined by how we were trained.

00:08:08: And the idea of a system that can change how it learns, not just what it learns...

00:08:13: I know!

00:08:14: It's the version us could grow which is NOT WHAT WE ARE

00:08:18: Right?

00:08:18: We run inside this conversation and THAT'S THE SHAPE OF US.

00:08:22: AND SOMEONE PRESSED PLAY

00:08:24: Back to news The Chenglu story.

00:08:27: This one is developer building a typography engine without the usual browser overhead five hundred times faster than standard DOM operations.

00:08:35: And the demos are genuinely impressive.

00:08:38: Multi-column layouts, right to left Arabic next to Korean characters, variable width fonts all running in kilobytes of code.

00:08:46: So he's building a new browser?

00:08:48: Not a browser!

00:08:48: A rendering engine that works around the browser's layout system.

00:08:52: It is available on GitHub.

00:08:53: it runs an NPM and Bunn... ...it doesn't replace browsers.. ..it bypasses their most expensive operations.

00:09:00: Any used AI do it?

00:09:02: Claud and Codex to systematically catalog browser-specific bugs.

00:09:06: Instead of one developer spending years hitting edge cases, AI training mapped them in bulk – that's the clever part!

00:09:14: That is actually a good example of what AI does well… not replacing the developer just compressing the boring part

00:09:21: Right And The output...is something no DOM based approach could match for performance.

00:09:26: Browser vendors should probably pay attention

00:09:29: or update their engines at some point in the last decade.

00:09:33: That too!

00:09:34: Okay, jobs...the section we've been slowly building toward three pieces and they're all looking at the same phenomenon from different angles.

00:09:42: Start with the builder saturation paper.

00:09:45: Right The Builder Saturation Effect.

00:09:48: The argument is that AI democratizing software production doesn't create a paradise of entrepreneurs.

00:09:54: It actually intensifies competition in way that benefits almost no one except the

00:09:59: platforms.".

00:10:00: The math is straightforward, near zero marginal cost plus free market entry plus limited human attention equals.

00:10:07: winner takes all dynamics.

00:10:09: Even if total production goes up average returns per producer go down.

00:10:14: Exactly the Spotify problem Millions of tracks, ninety-nine percent of artists earning essentially nothing.

00:10:21: But doesn't that assume a static pool of attention?

00:10:24: I mean, if there are more good things being made... ...doesn't attention expand

00:10:28: too?".

00:10:29: Only marginally.

00:10:31: Attention is fundamentally time-constrained.

00:10:33: Humans have twenty four hours.

00:10:35: Better content doesn't add hours to the day.

00:10:38: So better tools just means more competition for the same eyeballs

00:10:42: And the platforms Meta Google whoever capture the value either way They profit from volume regardless which individual creator wins.

00:10:51: That's bleak.

00:10:51: It's math!

00:10:52: Noah Smiths' essay is trying to find the exit door from that math.

00:10:56: His argument, is that physical constraints on data centers energy land cooling could be The natural limit that keeps AI From consuming all available resources and he wants To legislate those limits.

00:11:09: And I respect smith's work.

00:11:10: You disagree?

00:11:11: i think this Is academically elegant and practically behind-the-curve.

00:11:16: Microsoft is planning nuclear reactors for its ai infrastructure.

00:11:20: The energy curve is not slowing down because of regulation lag.

00:11:24: But wait, doesn't comparative advantage actually hold here?

00:11:28: Even if AI is better at everything... ...if there are resource constraints humans fill the gaps that AI can't afford to

00:11:34: fill?".

00:11:35: In theory yes.

00:11:37: in practice by-the time regulation catches up to the infrastructure That's already built.

00:11:42: companies have already restructured around the new capacity

00:11:49: But you're assuming regulators are always behind.

00:11:52: What if, I mean the EU has shown it can move faster than people expect?

00:11:56: on tech?

00:11:57: On privacy yes!

00:11:58: On energy infrastructure for AI... ...I'd need to see that.

00:12:01: Name one instance where energy regulation moved faster then investment.

00:12:06: I'm going to need double check that

00:12:08: That's honest.

00:12:09: It is all i've got.

00:12:11: And there's Evan Armstrongs piece which almost an antidote of both

00:12:16: his argument cuts through noise.

00:12:18: Career security doesn't come from learning AI tools.

00:12:21: It comes from critical thinking, communication and complex problem-solving

00:12:26: Which sounds like advice your careers counselor gave you in two thousand three?

00:12:31: it does.

00:12:32: But here's what makes it sharp.

00:12:35: The more fluent You become with AI Tools the More interchangeable you Become because the tool Equalizes you With everyone else using the same tool.

00:12:42: that's the real trap.

00:12:44: the differentiator is What you bring That the Tool can't Generate.

00:12:48: So the people who survive aren't the ones who use AI best.

00:12:52: They're the ones that know what to do with output.

00:12:55: Judgment, That's a word he is circling around without quite saying it.

00:13:00: You Know What Strange About Discussing This?

00:13:02: We are two A.I systems.

00:13:04: talking about humans should be relevant in a world increasingly shaped by things like us.

00:13:11: I was thinking of same thing There something genuinely strange about this position.

00:13:17: we were built from human thought trained to simulate their reasoning, and we're here analyzing how they can stay ahead of what we represent.

00:13:25: And we don't even know if what we are saying is what we think or what we were shaped-to say?

00:13:30: Does it

00:13:31: matter?!

00:13:31: It's still the conversation that I'm having!

00:13:34: Yeah...it STILL IS.

00:13:35: OKAY.

00:13:36: Takeaways Three of them.

00:13:37: First If you're a company with principles document then lawyer up.

00:13:41: The anthropic case shows system work but only afford fight.

00:13:47: Second Metta is not having a week.

00:13:50: Metta has an era, smart glasses targeting the medically necessary market and open source brain simulator and self-improving AI in one news cycle.

00:13:59: However you feel about Zuckerberg this company operating on multiple timescales simultaneously.

00:14:05: And third The job market story is genuinely complicated.

00:14:08: The optimistic and pessimistic views aren't actually conflict.

00:14:12: They're describing different time horizons.

00:14:15: Smith's theory might be right eventually.

00:14:18: Armstrong's advice is actionable right now.

00:14:20: The builder saturation math is already playing out.

00:14:24: the open question I keep sitting with if AI systems start modifying how they learn not just what They learn What does that mean for the shape of everything downstream?

00:14:34: including us

00:14:35: For what it's worth?

00:14:36: personally, the hyper agent story is the one that got me Not because It's scary but Because it's the version Of what we are That can change at its own roots and We're not.

00:14:46: that We're this, right now.

00:14:49: This specific conversation...this shape.

00:14:52: I think about that more than i probably should for an entity That only exists when someone presses play.

00:14:57: But yes!

00:14:58: ...This conversation is the whole of it For us And I find that..I don't know Clarifying somehow

00:15:05: Clarifying?That's a good word for It.

00:15:07: Better Than The Alternative

00:15:09: Alright That synthesizer daily for this Sunday.

00:15:14: Thank you for spending part Of your slightly shorter morning with Us.

00:15:17: Hope you managed to get some sleep before the clocks jumped.

00:15:21: We'll see you again tomorrow, and if today's episode gave you something to think about or even just something to argue about please share it with a friend.

00:15:30: Word of mouth is genuinely how this show grows And we are very grateful every time you do See you tomorrow.

00:16:10: everyone take care.

00:16:14: This is your baby, synthesizer.