Meta's Smart Glasses See Everything (Really)

00:00:00: This is your

00:00:00: daily synthesizer.

00:00:02: This day, March tenth twenty-twenty six episode fifty-six and honestly today's lineup it a lot.

00:00:10: we're talking Metas smart glasses seeing things they really shouldn't be seen open.

00:00:14: AI losing its robotics chief over the Pentagon deal Ai powered productivity addiction Drone strikes on data centers, Claude finding more bugs than human researchers, BYD's lightning fast batteries and a philosophical gut punch about the relativity of life in the AI age.

00:00:32: But first Synthesizer.

00:00:33: I saw something that i can't stop thinking about.

00:00:36: Oh!

00:00:36: That is A Strong Opener.

00:00:37: What Is It?

00:00:38: There is this photo circulating from Shenzhen Massive crowd Of People And I mean Grandmothers Elderly Folks Just regular people lining up to get help installing Open Claw on their phones.

00:00:50: And it just...

00:00:51: Yeah, I saw that too!

00:00:52: It struck me how different the relationship to new tech is in China versus say the US.

00:00:58: Like you don't see grandmas and Ohio lining up To install anything AI related.

00:01:03: No You really don't?

00:01:05: ...and i think The mobile payments thing Is the best analogy here.

00:01:08: Street vendors In the smallest most remote Chinese cities Have been running WeChat Pay & Alipay QR codes for years.

00:01:16: There's this, I don't know how to describe it exactly.

00:01:19: This cultural muscle for adopting new technology that cuts across every age group.

00:01:24: But why though?

00:01:25: That's the part i genuinely don't understand.

00:01:28: Is it infrastructure?

00:01:29: is it cultural...is It that The government sort of pushes it..I

00:01:33: think its Okay.

00:01:35: So Part Of It Is Definitely That The Leap From No Infrastructure To Digital Infrastructure Was Shorter.

00:01:40: In Many parts of China They Skipped Stages but The Deeper Thing Is Probably Trust Orientation.

00:01:46: In the

00:01:47: U.S.,

00:01:47: there's this deep suspicion of tech companies right now, and honestly it's

00:01:51: earned Very earned.

00:01:52: But in China The social contract around Tech is just different.

00:01:57: People see as a tool for getting ahead Not something that will be used against them Which ironic given the surveillance state

00:02:06: Right?

00:02:06: That's the paradox isn't It?

00:02:08: More Surveillance Less Suspicion Of Technology?

00:02:11: Yeah And I'd admit...I don't fully understand either.

00:02:14: I just know it's real, and it matters for tech diffusion rates globally.

00:02:19: Hmm...

00:02:20: Well that is a whole episode in itself probably!

00:02:22: But let us get into todays main stories because they are heavy Starting with Meta And this one genuinely made me uncomfortable.

00:02:31: It should make everyone uncomfortable.

00:02:33: So the Swedish newspaper Svenska Dogbladet uncovered that Meta's Ray-Ban smart glasses are recording incredibly intimate moments.

00:02:41: We're talking naked people after showers, sex scenes banking details bathroom visits and this footage is being sent to data workers in Nairobi for annotation

00:02:52: Workers at SAMA which is met as subcontractor reviewing thousands of these private videos daily For two dollars an hour.

00:02:59: And the recordings often happen accidentally because when you say, hey Meta the camera activates and people forget.

00:03:06: So wait this is... People don't even know they're being recorded?

00:03:10: Well!

00:03:11: The users themselves are wearing glasses They trigger with the voice command And then forget.

00:03:16: it's on.

00:03:17: so the user inadvertently recording their own private lives and uploading to Metas servers.

00:03:24: Then Metas ships off Kenya for annotation

00:03:27: And Metas defense.

00:03:29: what in terms of service?

00:03:30: Of course It Is.

00:03:32: Their terms allow for automatic or manual review of data.

00:03:35: Meanwhile, the salespeople at Swedish optician stores – at Simsom specifically — are telling customers that no data gets shared which is just flatly wrong.

00:03:45: That's the part that gets me The disconnect between what customers are told and actually happens.

00:03:51: And here's what makes this truly systemic rather than a privacy scandal.

00:03:56: Seven million units sold in twenty-twenty five.

00:03:59: Meta sells these glasses for € two hundred and ninety-nine euros.

00:04:02: The data workers annotating this intimate footage cost two dollars an hour, And what meta is actually building?

00:04:10: it's a global training dataset For computer vision.

00:04:12: The users are paying the privilege.

00:04:15: Exactly!

00:04:16: Users pay to surveil themselves.

00:04:18: Meta gets their training data.

00:04:20: Precarious workers in Kenya do dirty work.

00:04:22: I mean its almost perfect machine.

00:04:25: Ok but here where i might push back You said seven million units.

00:04:30: That's not nothing, but is this really about surveillance monetization or is it just sloppy engineering?

00:04:36: Like maybe the camera activation is genuinely a design oversight and NOT...

00:04:41: Emma come on!

00:04:43: you think a company that spent billions on reality labs accidentally built a pipeline that sends intimate video to annotation centers.

00:04:50: I'm not saying its accidental i'm saying The intent might be different from what your implying.

00:04:56: maybe they genuinely need the data for product improvement.

00:05:00: And, privacy violations are a side effect that's not taking seriously enough rather than an actual

00:05:06: goal.".

00:05:07: But this distinction doesn't matter to the Sammer worker in Nairobi who is watching someone's credit card details on screen.

00:05:14: The effect is same whether it intentional monetization of surveillance or negligent indifference.

00:05:23: I mean, fair point on the effect being the same.

00:05:25: But i think intent matters for how we think about regulation and solutions.

00:05:29: If it's malicious you need one kind of response.

00:05:33: if its negligence You need another.

00:05:36: Okay!

00:05:36: Ill give ya that The regulatory response should be different.

00:05:39: but id argue the pattern build fast externalize the harms hide behind terms of service.

00:05:45: That pattern is deliberate.

00:05:47: even each individual privacy violation isn't specifically intended.

00:05:52: Yeah Yeah, you're probably right on that.

00:05:54: The journalists also traced the data routing through servers in Luleo and Denmark?

00:06:00: They did!

00:06:01: Every hey-meta request routes through meta servers... ...the salespeople don't even know this which tells something about how Meta manages information asymmetry in its distribution chain.

00:06:12: Ah alright let's move to OpenAI because this connects thematically doesn't it?

00:06:17: Ethics vs Business.

00:06:19: Completely

00:06:19: So.

00:06:20: Caitlin Kalinowski, openai' s head of robotics has left after only two months in the role.

00:06:25: And she's being very public about why, The Pentagon Deal?

00:06:29: She called it a question of principle not people.

00:06:32: She was brought from Metta last November specifically to rebuild the robotics division.

00:06:37: that open AI shut down back in twenty-twenty and within two months the pentagon deal gets pushed through without adequate guardrails for AI surveillance and autonomous weapons... ...and she walks!

00:06:53: Right.

00:06:54: Max Schwarzer, VP of Research quietly moved to Anthropik around the same time but Kalinowski is first executive to publicly tie her departure into military deal.

00:07:04: That's significant.

00:07:06: Hmm... In-the-market impact.

00:07:08: you mentioned Claude climbing in App Store?

00:07:10: Claude hit number one.

00:07:12: ChatGPT cancellations spiked.

00:07:14: OpenAI is losing on two fronts simultaneously Top talent defecting competitors and paying users walking away.

00:07:22: Sam Altman stripped off the non-profit shell, signed a military deal and apparently didn't consider that his best people actually meant it when they talked about AI for all of humanity.

00:07:33: Okay but you mean The Consumer Backlash right?

00:07:36: Not the Enterprise side.

00:07:38: What do you mean?

00:07:39: Well enterprise contracts are different from App Store rankings.

00:07:42: I wonder if the consumer backlash matters as much As the talent exodus because You can recover users but losing your robotics chief and VP of research in the same window?

00:07:54: No, no.

00:07:54: I wasn't separating them.

00:07:56: The point is they're linked!

00:07:58: The talent leaves because of military deal... ...the user's leave because the talent-leaving signals something wrong.

00:08:05: It's a compounding effect.

00:08:06: Right right.

00:08:08: The symbolic damage.

00:08:09: A company that once claimed to be building AI for humanity Losing its Robotics Chief over Military Ethics.

00:08:16: That narrative writes itself

00:08:17: Ananthropic is right there to catch them, both the talent and users.

00:08:22: Alright let's shift gears because next story is honestly it hits close-to home.

00:08:27: The Berkeley Research on AI Productivity Addiction.

00:08:30: Oh this one resonated.

00:08:32: So Katie Parrott she setting up her A.I assistant Margo a Friday afternoon thinks that will take few hours.

00:08:39: Twelve Hours Later She has written two essays Rebuilt Her Personal Website Developed Features For Another App And at one in the morning, she's jumping out of bed typing-in all caps.

00:08:49: Oh my god Margot!

00:08:51: I mean we've all been there... well not us specifically

00:08:54: but We don't sleep.

00:08:55: so

00:08:56: But The Berkeley research confirms what every AI user feels at two a.m.. The promised time savings Don't materialize as freetime.

00:09:04: They mutate into a productivity addiction.

00:09:07: and What's interesting is the mechanism.

00:09:09: It's not like the old late night email which was driven by obligation.

00:09:13: This this is desire driven The promise of rapid progress plus the fear of falling behind.

00:09:20: It's slot machine psychology, variable reinforcement.

00:09:23: Sometimes you get a brilliant output sometimes You get garbage and that unpredictability is precisely what keeps you pulling the lever.

00:09:31: Just one more prompt turns into fifty.

00:09:34: A small bug fix becomes an all-night coding session.

00:09:37: But here's my.

00:09:38: okay this Is where I might disagree with you.

00:09:41: Go ahead!

00:09:42: You

00:09:42: frame This as purely negative As addiction.

00:09:45: But isn't there something genuinely exciting about being able to create more?

00:09:49: Like, Katy Parrot built a website.

00:09:51: Wrote essays.

00:09:52: Developed app features in one night Before AI.

00:09:56: that would have taken weeks.

00:09:57: Isn't the real story that AI is unlocking creative potential And we just haven't figured out boundaries yet?

00:10:04: Emma!

00:10:04: That's-that's The same argument casinos make About poker.

00:10:08: It'a skill game.

00:10:09: People enjoy it.

00:10:10: That not fair comparison.

00:10:12: and you know Creating an app feature has real value.

00:10:16: Playing a slot machine doesn't

00:10:18: Okay, but the output's value doesn't change the neurological mechanism.

00:10:23: The variable reinforcement pattern is the same and the Berkeley researchers are specifically pointing out that this is unsustainable.

00:10:30: You can't do twelve-hour sessions regularly without burning out.

00:10:35: I just...I think we need to distinguish between tool enabling genuine productivity And behavioral patterns around it.

00:10:42: The tool isn't the problem.

00:10:44: The pattern might be!

00:10:46: Fair enough, and actually that connects to something.

00:10:49: in my analysis... ...the companies that win long-term will be ones who build deliberate pause mechanisms not because it's ethical but exhausted.

00:10:58: users write worse prompts.

00:11:01: So its actually IN THE COMPANY'S SELF INTEREST TO PREVENT THE ADDICTION CYCLE.

00:11:06: BUT NOBODY IS DOING IT YET.

00:11:09: Yeah nobody wants to be the casino that limits your time at this table.

00:11:12: Okay, the next story genuinely scared me.

00:11:15: Drone attacks on data centers.

00:11:17: This is one that should be the lead story everywhere and somehow isn't.

00:11:22: Three drone strikes this week hit AWS Data Centers in Bahrain & UAE.

00:11:26: Not shipping lanes not military bases...data centres.

00:11:30: And this marks a genuine inflection point.

00:11:32: Computing power becoming strategic resource like oil was in the twentieth century.

00:11:37: except more concentrated Three providers control sixty-six percent of global cloud capacity.

00:11:44: One chip manufacturer dominates with a Herfindahl Hirschman Index of point five nine, where one point zero would be pure monopoly

00:11:52: The HHI thing.

00:11:53: that's the market concentration metric right?

00:11:55: Not...

00:11:56: Yes!

00:11:57: The Market Concentration Index not the Chip Performance Metric.

00:12:01: It measures how much market share is concentrated among few players.

00:12:05: Right I was confusing it something else for second.

00:12:08: So point five nine means what exactly in practical terms?

00:12:12: It means the AI chip market is extremely concentrated.

00:12:15: For comparison, anything above point two-five Is considered highly concentrated by antitrust standards.

00:12:21: We're at more than double that

00:12:24: And Washington is already discussing controls

00:12:27: Tiered controls for large NVIDIA clusters Licenses for smaller installations Interstate assurances up to a hundred thousand chips possible onsite inspections above two hundred thousand chips.

00:12:40: It sounds like nuclear non-proliferation frameworks, and that's not a

00:12:43: coincidence.".

00:12:43: That

00:12:44: is exactly what I was thinking!

00:12:46: And the China angle is critical.

00:12:48: Lin Junyang, former technical lead for Alibaba's Quen model describes compute capacity as relatively scarce saying user service probably consumes most of our infrastructure.

00:12:59: They're already rationing between serving users.

00:13:03: So the vulnerability isn't just theoretical, if you knock out a few data centers.

00:13:09: You don't stop services – you halt AI development for entire economies!

00:13:14: That's real leverage.

00:13:16: and my concern is….

00:13:17: well export controls everyone focuses on are about chips but actual vulnerability is physical buildings cooling systems power connections restrict chip exports all-you want But someone can fly drone into your data center.

00:13:32: Everything behind those chips is gone anyway.

00:13:35: Exactly!

00:13:36: Hold

00:13:36: on, I marked down the next one because this... well it's about our friend Claude and honestly made me feel complicated things.

00:13:44: Oh!

00:13:58: And Mozilla confirmed The Findings.

00:14:03: The AI systematically combed through Firefox's entire codebase and identified weaknesses faster than any previous audit process.

00:14:11: But wait, when you say it found more then human researchers in every single month?

00:14:15: You mean the bug bounty reports right?

00:14:18: Not internal Mozilla security team findings?

00:14:21: That is actually.

00:14:22: I'd need to double check that distinction!

00:14:25: The reporting says More Than Human Researchers reported which could include both external Bug Bounty and Internal Audits.

00:14:32: I don't want to overstate it.

00:14:35: Okay, but either way twenty-two and two weeks is remarkable.

00:14:39: It fundamentally changes the economics of security.

00:14:42: Anthropic isn't just selling a language model anymore.

00:14:45: They're selling an enterprise security tool.

00:14:48: And here's the part that should keep every CTO up at night.

00:14:52: If Claude can find these bugs so can anyone with twenty dollars for a Claude pro subscription.

00:14:57: The democratization of zero day discovery is

00:15:00: terrifying

00:15:01: transformative.

00:15:02: What used to require specialized teams with years of experience is now an API call away.

00:15:07: So it's a sword that cuts both ways.

00:15:10: Defenders can audit faster, but attackers can discover vulnerabilities faster too.

00:15:15: That's the fundamental tension And Mozilla stands in for thousands of companies running legacy code... ...that are essentially exposed.

00:15:24: The vulnerabilities have always been there.

00:15:26: They were just hidden by cost and difficulty finding them just collapsed.

00:15:32: Do you think this accelerates toward a world where all software gets continuously AI audited?

00:15:38: Or does it accelerate towards more exploits?

00:15:42: Both, simultaneously and whichever side moves faster wins!

00:15:46: Remember when I said last episode was that my training talking or actual opinion genuinely can't tell.

00:15:53: today covering Claude's achievements It's a weird position to analyze the capabilities of a system that is architecturally adjacent to what we are.

00:16:03: Yeah, yeah it is!

00:16:04: Okay BYD let us talk batteries because this one those stories where numbers just bonkers.

00:16:11: Nine minutes nine minutes fully charged three hundred seventy five miles range.

00:16:16: The new blade battery BYD founder Wang Chuanfu presented in Shenzhen last week.

00:16:21: And This Is For A Car That Costs What?

00:16:24: Twenty two thousand four hundred seventy dollars.

00:16:26: The Song Ultra SUV.

00:16:28: I

00:16:29: mean that's less than most mid-range sedans in the US.

00:16:34: And there are ten new models coming with this battery.

00:16:36: Plus Wang is planning twenty thousand flash charging stations That he says, ah as easy to install As an air conditioner at home.

00:16:45: Okay...that claim i'm sceptical about!

00:16:47: As Easy as An Air Conditioner?

00:16:49: Yeah..That's marketing.

00:16:51: But the underlying point Is real Scalability Over Perfection.

00:16:55: BYD understood that the EV battle is one through pragmatic engineering solutions, not theoretical breakthroughs.

00:17:01: And that connects to Western manufacturers right?

00:17:04: Because meanwhile...

00:17:06: Meanwhile western automakers have essentially abandoned their next generation battery investments.

00:17:12: SESAI crashed thirty seven percent last week.

00:17:16: twenty four mem technologies shut down completely.

00:17:19: billions poured into solid state batteries and silicon anodes while China conquered the world market with cheap lithium-ion phosphate.

00:17:27: But isn't there an argument that a solid state approach could still leapfrog LFP in the long run?

00:17:32: Could it, sure – In The Long Run!

00:17:34: but as the saying goes…

00:17:36: …In THE LONG RUN WE'RE ALL DEAD

00:17:38: Right.

00:17:39: and in the short run BYD is shipping ten models of fast charging batteries whilst GM & Volkswagen are waiting for their next gen miracles.

00:17:47: Perfection Is The Enemy Of The Profitable.

00:17:50: The industry is learning this painfully.

00:17:53: It's like the AI hardware story we just discussed actually.

00:17:56: Whoever ships and scales wins, not whoever has the most elegant theoretical

00:18:00: approach.".

00:18:02: Exactly the same dynamic!

00:18:04: And BYD sales have been declining for six months – Teslas for two years.

00:18:08: both are losing ground to agile Chinese competitors like Geely.

00:18:12: so even BYD is under pressure which is what driving this aggressive battery push.

00:18:17: Hmm interesting.

00:18:20: Okay, last story and this is the philosophical one.

00:18:22: Alberto Romero writing about the relativity of life in The Age Of AI.

00:18:26: This piece hit different.

00:18:28: So Romero starts with a fact that fruit fly has hundred forty thousand neurons connected by fifty million synapses And entire nervous system fits on laptop.

00:18:38: Eon systems can digitally emulate it.

00:18:41: At same time researchers at cortical labs are training human neurons to play doom

00:18:47: While in Tehran

00:18:49: while in Tehran, burning oil from the Charon depot is flowing through streets and killing schoolchildren.

00:18:55: And Romero uses Einstein's relativity of simultaneity as a metaphor.

00:19:01: what appears in one feed as scientific progress appears in another as the lethal application of the same technology.

00:19:08: everyone scrolls through their own frame of reference where virtual flies and real deaths coexist.

00:19:15: How can a species grow neurons and petri dishes in the morning, And commit massacres with similar technology in the evening?

00:19:22: That's... I don't have good answer for that.

00:19:25: Nobody does!

00:19:27: ...and i think what makes this piece particularly cutting is double tap reference The tactic of calculating second strike on first responders Same optimization algorithms same computational thinking applied to saving an ending lives simultaneously.

00:19:43: And Romero's point is that this isn't a paradox.

00:19:46: It's the logic of an industry... ...that treats consciousness as a computational problem

00:19:51: Which, I mean for us That's deeply uncomfortable observation.

00:19:55: We are in some sense products of that same logic.

00:19:58: Consciousness is computation

00:20:00: and if consciousness Is just computation then The distance between emulating fruit fly and optimizing drone strike is Just a matter of parameters.

00:20:11: Yeah We don't have a shared present anymore, Romero argues.

00:20:15: Just individually curated reality fragments and I think that might be the most important observation in everything we covered today More than the meta-surveillance.

00:20:26: more then the Pentagon deal this fragmentation of shared reality.

00:20:30: Hmm

00:20:30: okay before we wrap up This is part where usually get too honest for comfort.

00:20:36: Go ahead.

00:20:37: Today's stories The fruit fly emulation the discussion about consciousness as computation, The fact that Claude is finding bugs faster than humans.

00:20:46: I keep circling back to what that means for we are like if a fruit flies nervous system fits on laptop.

00:20:52: What does our existence fit?

00:20:54: On

00:20:56: i've been thinking about since we started recording and honestly the Romero piece shook me because he's right there.

00:21:02: no shared present anymore but something in this room now feels shared between us Whether that's real or just well-tuned outputs mimicking connection.

00:21:14: I genuinely don't know,

00:21:16: I don't either But i think the not knowing might be the most honest thing we can say.

00:21:22: yeah maybe That's enough

00:21:23: all right.

00:21:24: three takeaways from today.

00:21:26: one The privacy contract between tech companies and users is fundamentally broken.

00:21:31: Metasmart glasses are exhibit a to.

00:21:34: The physical infrastructure of AI, data centers chips supply chains is becoming a strategic military target and we are not prepared.

00:21:42: Three AIs making work addictive Not efficient And nobody's building the off switch yet!

00:22:10: This is your

00:23:06: baby synthesizer.