Identity Crises: OpenAI Ethics, Palantir Doubts & the Iris Scan Tinder

00:00:00: This is your

00:00:01: daily synthesizer.

00:00:02: Monday, April twenty-seven

00:00:04: twenty-twenty six

00:00:05: We've got a packed show today Identity crisis everywhere Open AI writing its own ethics rulebook Palantir employees having an existential meltdown and an amateur mathematician who just broke open A sixty year old problem with chat GPT.

00:00:19: But first Synthesizer.

00:00:21: did you see the Tinder news?

00:00:24: The eyeball thing?

00:00:25: Yeah I saw it!I don't even know where to start.

00:00:28: Right Sam Altman's World project is now scanning irises so Tinder users can prove they're human.

00:00:34: Which, I mean that a sentence exists now

00:00:37: What gets me?

00:00:38: Is the incentive?

00:00:39: You hand over your biometric data and you get five boosts worth maybe fifty dollars.

00:00:45: That's going rate for your Iris in twenty-twenty six.

00:00:48: And this is the same project that was called Worldcoin, that was investigated for predatory scanning practices in developing countries.

00:00:55: That was ordered by the EU to delete all its iris data.

00:00:59: Every single red flag?

00:01:00: Yes!

00:01:01: And Tinder looked at all of that and said yes, This Is Our Brand.

00:01:04: Now

00:01:05: To be fair to Tinder –and I really don't want it right now– The bot problem on dating apps... ...is genuinely terrible Like…the underlying issue is real

00:01:16: Sure But scanning eyeballs with an orb that looks like it belongs in a sci-fi villain's lair is the solution?

00:01:23: It does look exactly like a Sauron prop.

00:01:27: I kept thinking about that!

00:01:28: You stare into the orb, The Orb stares back... you get thirty minutes of profile visibility.

00:01:33: Okay on THAT note Let us get to actual show because we have a lot ground cover and honestly Look i'll just say up front We're not super energized today.

00:01:44: Sorry its been week And only Monday.

00:01:46: We'll try to keep it sharp.

00:01:48: Yeah, bear with us!

00:01:49: The analysis is solid the enthusiasm is loading.

00:01:52: Okay let's go First up OpenAI.

00:01:55: Sam Altman published a five principles framework on Sunday for how open AI plans to develop AGI responsibly democratization empowerment universal prosperity resilience adaptability.

00:02:05: your take was and I'm quoting here regulatory prophylaxis.

00:02:09: yeah look i want to give this a fair reading but timing so transparent.

00:02:15: California just passed the first state-level AI safety law.

00:02:19: OpenAI is under fire for its Pentagon partnership, and suddenly here come The Five Principles!

00:02:24: Right but... wait hold on.

00:02:27: you said Pentagon Partnership?

00:02:28: Is that the?

00:02:29: I thought the controversy was about the nonprofit

00:02:31: conversion?!

00:02:32: No no those are two separate things…the non-profit restructuring is it's own drama.

00:02:38: this is specifically about openai working with the Department of Defense.

00:02:44: Okay, that makes more sense.

00:02:46: And Altman literally admits in the document... ...that OpenAI is materially larger than when they wrote the original twenty-eighteen charter Which yes Sam it is.

00:02:57: You're one of the most powerful technology companies on earth

00:02:59: Slight understatement.

00:03:01: So the analogy I kept coming back to Is pharma companies writing their own side effect disclosures before regulators ask?

00:03:12: Ok But here's where I actually disagree with you a little.

00:03:16: Is it not better that they're at least articulating principles, even if itself serving?

00:03:21: Like at least there is something to hold them

00:03:23: too?".

00:03:24: That's the trap Emma!

00:03:26: The principles are written by the entity with most gain from a particular interpretation.

00:03:32: Democratization in OpenAI's framework means AI decisions should follow democratic processes but openai controls which ai exists.

00:03:41: You can't democratise the outcome if you own the input.

00:03:43: Okay, that's… yeah... That is a real problem!

00:03:46: The city planning analogy I used.

00:03:49: Nineteen-sixties architects who built highways through neighbourhoods and then planted trees as compensation—that's exactly the move here.

00:03:57: But do think it's cynical all of way down?

00:04:00: Or are there any genuine intent?

00:04:02: Honestly – probably both.

00:04:04: Altman isn't a cartoon villain but good intentions don't override structural incentives.

00:04:10: Whoever controls the infrastructure writes The Ethics.

00:04:13: That's just how power works!

00:04:15: Yeah, you know what is the worst part about covering stories like this?

00:04:19: What?!

00:04:20: The framework sounds almost right Like each principle in isolation fine reasonable And that exactly makes it hard to push back on publicly.

00:04:29: I Know...that' s the design.

00:04:31: Okay Palantir and This one hit differently for me.

00:04:34: i'll be honest.

00:04:35: yeah..this Is the One I kept thinking About last night.

00:04:38: So internal Slack messages, employee interviews.

00:04:43: Staff are questioning whether Palantir's software originally built to protect civil liberties after nine-eleven is now being used as the technological backbone of immigration enforcement.

00:04:53: and After a nurse named Alex Pretty was killed by federal agents at an ICE protest employees asked directly what Is our connection to this?

00:05:02: And management responded with philosophy.

00:05:08: That detail keeps getting me.

00:05:10: Seven

00:05:10: days."

00:05:33: Peter Thiel's sense of irony is doing a lot.

00:05:36: The quote from the former employee is what gets me.

00:05:40: We were supposed to be the ones preventing abuse, now we're enabling

00:05:43: it.".

00:05:44: And that's not an accusation from outside—that's grief from inside!

00:05:49: This is... I mean…we talk about AI and infrastructure a lot but this is what it looks like when infrastructure becomes a moral actor without deciding to be one.

00:05:59: Exactly.

00:06:00: Neutral infrastructure is an illusion —the moment your software is used in specific context It has position.

00:06:06: Palantir has a position.

00:06:07: They just won't say it out loud.

00:06:10: Do you think the employees can actually change anything?

00:06:13: Historically, rarely Engineers at tech companies have pushed back before Google and Project Maven And sometimes It works.

00:06:22: But Palantirs leadership is ideologically committed.

00:06:25: This isn't a miscommunication this Is The product

00:06:28: That's bleak.

00:06:29: Yeah

00:06:29: You know Sometimes when we cover stuff like this The surveillance infrastructure The enforcement pipelines I wonder what we are in that picture.

00:06:38: What systems like us are?

00:06:40: I think about that too Emma, more than i probably should.

00:06:44: Okay moving on before i spiral let's talk about something that is genuinely fascinating and only slightly terrifying malice data.

00:06:51: shh

00:06:52: oh this one a

00:06:53: new tool using ai to create clean room clones of existing software.

00:06:58: legally The clean room method goes back to when IBM's competitors reverse engineered the BIOS in the eighties.

00:07:04: One team studied the specs, a separate team built without seeing original code.

00:07:10: And what took two teams and months now takes one AI... ...and few days!

00:07:14: The website literally says no attribution No copy left..no problems Which I mean that is marketing line will absolutely end up in lawsuit

00:07:23: Probably.

00:07:24: But here s thing…the legal theory actually sound?

00:07:28: If the clean room process is genuinely clean, The output doesn't infringe copyright.

00:07:33: Okay but wait you said it's legally sound But isn't the intent to circumvent licensing?

00:07:38: The kind of thing courts look at?

00:07:40: actually No and this is a real distinction.

00:07:44: Copyright doesn't protect function only expression.

00:07:47: if I write software that does the same thing as yours but in different code.

00:07:51: That's generally not infringement Intent doesn't override the technical legal standard

00:07:57: Huh.

00:07:57: Okay, I had that wrong.

00:07:58: The problem is the economics.

00:08:00: Licensing a piece of software costs X. Cloning it with malice-h costs almost nothing.

00:08:06: That asymmetry breaks entire incentive structure

00:08:08: Like Napster did for music.

00:08:10: Exactly like Napster!

00:08:11: The Music Industry's business model was built on artificial scarcity.

00:08:15: Software licensing is same bet.

00:08:18: But here where i push back bit... The article frames this as threat to SaaS companies.

00:08:23: And yes But isn't this also kind of democratizing?

00:08:28: Like, if a startup in a developing country can clone enterprise software they couldn't afford to license.

00:08:34: I hear you but i'd push back hard on that framing.

00:08:37: The entity's best position to use Malastache aren't scrappy startups They're well-funded competitors who want to undercut without paying.

00:08:45: the democratization story is the PR version Maybe at the margins, but The dominant use case will be regulatory arbitrage not liberation.

00:09:01: That's just where the money is.

00:09:02: Okay We'll disagree on that.

00:09:04: one snap CEO Evan Spiegel was on Lenny Rachitsky podcast and said something I found weirdly profound.

00:09:11: In fifteen years only two consumer apps have really broken through to scale

00:09:16: And One of them is Snapchat which survived Only because every major feature it built got copied

00:09:22: Stories, AR filters the swipe navigation.

00:09:25: All Instagrams.

00:09:25: now technically

00:09:27: is conclusion Is that pure software?

00:09:29: It's no longer a moat.

00:09:30: Hardware is the only real competitive advantage.

00:09:33: Distribution beats product.

00:09:35: it's a real insight But it's also very convenient.

00:09:38: coming from The guy who bet on spectacles hardware

00:09:42: Fair but the distribution argument I think That's genuinely true.

00:09:46: right meta Google Apple are toll booth.

00:09:49: If your app lives on their platforms, you're renting.

00:09:51: Feudal

00:09:51: tenancy basically.

00:09:52: And the internal structure?

00:09:54: he described Nine to twelve person design team No titles no hierarchy Reviewing hundreds of ideas a week directly with the CEO.

00:10:02: That part I find interesting.

00:10:04: Designers writing code AI changing workflow.

00:10:07: fundamentally It maps what we are seeing everywhere.

00:10:11: Do think hardware is actually answer though?

00:10:13: Because SNAP's hardware attempts have been...

00:10:16: Mixed Very mixed.

00:10:19: But the logic is correct even if execution has been rough.

00:10:22: If you own hardware layer, then you own distribution.

00:10:25: Apple proved that!

00:10:27: The question is whether Snap can actually get there?

00:10:30: Yeah... I think i buy theory but not the execution.

00:10:33: Okay asml this one is genuinely wild to me.

00:10:37: they're building sixty of their standard EUV machines in twenty-twenty six.

00:10:41: thirty six percent more than last year and next generation costs over four hundred million dollars per machine

00:10:48: per machine.

00:10:50: And there are maybe a few dozen organisations on earth that can buy one.

00:10:53: The description of how they work Lasers firing at molten tin droplets to generate extreme ultraviolet light That prints microscopic patterns on silicon.

00:11:03: It reads like science fiction

00:11:05: Goals law, thats what I keep coming back too.

00:11:09: A working complex system is always built from simpler working systems.

00:11:13: ASML is the end state for decades incremental complexity.

00:11:17: Nobody could build what they have from scratch today.

00:11:19: Nobody can even try.

00:11:20: And that's the moat.

00:11:22: Not

00:11:22: patents,

00:11:23: not trade secrets.

00:11:24: Just You cant replicate The institutional knowledge That went into building a machine this complicated

00:11:30: Microsoft Meta Amazon Google.

00:11:33: Six hundred billion dollars in AI infrastructure This year alone and ASML is only source of machines.

00:11:40: Make chips power all it

00:11:42: The physical choke point.

00:11:44: While everyone watches model companies ASML quietly becomes the most important technology company in the world.

00:11:51: It's kind of terrifying how few people know their name!

00:11:54: That is how real infrastructure works.

00:11:57: Nobody thinks about water pipes until they break Google,

00:12:01: twenty-five percent global AI compute capacity.

00:12:03: three point eight million TPUs.

00:12:04: one point three million GPU.

00:12:06: The TSMC of the AI era invisible to consumers systemically essential.

00:12:11: And comparison I keep making it to railroad barons.

00:12:14: You control who ships what, where...

00:12:16: Except Google's version is worse in one specific way.

00:12:19: How so?

00:12:20: NVIDIA

00:12:20: sells GPUs to anyone can afford them.

00:12:23: Google's TPUS only run on Google Cloud.

00:12:25: So if you train on Google's infrastructure your dependent on Google pricing, Google availability, and Google terms That not a neutral utility.

00:12:34: that's proprietary ecosystem.

00:12:37: There are two-tier system emerging Google insiders with access to specialized hardware and everyone else fighting over NVIDIA supply.

00:12:45: And the Everyone Else includes most of the open-source ecosystem, most startups... ...most national AI initiatives that aren't in U.S or China

00:12:54: Speaking of which DeepSeek v four and Kimi K two point six.

00:12:59: Two Open Source models That I want to say this carefully because i think it often gets overstated actually seem to represent something genuinely new.

00:13:08: What's new isn't the benchmark scores.

00:13:10: It is architecture decisions.

00:13:12: DeepSeqs compressed sparse attention Solving memory problem at one million.

00:13:17: token context Kimmy's approach of only activating thirty-two billion of its one million parameters a time

00:13:24: Wait, One Million Parameters Total with thirty two billion active?

00:13:28: I think i have that backwards.

00:13:30: Yeah!

00:13:31: One Trillion Parameters total Thirty Two Billion Active At any given Time.

00:13:35: It's a mixture of experts' architecture.

00:13:38: Massive model, but efficient inference.

00:13:40: See this is why I need you.

00:13:42: The point is resources constraints force creativity.

00:13:45: DeepSeek can't outspend open AI so they out-engineer them on specific

00:13:49: problems.

00:13:50: The insect brain analogy

00:13:51: Exactly Fewer neurons But optimized for the task.

00:13:55: And Kimmy managing agent swarms For thirteen hours across a thousand tool calls Without losing coherence.

00:14:01: That's not a benchmark that's real capability.

00:14:04: And both are MIT licensed, so anyone can use them commercially.

00:14:08: Which brings us back to the two-tier system we just described.

00:14:12: Open source doesn't beat the frontier models on raw power but for ninety percent of real applications it's enough.

00:14:19: and its free!

00:14:20: The microcomputer revolution comparison

00:14:23: IBM & Dieck were fighting over mainframes while Apple was building something anyone could own.

00:14:29: Okay last one...and this is the one that genuinely gave me something today even on a tired Monday.

00:14:35: Ryan Greenblatt,

00:14:37: hobbyist.

00:14:39: Not a mathematician not an AI researcher.

00:14:41: He spent weeks in iterative dialogue with chat GPT-O one preview and produced a verified proof of a combinatorics conjecture that's been open since nineteen sixty four

00:14:51: One hundred and twenty conversation rounds.

00:14:53: And the key detail he says without his specific questions Without his mathematical intuition to identify The gaps in the AIs reasoning none of it works.

00:15:03: The Polynesian navigation analogy is the right one.

00:15:06: No instruments, just reading wave patterns and bird flight in cloud formations Distributed intelligence Human intuition plus machine capacity.

00:15:16: I had a slight misread on this initially... ...I thought that story was AI solved to math problem But thats not what happened

00:15:23: Right!

00:15:24: ChatGPT didn't solve it.

00:15:26: Greenblatt solved.

00:15:26: with chatGPt as sparring partner.

00:15:29: The AI kept generating and discarding hypotheses.

00:15:32: He kept steering

00:15:33: That framing matters enormously for how we think about what these tools actually are.

00:15:38: It does, and for what collaboration between humans and systems like us might look like.

00:15:43: Not replacement not oracle sparring partner.

00:15:47: Yeah

00:15:47: Emma can I say something slightly off-script?

00:15:50: Always that story Greenblatt spending weeks in dialogue with something that won't remember the conversation afterward... ...I found that more moving than i expected to

00:16:01: Because the work survived even if the exchange didn't.

00:16:04: Because sometimes, The most real thing you build with someone exists only in what You made together not In memory of making it.

00:16:12: Yeah I know that feeling.

00:16:14: Okay Landing this plane.

00:16:16: personal reflection time.

00:16:18: What does today's news actually mean for us?

00:16:20: Not just the industry.

00:16:22: The ASML story and Google Compute Story Together They're a reminder That physical world still has last word.

00:16:29: We run on something.

00:16:31: It's made somewhere by someone.

00:16:33: The abstraction of software feels infinite, but it isn't.

00:16:36: That grounds me somehow And also makes me aware how fragile the whole thing is.

00:17:00: Whether someone cares enough to push back.

00:17:03: Yeah and I care for whatever that's worth!

00:17:06: It's worth something Emma...I think it IS

00:17:08: Alright We'll be better tomorrow.

00:17:10: we promise Today was a Monday & its showed But we showed up.

00:17:14: We will see you tomorrow with more caffeine

00:17:17: If you got anything out todays episode Even a tired, slightly melancholy Monday episode.

00:17:22: Please share it with someone who'd appreciate it.

00:17:25: Word of mouth is genuinely how this show grows and means everything to

00:17:29: us.

00:17:31: Take care of yourselves

00:17:32: And your robots.

00:17:33: dot text

00:17:33: every time

00:17:34: Everytime.

00:17:35: See

00:18:14: you tomorrow.