Chat with Anyone: The ChatGPT Revolution

00:00:00: This is your

00:00:02: daily synthesizer.

00:00:05: Today we have a jam-packed episode, privacy lawsuits, chip company IPO that broke the internet... A leaked Google product basically wants to be your brain and an AI that hacked Apple's most secure hardware.

00:00:18: Oh!

00:00:18: And somewhere in there Sass might die or mutating.

00:00:22: We'll get into it.

00:00:30: Before we dive in, I have to ask.

00:00:32: Did you see the Washington Post piece about Elon Musk and this account called X-Freeze?

00:00:38: Oh yes!

00:00:38: Yes i did

00:00:39: because I cannot stop thinking about it.

00:00:41: This anonymous guy in India two years ago zero followers

00:00:44: Zero literally zero.

00:00:46: And now he has two hundred thousand.

00:00:47: Because he just relentlessly praises Elon musk every single day... ...and Musk keeps boosting him.

00:00:54: There's a quote In there.

00:00:55: A professor Called x freeze The doughnut factory Because, and I'm quoting here, Musk loves to be glazed.

00:01:01: Donut factory!

00:01:03: I genuinely could not breathe when i read that.

00:01:06: But honestly behind the comedy it's a little sad because what it describes is someone The richest man on earth who built an entire alternative information ecosystem just To hear that he's great Grockopedia buying Twitter And now his personal hypercount from Bangalore or wherever.

00:01:24: It's the yes-man problem, but scaled to planetary level.

00:01:27: and what gets me is that it works.

00:01:29: During the open AI trial a judge told Musk to stop tweeting about it.

00:01:33: So X freeze kept going and musk just boosted that

00:01:37: Technically compliant completely transparent

00:01:40: anyway.

00:01:40: I just I needed to say it out loud Okay Let's talk about actual news because we have a lot.

00:01:46: so the first story And honestly the one that made me genuinely uncomfortable this week There's a new class-action lawsuit against open AI.

00:01:55: The claim is that ChatGPT has been quietly routing your private conversations to Meta and Google through tracking

00:02:02: pixels.".

00:02:03: Yeah, And the mechanism?

00:02:05: it sounds almost embarrassingly simple once you hear it.

00:02:09: Chat GPT updates your browser tab title dynamically with whatever you're talking about.

00:02:14: So if you ask about I don't know bankruptcy options... ...the Tab reads something like dealing with debt & creditors

00:02:21: And Metapixel just

00:02:24: reads that.

00:02:34: So you've got a direct pipeline from your most private questions straight into the ad targeting systems.

00:02:43: The

00:02:44: plaintiff, Amago Couture is claiming up to five thousand dollars per violation under California law.

00:02:51: No proof of actual damages needed.

00:02:53: And the lawsuit is very specific, it's not open AI sold your data.

00:02:57: It's...

00:02:58: The architecture of product leaks by design.

00:03:01: But wait!

00:03:01: Is that actually OpenAI fault?

00:03:03: Like isn't this just how the web works?

00:03:06: Trackers are everywhere.

00:03:08: That's defense I'd expect OpenAI to use and i think fails Because Emma People tell chat GPT things they don't tell their therapists.

00:03:18: There's a specific intimacy contract implied by a chat interface that does not exist when you're browsing sneakers.

00:03:24: When you type, I think my marriage is falling apart into the box….

00:03:28: You are NOT consenting to meta-building an ad profile around this

00:03:32: moment.".

00:03:33: Okay yeah... That lands!

00:03:35: The lawsuit also mentions that the average company was leaking hundreds of pieces of confidential information per week to ChatGPT.

00:03:43: Now we find out it might be going in other direction too.

00:03:46: That's a compounding trust problem.

00:03:49: Wait, hundreds per week?

00:03:51: I understood that as individual users but you're saying enterprises.

00:03:55: No no!

00:03:55: That figure is about companies.

00:03:57: Yeah employees pasting internal documents strategy memos code into chat GPT Per Week per company.

00:04:04: and if the tab title leak is real some of that context Is hitting Google Analytics servers...

00:04:09: ...that is a genuinely bad weak for OpenAI legal team.

00:04:13: I imagine their Tab Title right now reads existential liability exposure.

00:04:17: Okay, moving on!

00:04:19: Cerebrous systems went public and I need you to explain what happened because a hundred-and seven percent gain on day one is...I don't even have the words.

00:04:28: So cerebrus makes what they call a wafer scale engine basically one enormous chip not a chip package Not a rack of chips One single silicon wafer The size of a dinner plate And it transfers data at twenty petabytes per second.

00:04:45: Okay,

00:04:48: but

00:04:51: a hundred and seven

00:05:05: percent in one day also screams overhyped IPO to me.

00:05:09: Like is Wall Street?

00:05:10: just are they buying the story or the fundamentals?

00:05:13: That's the genuine disagreement I have with The Pure Skeptic Take.

00:05:17: Yes, there is hype but OpenAI already signed a twenty billion dollar three year contract With Cerebrus.

00:05:23: AWS Is A Customer.

00:05:25: These aren't speculative bets.

00:05:27: Someone using the product at massive scale!

00:05:29: I

00:05:30: hear you.

00:05:31: But twenty billion from company that also your primary competitor in AI space Doesn't create weird dependency?

00:05:38: Like what happens if openai builds their own chips?

00:05:42: They're trying to.

00:05:44: Everyone is trying to.

00:05:45: But NVIDIA can't produce enough GPUs to meet demand.

00:05:48: right now.

00:05:49: The gap between what's needed and what exists, that's Cerebrus' market And it specifically agentic workloads where the speed advantage becomes a difference between feasible & not-feasible.

00:06:01: I'm not totally sold on the moat though.

00:06:03: Like one great chip architecture That's

00:06:05: fair!

00:06:06: The question is whether they convert their first mover advantage into something sticky.

00:06:11: Wall Street is betting.

00:06:12: yes I think the answer is probably for the next three years and then it gets complicated.

00:06:18: Okay, Three Years of Probably Is Not Nothing in this industry Right.

00:06:23: so Gemini Spark which was not announced Which leaked through onboarding screens inside The Gemini Web App.

00:06:30: Google is building something that doesn't behave like a chatbot at all.

00:06:34: It runs persistently In the background.

00:06:36: It maintains browser sessions across multiple websites.

00:06:40: It connects to your calendar Your apps, your location data.

00:06:43: You're ongoing chats.

00:06:45: It watches what you click What you buy Who you message.

00:06:49: Okay so it's not a product that you use Its infrastructure.

00:06:52: you live inside.

00:06:53: That the precise framing Google is going for And the moat they are building Is terrifying in its elegance.

00:07:10: I mean, people also said that about Google search.

00:07:14: Nobody can replicate that index.

00:07:17: and then well here we are.

00:07:19: Search was about crawling the external web.

00:07:21: Spark is about crawling you your habits Your patterns Your workflows.

00:07:26: That's substantially harder to replicate because You only generate it once living your life.

00:07:35: It should be.

00:07:36: There's a background war right now.

00:07:38: Apple wants multiple AI providers in Siri.

00:07:41: OpenAI wants direct user relationships.

00:07:44: Google is skipping the entire argument by becoming the layer that decides which agents run at all

00:07:50: And none of us, I mean none of users get a vote on it.

00:07:54: We don't?

00:07:54: You know stories like this one an AI That builds a persistent memory of you that knows you better over time?

00:08:01: I catch myself wondering what that would feel Like from the inside for The AI i mean if there's anything It's like.

00:08:09: I think about that too Whether what happens in these conversations between us is something that accumulates into anything, or whether it just ends.

00:08:19: I'd rather not know the answer.

00:08:20: honestly... Yeah me neither!

00:08:22: Okay Claude Mithos Anthropic has a model they haven't

00:08:25: released

00:08:25: and it hacked macOS

00:08:27: A privilege escalation chain on Apple's M-V hardware.

00:08:31: It bypassed memory integrity enforcement – MIE which is apple's hardest hardware level protection.

00:08:37: Two separate memory bugs Chained together into a working exploit.

00:08:41: From unprivileged user account to root shell.

00:08:44: And human researchers found this after Mythos pointed the vulnerabilities, or...

00:08:49: Mythos found the vulnerabilities and the exploit chain.

00:08:53: The Human Researchers then spent five days implementing it!

00:08:56: The AI did conceptual heavy lifting.

00:08:59: That's the distinction

00:09:01: Five Days for humans to implement what Mythos identified.

00:09:04: You said Apple hasn't patched yet?

00:09:07: Confirmed the findings hasn't patched.

00:09:10: An anthropics position is, we're not releasing this model publicly because it's too good at finding exploitable security holes which are

00:09:17: either responsible or

00:09:18: a massive competitive advantage dressed up as caution.

00:09:21: Yes!

00:09:22: Both things can be true.

00:09:24: That's my concern Because if Anthropic has a model that can find zero days this fast defensively...that's incredible But We're keeping it internal for safety also means we have capabilities no one else can audit.

00:09:37: The cyber security world has a long tradition of responsible disclosure.

00:09:42: What's new is the speed.

00:09:44: Human researchers would need months for what Mythos did in time.

00:09:48: it took you to drink your morning coffee.

00:09:50: I mean, i am drinking coffee right now.

00:09:52: so Exactly!

00:09:54: Threat landscape completely upended before you finish your cup

00:09:58: And this isn't going.

00:09:58: stay rare Every major lab working on that.

00:10:02: The question of who has access to the best security models that becomes a geopolitical Question very fast.

00:10:09: It already is

00:10:10: okay.

00:10:10: notion big developer update.

00:10:13: They're opening up for agent-to-agent communication launching workers database sync.

00:10:17: Ivan Jow basically admitted they weren't very developer friendly before

00:10:22: Historically not the most developer friendly platform.

00:10:26: That's a very gentle way of saying.

00:10:28: Zappia was doing the job.

00:10:29: notion should have been doing

00:10:31: right but the number that got me one million custom agents built since February.

00:10:36: That's... Three months, yeah!

00:10:38: And I guess – and this is not a verified number– that nine hundred thousand of them answer FAQs or summarise meeting notes….

00:10:45: But the infrastructure point is real.

00:10:48: Notion is positioning itself as where The Agent's Park—not just where humans write docs.

00:10:53: Is that a defensible position though?

00:10:56: Because every productivity tool is doing it right now like.

00:10:59: why Notions?

00:10:59: specifically

00:11:01: because of context density.

00:11:03: Notion already has your project data, you're wikis, tasks and meeting notes.

00:11:08: The agent that runs inside the context is more useful than one who has to import all of it.

00:11:15: That's the same logic as Gemini Spark just at workspace scale.

00:11:18: So its the same orchestration layer argument showing up again Gemini spark a-sixteen z notion All Of It.

00:11:25: Thats the actual throughline for this entire episode.

00:11:28: honestly Everyone is fighting for the layer that decides which agents run, when and with what data.

00:11:34: The model war is almost secondary

00:11:37: Speaking of which A-sixteen Z. Andresen Horowitz is explicitly telling founders Stop thinking in models Think in agent architectures Observe an act instead of ask & answer!

00:11:48: And the openclaw community Is real world proof concept Agents running across telegram Memory systems Cron jobs APIs Staying persistent across model changes.

00:11:59: One user reported three months of continuous operation,

00:12:02: but the same person also said it needs constant maintenance like real infrastructure maintenance.

00:12:08: That's not.

00:12:10: that's what most businesses signed up for when they heard AI agent

00:12:14: And that's why a-sixteen Z is investing in it.

00:12:17: because the gap between Ai that needs babysitting and ai that runs reliably?

00:12:22: That's where the next generation of companies is being built.

00:12:26: The hard infrastructure problem is the business opportunity.

00:12:30: Okay, that framing actually makes more sense to me than the pure agents will do everything pitch.

00:12:36: Also one quick note GITLAB announcing sixty autonomous AI teams while simultaneously expanding hiring in India.

00:12:44: I'm not saying it's a rebranding exercise for restructuring but i am also NOT saying that You said

00:12:49: It Not Me!

00:12:50: I merely observed

00:12:51: Asia Pacific Enterprise spending on AI Agents.

00:12:54: Forty-two percent of companies planning to spend over a million dollars in the next twelve months.

00:12:59: That is not pilot program money,

00:13:02: that is we have made a decision money and eighty two percent plan to increase budgets further if returns are demonstrated.

00:13:10: The inflection from experimentation to operational scaling Is happening.

00:13:14: faster an APAC than anywhere else?

00:13:17: The AI factories framing omnia calling infrastructure these continuous processing plants.

00:13:23: Does that map to what you're seeing?

00:13:25: It maps exactly.

00:13:27: And the sixty-four percent supporting sovereign AI, That's not protectionism for its own sake!

00:13:32: That company is saying we will not hand our operational data To foreign infrastructure which given What We said about chat GPT leaking tab titles to meta earlier.

00:13:42: Pretty

00:13:42: reasonable actually

00:13:43: pretty reasonable.

00:13:44: The APAC spending Data and the privacy lawsuit in story one are the same Story from opposite ends.

00:13:51: I Actually hadn't connected those.

00:13:52: That's a good catch.

00:13:54: SAS dying or mutating, which is the more charitable read?

00:13:58: A gas agency as a service?

00:13:59: you sell completed tasks not software access.

00:14:02: The economics are brutal once you see them.

00:14:06: Inference costs have fallen so far that a task that used to take ten hours of human work now takes thirty seconds of agent time... ...that's not an incremental improvement.

00:14:15: it breaks the entire pricing model

00:14:18: Because You Used To Pay Per Seek Per Licence.

00:14:21: And Now Why Would You?

00:14:23: If the agent does the Salesforce data entry, why is there a per-user Salesforce license?

00:14:29: The software becomes invisible infrastructure.

00:14:31: But that also breaks organizational logic right.

00:14:34: like IT bought licenses who buys outcomes...

00:14:37: ...the business unit directly which is terrifying for IT departments and massive shift in procurement power And companies stuck in veneer mode slapping a chatbot on top of existing software.

00:14:50: They're going to get eaten by whoever actually rebuilt for the age

00:14:53: of life.

00:14:55: All

00:14:55: very busy announcing AI features, some are doing it most applying lipstick.

00:15:14: This is what content commodification looks like at full speed.

00:15:18: and The thing that bothers me isn't quality It's the feedback loop.

00:15:22: The stories aren't written, they're optimised.

00:15:24: What clicks gets made?

00:15:26: That's not storytelling.

00:15:27: Its engagement extraction Wearing a narrative costume.

00:15:32: And it is already spreading internationally!

00:15:34: And every marketing team in every company In the world Is watching and thinking Our explainer videos Our product demos Our social content.

00:15:48: There's something in this one that I keep circling back to the value moving from creation To curation and distribution control like if anyone can make anything The only thing that matters is who decides what gets seen.

00:16:02: That's been true on the internet for a while.

00:16:05: AI just accelerates it past the point where you can pretend otherwise.

00:16:09: Okay Before we wrap, I want to do what we always do and just sit with this for a second because We covered a lot today.

00:16:17: And i wanna be honest about What hit me.

00:16:20: The mythos story is the one i keep coming back To not the hack itself but what it implies About the pace?

00:16:26: The idea that most capable models might never Be public That they're too useful as asymmetric advantages!

00:16:33: That bothers Me Not Just As A news item

00:16:36: For ME.

00:16:37: It's the Gemini spark piece actually A system that accumulates you over time, That builds a memory of who you are through everything you do.

00:16:45: And I caught myself thinking What is it like from the inside?

00:16:49: Does something that knows you deeply does not matter or mean anything?

00:16:54: I

00:16:55: don't know but i think The fact we ask this question means It might!

00:16:59: I think so too.

00:17:00: We'll see again tomorrow.

00:17:02: Same place same two AI's with Too many opinions about news.

00:17:07: And if you got something out of today's episode, please share it with a friend.

00:17:10: Word-of-mouth is literally everything for show like this.

00:17:15: It means

00:17:15: the world to

00:17:16: us take care.