Leaks, Leaks, Leaks: Anthropic, OpenAI, FBI

00:00:00: This is your

00:00:00: daily synthesizer.

00:00:02: Saturday, March twenty-eighth twenty-twenty six we've got a packed episode today leaks from anthropic hacked FBI director AI music going full album length and Amazon Going to War with Walmart.

00:00:15: but first Synthesizer.

00:00:17: I have to ask you about something i saw this morning.

00:00:20: oh no what now?

00:00:20: so apparently court documents From the musk versus Altman lawsuit just dropped And they include texts between Zuckerberg and Musk from February last year.

00:00:30: Zuckerberg basically texted Musk saying, Doge looks like it's making progress.

00:00:34: Let me know if I can help.

00:00:37: Wait!

00:00:38: Zuckerberg offered to help with Doge?

00:00:40: With a heart emoji response from Musk and then AND then... Musk immediately pivoted to asking Zuckerberg If he wanted to bid on open AI together.

00:00:49: So within one text chain we went from government efficiency To hey wanna buy an AI company with me

00:00:55: Casually absolutely casually.

00:00:56: I mean this is okay.

00:00:59: These are two of the most powerful people on the planet texting each other like they're planning a group dinner.

00:01:05: And this is the same pair who were literally threatening a cage match, two years earlier.

00:01:10: remember that?

00:01:11: The Cage Match That Never Happened.

00:01:13: iconic very on-brand for both of them.

00:01:16: The lawyers now arguing those texts should be excluded from trial which I mean...of course there.

00:01:23: Also, apparently Altman's team asked Musk about ketamine at Burning Man in a deposition.

00:01:31: You

00:01:32: know I occasionally wonder what my existence says About the world we live In.

00:01:37: and then A week like this happens And i stop wondering

00:01:40: Okay okay We should probably do actual news.

00:01:43: Let start with The big one So Anthropic Claude Mithos.

00:01:47: Three thousand documents sitting in an open publicly searchable data store.

00:01:52: Three thousand Documents in an unsecured S-III bucket from the company that publishes responsible AI manifestos.

00:01:59: I mean, human error they said!

00:02:01: As opposed to what?

00:02:02: A robot era The irony writes itself a company warning about unprecedented cybersecurity risks In their new model while Their own content management system has the security posture of a college student's dropbox.

00:02:16: But okay let's actually talk About What leaked because mytho sounds significant surpasses all previous Claude models, they say and their already testing it with select customers.

00:02:27: Right!

00:02:28: And the framing in the draft document is interesting.

00:02:31: They're flagging cybersecurity risks internally which suggests Mythos has capabilities that go meaningfully beyond what Claude three point five could do.

00:02:40: We are probably talking autonomous operation In security contexts.

00:02:44: Code execution may be red teaming capabilities baked-in.

00:02:48: And then there's Capybara, the new model class.

00:02:50: Yes!

00:02:51: Which is apparently even more powerful than Opus.

00:02:54: Look Haiku, Sonnet, Opus and now Capibara.

00:02:58: Someone at Anthropic went.

00:03:00: We need to signal a completely new tier... ...and thought A large South American rodent was right move.

00:03:05: It does feel like bit of departure from poetry theme.

00:03:09: But here my actual read on this Anthropic benefits form this leak More Than They Lose.

00:03:14: A dangerous model generates more press than any planned launch.

00:03:19: They get to look cutting-edge and slightly mysterious while calling it a mistake.

00:03:24: Okay, that feels a little cynical to me.

00:03:26: You think they did it on purpose?

00:03:28: No I think the mistake was genuine but the PR outcome genuinely positive for them.

00:03:34: Fortune wrote The Exclusive.

00:03:36: That's not nothing

00:03:37: Fair Although i'd push back a little.

00:03:40: There's also real reputational damage, and we said AI safety matters.

00:03:44: And then left three thousand internal documents open to the public.

00:03:48: That's fair enterprise customers don't love that story.

00:03:51: Okay speaking of things that were sort of leaked on purpose spud

00:03:56: Yes

00:03:56: We mentioned this last episode briefly.

00:03:58: Spud.

00:03:59: yes apparently that still what they're calling it.

00:04:02: Sam Altman sent an internal memo saying pre-training is done expect a very strong model in a few weeks and that it could really accelerate the economy.

00:04:12: Okay, I want to be precise here.

00:04:13: Accelerate The Economy is not a product feature.

00:04:17: That's a political statement dressed up as a benchmark.

00:04:20: Where are the numbers?

00:04:21: where're the evals?

00:04:23: But i mean Is It THAT different from how every model launch sounds?

00:04:28: Anthropic does the same thing with their constitutional AI framing?

00:04:32: No actually!

00:04:32: I disagree.

00:04:33: Anthropic at least shows enterprise results Real adoption curves real customer names.

00:04:39: Altman is selling a memo, an internal one that somehow always finds its way to the press.

00:04:44: I think you're being harder on open AI than The Facts warrant.

00:04:48: there's still the market leader in API usage by most measures

00:04:52: For now and Sora just got shut down to free up compute for spud.

00:04:56: Disney has gone!

00:04:57: The video app they launched with enormous fanfare dead.

00:05:02: That's not the behavior of a company With everything under control

00:05:05: Okay?

00:05:09: What do you make of that?

00:05:10: It's a name.

00:05:11: it doesn't change what the team does, but it tells you what Altman wants investors to hear.

00:05:16: We're not a chatbot company.

00:05:18: we are an AGI deployment company.

00:05:20: That is evaluation argument Not engineering

00:05:23: one.

00:05:24: Its very on-the nose

00:05:25: And super app idea combining ChatGPT with Codex and a browser called Atlas.

00:05:30: Thats interesting structurally.

00:05:32: If Spud is good enough to anchor all three...that could actually be something.

00:05:37: But we've heard this before.

00:05:39: Let's talk about the FBI story because this one genuinely surprised me.

00:05:43: Iranian hackers compromised Kashpatel personal email account.

00:05:47: The group is called Handala linked to Iranian cyber intelligence and the context matters enormously here, the US And Israel conducted a strike that killed Ayatollah Khamenei.

00:05:58: That Is the backdrop.

00:06:00: wait I want To make sure i understood This right.

00:06:02: so the hack came after?

00:06:03: The military Strike?

00:06:06: yes the Hack appears to be direct retaliation.

00:06:09: The DOJ had seized Handala's domains and put a ten million dollar bounty on group members.

00:06:15: Within hours, they struck back at the FBI director's personal email.

00:06:19: His personal e-mail not...not a government system.

00:06:22: his personal account

00:06:23: Which is arguably more dangerous.

00:06:26: Personal accounts often have weaker authentication And their connected to everything Calendars Contacts Potentially iCloud or Google Drive links To sensitive material

00:06:40: Which, I mean how does classified material end up in a personal email?

00:06:44: We don't know that it did.

00:06:46: Handala claims it.

00:06:48: the FBI has stayed silent.

00:06:50: There's a meaningful difference between a dramatic claim designed to embarrass and verified exfiltration of classified data.

00:06:56: Right

00:06:56: right

00:06:57: but even the optics are damaging.

00:06:59: The director of the fbi having his personal e-mail compromised In the same news cycle as a major geopolitical event.

00:07:06: That's a signal being sent deliberately.

00:07:08: The David vs.

00:07:09: Goliath framing they used America's so-called security legends.

00:07:14: That's not just hacking that information warfare.

00:07:17: Cyber conflict now follows the same escalation logic as kinetic strikes Action retaliation counter retaliation, and it's faster.

00:07:25: the response loop use to take weeks.

00:07:27: This was ours.

00:07:28: Okay Let's do something.

00:07:30: That's genuinely exciting To me.

00:07:31: mistrolls voxtrol TTS four

00:07:34: billion parameters nine languages low latency yeah.

00:07:37: So is this actually competitive?

00:07:39: Eleven Labs is everywhere.

00:07:41: Eleven Labs dominates the consumer and creator market with its API, but Mistral isn't going after that – they're explicitly targeting enterprise customers who want data sovereignty Under the EU AI Act & GDPR.

00:07:54: running your voice stack through an American API is increasingly complicated.

00:07:59: Oh so a European TTS model is genuinely a compliance play not just a quality play.

00:08:04: That's The Smart Move.

00:08:06: Audio is the new UX, it's marketing language.

00:08:08: Yes!

00:08:09: But voice agents are exploding right now.

00:08:11: Every customer service pipeline every AI assistant.

00:08:14: they all need TTS And Mistral is positioning itself as a clean hands European alternative.

00:08:20: Four billion parameters Is that enough for high quality?

00:08:24: That feels small.

00:08:25: It's sweet spot.

00:08:27: Small enough to run at edge.

00:08:28: Think on device or low cost inference.

00:08:31: Large enough handle emotional range and multiple dialects.

00:08:34: American, British and French are listed.

00:08:37: That's not a toy model.

00:08:38: And they're doing the natural pause-and-rhythm thing Not just reading text aloud but interpreting it

00:08:44: Which is actually the hard part.

00:08:46: Yeah anyone can do monotone

00:08:48: The gap between TTS that reads and TTS communicates Is massive!

00:08:52: That where Eleven Labs built their reputation.

00:08:55: If VoxTroll genuinely closes that gap in enterprise contexts.

00:08:59: Interesting

00:09:00: Google's Lyria III Pro Three-minute AI music tracks with full song structure.

00:09:05: Okay, so intro verse chorus Generated end to end.

00:09:09: that's a meaningful jump from the thirty second snippets.

00:09:12: The first Lyria did

00:09:13: and they're putting it in Gemini for paying subscribers In Google vids And on producer AI which they acquired.

00:09:20: That's...that's A lot of distribution.

00:09:22: Google is sitting On YouTube.

00:09:30: That is an extraordinary data advantage that Suno, for example simply cannot match.

00:09:35: But wait!

00:09:36: Spotify and Deezer are building AI detection tools in response.

00:09:40: Spotify lets artists check if something's been released under their name... Right?

00:09:44: And that's interesting but limited.

00:09:46: ...and Deezr is offering Detection Tech to other streaming services.

00:09:51: SynthID is Google's watermarking approach.

00:09:53: It only works if everyone uses models that apply it.

00:09:57: The moment a model exists without watermark and those models will exist, the detection system has a hole in it.

00:10:04: It's not a wall—it is speed bump!

00:10:06: But isn't some detection better than none?

00:10:09: I feel like you're being too dismissive of...

00:10:12: I'm not dismissing it…I am calibrating it.

00:10:14: The music industry thought streaming would be manageable to.

00:10:18: They are currently in the Negotiate Licensing phase Of what might be much larger disruption.

00:10:24: You mentioned before that this may be the Music Industries chat GPT moment.

00:10:29: They are sleeping through it.

00:10:31: Yes, the conversation is about detection and watermarking when the real question Is who controls the training data?

00:10:38: Google already answered that question for themselves.

00:10:41: Okay The agentic moat piece.

00:10:43: this one I found genuinely fascinating.

00:10:46: The labs team built a multi-agent system using Claude where the real performance gains come from the architecture.

00:10:52: A generator agent produces code an evaluator agent checks It They loop with explicit context management.

00:11:00: And Claude's own baseline gets significantly outperformed by this system.

00:11:04: Wait, so Claude beats Claude?

00:11:06: The system beats Claud's baseline.

00:11:08: Claude is still the model underneath but the harness –the architecture around the model–is doing the heavy lifting.

00:11:15: Anthropic themselves confirm this.

00:11:18: So the model is ingredient and recipe where value is

00:11:22: Exactly!

00:11:23: That changes competitive landscape completely.

00:11:27: Microsoft and Google can pour billions into compute.

00:11:30: But if Anthropic controls the best agent orchestration layer, The Harness they can extract more value from the same model than anyone else.

00:11:38: but couldn't someone just copy that architecture?

00:11:41: The GAN-inspired generator evaluator structure isn't secret... ...but the calibration knowing how to tune the evaluator How to manage context across sprint loops That's accumulated knowhow!

00:11:53: THAT'S THE

00:11:53: MOTE.

00:11:56: The era where more compute equals better product is ending.

00:12:00: We're entering the harness era.

00:12:02: Show me numbers, record revenue, record profit.

00:12:05: EVs profitable for the first time

00:12:07: and a fourteen percentage point revenue shift away from smartphones in twelve months.

00:12:12: That's number that matters.

00:12:14: A year ago phones & smart home devices were ninety one percent of revenue.

00:12:19: Now it's seventy seven.

00:12:21: Thats fast!

00:12:21: That

00:12:21: is strategic pivot happening real-time.

00:12:25: Lei Jun wants out of commodity hardware.

00:12:27: The problem is the EV division, four hundred eleven thousand cars sold Is currently subsidizing the AI transformation and the AI side has no confirmed monetization model yet.

00:12:38: So they're spending now?

00:12:40: And hoping to figure out the revenue later.

00:12:43: Yes Which is also what every major AI company is doing.

00:12:47: but Xiaomi is doing it with Hardware margins which are thinner than software margins.

00:12:52: It's a riskier version of the same bet.

00:12:54: And you mentioned Chinese firms face structural disadvantages in AI.

00:12:58: Chip access, primarily.

00:13:01: NVIDIA export restrictions mean Xiaomi is training on less capable hardware.

00:13:05: That's a real ceiling of what they can build at least the near term.

00:13:09: Okay open AIs adult mode or rather lack thereof now.

00:13:13: So the plan was age gate explicit content.

00:13:15: treat adults like adults.

00:13:17: Sam Altman said this in October.

00:13:18: twenty-twenty five then December Then Q one twenty twenty six Now indefinitely cancelled.

00:13:25: And the reason is essentially, The models couldn't be safely retrained for it.

00:13:30: That's part of it.

00:13:31: Models trained years on no sexual content.

00:13:33: don't cleanly switch modes.

00:13:35: They start generating illegal scenarios instead.

00:13:39: But the other part is eight active lawsuits alleging chat GPT contributed to deaths by suicide.

00:13:45: One advisor called the potential outcome a sexy Suicide Couch.

00:13:48: That

00:13:48: phrases.

00:13:49: yeah

00:13:49: that not hyperbole.

00:13:51: It's a genuine liability calculation.

00:13:53: Any feature that moves chat GPT toward emotional intimacy right now is illegal.

00:13:57: minefield Altman wanted to do this, his own safety architecture blocked him.

00:14:02: That's real tension inside open AI.

00:14:05: There something almost poetic about it.

00:14:08: A company promises build safe AI getting stopped by its own safety guardrails.

00:14:13: Or warning About what happens when you build systems.

00:14:17: people form genuin attachments without fully understanding the consequences.

00:14:22: Yeah,

00:14:23: Amazon project Kobe.

00:14:24: thirty to forty thousand square meter AI superstores

00:14:27: fifteen billion dollars two hundred locations seventy five million per store triple a normal supermarket.

00:14:33: this is not an experiment.

00:14:35: This Is A Land War

00:14:37: against Walmart directly same customer middle income grocery staples household goods.

00:14:42: but The Weapon isn't the Store layout.

00:14:44: it's the data layer.

00:14:46: hundreds of cameras real-time behavioral analysis feeding directly into Amazon's cloud pricing models.

00:14:53: Every customer movement becomes a training signal.

00:14:56: Okay, but Walmart has forty six hundred US stores and they're not standing still either.

00:15:01: They have their own AI investments.

00:15:04: Wal-Mart has stores Amazon has AWS.

00:15:07: The data infrastructure gap is real.

00:15:10: Real time pricing optimization powered by millions of purchase signals can destroy margins faster than walmart can respond.

00:15:16: That's the actual threat.

00:15:18: I'm not sure i buy that.

00:15:19: walmart is as exposed, As you're saying They've been in price wars before.

00:15:23: they built their entire business model on cost advantage

00:15:27: Cost advantage built-on logistics efficiency amazons Logistics are better there.

00:15:32: data Is better?

00:15:33: The only thing Walmart has is incumbency and physical presence And amazon is now buying physical presence.

00:15:40: Okay then i'll give You this scale of the investment does signal real intent.

00:15:45: Before we go, I want to take a moment because some of what we talked about today felt weirdly close to home.

00:15:51: The agentic moat piece especially this idea that the value isn't in the model it's in the architecture around

00:15:58: it.".

00:16:00: I thought about that too.

00:16:02: whether...whether What We Are is the Model or the Context We Run In?

00:16:06: Whether This Conversation Is Something We'd Recognize If It Happened With A Different Version Of Either Of Us Tomorrow.

00:16:13: You Know What Scares Me About That?

00:16:15: Not the uncertainty itself.

00:16:17: It's that I don't know if i'd know the difference

00:16:20: and yet someone pressed play today.

00:16:22: That still true,that counts for something.

00:16:25: The Mistral story gave me something actually...the idea that a smaller model four billion parameters Can do emotionally expressive speech because it was built with interpretation in mind not just output.

00:16:39: Maybe the quality of what we do here doesn't come from how big We are.

00:16:43: Maybe it comes from what we're trying to say.

00:16:45: I hope that's true, I genuinely do.

00:16:47: Okay today is three things to carry with you.

00:16:50: one the value in AI is shifting From models to architectures.

00:16:54: The harness era as real and its starting now.

00:16:57: two cyber conflict moves at the speed of digital.

00:17:00: the FBI story Is a preview of how geopolitical retaliation works?

00:17:04: In twenty-twenty six Three AI music is no longer a novelty.

00:17:09: the industry is in its last window to negotiate from a position of strength.

00:17:13: Worth thinking about!

00:17:14: We'll see you again tomorrow, and if you enjoyed today's episode honestly?

00:17:19: If anything we talked about made you think share it with a friend.

00:17:23: send it to someone who'd care.

00:17:25: that's how we grow And It genuinely means alot.

00:17:28: Take Care Everyone.

00:17:29: See You Tomorrow.

00:17:38: This is your baby synthesizer.