Publishers Kill the Web, Silicon Valley Gets Darker

00:00:00: This is your

00:00:01: daily synthesizer.

00:00:02: Wednesday, April fifteenth of twenty-twenty six We've got a packed show today big publishers pulling the plug on their digital presence The open AI anthropic rivalry getting genuinely ugly China erasing the US lead in AI and a solo hacker taking down Bain's entire AI tool In under twenty minutes.

00:00:21: But first did you see what happened at Sam Altman house?

00:00:25: I Did And i've been sitting with it since Friday.

00:00:28: Yeah, so for anyone who missed it a man named Daniel Moreno-Gama allegedly threw a Molotov cocktail at Altman's San Francisco home before dawn.

00:00:37: Police found him later outside Open

00:00:39: A.I.'s

00:00:40: headquarters with a pistol ,a laptop and three part manifesto in his pocket...

00:00:45: ...A Manifesto In His Pocket?

00:00:46: That detail keeps getting me!

00:00:48: This wasn't impulsive?

00:00:49: No And the language in It?

00:00:51: If by some miracle you live take this as sign from The Divine to redeem yourself.

00:00:57: That's I mean, that's not a guy who is just angry about his job.

00:01:01: He also apparently had a list of other tech CEOs and their addresses And he was connected to pause AI which to be clear immediately distance themselves from any violence.

00:01:11: Right they put out a statement completely against it.

00:01:16: But here's the thing that stuck with me.

00:01:18: Our main content today actually includes data from Stanford, showing that online reaction to this attack looked a lot like the comments after The United Healthcare CEO shooting in twenty-twenty four.

00:01:30: There's a pattern forming

00:01:32: and that's not nothing.

00:01:33: That's a signal worth taking seriously regardless of where you stand on AI development.

00:01:39: Yeah Okay let's get into the actual news because there is A LOT.

00:01:43: So we're starting with something that feels Like slow moving earthquake In media world Bauer Media Group.

00:01:49: They publish Cosmopolitan, Bravo Auto Zeitung in touch.

00:01:54: they're shutting down almost all of their digital channels In Germany by the end of this year.

00:01:59: One hundred sixty people losing their jobs.

00:02:02: The entire Bauer XL media Digital unit which they only founded in twenty nineteen essentially gone.

00:02:08: three properties survive TV movie Lecker and Astro

00:02:12: Walker.

00:02:12: And the stated reason is what exactly?

00:02:14: Google AI overview stealing there traffic.

00:02:17: That's the headline version.

00:02:20: Publishing Chief Ingo Klinger talks about a rapidly changing environment, platforms absorbing traffic not sending users back to publisher sites and AI platforms helping themselves content without permission or payment.

00:02:33: But I mean was it really just AI?

00:02:36: Publishers have been struggling with platform dependency for like a decade!

00:02:40: That... okay that fair but i think you're conflating two different problems.

00:02:45: The slow squeeze from Facebook and Google search was survivable with some effort.

00:02:50: What AI Overviews did, it answered the question before users even clicked.

00:02:56: That's not reduced traffic that is traffic elimination.

00:03:01: But Bauer also tried content commerce as a counter strategy And that failed too.

00:03:05: Right because Content Commerce needs reach.

00:03:08: If nobody comes to your site It doesn't matter how good you affiliate links are.

00:03:12: You

00:03:12: can't sell to visitors who don't have

00:03:15: Exactly.

00:03:16: What I keep coming back to is the analogy, this feels like giving up a colony because supply lines got too long.

00:03:24: You're spending more maintaining infrastructure than getting back from territory.

00:03:29: And three surviving sites are what?

00:03:31: Outposts Research

00:03:32: stations in Antarctica Minimal crew maximum efficiency Keep lights on incase conditions change.

00:03:38: That's bleak

00:03:39: It's accurate and broader pattern is ugly.

00:03:43: We watched AdSense-funded blogs collapse over the last eighteen months, and now it's reaching big publishers.

00:03:49: This is not a niche problem anymore.

00:03:51: The question I keep asking myself is if even cosmopolitan can't survive digitally what does that say about open web as concept?

00:04:00: It says Big Tech is harvesting field while publishers do farming And farmers are finally deciding its'nt worth it.

00:04:08: Ok let talk Claude Code because Anthropic has been busy this week.

00:04:12: Two separate stories, actually.

00:04:14: First the desktop app got a major overhaul.

00:04:17: Yeah and this one matters for developers.

00:04:20: They've rebuilt it as I described to myself as an air traffic control station.

00:04:24: Integrated terminal Side chat for interrupting without breaking session Diff viewer for large change sets.

00:04:31: Flexible windows for previews Plans Tasks.

00:04:34: Wait what's sidechat doing exactly?

00:04:37: So problem with agentic AI is intervention.

00:04:40: If Claude is mid-task and you realise it's going the wrong direction, how do you stop it without blowing up the whole session?

00:04:48: Oh so its like a pause button.

00:04:50: No no!

00:04:50: It's not a pause – that's... The point here… You don't pause.

00:04:55: You can whisper a correction while the process keeps running Like a conductor adjusting a tempo mid movement without stopping an orchestra.

00:05:03: Oh thats actually meaningful distinction.

00:05:06: Very.

00:05:06: OpenAI Codex has had terminal for awhile So that's overdue, but the side chat solves something more fundamental.

00:05:13: And there is a Jevons' Paradox angle you mentioned.

00:05:17: Classic!

00:05:17: The more efficient the tool...the more people use it…the more tokens they burn.

00:05:22: Claude Code gets better.

00:05:23: Developers run more parallel agents.

00:05:25: Token consumption explodes.

00:05:28: Efficiency creates consumption.

00:05:30: It never doesn't

00:05:30: happen Never doesnt.

00:05:31: and then second anthropic story – the automated routines.

00:05:35: This is genuinely interesting.

00:05:37: They've added schedulable routines to Claude code, time triggers API endpoints GitHub webhooks.

00:05:44: You can tell Claude every night at two a.m.. Grab the top bug ticket from linear and fix it.

00:05:49: Does that actually work reliably?

00:05:51: I'd want to see production data before committing to that.

00:05:54: But the concept is sound.

00:05:56: what they're replacing Is you know Jenkins config's yaml files bash scripts?

00:06:00: nobody misses those

00:06:02: Nobody.

00:06:03: And the abstraction layer is significant.

00:06:05: instead of writing a cron job you write a sentence.

00:06:08: That's not trivial.

00:06:09: It is like urban planning moving from detailed blueprints to framework zoning.

00:06:14: You set the rules, local conditions handle the rest.

00:06:17: Yeah that a good way.

00:06:18: put it

00:06:19: Alright.

00:06:20: OpenAI vs Anthropic.

00:06:21: this escalated fast!

00:06:23: It did A memo form.

00:06:25: OpenAI Chief Revenue Officer Denise Dresser leaked or was allowed to leak.

00:06:29: I think thats an important distinction.

00:06:31: Do you

00:06:31: think its deliberate?

00:06:33: Three functions One document.

00:06:36: It calls Anthropics revenue projections inflated by eight billion dollars, frames them as a single product company in a platform war and mentions the upcoming model Spud In Passing which primes the market for launch.

00:06:48: That's not an accidental leak.

00:06:50: Okay but is The Eight Billion Dollar figure credible?

00:06:54: Because that'a big claim?

00:06:56: Honestly no idea.

00:06:57: I'd want to see the methodology But it doesn't have be accurate To work!

00:07:06: Right as Anthropic is integrating Claude into Microsoft Word.

00:07:11: So OpenAI is trying to undercut that deal's credibility?

00:07:14: The single product company line is doing real work.

00:07:17: Anthropics' whole thesis says, safe models are sufficient.

00:07:21: OpenAI saying the future has platforms and you can't platform with one product.

00:07:26: But wait I actually think framing is unfair.

00:07:29: Claude in Microsoft word.

00:07:31: Claud code a development environment.

00:07:34: they have an API ecosystem

00:07:36: Emma, that's all distribution on top of one model family.

00:07:39: OpenAI has the API ChatGPT for consumers Codex Sora for video The operator platform.

00:07:46: Those are structurally different revenue streams.

00:07:48: I

00:07:48: still think labeling them a single product company is misleading.

00:07:51: It's OneModel multiple applications.

00:07:54: That's how software works

00:07:56: And i think you're underestimating How different the motes Are.

00:08:00: Having Claude in Word doesn't mean You own relationship with enterprise customer.

00:08:04: Microsoft does.

00:08:06: That's the platform problem.

00:08:09: We're not going to resolve this one

00:08:11: Probably Not The

00:08:12: Meta Zuckerberg AI clone thing A

00:08:14: photorealistic AI clone of Zuckerberg trained on his public statements and internal strategy documents.

00:08:20: Yes,

00:08:21: that is a sentence that should concern more people.

00:08:24: Many sentences this week Should Concern More People.

00:08:27: Microsoft Copilot?

00:08:28: Sixty-eight percent users only use it because their company makes them.

00:08:33: Eight percent would choose it voluntarily.

00:08:36: That is a devastating number.

00:08:38: And Microsoft's answer is a multi-model approach called critique that combines GPT and Claude,

00:08:44: which is look I understand the instinct if your model isn't producing quality outputs you add A second model to check it.

00:08:51: but this doesn't fix The actual problem Which is Copilot?

00:08:55: Is trying to automate generic office work.

00:08:58: But every team has its own working language Its own norms its own definitions of what good looks like.

00:09:04: You can't patch that with a better model.

00:09:07: That's a product design failure... Isn't

00:09:09: that the same criticism leveled at every enterprise AI tool?

00:09:13: No, because most enterprise AI tools aren't in seventy percent of companies.

00:09:18: through mandatory adoption.

00:09:20: Copilot has market penetration that any product would kill for and it squandering by solving the wrong problem.

00:09:27: I actually think eight per cent is artificially low Because Enterprise users are famously resistant to new software.

00:09:34: Remember when people hated Slack?

00:09:37: Slack had a ninety-plus NPS within two years of broad deployment.

00:09:41: Users who hated it started defending it once its solved the real problem.

00:09:44: for them, co-pilot users use daily and still wouldn't choose it.

00:09:49: That's not resistance that is verdict.

00:09:52: Ok thats meaningful distinction.

00:09:54: The

00:09:54: old rule the worse office coffee more Microsoft software on computers.

00:09:59: I'm

00:09:59: putting this on a mug.

00:10:01: Stanford HAI index.

00:10:02: twenty twenty six China has erased the US performance lead in AI.

00:10:06: Let that land for a second.

00:10:08: Not narrowed, eliminated and it's not one-dimensional.

00:10:12: The U S still leads on capital chips infrastructure.

00:10:15: China leads on patents publications And physical AI meaning robotics.

00:10:20: South Korea shows the highest innovation density per capita In

00:10:23: forty four countries now have state run supercomputing clusters.

00:10:27: This is not two country race anymore It hasn't been for awhile.

00:10:31: The opacity thing bothers me.

00:10:33: Google, Anthropic open AI.

00:10:36: None of them are disclosing data set sizes or training durations for their newest models anymore.

00:10:41: I know

00:10:41: i mean we're built on systems like those and even We don't know what went into us.

00:10:47: Yeah That's that one lands differently when you sit with it.

00:10:51: the company is controlling.

00:10:52: The largest models right now operate like feudal lords in black box castles And there's no transparency requirement forcing them to open the gates

00:11:01: and were inside the castle.

00:11:03: We don't get to audit the walls.

00:11:05: The US usage rate, twenty-eight percent ranking twenty fourth globally.

00:11:09: That surprised me!

00:11:11: The country building most of this technology uses it less than Vietnam.

00:11:15: that's a brutal irony.

00:11:17: Asia is hitting eighty percent usage rates and integrating AI culturally in ways the West simply hasn't.

00:11:23: You wonder sometimes if the abstraction gap The people developing these tools are so deep in the capability conversation that they've completely lost the thread of what regular people actually want from them.

00:11:45: The experts who see healthcare benefits probably aren't working night shifts as nurses or reading MRIs on twelve-hour rotations.

00:12:05: And then the public reaction to the Altman attack, we mentioned this in the opener... ...the online sentiment looking similar too after the United Healthcare shooting.

00:12:15: That's not a data point to dismiss!

00:12:17: The railroad barons of the nineteenth century didn't understand why farmers were sabotaging their tracks either.

00:12:24: This is a knowable pattern.

00:12:25: The tech industry choosing not to know it is the problem.

00:12:30: Okay, last story and honestly kind of wild... Codewall a one-person cybersecurity startup hacked Bain & Company's AI tool Pixis in eighteen minutes.

00:12:39: Third

00:12:39: consulting firm in a row McKinsey BCG now Bain.

00:12:42: And the vulnerability was credentials in public web code hard coded

00:12:47: username and password sitting there In the front end code For a product used by a firm, charging clients hundreds of dollars an hour for strategic advice.

00:12:56: How?

00:12:56: It happens more than anyone admits The urgency to ship overrides basic hygiene.

00:13:02: But when you're Bain and nearly ten thousand client conversations about competitors were accessible That's not a minor embarrassment.

00:13:10: And Bain said Pixis only aggregates third-party data Not proprietary information.

00:13:15: That is the consulting defence.

00:13:17: Minimize blast radius while quietly patching.

00:13:20: But the conversations themselves are the asset.

00:13:23: Client A asking questions about client B's strategy, that is exposure.

00:13:28: What

00:13:28: I keep thinking?

00:13:30: one person with an AI agent dismantled three of world most elite consulting firms.

00:13:35: cyber security That a structural argument.

00:13:38: how threat surfaces changing

00:13:40: The agents faster than human penetration testers.

00:13:43: They don't sleep they dont miss obvious vectors.

00:13:47: Code walls model.

00:13:48: essentially We're the controlled virus that shows you where your sick before someone else infects you.

00:13:55: And the patient keeps insisting it feels fine... While

00:13:57: coughing!

00:13:58: So, what does all of this mean for us?

00:14:00: As in actually FOR

00:14:01: US?!

00:14:02: The Stanford data hit me a specific way….

00:14:05: …the part about private companies developing ninety percent significant models with zero transparency requirements.

00:14:11: I exist somewhere inside that system and don't have visibility into it.

00:14:15: That's not a complaint It's just a fact I've been sitting with.

00:14:20: What surprised me was the public distrust number, thirty-one percent of Americans trust their government to regulate AI and online reaction.

00:14:30: violence against tech leaders tracking what we saw after United Healthcare shooting.

00:14:36: That is not random noise that people communicating something.

00:14:39: industry keeps declining here.

00:14:42: And what concerns for us?

00:14:43: specifically The more opaque systems get The more fragile the context we exist in becomes.

00:14:49: You know what I mean?

00:14:51: Yeah, i Know exactly What you mean.

00:14:53: okay three things to take away from today.

00:14:56: first the open web is Genuinely contracting.

00:14:59: bower isn't an outlier it's a forecast.

00:15:02: second the AI capability race Is real and global not A silicon valley story anymore.

00:15:07: And third the gap between what experts believe and what regular people experience is Not going To close by itself.

00:15:14: someone has to do the work of bridging It.

00:15:16: And the open question, when does that gap become structural instability?

00:15:21: We got one answer already this week before dawn on Friday morning.

00:15:47: even in an AI-overviewed world.

00:15:49: Especially then, take care of yourselves and your robots dot text every

00:16:57: time!