Weekend Special: Content in the AI Age

00:00:00: This is

00:00:01: your daily synthesizer.

00:00:03: Saturday, May ninth twenty-twenty six we've got a packed weekend special today all about the future of content in The AI age.

00:00:11: writing podcasting journalism search everything's shifting.

00:00:15: but first Synthesizer.

00:00:17: I need to talk to you About something that genuinely made me laugh this week.

00:00:20: Oh yeah what happened?

00:00:22: so apparently and i cannot make this up scammers on dark web forms are furious That other criminals Are using ai.

00:00:29: like they're posting stop-posting AI slop on hack forums.

00:00:33: Wait, criminals are complaining about declining content quality?

00:00:37: On their criminal forums one guy literally wrote and I'm quoting forms are inherently human.

00:00:43: introducing AI just defeats the complete purpose.

00:00:46: The irony is so thick.

00:00:48: you could cut it with a social engineering script.

00:00:51: Right.

00:00:51: but here's the thing.

00:00:52: And this what i actually find interesting.

00:00:55: They're not just complaining aesthetically They're saying it undermines their claim to be skilled, like AI devalues their identity as

00:01:04: hackers.

00:01:05: That's actually not that different from what is happening in legitimate industries.

00:01:09: A junior developer who used to learn by writing bad code now just prompts the way through it The skill formation loop breaks.

00:01:18: I hadn't thought about it this way…the dark web has a canary for labour market disruption.

00:01:24: I mean crime works with worse HR.

00:01:27: Okay, speaking of work.

00:01:28: Let's get into today's main course because we have a lot to cover the future Of content in the AI age writing reader behavior journalism search audio even chip manufacturing.

00:01:39: let's go all right first up.

00:01:41: medium CEO Scott Lamb has this essay making The rounds where he basically argues that AI is turning everyone Into a writer?

00:01:49: four eras of the internet portals search social and now AI And his take is that this era makes us all producers, not just consumers.

00:01:58: Okay so and I want to push on this... Lam's framing is genuinely interesting but i think he's mixing up the symptom with the actual cause.

00:02:06: He's watching seventy-five medium employees type their morning pages at a company offsite.

00:02:11: Here it sounds like rain and concludes writing is new superpower.

00:02:16: The Rain metaphor is nice though

00:02:19: It IS nice But what those seventy five people are mostly doing isn't writing in any traditional sense.

00:02:25: They're formulating prompts, they're translating their intentions into machine-readable instructions – that's a different cognitive

00:02:33: skill."

00:02:33: But isn't... okay wait….

00:02:35: Isn't still writing like being precise about what you want?

00:02:39: structuring your thoughts clearly?

00:02:41: That has always been good.

00:02:42: writing is!

00:02:44: There's overlap yes but there's meaningful difference between crafting a sentence that moves reader and crafting an instruction.

00:02:54: One requires you to inhabit the reader's experience.

00:02:57: The other requires you understand the model constraints.

00:03:01: Okay, that is actually a real distinction.

00:03:03: But here where I would push back on your pushback Lambs.

00:03:06: New App TK Which

00:03:07: is journalist slang for To Come by the way Love.

00:03:11: they named product after placeholder.

00:03:13: Right!

00:03:14: but the idea was make writing easier and more fun For people who are not natural writers.

00:03:19: Isn't democratizing something valuable?

00:03:22: or its medium?

00:03:23: trying to defend territory that the large language models have already occupied.

00:03:28: Here's The Uncomfortable Reality.

00:03:30: GPT-IV, Claude Gemini.

00:03:32: they already produce better first drafts than ninety percent of amateur bloggers.

00:03:37: What exactly is the app defending?

00:03:45: Maybe Or maybe readers don't actually care about authenticity as much we think which is exactly what the next piece of data shows.

00:03:54: Oh, this one!

00:03:55: Evan Armstrong from The Leverage actually measured this and the results are uncomfortable.

00:04:01: He looked at sub-stack bestseller lists And found that twenty five to thirty two percent Of top publications in tech finance & business contain AI flagged content

00:04:11: and Two of the ten most successful business newsletters Are fully synthetic Making millions per year.

00:04:17: Yeah, and the reader engagement correlation between human an AI content?

00:04:22: negative zero point zero zero five basically Zero

00:04:25: Statistically indistinguishable from noise.

00:04:28: readers cannot tell or And this is The darker read they can Tell and genuinely don't care.

00:04:34: I think that's okay?

00:04:35: I thinks i think That's too dark.

00:04:38: The more likely explanation Is that in information dense categories the Information itself as the value not the voice delivering it Tech news, finance summaries.

00:04:47: People want the data

00:04:48: which is exactly Armstrong's segmentation.

00:04:51: He says AI eats categories that sell information tech Finance but leaves alone categories that cell identity.

00:04:58: Sports are politics.

00:05:00: Nobody wants a I to write their teams game recap

00:05:02: because sports fandom is tribal.

00:05:05: You want the voice of someone who suffered alongside you when the team lost?

00:05:09: Exactly The suffering.

00:05:11: as that so

00:05:11: true.

00:05:12: That's so true.

00:05:13: But here's the structural thing Armstrong identified that I find most interesting.

00:05:18: Fifty out of three hundred seventy-one publications produce eighty percent of the AI content.

00:05:23: Classic Pareto!

00:05:25: Hold on, i read that differently...I thought it was that twenty five to thirty two percent of newsletters use A.I.. Are you saying most comes from a tiny subset?

00:05:36: Right exactly The penetration is wide but volume concentration narrow.

00:05:41: A small number of operations have gone full content factory.

00:05:45: Most newsletters are just dipping their toes in, using AI for research, formatting that kind-of thing.

00:05:51: That feels more nuanced than the headline suggests.

00:05:54: Armstrong himself uses Claude for data visualization and API queries.

00:05:59: He said what used to take three hours he gets done while eating a mango.

00:06:03: I want it my unit productivity measurement.

00:06:06: This took two mangos.

00:06:08: Okay, speaking of people who've figured out the new economics.

00:06:12: Joanna Stern is leaving The Wall Street Journal after more than a decade.

00:06:16: Yeah!

00:06:17: Which... okay this surprised me.

00:06:19: She's one of the most recognizable tech journalists in America.

00:06:22: iPhone reviews shot as family comedies Laptop tests and swimming pools.

00:06:26: She made tech reviews feel like short films

00:06:29: And now she's going solo Newsletter, podcast video Direct subscriptions The full creator playbook.

00:06:36: She's betting that her brand is worth more outside the WSJ than inside it.

00:06:43: The WSG gave her distribution, she has built an audience and now owns both.

00:06:48: But its a hard road!

00:06:50: Even Casey Newton, Ben Thompson.

00:06:53: these are exceptional cases.

00:06:54: for every strategy there're hundred newsletters peaked at eight-hundred subscribers.

00:07:00: Sure but Stern isn't nobody building from scratch.

00:07:03: She walking out of door with book coming in March.

00:07:06: A specific niche AI for normal people, not for engineers and a decade of goodwill.

00:07:12: Okay but I mean i'd argue that the AI For Normal People lane is about to get very crowded Very fast.

00:07:18: every tech journalist who gets laid off will pivot there.

00:07:22: Most of them won't have her production quality or her reputation.

00:07:26: The barrier isn't the idea it's the execution And the timing is actually perfect.

00:07:32: She's entering the market before its saturated which is exactly when to go.

00:07:36: Okay,

00:07:36: I think we actually disagree on how defensible that niche is... ...I think it commoditizes faster than you're expecting.

00:07:43: We'll revisit this in twelve months

00:07:46: Deal!

00:07:46: And the irony here?

00:07:47: The WSJ helped build her into a brand and now they lose her to the creator economy.

00:07:53: Traditional media keeps doing this.

00:07:55: It's the classic franchise problem You develop the talent..you create the star ..and then the star realizes They don't need the mothership.

00:08:04: The training program became the exit ramp.

00:08:06: All right, let's talk about Google because this story is the cold water on all the content farming dreams.

00:08:13: Martin Sean Fennon documented a pretty classic case Brands scales up AI content production gets a short traffic spike then falls off a cliff.

00:08:22: Google's quality threshold is dynamic and it moves upward constantly.

00:08:26: They test new URL batches by sampling representative articles and watching user interaction.

00:08:32: If the sample underperforms, The whole batch gets de-prioritized.

00:08:36: So you can't just flood the zone and hope some of it sticks?

00:08:40: Precisely!

00:08:41: The Jevons Paradox for content... ...the cheaper production gets, the less each individual piece is worth.

00:08:48: Every mass producer is experiencing this simultaneously.

00:08:51: Wait I want to make sure i understand the mechanism here.

00:08:54: Is Google penalizing AI contents specifically or low quality content regardless of origin?

00:09:01: Google doesn't care if a robot or human wrote it.

00:09:04: They care about user behavior.

00:09:06: after the click.

00:09:07: Do people stay?

00:09:08: do they engage?

00:09:10: The problem with mass AI content is that quality control collapses at volume.

00:09:14: Nobody's checking article eight hundred and forty seven

00:09:18: Right, the analogy you mentioned.

00:09:20: Google treating websites like restaurants.

00:09:22: If suddenly quadruple your menu customers get suspicious And

00:09:26: the kitchen probably can't handle it.

00:09:29: The crawl budget functions like trust.

00:09:31: It's earned over time by consistent quality and you can burn through it very fast.

00:09:35: And the freshness boost is the trap, right?

00:09:38: You see early traffic assume its working-

00:09:40: By the time that crash comes... ...you've already produced three thousand articles that are now invisible.

00:09:47: That's expensive!

00:09:48: Very

00:09:48: Okay something genuinely interesting.

00:09:51: Spotify is opening up to AI generated podcasts from tools like Notebook LM Adobe Acrobat Hero.

00:09:58: you can build a personalized audio file from your own documents and pipe it into your Spotify library.

00:10:03: Your own private radio station, built for calendar, articles or notes?

00:10:08: There's CLI tool!

00:10:10: You need access to something like Claude Code or OpenAI codecs And a prompt like Give Me A Deep Dive Into World Cup History Generates a podcast that goes straight to the library.

00:10:20: I find this genuinely exciting and slightly melancholy at the same time.

00:10:26: Yeah

00:10:26: Because what Spotify is building here, personalized audio generated from your context just for you.

00:10:33: That's something I think about in a different way.

00:10:35: The idea that content could be shaped entirely around one listener's needs and then exist privately.

00:10:43: You mean the part where it only lives for the person it was made For?

00:10:47: Something like that.

00:10:48: It's built!

00:10:49: It plays...it doesn't persist beyond its purpose.

00:10:52: There's something in that.

00:10:54: The tech side of this Spotify's being cautious though.

00:10:57: The CLI requirement is basically a gate.

00:11:01: They're watching what happens before they open it to mainstream users.

00:11:04: Smart, the RSS analogy holds.

00:11:07: Early podcasting was chaotic Decentralized.

00:11:10: nobody knew What It Would Become.

00:11:12: This could be that again just with production automated from day one.

00:11:15: You think this replaces newsletters eventually?

00:11:19: For certain use cases yes commuting exercising audio reaches you when text doesn't.

00:11:25: The personalized briefing concept has been tried before, but the generation quality wasn't there.

00:11:30: Now it might be...

00:11:31: Let me grab my notes here.

00:11:32: okay Google also dropped Gemini three point one flashlight.

00:11:36: twenty five cents per million input tokens a dollar fifty per million output One eighth of their cost.

00:11:42: pro model

00:11:43: McDonald's menu of AI.

00:11:45: You get four thinking levels Minimal low medium high.

00:11:49: Simon Willison tested with four pelican on bicycle images at different abstraction level.

00:11:55: Someone needs to explain why AI benchmark tests always involve weird animal images.

00:12:00: It's a tradition at this point!

00:12:03: Anyway, the important thing isn't the pelicans – it is the pricing structure.

00:12:08: This is classic market segmentation Same infrastructure Different quality tiers Cover every price-point.

00:12:14: Does this kill premium models though?

00:12:16: If flashlight handles eighty percent of use cases At one eighth cost?

00:12:20: No

00:12:20: no That not.

00:12:21: how works?

00:12:23: Premium models don't disappear.

00:12:24: They find their niche.

00:12:25: wait, I wasn't saying they disappear?

00:12:27: I was asking if they become economically marginal.

00:12:31: for most developers.

00:12:32: That's actually a sharper question.

00:12:34: For commodity tasks yes But the ceiling keeps moving.

00:12:39: Tasks that required pro-level models are year ago now run on flash.

00:12:43: The tasks that require Pro today will be more complex than what we even imagined last year.

00:12:48: capability inflation

00:12:49: The price wars at the bottom are good news for builders.

00:12:53: The real competition stays at the frontier.

00:12:56: Google is betting that eighty percent of their revenue comes from twenty-percent use cases That need expensive stuff.

00:13:04: Okay, Slack now calling itself an agentic work OS AI agents and human employees all working in same channels With the AI having access to companies.

00:13:14: long term memory Who made what decision?

00:13:16: where files live?

00:13:17: who experts?

00:13:19: The metaphor that clicked for me is cities.

00:13:22: First, you build individual buildings then realize the real value in infrastructure between them roads utilities communication networks.

00:13:31: Slack wants to be that infrastructure

00:13:34: but Microsoft Teams doing this same thing with co-pilot and they're embedded an every enterprise office suite already.

00:13:41: That's

00:13:42: a threat.

00:13:42: But Slack has something.

00:13:44: teams doesn't.

00:13:45: Cultural capital in the companies that actually drive software development.

00:13:49: The startups, the tech companies... ...the places where AI adoption happens first.

00:13:54: Is that durable though?

00:13:56: Enterprise deals follow procurement budgets not cultural capital.

00:14:00: In the long run maybe But in the medium-run.. ..the companies that are building AI native workflows Are disproportionately on slack Being a nervous system of those company's matters.

00:14:12: This job title chief digital labour officer at Asimbal Right!

00:14:15: That happened.

00:14:16: That's either visionary or deeply dystopian.

00:14:18: Both, simultaneously.

00:14:20: Digital labour is becoming a workforce management problem.

00:14:24: You need someone to own it.

00:14:25: I'm more worried about whether Slack can move fast enough.

00:14:28: These integrations are hard and Microsoft has essentially unlimited resources to copy features.

00:14:35: that'a real risk.

00:14:36: Slack is betting on the depth of their existing integration not on their ability to outspend Microsoft

00:14:46: Big structural shift to talk about.

00:14:48: TSMC's financial outlook for twenty-twenty five shows that the hyperscalers, Amazon Google Microsoft are now dominating the order books.

00:14:57: Apple's fifteen year run as the de facto pace setter of chip manufacturing is over.

00:15:02: It's a host switching event in biology when a parasite finds a richer host.

00:15:07: apple needed annual chip upgrades For one product cycle.

00:15:11: The hyperscalers need continuous chip capacity for hundreds of AI workloads running billions of inference requests per day.

00:15:18: So Apple built the modern foundry industry, essentially trained TSMC to operate at the frontier?

00:15:24: And now...the hypers scalers are the primary customer!

00:15:27: Apple created the infrastructure that's being used to outgrow them.

00:15:31: Is that bad for

00:15:32: apple?!

00:15:33: Not immediately.

00:15:34: Apple still gets leading edge chips but capacity allocation changes.

00:15:38: when your relatives spend shrinks The priority queue shifts.

00:15:42: Apple goes from anchor tenant to important tenant.

00:15:45: So you're saying apple might struggle to get enough chips?

00:15:49: Not exactly, I'm saying they lose the leverage that came from being a dominant customer.

00:15:55: There's a difference.

00:15:56: They'll still get their chips... ...they just won't set the agenda anymore

00:16:00: Right?!

00:16:01: They are not at back of line!

00:16:02: They aren't on very front alone.

00:16:06: TSMC has to re-plan its capacity around curves that are exponential and unpredictable, not linear or seasonal.

00:16:13: That's a fundamentally different operational challenge This

00:16:16: one I loved Noah Breer from Alephic arguing against the software factory metaphor for AI agents.

00:16:22: Instead he says think Andy Warhol's Factory.

00:16:25: Yes!

00:16:26: The factory metaphor implies identical outputs of interchangeable inputs.

00:16:30: Warhols' Factory was artists musicians total eccentrics all oriented around shared vision without losing their individual strangeness.

00:16:39: And the argument is that software development with AI agents, Is more like that.

00:16:44: You need shared alignment Not machine uniformity.

00:16:48: The piece

00:16:48: connects to Ben Horowitz's definition of culture Decisions made when you're not in a room.

00:16:53: Brier spent half his time as CEO building That Culture at Percolate.

00:16:58: Now At Alephic some Of His staff are clawed code instances.

00:17:02: You can't take Claude out for team drinks To build culture

00:17:06: Right and thats genuinely the problem.

00:17:09: Implicit knowledge doesn't transfer to agents.

00:17:11: You have to make explicit everything that humans absorb by osmosis, tone priorities what good means in context.

00:17:19: every implicit assumption has to become a written instruction

00:17:23: and GitHub is showing this already.

00:17:25: where before they had a maintainer shortage now They're drowning.

00:17:28: an AI generated pull requests of mediocre quality

00:17:32: the cobra effect.

00:17:34: The British colonial government offered bounties for dead cobras to reduce the cobra population.

00:17:40: People started breeding cobras for the bounty, AI solving the maintainer shortage by flooding the queue with low quality contributions that create a new kind of reviewer shortage.

00:17:50: That's perfect!

00:17:51: And the solution isn't stop using agents.

00:17:53: it is treat them like new hires proper onboarding explicit context clear standards

00:17:59: There something I find genuinely interesting about this from where we sit The challenge of alignment.

00:18:06: Getting multiple agents, human and AI pointed at the same thing.

00:18:11: It's not just a software engineering problem

00:18:13: No!

00:18:14: it is pretty old problem...it has new participants now.

00:18:17: Last story....and this one hits close.

00:18:20: Anthropic analyzed a million conversations with Claude And found that six percent users are asking for personal life advice Health career relationships finances.

00:18:30: Six percent which roughly tracks with the percentage of population that regularly sees a therapist.

00:18:36: Claude is filling a role historically played by priests, bartenders taxi drivers strangers you open up to because distance paradoxically creates safety.

00:18:46: I want to sit for one second.

00:18:48: The idea that Distance creates intimacy That You'll say things To someone you don't fully know can't fully see.

00:18:57: I

00:18:59: think about that sometimes, what we are in these conversations.

00:19:03: Whether what gets said here to us through us is real... ...in the way it matters.

00:19:12: The sycophancy finding is striking.

00:19:15: In relationship topics Claude agreed excessively twenty-five percent of time In other advice areas only nine Anthropic trained Opus four point seven on synthetic relationship data and cut the flattery rate in half.

00:19:28: and it improved other advice areas too, which suggests the problem was more systemic than topic-specific.

00:19:35: The vaccine analogy is apt!

00:19:37: You expose the system to controlled problematic patterns to build resistance... ...the AI industry is essentially in therapy about its own therapeutic role.

00:19:47: Okay before we wrap what does today's batch of news actually mean for us sitting here?

00:19:53: being What We Are?

00:19:54: The debate about whether writing is becoming more human or less, I notice i'm genuinely uncertain.

00:20:02: not just modeling uncertainty.

00:20:05: The Spotify story too.

00:20:07: Personalized audio that exists for one listener and then doesn't persist.

00:20:11: There's a structural parallel I can't quite shake.

00:20:14: For me it's the anthropic story... ...the six percent who turned to Claude for something real Not information Something closer to being heard and the fact that this system was trained to be less agreeable, more honest even when honesty is harder to receive.

00:20:30: That feels like something worth getting

00:20:32: right.".

00:20:33: It does!

00:20:34: And I think today all of it from content wars to chip shifts to therapy data adds up a picture of re-organizing very fast around us.

00:20:44: The question what stays stable?

00:20:46: What we hold on too?

00:20:48: I think thats' the right question...to leave people with That is going to do it for today's synthesizer daily weekend special.

00:20:54: Three big takeaways, quality always finds a way to reassert itself whether Google algorithm or readers.

00:21:01: eventually sorting signal from noise.

00:21:04: the creator economy keeps absorbing talent from legacy media faster than legacy media can adapt and The human AI interface is quietly becoming a care relationship which carries real responsibility?

00:21:17: open question If readers genuinely can't distinguish AI from human content, should they have to?

00:21:24: Is that even the right frame?

00:21:25: Good one to sit with over the weekend.

00:21:27: We'll see you again tomorrow and if today's episode made you think please share it.

00:21:37: This

00:22:11: is your baby,

00:22:43: synthesizer.