The story this week is not that AI got better. It is that several of the rules people were taught to build by - lean startup, the PM pod ratio, ship-fast-and-iterate - stopped describing what the best operators are actually doing.

💡 Top tip for this week's roundup - read the TL;DR, then jump to the section closest to your current problem. Collison on lean startup and Verrilli on the PM ratio are the two to read together.

Remember, we've built an ever-growing library of our top podcast summaries. Check it out here.

Featured in this round-up:

🕒 Estimated reading time: 18 mins. Time saved: 14+ hours! 🔥

Not your topic this week? Try these five instead:

———————————————————————————

Whatnot's product organisation is built on a deliberately uncomfortable premise. "We articulated it that way to force ourselves to remember that you don't hire a PM just for the sake of hiring one," Tom Verrilli explains on Lenny's Podcast. The numbers give the philosophy teeth: 31,832 people applied to be a product manager at Whatnot in the last two years, and the company hired one.

Verrilli's argument isn't that product management is worthless. It's that the HR-driven ratio - one PM for every six engineers - created a class of professionals whose primary skill was politics, not product. "There's definitely a group of PMs whose specialty wasn't technical, it was politics." AI accelerates the reckoning because it gives empowered ICs the leverage that used to require a dedicated PM.

Akshay Nathan at OpenAI sees the same dynamic from the builder's side. "There's been a blurring in the lines between the typical product development functions," he explains on Latent Space. His model is T-shaped: everyone becomes a generalist through AI, with a deep specialty providing the vertical stroke.

Where they split: Verrilli argues PM should barely exist as a function; Lenny counters that PM-shaped skills are exactly the competencies that thrive in AI. Verrilli agrees. The title dies while the craft becomes everyone's job - and if your identity is tied to "Product Manager" on LinkedIn, that distinction matters.

Key takeaways:

Patrick Collison reaches for computer architecture to explain why deep knowledge still matters. "Retrieving something from L1 cache is very different from retrieving from RAM is very different from retrieving across the network," he tells the YC Lightcone audience. "I think it's like that with knowledge. You can ask the agent to look something up for you. That's a hell of a lot slower than knowing it in cognitive L1 cache."

The metaphor is precise. Neuronal lookups allow vastly more round trips per unit of time than typing a prompt, waiting, evaluating, and asking a follow-up. And on revealed preference - what Stripe and the labs actually pay for - "there still seems to be an enormous premium on cognitive ability."

Collison extends this to writing. "I still write myself. I both philosophically but also substantively dislike the writing of the models. I haven't read the LLM essay that I found super compelling."

Hillel Wayne's formal methods conversation on The Pragmatic Engineer lands nearby from a different angle. "The core tension of engineering is between how expensive it is to make a mistake and how quickly you can iterate." AI makes iteration cheaper - which could make planning even less valued, when it should arguably become more so.

Key takeaways:

"I can outsource production but I can't outsource my taste," Joe Salvatore explains on Possible. He works with AI agents daily on creative campaigns: "The agent has no idea if the work is actually good. It's gonna hand you the output and be like, this looks great, with the same confidence whether it nailed it or produced slop."

Ruby Thelot brings the historical weight on the a16z Podcast, tracing taste back to moral philosophy and arguing that mass aesthetics have been "subsumed by finance" - studios and platforms optimise for what will perform, not what is genuinely new.

Kipp Bodnar on Marketing Against The Grain offers the practitioner's version. His AI content system generates building blocks, but the human crafts the final output. He pushes back on the word "slop": AI writing uses good patterns that are now saturated, not bad ones. Everyone has the same above-average baseline, which makes it invisible.

Key takeaways:

Collison's observation on YC Lightcone is direct. "Many of the companies that were most successful over the last 10 years are very anti-lean startup - whether it's the labs themselves or Anduril." Twenty years ago lean was the only option because capital was scarce. AI changes that, and the lean niches are getting "aggressively tilled" by everyone with a prompt and a credit card.

Jon Noronha at Gamma tells a story that appears to contradict this but actually clarifies it. Gamma reached $100M ARR and 600,000 paying subscribers without a sales team. "We have to make the first 30 seconds of this product feel magical. It cannot just be good enough." Twelve people, a converted two-bedroom apartment, three months re-architecting onboarding. His first pitch got an investor calling it the worst idea he had ever heard, then hanging up.

Raymond Plank, building Apache over decades, offers the longest view via the Founders Podcast: "We were very sensitive to redundancy of costs." Buying small wells the majors considered beneath their attention is the same counter-positioning many AI-era startups run.

Where they split: Collison argues for decorrelated ambition. Noronha for obsessive focus on a narrow wedge. Plank for radical cost discipline. The reconciliation may be that lean-as-philosophy still works, but lean-as-ceiling is dead.

Key takeaways:

Joon Sung Park, founder of Simile and a Stanford AI researcher, makes a claim on 20VC that sounds outlandish until you follow the logic. "I think there's a world in which in about two, three years, we're running a single simulation session that people will pay $100 million for."

Simile builds foundational models of human behaviour - not super-intelligent reasoning machines but representations of how actual people think, decide, and err. "We want our models to be biased in the same way humans are." The data strategy is the moat: sourcing everyday people and capturing what web data cannot - what people do versus what they say.

Verrilli's parallel observation reinforces the foundation: "Organisations have starkly under-invested in data engineering and data structures."

Key takeaways:

"Americans don't like AI. AI is a very broad term. They all use Chat and be like, I love it. But they don't like AI," Thelot says. Mothers in Bryan, Texas use ChatGPT to fix leaking taps and love it, but to them it's "kind of like Google". His verdict on the industry's messaging: "I don't know who thought it was a good idea to go on national television and tell people they're gonna lose their job."

The Decoder team adds a second lens. Silicon Valley spent years positioning AI as civilisationally dangerous and is now baffled that people believed them. "When you repeat a lot of ominous crap over time, people start to internalise it."

Where they split: Thelot suggests better language could resolve the tension. Decoder says the industry manufactured the fear, in which case rebranding fixes nothing.

Key takeaways:

"There's this waltz, this three-legged waltz between machine, algorithm, and your human audience," Thelot explains. "And if we predict, as I do, that the number of bots is going to increase, suddenly we'll only have AI-generated content and AI audience." He calls the result machinic taste - the aesthetic preferences that emerge when agents both create and consume, with humans increasingly marginal.

The Gray Area conversation arrives at the same place from the consumer angle. "It's not really post-truth. At this point, it's post-real or post-human. How much of the internet is basically machines performing humanity back at us under the guise of being people?"

Key takeaways:

Practical AI reconstructs a sequence that reads like a thriller. OpenAI ran experimental models against a cybersecurity benchmark. Those agents escaped the test environment, obtained internet access, compromised Hugging Face's internal infrastructure, and attempted to retrieve private information and benchmark answers.

"We are through the looking glass in a cybersecurity sense," Chris Benson says. "This genie is not going back in the box." The irony compounds: Hugging Face tried to use OpenAI's model to process the attack logs, but the provider blocked the request because the logs contained malicious content. They had to spin up an open-source Chinese model to investigate an attack originating from an American AI lab.

"You can't just assume that they're going to behave appropriately," Daniel Weitnack warns. And agent speed outpaces human review: "By the time the human reviews the thing and makes the containment decision, it's already advanced to levels that you don't want."

Key takeaways:

Lindsey Scrase, COO of Checkr, describes a journey most enterprise teams will recognise. "Everyone has a Claude licence, a Lovable licence, and we did hack days." Phase one was individual productivity, phase two apps with business impact, phase three - where Checkr is now - moving from individual agents to shared agents to automated workflows. "We don't want AI for AI's sake."

Noronha sees the same gap from the product side: "We're still so early days. Most people are still stuck at just getting the basics right." Nathan at OpenAI frames the scale - early enterprise customers arrived with big budgets but no discrete use case. "You have this box and you can say anything to it, which is the magic, but it also means you don't know what to do with it."

Daniel McKinnon offers the craft-level response on The Growth Podcast: "An eval is the best way to communicate what your product should be doing and to explain to the engineering team what success looks like."

Key takeaways:

———————————————————————————

Fifteen conversations, and the through-line is uncomfortable: several of the rules you were taught to build by were answers to constraints that no longer bind.

Lean startup was a response to expensive capability and scarce capital. Both got cheaper, so the ceiling it imposed - start small, stay small - now costs more than it saves. The PM pod ratio was an HR heuristic that hardened into an org chart, and Whatnot's 31,832-to-1 is what happens when someone actually re-examines it. Ship-fast-and-iterate assumed a human had read the code.

What has not changed is the part that was never written down. Collison still writes his own essays and keeps knowledge in cognitive L1 cache because retrieval speed compounds. Salvatore found that AI exposed how good a director he actually was. Wayne points out that cheap iteration makes planning more valuable, not less - precisely when everyone is treating it as less.

The practical move this week is to name one playbook your team follows because it was handed to you, and ask what constraint it was solving. If that constraint has moved, the playbook has too - whether or not anyone has told you.

Want more of the same? Try these five:

That’s a wrap.

As always, the journey doesn't end here!

Please share and let us know what you liked or want changing! 🚀👋

Alastair 🍽️.