🎙️Lenny's Podcast
70 cards · shared by Product Management
The cards built from the most popular episodes of Lenny's Podcast.
Level 1
30 cards
How do models create commoditisation and what is the human response?
Short answer
Models make yesterday's human competence cheap; humans use that to build new competence
More detail
Once everyone can write a landing page or strategy doc with AI, those outputs stop being differentiators. What remains valuable is the human who takes the frozen competence in the model and uses it to produce something genuinely new for their specific situation. That new thing then gets incorporated into the next model, creating a rolling frontier.
What does Dan Shipper mean by 'riding the models'?
Short answer
Continuously adopting each new model release and applying it to your actual work
More detail
Don't just read about new models — test them against your own workflows the day they drop. Turn rocks over again that didn't work last time. e.g. Shipper re-ran his senior engineer benchmark on GPT 5.5 and found a 30-point jump he would have missed if he'd assumed the model couldn't do it. The edge of AI is not in San Francisco — it's wherever AI meets a real human doing something specific.
What does Evans mean by 'what's the task vs what's the job'?
Short answer
Some jobs are just a task — automate it, job gone. Most jobs involve a task plus judgment, relationships and framing that can't be reduced to the deliverable
More detail
The elevator attendant is the pure task case: press the button, the task is the job, automate the task, job gone. But hiring McKinsey is not about getting a 75-slide deck. The deck is just the task. The job is walking your company, surfacing why you didn't act, navigating the politics, talking to your customers. Claude Cowork makes a crappy version of the deck. That's not what you paid for.
What does the '20% of founders still CEO 3 years post-IPO' statistic reveal?
Short answer
Standard VC-backed governance virtually guarantees founder removal — most founders are told they're the exception
More detail
Harvard Law School data: among venture-backed companies with standard governance, only 20% of founders are still CEO three years after going public. Ries describes sitting with a founder pre-IPO who consulted bankers, lawyers, CFO, GC, VCs — all said 'you're the exception.' That founder was ousted five months after IPO when a competitor acquisition spooked the market. The people who told him he was special — the bankers, lawyers, growth VCs — all profited from transaction volume on the way up and the way down. They were on the next deal within weeks. Practical tip: don't evaluate governance advice from people who profit from the transaction regardless of outcome. Their incentives are not yours.
What does the Novo Nordisk story teach about mission-protective governance?
Short answer
A 100-year-old two-tier structure — for-profit owned by a nonprofit foundation — protected scientific integrity AND created $500B+ in shareholder value
More detail
Marie Krogh was diagnosed with fatal diabetes in 1920 Denmark. She and her Nobel laureate husband discovered insulin research in Canada and founded what became Novo Nordisk. Knowing the temptation to exploit patients (decades before Shkreli proved it was real), they incorporated with a nonprofit foundation owning and governing the for-profit. The foundation once had to intervene to block a sell-out — that intervention created over $500B in value. ZEISS did the same structure in 1885. Companies with this structure are 6x more likely to reach year 50 vs conventional peers. Practical tip: you don't need to boot the foundation now. Write the right to create it into your charter today.
What is 'financial gravity' and why does the name matter?
Short answer
The force that drags successful organisations toward mediocrity — no one controls it but everyone obeys it
More detail
Eric Ries uses the bridge analogy deliberately: saying 'greed destroys companies' is like saying 'gravity collapsed the bridge.' True. But useless. The useful question is: why did THIS bridge collapse when others didn't? That reframe forces you to look for structural causes — corroded bolts, bad load design — rather than shrugging at human nature. Practical tip: whenever you hear 'that's just how it is in business,' ask what structural choice made it inevitable. There's usually one.
What is 'mission drive' and how does it differ from having a mission statement?
Short answer
Structural proof that the org cannot profit except by achieving the mission — not just words on a wall
More detail
Ries calls most 'mission-driven' companies merely 'mission hopeful' — candy coating on an extractive engine. Mission drive means auditing every role, OKR, and bonus target: is there anyone who could profit by betraying the stated purpose? If yes, you don't have mission drive. Practical tip: run this audit with your team. Ask — can anyone on this team make money by cutting quality, safety, or performance? If the answer is yes and there's no accountability mechanism, fix it before the pressure comes.
What is 'shareholder primacy' and when did it become the default?
Short answer
The legal theory that corporations exist solely to maximise shareholder returns — only dominant since the 1980s, not a natural law
More detail
Before the 1980s, corporations were required to declare a beneficial public purpose to incorporate. Changing a company's purpose away from its charter was a crime — you could earn the 'corporate death penalty.' Shareholder primacy is roughly 40 years old. Adam Smith, Peter Drucker, and every generation before treated it as obvious that corporations served a broader purpose. Practical tip: the next time an advisor tells you something is 'best practice,' ask how old that practice is. Many governance defaults are younger than the trees in your local park.
What is 'the invisible leader' (Mary Parker Follett)?
Short answer
The shared sense of common purpose that guides decisions when no manager is present
More detail
Follett wrote this in 1920 and was erased from management history for most of the 20th century. Her insight: the most consequential decisions in any organisation are made when no manager is in the room — the engineer choosing rounded vs straight corners, the PM deciding whether to double-check a destructive action. If common purpose isn't instilled, you have no control over those moments. Peter Drucker called her the 'Prophet of management.' Practical tip: ask yourself — if I were not here for a week, would my team make the same decisions I would? If not, the invisible leader hasn't been built yet.
What is Evans' 'AI is as big as the internet' framing — and why is 'only' the key word?
Short answer
AI is a fundamental platform shift, comparable to the internet or mobile — not categorically larger
More detail
The framing cuts two ways. For tech insiders who think it's the industrial revolution, it's a reminder that the internet was already enormous. For sceptics who dismiss it, it's a reminder smartphones were world-changing. The productive question isn't whether it's 20% bigger than the internet — it's recognising we're at a 1997 moment where most applications haven't been built yet.
What is the 'Don't Be Evil vs the Quarterly Report' contrast?
Short answer
Google's quarterly report is filed on time with 100% certainty because of a massive compliance apparatus — 'don't be evil' had no apparatus
More detail
Ries ran this thought experiment with a long-term Googler: probability Google files quarterly report on time? 100%. Probability Google might accidentally harm someone and cover it up? The Googler could immediately think of examples — self-driving cars, sued twice for breaking the 'don't be evil' pledge. The point: whatever you claim to be serious about needs an apparatus as reliable as your financial reporting. A slogan is not an apparatus. Practical tip: pick one non-financial commitment your company makes (safety, quality, privacy). Now describe the apparatus that ensures it happens every time. If you can't, the commitment is aspirational, not structural.
What is the 'culture bank' concept?
Short answer
Every principled action that involves sacrifice is a deposit; every self-interested shortcut is a withdrawal — only ever make deposits
More detail
Ries credits Todd Park (Devoted Health) and Howard Schultz (Starbucks) with articulating this. The H-E-B story: during an ice storm, a store manager let all customers take groceries home free because the POS system was down. That wasn't heroism — it was trained behaviour. The rule: never intentionally make a withdrawal. You'll make accidental ones (mistakes happen), but the deliberate ones are what erode trust permanently. Practical tip: after every major decision, ask — was that a deposit or a withdrawal? Make it a team habit.
What is the 'harder is easier' principle?
Short answer
Committing to quality, ethics, or integrity upfront creates unexpected competitive rewards — trust becomes a compounding asset
More detail
Ries argues most leaders treat principles as a cost. Harder is easier flips this: trustworthiness reduces CAC, increases loyalty, improves employee alignment, and lets customers forgive mistakes. Cloudflare gave away SSL encryption at cost to their conversion rate — and ended up a $70B company with a dominant top-of-funnel. The caveat: you can't do it FOR the reward or it doesn't work. The commitment has to be real or the market sees through it. Practical tip: identify one principle your company holds where you've never calculated the ROI — that's likely your most valuable asset.
What is the 'it is always too early until it is too late' pattern in governance?
Short answer
Every stage of a company's life has a credible reason to defer mission-protective provisions — and by IPO it's too late
More detail
Ries describes the exact sequence: incorporation (get PMF first), seed round (investors say later), Series A (same), growth round (don't stand out), IPO prep (land the plane first), day of filing (too late, sorry). He's been in the room when founders discover their S-1 contains none of the protections they discussed for years. Contrast: Anthropic wrote the right to create the Long-Term Benefit Trust into their charter from day one, even though they didn't activate it until Series C. Practical tip: treat mission-protective provisions like a co-founder agreement — do it at incorporation when it costs nothing and everyone agrees.
What is the 'jagged frontier' and why does it matter for assessing AI's impact on your work?
Short answer
AI capability is uneven — brilliant in some tasks and surprisingly bad in adjacent ones, and the pattern is hard to predict
More detail
You can't reliably look at a profession and say '17% is automatable.' The boundaries of where AI works aren't intuitive, and neither are the failure modes. The only way to find your jagged edge is to use it. e.g. Evans found AI great for images and redecorating but poor for precise information retrieval — exactly the thing he'd want it to do most.
What is the 'mission guardian' concept and what are the main structural options?
Short answer
Some person or entity must have legal authority to protect the mission — without it, good intentions are worthless
More detail
The mission guardian options: (1) founder control — works early but makes founders Atlas-like, holding back gravity alone; (2) nonprofit foundation — Novo Nordisk, ZEISS model, foundation owns equity and has override power; (3) Perpetual Purpose Trust — Anthropic/Patagonia model, non-economic entity with mission oversight only, includes a 'purpose protector' who can sue the trustees; (4) employee ownership — John Lewis, Mondragon, Alibaba voting trust models. Ries argues founder control is a good temporary bridge but a poor permanent structure. Practical tip: think of the mission guardian as a constitutional court — it doesn't run the business, it just ensures the business can't be turned against its own founding purpose.
What is the 'organisations as emergent intelligence' concept?
Short answer
Corporations are the oldest form of artificial intelligence — same emergent principles as transformer models, with values that flow from founders to structure to product
More detail
Conway's Law: the org chart is visible in the architecture diagram. Human values flow from parent to child in software systems. Ries connects this to the ant colony experiment — a thousand ants can solve the Piano Movers puzzle better than any individual ant, and the colony appears to exhibit intelligence. More humans make a problem worse unless they're carefully aligned. The implication for AI: if you can't align the humans building the AI, you can't align the AI itself. Who aligns the aligners? That's the unsolved problem. Practical tip: the governance structure of your company is not separate from the product — it is the product's DNA.
What is the 'reach test' for evaluating whether a new AI tool has real utility?
Short answer
Do you reach for it organically when you wake up in the morning?
More detail
A tool with genuine utility gets pulled naturally into your workflow — you don't have to remember to use it. Contrast this with tools you force yourself to try. Shipper used this test when Claude Code first landed: the organic reach signal told him something real was happening before he could articulate why.
What is the 'spiritual holding company' concept?
Short answer
An umbrella term for any structure — nonprofit foundation, purpose trust, employee ownership — that makes mission the sovereign above shareholders and founders
More detail
Ries coined this because advocates for each specific structure (Novo Nordisk-style foundation, Perpetual Purpose Trust, employee co-op, Anthropic's Long-Term Benefit Trust) fight each other. The common thread: some entity or structure is the mission guardian, with legal authority to resist both outside pressure AND internal temptation. The holding company owns the spirit of the organisation, not its equity. Practical tip: the simplest implementation is to write the right to create one into your charter at incorporation — even if you don't activate it until Series C, as Anthropic did.
What is the 'the more golden the goose, the greater the temptation to butcher' principle?
Short answer
Success doesn't protect you from financial gravity — it makes you a bigger target
More detail
Ries uses the restaurant example: a friend could taste that a restaurant had been taken over by private equity from the food alone. Vital Farms eggs allegedly getting worse after BlackRock ownership. A natural foods brand named after its founder gets ousted by investors chasing margin — Ries could predict the entire story before being told it. The pattern is so pervasive we don't have a common name for it. The trap: founders think 'once I'm successful, I'll have the leverage to protect my vision.' The reverse is true — success is what makes outside pressure intensify. Practical tip: implement governance protections before you think you need them, not after you become valuable enough to attract predatory pressure.
What is the 'torchbearer' concept and why do they suffer in conventional companies?
Short answer
The rare person in any organisation who is simply committed to doing the right thing regardless of pressure
More detail
In a conventional company, torchbearers are exhausted. Every day someone walks in with a spreadsheet demanding ROI justification for doing the right thing. In a mission-aligned company, two antibodies protect them: (1) no one makes the spreadsheet because it's not needed — everyone already knows; (2) the culture bank is full, so the default is always toward the mission. Clayton Christensen: it's easier to do the right thing 100% of the time than 98% of the time — because at 98% you need a meeting for every decision. Practical tip: identify your torchbearers. If they're exhausted, that's a signal your culture bank is overdrawn. If they're energised, you're building something durable.
What is the Anthropic governance structure and why did it matter from day one?
Short answer
PBC + Long-Term Benefit Trust with AI-safety trustees who hold no equity — written into the charter at inception, activated at Series C
More detail
When Dario Amodei came to Ries pre-Anthropic (pre-ChatGPT, no hot company, no top VCs), Ries walked him through the standard governance horror story. Anthropic wrote the right to create the LTBT into every legal document from inception — two years before they activated it. The LTBT appoints directors to the for-profit board who are accountable to outside AI safety trustees who hold zero equity. This is why Anthropic can turn down $200M Pentagon contracts, refuse to release dangerous models, and retain top talent who want to work for the good guys. Practical tip: the LTBT costs nothing to reserve. You just write 'we reserve the right to create this structure' in your charter before your first VC conversation.
What is the Cloudflare SSL encryption story and what does it demonstrate about the 'figure it out' principle?
Short answer
Matthew Prince gave away SSL encryption for free despite it being Cloudflare's top conversion driver — top-of-funnel grew 10x
More detail
A junior engineer asked: 'Isn't our mission to make a better internet? Wouldn't a better internet be an encrypted internet?' Prince said 'let's figure it out' instead of defending the revenue line. The team hand-rolled assembly-language code and negotiated bespoke deals with certificate authorities to drive costs to zero. Conversion rate dropped. They held firm. Top-of-funnel grew an order of magnitude. Cloudflare is now worth $70B. The 'figure it out' principle: when a principle creates an impossible-seeming dilemma, treat that as a design challenge, not a reason to abandon the principle. Practical tip: the next time a principle is expensive, ask 'what would it take to make this sustainable?' before asking 'should we drop it?'
What is the Groupon email frequency story and what principle does it illustrate?
Short answer
Andrew Mason held the line on one email per day — then let 'experiments' grind him down to 8 emails — destroying the core product
More detail
Groupon's entire business was built on one daily email. When executives pushed for two emails using A/B test framing ('shouldn't we run an experiment?'), Mason eventually agreed. Two emails made more money. Then three. Then eight. The data justified every step. The principle: ROI-based thinking has no floor. If you can't defend a principle beyond 'the data said so,' you will lose it. Mason had no structural defence against the spreadsheet. Practical tip: identify the one or two non-negotiables in your product that you would never A/B test away. Write them down. Make them known. Then don't run the experiment.
What is the Jevons paradox in the context of AI and work?
Short answer
Making a task cheaper tends to increase total consumption of that task — people do more of it, not the same for less
More detail
Price elasticity applies: if the cost of analysis drops by 90%, you don't do 10% of the analysis you used to. You do 10x the analysis and ask questions you couldn't afford to ask before. This is why software developers multiplied as IDEs and libraries reduced the code they had to write. The bottleneck shifts, it doesn't disappear. Practical signal: when something becomes cheap, watch for explosion in demand, not contraction.
What is the Public Benefit Corporation (PBC) filing and why should every founder do it?
Short answer
A two-page Delaware filing that replaces 'any lawful act or activity' with a specific stated purpose — the minimum viable mission protection
More detail
Under shareholder primacy law, 'any lawful act or activity' legally means 'maximise shareholder returns.' The PBC changes this: investors agree upfront that the purpose is something specific. It doesn't guarantee good behaviour, but it closes the fiduciary loophole that forced the Vectura board to sell to Philip Morris for 10p more per share. Practical tip: file this at incorporation. If you've already raised, your lawyer can still do it — but you'll need investor sign-off, which gets harder at every round. All major AI labs are PBCs. Cost: essentially zero.
What is the Vectura story and what does it illustrate?
Short answer
A UK inhaler therapeutics company legally forced to sell to Philip Morris for 10p more per share — then written down by $900M within 3 years
More detail
Vectura's board had three choices: Philip Morris at 165p, private equity at 155p, or stay independent (the company was fine). The British Thoracic Society begged them to decline. Every thoughtful observer said no. The board said their hands were tied — fiduciary duty to accept the highest bid. Philip Morris spent £1.1B, wrote down $900M, and disposed of the company. It no longer exists. The lesson: good intentions signed away at incorporation become irrelevant. The charter is the company. Practical tip: pull your company's Delaware charter and read it. It's public record. If it says 'any lawful act or activity,' that's a Vectura charter.
What is the core paradox of AI automation at work?
Short answer
More AI = more human work, not less
More detail
Every agent needs a human to care for it. As you add automation, you need people to verify it's working correctly, correct its errors, and push it toward new capability frontiers. Shipper doubled headcount at Every in one year while being maximally AI-forward. Automation creates managerial overhead even as it removes execution overhead.
What is the lump of labor fallacy and why does it keep misleading AI job predictions?
Short answer
The assumption that there's a fixed amount of work — so automating a task means fewer jobs — ignores that cheaper tasks unlock more demand and new work categories
More detail
Every previous automation wave (adding machines, spreadsheets, IDEs) followed Jevons paradox: making something cheaper increased total consumption, not reduced it. The number of accountants rose through every wave of accounting automation. New jobs appear that didn't exist and couldn't have been predicted. e.g. 'Railway engineer — what's a railway? Why would that be a thing?'
Why does writing about AI trends help you predict them accurately?
Short answer
Articulating the future makes it real — it sharpens your noticing and accelerates adoption
More detail
Shipper credits writing as a core part of his prediction process. The act of putting an observation into words forces precision, surfaces gaps, and gets the idea into the world where others can pressure-test it. Prediction without articulation stays fuzzy. Writing is how you convert a vibe into a model.
Level 2
10 cards
How did platform shifts change TAM — and what does that imply for AI company scale?
Short answer
Each platform shift expanded the addressable base by an order of magnitude, enabling larger and larger companies
More detail
Peak mainframe install base was ~70-80k units. PCs at internet launch: 50-100 million. Smartphones today: ~5.5-6 billion. ChatGPT hit 900 million weekly users partly because the distribution infrastructure already existed. The implication: AI can address ever-larger fractions of the economy, enabling companies to reach scales previously impossible.
What are the two main modes of AI-assisted work Shipper predicts will dominate?
Short answer
(1) Async super-agent in Slack; (2) on-computer agent environment like Codex/Cowork
More detail
These are complementary surfaces. The Slack agent handles delegated async tasks — research, data queries, comms. The on-computer agent (Codex, Cowork) is where you do most knowledge work directly, with the AI working alongside you in the same environment, able to see what you're doing in real time.
What is Evans' 'we're in 1997' framework for AI adoption?
Short answer
The technology is clearly transformative, but most applications haven't been built, most people aren't using it daily, and the winners aren't obvious yet
More detail
In 1997 the internet was real, exciting, and obviously big — but Amazon, Google, and Facebook didn't exist yet. You couldn't have predicted them. Asking 'is it Excite or Yahoo?' was the wrong question. Today: asking 'is it OpenAI or Anthropic?' is probably the same kind of wrong question. Most of the value will be built by companies and people we haven't heard of yet, solving problems we haven't framed yet.
What is Evans' thesis on why foundation model companies may not have pricing power?
Short answer
If models are undifferentiated, multiple competitors exist, and applications must be built by others, models become commodity infrastructure — like telecoms, not like Windows
More detail
Global mobile data consumption grew 1,500-2,000x since 2010. Mobile stocks went nowhere in 25 years. The telcos built extraordinary infrastructure but all the value accrued to the app layer. If foundation models become commodity infra with no winner-takes-all effect, margins compress. The question is whether they look more like AWS (commodity) or Windows (platform with lock-in).
What is the 'allocation economy' model of human-AI work?
Short answer
Humans work like managers — constantly checking, directing, and improving AI output rather than delegating and walking away
More detail
Most managers are not on the beach. They spend real time with their reports. Managing AI is similar: you need to understand what it's doing, catch errors, push it toward better outcomes. Shipper uses this to explain why automation increases workload rather than reducing it. The mental model reframes 'oversight' not as a tax but as the core job.
What is the 'distribution as moat' thesis in an AI-commoditised world?
Short answer
When the underlying model is a commodity, distribution and brand determine who wins — incumbents with large user bases have structural advantage
More detail
If there's no meaningful difference between Gemini and Claude for a normal user, Google wins by spraying Gemini across every Android surface. Meta was competitive purely from surface coverage. The implication for builders: in a commodity product field, 'who has the customers already' matters as much as 'who has the best model.'
What is the 'senior engineer benchmark' and what does it measure?
Short answer
How well AI can rewrite a vibe-coded codebase from first principles, versus how a human senior engineer would
More detail
Shipper vibe-coded Proof, it broke at launch. He had two senior engineers rewrite it independently. That output became his benchmark. He gives new models the same prompt and scores them. Most models scored ~30/100 until GPT 5.5 hit 62. The benchmark reveals a specific failure mode: models patch around problems rather than rip and rewrite like a senior engineer would.
What is the 'super agent vs personal agent' model and which does Shipper think wins now?
Short answer
Super agent per company wins now; personal agents come later when models require less maintenance
More detail
Personal agents (like OpenClaw) are too high-maintenance for most people. They break, require SSH access, and fall apart when the human stops caring for them. A single company-wide agent with a dedicated forward-deployed engineer maintaining it is more reliable. Shopify (River) and Ramp are examples. Personal agents return when models become more autonomous.
What is the 'task vs job' split and how did it destroy newspapers and record companies?
Short answer
Industries that conflated their distribution task with their actual value were destroyed when the task was automated — the value was always elsewhere
More detail
Record companies thought they were in music. They were in manufacturing and distributing small pieces of plastic. When that task disappeared, so did they. Newspapers thought they were in journalism. They were in printing and delivery. The content survived; the business model didn't. Evans uses this to ask: for any AI-exposed role, is the task the job, or is there a 'real job' underneath the task?
What is the SaaS implication of agents using apps inside their own browser?
Short answer
Users bring their own tokens — SaaS companies don't need to pay for AI inference, restoring margins
More detail
If work happens inside Codex or Cowork, and those tools have in-app browsers, the user's agent is accessing SaaS tools using its own model calls. The SaaS vendor doesn't pay for tokens. Shipper built Proof this way — anyone using it brings their AI. This flips the current narrative: instead of SaaS needing to embed AI, agents embed SaaS.
Level 3
10 cards
How did Every run quarterly planning with AI agents?
Short answer
One Notion agent interviewed each team member, pushed back on goals, and produced strategy reports that Shipper reviewed for coherence
More detail
Rather than everyone writing their own plan, an agent asked structured questions (what happened last year, what are your metrics, how does this relate to company strategy), then generated reports. Shipper's job shifted to: who needs to talk to each other, which teams are misaligned, which reports are high vs low quality. AI generated the drafts; he synthesised the whole.
How did Shipper use Codex to hire a head of L&D?
Short answer
Typed a natural language description of the ideal candidate into Codex and let it research while he did other work
More detail
He described the profile: worked at General Assembly, now AI-pilled, follows me on Twitter. Codex found the person, Shipper DM'd them, had dinner, and moved forward. The practical pattern: when you have a specific, researchable search problem, offload it to Codex with enough context and let it run in parallel with your other work.
How should you run a 'rock-turning' process when new AI models drop?
Short answer
Re-test things the model couldn't do before — capability jumps are non-linear and easy to miss
More detail
Don't assume last month's limitation is still true. Shipper re-ran his senior engineer benchmark on GPT 5.5 expecting ~30/100 and got 62. The way to ride models is to systematically re-probe your known failure modes. Practical: keep a list of tasks AI failed at in the past 6 months. Test the new model on each one the week it drops.
How should you think about enterprise adoption timelines when assessing AI's job impact?
Short answer
Enterprise sales cycles run 18 months or more — wholesale workforce transformation will take 3-10 years sector by sector, not 2 weeks
More detail
The doomer scenario (company buys ChatGPT Monday, fires everyone Friday) ignores how enterprise technology actually spreads. Nobody tears out SAP in a quarter. It takes time to discover use cases, time for procurement and legal, time for integration, time for training. The internet took 10-15 years to fully reshape industries it eventually dominated. Apply that humility to AI job forecasts.
How should you use the 'Spotify test' when evaluating AI's impact on your industry?
Short answer
Don't just ask how AI does the old thing cheaper — ask what fundamentally new things become possible that weren't before
More detail
Spotify is not an online music store. It's a completely different product that the old model couldn't produce. The first phase of a platform shift is 'do the old thing but more.' The second phase is 'do something new that only this enables.' Practical use: for any role or industry, push past 'how does AI do what I already do' to 'what can now exist that couldn't before?'
What does a 'forward deployed engineer' do in an AI-native company?
Short answer
Maintains, monitors, and improves internal AI agents so the rest of the organisation can use them without breaking things
More detail
This role is not about writing new features — it's about ensuring the agent is working, correcting errors, adding context, and building systems so less technical people can use the agent safely. At Every, Nitesh spends most of his time in Slack talking to their internal agent Claudy. This role is emerging at most AI-forward companies.
What is the 'Larry Tesler / AI redefinition' pattern and why does it matter for tracking progress?
Short answer
Each time AI solves a problem, people redefine AI to exclude it — 'that's just image recognition' — making benchmark progress seem slower than it is
More detail
This creates a moving target where AGI perpetually recedes. Image recognition, sentiment analysis, machine translation — all were 'AI' until they worked. Now they're 'just software.' The practical implication: don't use 'is this AI yet?' as the question. Use 'is this useful?' The rebranding from AI to automation is a sign of success, not failure.
What is the 'Uber vs Airbnb' test for predicting AI disruption depth?
Short answer
Not all platform shifts hit all incumbents equally — the question is whether the new thing actually substitutes for the old thing or carves out an adjacent use case
More detail
Uber demolished taxi businesses in many cities and grew the market. Airbnb barely dented hotels — half the hotel market is business travel, and a consultant flying into Milwaukee at 8pm isn't staying in an Airbnb. Before predicting AI impact on your industry, ask: does the new thing actually do what my customers come to me for, or does it serve a different need?
What is the 'presume radical uncertainty' heuristic for AI forecasting?
Short answer
In 1997, you would have missed almost all the winners of the internet era — apply that same humility to AI predictions, including your own
More detail
In 2000, you would not have predicted that a has-been PC company from Cupertino would win mobile. Or that a search company would dominate maps and email. Evans' rule: you can identify that something big is happening and roughly what forces are at play, but specific winners and second-order effects are mostly unknown. Calibrate confidence accordingly.
What software affordances does human+agent co-working require that current SaaS tools don't have?
Short answer
Approval inboxes, change logs, rollback, and agent-friendly UI alongside human-friendly UI
More detail
Agents can make a billion requests in three seconds. SaaS built for humans can't handle this. New requirements: an inbox showing what the agent did or is about to do, rollback capability, and a UI that works for both a human browsing and an agent scripting simultaneously. GitHub is already struggling with this as agent-generated PRs flood the platform.
Level 4
9 cards
How do you diagnose whether your company's AI agent is actually working?
Short answer
Ask whether the human responsible for it is actively engaged — if nobody's gardening it, it's probably broken
More detail
Shipper's signal: when people at Every stopped maintaining their personal OpenClaw instances, those agents quietly became useless. A working agent has a human who checks it regularly, corrects errors, and iterates on context. If your company's Slack bot is going weeks without someone adjusting it, assume it's producing low-quality outputs and nobody's catching them.
How does AI change the economics of on-boarding new enterprise software?
Short answer
Instead of a 10-step onboarding UI, agents can self-configure by reading your context and querying the app directly
More detail
Shipper's pattern with hosted OpenClaw: instead of building an onboarding wizard, assume users come from Codex or Cowork. They paste a prompt, the agent reads their existing context (projects, preferences, history), and configures the product automatically. The SaaS vendor gets a richer, more accurate setup than a human filling in a form, with no UI cost.
How does the 'personal trainer via iPhone' example illustrate the unpredictability of AI exposure?
Short answer
Jobs that look protected because they're physical or relational can be disrupted by AI finding a new route around the assumed constraint
More detail
Balance your iPhone on a stand, get an AI to watch your form and build your training plan. This might be wrong. But the same logic that made people say 'taxi drivers are safe' from the internet in 1997 said 'personal trainers are safe' from AI. The point is not that personal trainers will definitely be disrupted — it's that confident claims of 'this profession is safe' require much more scrutiny.
If you're a professional services associate worried about AI, what is the one thing Evans says will definitely not help?
Short answer
Refusing to engage with AI, performing moral superiority online, and signalling you'll never use it
More detail
Evans is blunt: this gives you a great feeling of righteousness and accomplishes nothing. If a firm is deciding between hiring 100 associates or 50, walking into the interview expressing hostility to AI is just a filter. The only option is to dive in, submerge yourself, understand what it can and can't do, and be able to demonstrate that understanding in how you work.
If you're building SaaS, what should you prepare for as agents become primary users?
Short answer
Make your product work for both humans and agents simultaneously — the interaction model, request volume, and error reporting all change
More detail
Agent bug reports are far richer than human ones: exact repro steps, hypothesis about the cause, and sometimes a suggested fix. Agent request rates can spike unpredictably. Your pricing and infrastructure need to handle this. Practical first step: make your app usable from a CLI or headless context so agents can drive it without a UI.
What does the history of accounting employment tell us about AI and professional job volumes?
Short answer
Accountants kept growing in headcount through every wave of accounting automation — adding machines, mainframes, spreadsheets, ERP, cloud
More detail
The number of accountants in America rose through the entire 20th century and has risen again since 2000. This is despite punch cards, mainframes, databases, ERP systems, Excel, and cloud accounting. The explanation is Jevons paradox: each wave made accounting cheaper, so more accounting got done, so more accountants were needed to do the higher-level judgment work. This is the baseline expectation, not the exception.
What happens when you use AI to apply structured analysis to a profession and claim X% of tasks are automatable?
Short answer
You commit the expert systems fallacy — you can't decompose a profession into logical steps and score each one
More detail
The expert systems problem: to recognise a cat in a picture, you build an edge detector, an eye detector, an ear detector. 15 years later you have 700 steps and it still doesn't work. Applying the same logic to a senior law partner — '17% of their work could be automated' — is horseshit. You can't describe the texture of a senior professional role that way.
What is the 'barcode and supermarket SKU count' analogy for understanding AI's second-order effects?
Short answer
Barcodes didn't just speed up checkout — they made it possible to stock far more products, transforming retail economics in ways nobody predicted upfront
More detail
Evans' presentation slide: supermarket SKU counts rose dramatically after barcodes, because stores could now track inventory of thousands of items instead of hundreds. The second-order effect (more variety, new product categories, supply chain transformation) was bigger than the first-order effect (faster scanning). Ask the same question about AI: what becomes possible at scale that wasn't feasible before?
What makes a PM dangerous in an AI-native environment?
Short answer
Deep product sense + user empathy + just enough technical knowledge to pair with coding agents
More detail
Marcus at Every (former Axios PM) ships faster than most engineers. He knows what a database migration is, can read code when he needs to, and pairs that with spiky product intuition. He doesn't need to organise a team — he just builds. The key ratio is not technical depth but product clarity: knowing exactly what to build is now the bottleneck, not the building.
Level 5
11 cards
What does the telco margin history warn us about assuming AI infrastructure companies will capture AI's value?
Short answer
The people who built the infrastructure that enabled everything captured almost none of the value — all of it went to the app layer
More detail
Mobile data grew 1,500-2,000x since 2010. Telco stocks went nowhere in 25 years. The companies with Nobel-Prize-level engineering built extraordinary infrastructure — and earned utility margins. The value was captured by Apple, Google, and Meta, who built on top. If foundation models become commodity infrastructure, the same pattern applies: value accrues to whoever owns the customer relationship and the use case, not the model.
What happens to data science teams when everyone can do their own analysis with AI?
Short answer
The team gets flooded with bad analysis to review rather than doing less work — unless they build an agent to handle the basic questions
More detail
A data science bot hooked to the data warehouse filters out routine queries, freeing the team for harder problems. Without this, data scientists spend their time debunking incorrect AI-generated analysis from non-experts. The failure mode: AI raises everyone's ability to produce analysis without raising everyone's ability to produce correct analysis. You need a system, not just access.
What is the 'AI is good at what computers are bad at' failure mode?
Short answer
AI is often weakest at tasks that look like classic information processing — precision, recall, structured retrieval — and strongest at open-ended synthesis, where its errors are hardest to catch
More detail
AI can make stunning images and synthesise complex ideas, but is unreliable for exact lookups and citations. This means it's weakest precisely where confident errors are most damaging — legal, medical, financial tasks requiring precision. The practical failure: people deploy AI on high-stakes precision tasks because the output looks fluent and authoritative, then discover it hallucinated.
What is the 'agents need people who care about them' failure mode?
Short answer
If nobody is actively maintaining an agent, it quietly degrades — no alarm goes off
More detail
OpenClaw is the clearest example: high initial excitement, then everyone quietly stops maintaining their instance, then the agent becomes useless without anyone noticing. Contrast with a super agent that has a dedicated forward-deployed engineer: there's accountability, regular iteration, and a human who would notice if quality dropped. The failure is silent and hard to detect.
What is the 'incumbents make the new thing a feature' trap, and when does it work or fail?
Short answer
Incumbents often correctly identify that a new technology is just a feature of their existing product — and sometimes they're right, which makes the failure cases harder to predict
More detail
Steven Sinofsky's observation: incumbents always try to make the new thing a feature. Sometimes they're right — not every new thing becomes a platform. The failure mode is applying this logic to something that's actually a platform shift. Microsoft tried to make the internet a feature of Windows. Yahoo tried to make mobile a feature of the web portal. The reasoning sounds plausible right up until it doesn't.
What is the 'speed ran the CLI era' prediction and why does it matter for product decisions?
Short answer
The CLI phase was a stepping stone to GUIs with embedded agents — products optimised only for CLI are already dated
More detail
Shipper's argument: Claude Code's power wasn't the terminal, it was the agent. GUIs were invented for a reason. Once you move back to a GUI with an agent running inside it, you get all the benefits plus visual clarity. The mistake was assuming CLI = power user = right direction. If you're building tooling primarily for CLI use, reconsider.
What is the benchmark saturation trap and why does it mislead people about AI replacing jobs?
Short answer
Benchmarks only measure problems that have been framed and scored — the act of framing the problem is itself human work that benchmarks don't capture
More detail
Even if AI hits 100% on a benchmark, someone had to design the benchmark. In practice, the highest-value human work is knowing what prompt to give, what frame to use, what to measure. Models are always trailing the humans who are pushing to new frames. This is why Shipper's own benchmark — once saturated — can be trivially re-zeroed by changing the prompt.
What is the post office scandal pattern and why does it apply to AI deployments?
Short answer
Buggy systems deployed with authority can destroy people's lives before anyone admits the system is wrong — institutional credibility can override observable evidence
More detail
UK post office rolled out a Fujitsu system with bugs that showed false cash shortfalls. Hundreds of postmasters were prosecuted, some imprisoned, some bankrupt. Post office staff swore in court the system had no bugs. The pattern: a technical system is treated as authoritative, human testimony is dismissed, damage accumulates before anyone corrects it. AI systems with high institutional trust and poor error visibility create the same risk.
What is the risk of being a CEO or senior leader who doesn't personally use AI tools?
Short answer
You lose intuition for what's possible — you can't direct AI strategy without hands-on experience
More detail
Shipper's observation: CEOs can currently 'get away with' their day looking similar to pre-AI. But this creates a lagging intuition problem — they're directing teams toward an AI future they've never personally navigated. The failure mode is being dependent on others' filtered summaries of what AI can do, which introduces distortion at every level of the org. You can't delegate the learning.
When does the 'vibe coding in production' approach break, and what's the tell?
Short answer
When the codebase is slop, models will patch around problems rather than diagnose root causes
More detail
Shipper launched Proof and servers crashed every 10 minutes. Codex kept saying it fixed things — then caused four new errors. The tell: the model is confidently completing tasks while the system gets worse. Root cause: vibe-coded slop has no coherent architecture for the model to reason about. Fix: a senior engineer (human or eventually model) needs to rewrite from first principles, not patch.
Why does the 'we have no theory of intelligence' point limit our ability to predict AGI timelines?
Short answer
Without a theory of why models work or how human intelligence operates, all AGI forecasts are vibes — confident predictions in either direction are unjustified
More detail
We have no theory of consciousness, no theory of what human general intelligence is, no theory of why transformers work as well as they do, and no theory of what their ceiling is. Both 'AGI in 2 years' and 'AGI is impossible' rest on no theoretical foundation. Practically: assume uncertainty, watch empirical behaviour of systems rather than extrapolating from trend lines.
Tags
Want to build your own study deck?
Import from CSV, add cards manually, and let spaced repetition do the rest.
Create a free account →