Fences, flagpoles, and the comeback of the generalist

The three questions that tell you if AI will be disruptive

Is AI actually a big deal, or just another hype cycle? Eric and John apply a three-matrix framework to cut through the noise and find a clear answer. SUMMARY John opens with a hot take that’s on everyone’s mind: is AI as big a deal as everyone says it is? Instead of swapping opinions, Eric proposes a framework: three 2x2 matrices used to evaluate any technology's real-world impact, then walks through historical examples before applying all three to AI. Matrix one is breadth versus depth: does a technology affect one area deeply, many areas broadly, or both? Matrix two is rate of improvement versus rate of adoption: how fast does the technology get better, and how quickly can people actually access those improvements? Matrix three is novelty versus precedent: is the technology truly new, and does it feel familiar enough to adopt quickly? GPS scored high on depth first, then breadth later. The iPhone scored high on precedent and breadth but was barely novel. Most technologies land high on one or two axes but rarely all three. AI, Eric argues, is high on all three simultaneously and in the first years of its existence, which is historically unusual. The conversation ends with personal examples: a presentation Eric built in two hours that would have taken weeks before, and a best man speech John polished with voice AI coaching he never would have sought otherwise. Their conclusion is quiet but firm: AI will produce an unleashing of human creativity unlike anything we have seen before. KEY TAKEAWAYS Breadth plus depth is the bar for technologies that change everything: a technology that only affects one industry or user deeply rarely reshapes society. The ones that go broad and deep, across industries and users, tend to be the transformative ones. Rate of adoption can lag rate of improvement by decades: fiber internet is the clearest example. The technology is unambiguously superior, but capital cost means most people still don't have it. AI is nearly the opposite: improvements are immediately available to anyone. Novelty alone is not enough, and neither is precedent alone: GPS was truly novel and took decades to reach consumers. The iPhone was barely novel but was adopted almost instantly because it wrapped familiar behaviors in a better form. AI is rare in being genuinely high on both axes at once. The thing that looks like a better search engine is actually something else entirely: many people are using AI as a smarter Google. That framing is not wrong, but it undersells what the technology is capable of by a wide margin. AI's novelty goes all the way down to hardware: Andrej Karpathy's observation that GPUs and TPUs are replacing CPUs as the baseline compute layer illustrates that this is not just a software shift. The infrastructure of computing itself is being redesigned around it. The most underrated use of AI is learning: amplifying skills you already have gets most of the attention, but using AI to rapidly acquire skills you don't have is arguably more powerful and less discussed. AI enables things people simply would not have done before: John's use of voice AI to rehearse and refine a best man speech is not productivity. It's a category of effort that just didn't happen before the tool existed. NOTABLE MENTIONS AND LINKS GPS is used as the primary historical example for the breadth-versus-depth matrix: it started with extremely deep impact in military and industrial applications, then spread broadly to consumers over decades as consumer devices caught up. ... (Read more at the episode page)

You're probably paying too much for AI

Most businesses are spending on AI without measuring the return. Eric and John break down the three factors that determine whether AI actually earns its cost. SUMMARY Eric and John open with a question John raised over lunch: is AI actually too expensive for some businesses? It sounds simple, but the answer turns on three distinct problems most companies never separate: whether people actually know how to use AI well, whether you can honestly measure the return, and what you are actually paying versus what you think you are paying. They work through each one in order. On the usage side, they argue that buying licenses and hoping for adoption is a recipe for low ROI. Power users are rare, and the gap between someone who uses AI constantly but ineffectively and someone who uses it to think better about hard problems is enormous. On the ROI side, they draw a sharp line between cost savings (which are measurable) and revenue attribution (which is often fuzzy), and point to prospect research and faster creative iteration as two of the clearest paths to a direct revenue connection. The conversation lands on the cost structure itself. Most businesses default to the most powerful and expensive models for every task, without realizing that cheaper models handle routine work just as well and can cost orders of magnitude less. John's story about using a flagship model to rewrite prompts for a cheaper one captures the whole episode's argument: with the right approach, AI is rarely too expensive. Without it, you are paying full price for a fraction of the value. KEY TAKEAWAYS AI without adoption is just a sunk cost: Buying licenses does not create leverage. Most employees will not use AI well without deliberate training and incentives, and the power users tend to already be power users of other software. Using AI to think is the highest-leverage move: The biggest gap is not between people who use AI and people who don't. It is between people who use it to execute tasks and people who use it to think through bigger, harder problems. ROI has two sides, and cost is the easier one: Measuring hours saved and seat count reductions is straightforward. Attributing revenue gains to AI is harder because process improvements and business discipline often deserve as much credit as the tool itself. Start ROI tracking with use cases that have a clear line to revenue: Prospect research, faster creative iteration, and personalized sales demos are examples where the connection between AI effort and business outcome is concrete enough to measure. The default model is almost always the most expensive one: AI providers set flagship models as the default, and most business users never change them. Simpler tasks like reading a PDF or summarizing text work fine on models that cost a fraction of the price. You can use a smarter model to optimize for a cheaper one: If a task fails on a lower-cost model, asking the expensive model to rewrite the instructions for the cheaper one often solves it, and then you run all future instances on the cheaper version. Businesses on prosumer plans are sitting on a narrow window: Individual and small-business tiers are still heavily subsidized by providers preparing for IPO. That subsidy will shrink as these companies move toward profitability. NOTABLE MENTIONS AND LINKS Klarna is the go-to example of high-profile AI cost savings: the company announced its AI assistant had replaced the equivalent of 700 customer service roles, then later reversed course and began rehiring human workers, illustrating how easy it is to overclaim AI ROI. ... (Read more at the episode page)

The honest scorecard for what AI can actually do

Eric and John rate five AI use cases on a scale from 1 to 10: deep research, running an autonomous company, creative work, coding, and voice. The results are not what most people expect. SUMMARY Eric and John open with a question they get constantly: what can AI actually do? It sounds simple, but the honest answer swings wildly depending on who's asking and what they're trying to accomplish. Before scoring anything, they work through how AI actually works, using Google Translate as an accessible entry point into why context is everything. Then John runs five use cases and asks Eric to react with a live score before he weighs in. Deep research scores an 8 from both. Running a fully autonomous company scores a 2. Creative work splits them. Coding lands at a 7. And voice, which almost nobody is using to its potential, scores a 9. The episode closes with an observation that cuts against most AI coverage: the most impressive capability on the list is also the most underutilized, and the use case everyone talks about, the autonomous AI company, is the one that works almost nowhere in the real world. KEY TAKEAWAYS AI's power scales with how specific your context is: the Google Translate analogy shows why; a vague prompt draws on everything, a specific one draws on exactly what you need, and the results are dramatically different. Deep research is genuinely an 8 out of 10, but only if you pay: the capability is there, but it requires a paid tier and an intentional mode most people forget to activate. The autonomous company works for one-dimensional content businesses and almost nowhere else: AI handles research-to-publish pipelines remarkably well, but real businesses are multi-dimensional, and context shifts too fast for full automation. AI raises the floor on creative and software work, not just the ceiling: the average quality of design and code will improve because AI lets skilled people iterate through more options faster, even if the best human work remains out of reach. Voice is the most underrated capability on the list: talking to AI while driving, walking, or thinking out loud is a 9 out of 10 experience that most people still haven't tried, and it is likely to become the dominant way people interact with AI. Your plan tier changes what AI can actually do for you: deep research, voice integrations, and enterprise features are meaningfully better at paid and enterprise levels, which means people on free tiers often form impressions based on a limited version of the tool. NOTABLE MENTIONS AND LINKS Google Translate opens the episode as Eric's preferred analogy for explaining how AI works: predicting the next word from an enormous dataset, which is accessible, accurate, and extends naturally to explain why context makes results better. The MacBook Neo is Eric's hypothetical research example, illustrating how an AI model issues 30 to 40 web searches, visits each page, reads the content, and returns a cited summary instead of making you do it yourself. ChatGPT and Claude are the two tools Eric and John use daily and reference throughout as the primary benchmarks for each use case scored in the episode. Grok gets a specific mention for releasing a new voice model the week of recording, which John calls out as genuinely good even though GPT remains his preference for voice. WhisperFlow is mentioned as a tool that can bridge some of the voice integration gap by cleaning up spoken input and feeding it directly into an AI model as a prompt. The reddit post about an AI-generated Monet which got millions of views and hundreds of comments critiquing what made it inferior to the original, only to turn out to be an actual Monet, becomes the episode's clearest illustration of how close AI image generation has gotten to professional-grade creative work.

Can AI actually replace an employee?

The headlines say AI is replacing workers. Eric and John dig into what's actually working, what isn't, and where the real ceiling is right now. SUMMARY Eric opens with a viral post from David Cramer, founder of Sentry, pushing back on the idea that people are running fleets of AI agents doing real work overnight. John responds from firsthand experience, explaining that his company has run dozens of internal experiments, and the honest answer is that almost none of them are used to do real client work. They map the landscape by use case, from personal productivity tools to team-wide deployments, and find that the team tier is where almost everyone stalls. The tools are developer-focused, the adoption problem is real, and getting AI to work reliably across a group requires far more investment in guardrails and oversight than the demos suggest. The episode ends with guidance on what’s practical today. The most compelling near-term model is not a zero-person company but a "co": a single AI assistant that one person owns, trains over time, and stays responsible for. KEY TAKEAWAYS Impressive demos and production deployments are two different things: most agent experiments stay internal, and the gap between "kind of works" and "works with real clients" is larger than most AI coverage admits. What works at home does not automatically work at work: personal AI tools, team tools, and company-wide deployments each have different friction points, and almost everyone has figured out the personal tier and almost no one has figured out the team tier. AI tools are built by developers, for developers, and it shows: most frameworks default toward building and generating, with not enough support for planning, quality checks, and oversight, which limits what they can reliably do. AI will try to answer even when it shouldn't: agents respond by default even without enough context to be accurate, and building the guardrails to prevent that is harder and more expensive than it looks. Owning a single AI assistant beats managing a fleet: a one-to-one "co" that you prompt carefully, iterate over time, and stay responsible for is more practical and more trustworthy right now than trying to orchestrate autonomous teams of agents. AI helps analysts work faster, but it cannot replace what they know: giving AI access to data and asking it to run queries works well when a skilled human with domain knowledge is in the loop; without that, the answers are unreliable. NOTABLE MENTIONS AND LINKS David Cramer's post on X is the episode's opening provocation, in which the founder of Sentry argues that nobody doing serious work is running 20 agents overnight, and that the real benchmark is whether you can reliably ship one production-quality fix at a time. Block, Inc. is the financial services company behind Square and Cash App, and its high-profile layoff of over 4,000 employees in February 2026 became a recurring example in the AI-is-replacing-workers news cycle that frames the episode. OpenClaw is an open-source personal AI assistant that runs on your own hardware, connects to messaging channels like iMessage and Telegram, and can be given broad access to your computer, including, for those who push it furthest, credit cards and prediction markets. Zo Computer is described as a middle ground between OpenClaw and a consumer app: AI running inside a secure cloud computer with built-in limits, more powerful than a chat interface but without the security exposure of a fully local setup. Poke is a consumer-facing personal agent that works entirely through existing messaging apps like iMessage or Telegram, with no separate interface of its own. ... (Read more at the episode page)

Fences, flagpoles, and the comeback of the generalist

AI is removing the barrier of specialization, giving generalists the ability to span more domains and solve the most important problems faster. SUMMARY Eric and John unpack a shift many knowledge workers can already feel: AI is changing which kinds of people create the most value. Their frame is the “fence-shaped” generalist, someone with broad range and multiple usable areas of depth, rather than one towering specialty. That kind of operator has always been valuable in startups and zero-to-one work, where bottlenecks move constantly and dependencies kill speed. But they also explore the risk of burning out, topping out, and getting trapped by taking on too many responsibilities. They land on the real shift: AI lets generalists execute across more domains without waiting on specialists, shrinking the gap between seeing the bottleneck and solving it. KEY TAKEAWAYS Breadth matters most when bottlenecks move: the best generalists keep shifting toward the current constraint instead of clinging to yesterday’s valuable work. The trap is taking on too much: range becomes a liability when a generalist spreads effort across many useful tasks instead of the highest-value one. AI deepens adjacent skills: tools now let broad operators reach workable depth in coding, analysis, and research without full specialization. Depth still matters for trust: organizations still reward visible expertise, even if AI lowers how much specialist help is needed to get real work done. Context beats syntax: AI can write SQL or Python, but knowing what to ask, what to filter, and what to trust remains the durable edge. NOTABLE MENTIONS AND LINKS T-shaped skills describe broad cross-functional awareness plus deep expertise in one domain, and they give the baseline model Eric and John are reacting against in this episode. X-shaped skills extend the older metaphor toward leadership and people skills, and they come up as an example of how organizations keep inventing new shapes to explain modern work. Zero-to-one projects inside larger companies also favor generalists because they can move quickly with fewer dependencies and get new initiatives off the ground. Regression analysis is the episode’s clearest example of adjacent expertise, because AI now helps non-specialists do work that previously required more dedicated technical support.