Writing the Revisiting AI + Shopping series was fun and thought-provoking. It’s incredible how much has changed in AI and agentic commerce over the past year, so it was worth reevaluating the landscape. What’s even more interesting is how much hasn’t changed when it comes to AI’s impact on shopping. If you’re new here (or would like to give it another read), I’ll link each part at the bottom of the article. You’ll get the rundown on agentic commerce’s fundamental flaws, data collection challenges, and the companies most likely to win the race.

Now, onto regularly scheduled programming.

We are back with our 9^th edition of Hot Takes in AI. I’ve had a few queued up and I’m looking forward to diving in.

Hot Take #1: There is not a clearer signal of the physical AI narrative gaining steam than Anthropic hiring a Human Data Platform product manager.

I’ve said it before, and I’ll say it again: you get a peek at a company’s long-term strategy by combing through their active job openings. And if you are looking at the Careers portal at a company like Anthropic, you get a peek at what the future of technology holds.

I’d like to highlight a few sentences from the product manager role above.

“Builds systems designed to collect data to improve our models. This includes infrastructure to simulate real-world environments and tasks”

“Good instincts and an eye for intuitive user experiences, particularly those involving complex UI interactions or annotation workflows”

We’ve been highlighting the need for AI labs to acquire offline data to improve their models, as there is a relatively finite amount of information online. This data will be crucial for the robotics industry to power the next wave of “autonomous” everything.

My interpretation of this job posting is that Anthropic wants to productize the collection of offline data and is looking for someone who can navigate messy workflows, which sounds a lot like how getting data from people recording themselves will go.

I have a hunch this role will work closely with external data providers of real-world data. Companies like DoorDash, or Uber, and even Shift. Perhaps even a marketplace like Mercor that acts like TaskRabbit but for white collar work. Companies will view their real-world data collection as a monetizable revenue stream and ultimately sell it to AI labs.

Calling back to a previous Relentlessly Curious article: can you imagine how messy the data will be when maids on Shift submit their cleaning recordings? Of course, AI will take a first pass, but someone on the AI labs side will have to find a way to structure this data in a format that can inform an LLM. A difficult task, but incredibly interesting work.

Expect more job postings like this from OpenAI, Google, and even xAI. The robotics industry needs offline data, and the frontier labs would likely want to capitalize on this opportunity because real-world data improves their models and creates another big business opportunity.

Leave a comment

Hot Take #2: Meta’s stock has been punished over the past year because the market doesn’t believe it should deviate from being an advertising business.

Source: Yahoo Finance

Not considering a recent uptick, Meta’s stock price has been down a double-digit percentage year-over-year. The market has not been kind to Meta, as the market is asking a very fair question: does Meta really need to spend $135B this year on AI capex in the name of making their ads more efficient?

99% of Meta’s total revenue stems from advertising. Let’s not shy away from the numbers: Meta is an ad-tech company. It sells ads.

However, Meta does have a few side bets in the works, including an internally developed frontier model (codenamed “Watermelon”), as well as the recent announcement of a cloud computing business.

It sounds like they are trying to encroach on every Big Tech player’s turf (cloud, frontier model). But Meta is late to the party. The market has already discounted its juggernaut advertising business (as reflected by the stock performance) for their side bets. And if they want any sort of shot to build their cloud book, they’ll need to steal share from Google, Amazon, and Microsoft. The only wedge I could see working is if Meta commits to underpricing their competitors. Which is a bold bet that I doubt the market will tolerate given the time it will take to build the tens of billions of dollars in cloud revenue needed to compete with the incumbents.

And who knows if Meta will try for enterprise adoption of Watermelon. Good luck gaining a sliver of the market from OpenAI, Anthropic, Google, and xAI. Not to mention the Chinese models.

I get that Meta is a Big Tech player and needs to have a massive AI story. But its advertising business is the most efficient on the planet and keeping the focus there should be the top priority.

Playing devil’s advocate, Meta’s current market capitalization is $1.7T. If Zuckerberg needs a story to get Meta to $3T, further optimizing ads is the play. But if the goal is $10T, I get it. Meta will need to make company-altering, strategic bets to achieve exponential growth from here. Given Meta’s massive AI capex budget, my take is that Zuckerberg is more focused on the $10T figure, instead of the $3T figure (all speculative). He built a multi-trillion-dollar company from scratch, so perhaps the market should give him slack to pursue a few moonshot bets.

But it’s going to be a long, bumpy road for shareholders since Meta will be playing catch-up with foes who are spending just as much if not more on AI capex.

Subscribe now

Hot Take #3: AI labs’ new model release strategy is weird.

Call me a luddite, but I really don’t understand the emphasis on AI model benchmarks.

Last week, OpenAI announced the release of their flagship model, GPT 5.6. With major improvements for cybersecurity and agentic coding, it’s expected to pose as a formidable competitor to Anthropic’s Opus 4.8 and Fable 5.

Regardless of the model, each AI lab’s press release is a rundown of model performance versus benchmarks.

Here’s the thing: do model benchmarks make sense to be the bulk of a new model release, if the focus for AI labs is to show blistering revenue growth?

For what it’s worth, the AI labs are growing quite quickly. Specifically, Anthropic is setting records each month, with its annual recurring revenue (ARR) expected to breach $70B by month’s end. Keep the revenue machine humming, but maybe the model release should be a tactical use case demonstration. Pair what works well about your new model against your competitors’ weaknesses. Highlight an early beta partner client for the model release and the value it drove for them. Literally anything besides showing a leaderboard on various exams that mean nothing to almost every consumer and most enterprise buyers. I haven’t come across anyone in my work that claims they use one model over another because it’s atop the model leaderboard.

Call me a luddite for not being entertained by the model benchmarks. I’m not an AI researcher.

Just show me what the model can do and how it fits into my workflow. Then I’ll decide if I want to use it or not.

Revisiting AI + Shopping Series

Part I

Part II

Part III

Part I: Agentic commerce is best suited for routine, monotonous purchases. However, Amazon or DTC websites have already solved this problem through subscription offerings.

Part II: Agentic commerce needs offline, real-world data to attempt to serve as a reasonable alternative to in-store shopping.

Today, we’ll tie a bow on the series by making a clear statement on where agentic commerce fits into a shopper’s journey and where investment in agentic commerce is justified. It’s worth asking: what type of commerce environment is truly compatible with AI agents today?

That’s easy. One with clear rules and low ambiguity. In other words: marketplaces.

Here’s what we know about agentic commerce so far: you can either delegate making a purchase with predefined guardrails to an agent, or you can have an agent help you think through and identify what you’d like to buy. The definition continues to evolve rapidly.

See, commerce exists on a spectrum from deterministic to probabilistic. And agentic commerce realistically only works on the deterministic side of the spectrum.

If you’re buying the same toothpaste every month or replacing the same printer ink you’ve always used, those are deterministic purchases. But if you’re not sure what you want, you’re leaving much more discretion to the AI agent, making the outcome probabilistic.

Think about what data you’d provide an AI agent to make a purchase on your behalf. It may include color, style, price range, as well as any specific constraints. This could range from buying toothpaste to, perhaps one day, purchasing a car (yes, in the distant future). If you said, “Go buy a new t-shirt”, you’ve left your agent an extremely difficult task of figuring out what kind of t-shirt to buy based on the data it can pick up about you. I doubt you’re going to like what arrives in the mail.

By providing requirements to an AI agent (along with a form of payment), you’re establishing a set of rules for the agent to operate within. It’s essentially a dynamic filtering system, like that of a website. To satisfy your request, the agent needs countless data points like inventory availability, size, color, shipping timeline, and return policy to start narrowing down what it is the agent may buy on your behalf. Marketplaces like Amazon or Walmart have the infrastructure to meet these inquiries, which include shipping windows, returns processes, pricing, reviews, and fulfillment expectations. The marketplace has already standardized the rules of commerce. AI agents simply operate within those rules. This level of rigidity exists on a much larger scale than what DTC brands (using Shopify as the reference point) operate within.

Share Relentlessly Curious

Another key reason why marketplaces are better positioned to win agentic commerce than DTC brands is because of economic incentives. Marketplaces (and AI agents) optimize for conversion, while DTC brands look to maximize lifetime value (LTV). Marketplaces make money when transactions happen. Brands make money when customers buy more over time. This comes alive in a few different ways for DTC brands that don’t fit well into how agents operate.

Shipping Thresholds

Another point of confusion (or even failure) for AI agents is free shipping thresholds. When a DTC brand offers “Free shipping above $100”, but you programmed your AI agent to not buy a long-sleeve button-down shirt for more than $80, the agent may check out at that brand’s website and pay the shipping fee (say the shirt is $60, and the shipping fee is $10). There will be ambiguity around if a price includes free shipping and leaves it up to the agent to make the call on whether the price range (if not specified by the user) includes a potential shipping fee or not. If the agent must ask the user, then that is an extra step for the user to have to decide whether it’s worth adding another item from the brand to avoid the shipping charge.

This introduces unnecessary friction and violates the principles of agentic commerce, as it increases the amount of time someone spends shopping (instead of full delegation). It’s a worse experience for the customer, who may look negatively upon the brand for charging for shipping when the customer didn’t think shipping was included in the total price the agent told them.

Cross-Sells and Up-Sells

Think about every time you’ve reached checkout on a DTC website and encounter “shoppers like you also bought”. Or “bundle and save” offers. AI agents are less susceptible to emotional merchandising tactics than humans because they’re optimizing against user-defined objectives rather than impulse. They are provided with clear instructions on what can be bought, and for what price range. Purchasing a commonly added item to a cart goes against the ethos of a hyper-efficient agent. Also, say that you put in a trigger to your agent to alert you of any cross-sell or up-sell offers. I would argue that there should not be any human-in-the-loop steps for a shopping agent (unless it fails to execute the transaction), otherwise, it would be more efficient for you to check out on your own. Brands are likely to lose out on increasing their average order value (AOV) thanks to agents.

Because of these reasons, agentic commerce is a better fit for those that own the marketplace infrastructure already. I believe Amazon will continue to lead in agentic commerce, establishing what frontier can look like in this space. They have decades of commerce-specific data, hundreds of billions of gross merchandise value (GMV) each year, and capital to invest in building the AI necessary to turn Amazon.com into an agentic marketplace.

My bet is that the home page of Amazon.com looks very different a year from now. I believe Amazon’s homepage will primarily become a conversational search experience, where conversations become a first-class interface alongside browsing and search. Alexa will be your personal shopper assistant, regardless of the category you are looking for (previously a role reserved for in-store associates at luxury brands).

It’s time to bring us home. See, agentic commerce doesn’t replace marketplaces; it amplifies them. Marketplaces transformed shopping into a rules-based system decades ago. Now it’s time for a new phase of marketplaces. One that will increasingly be operated by AI agents.

Integrating AI into shopping is hard. Just ask Andy Jassy, CEO of Amazon:

“I think that [AI] has the chance to make it easier for customers to find what they want. If you know what you want it’s pretty hard to find a better experience than popping onto Amazon and searching and finding it. But the one place still where physical retail has some advantages in my opinion is the ability to go in, not know what you want, ask questions, refine those questions, have somebody point you to different things. And I think agents are going to help customers with that type of discovery” – Andy Jassy, being interviewed at the World Economic Forum in January, 2026. Original post from Juozas Kaziukenas here.

Five months ago may as well be an eternity in AI. But there hasn’t been a breakthrough moment in agentic commerce compared to what we’ve seen in areas like agentic coding. It’s not as though OpenAI’s GPT-5.5 has materially improved agentic commerce (in fact, OpenAI sunset their ChatGPT shopping module, Instant Checkout, back in March).

Let’s revisit Jassy’s quote. Particularly this line:

“But the one place still where physical retail has some advantages in my opinion is the ability to go in, not know what you want, ask questions, refine those questions, have somebody point you to different things.”

I believe what Jassy is referring to, and what is so difficult to replicate digitally, is the concept of discovery commerce. He’s not referring to the $7 latte you pick up at your local coffee shop (however, the prices of coffee in NYC are getting to the point where it may as well be a high ticket purchase). He’s referring to purchases where preferences are latent, evolving, or difficult for consumers to articulate. Typically, emotional or higher ticket transactions fall into this bucket, like buying a handbag or a car. But buying clothes can also fall under discovery commerce if the consumer is unsure of what they want to buy.

If Jassy says Amazon hasn’t cracked discovery commerce, then that means OpenAI may have a chance to catch up in the agentic commerce race. If I were advising OpenAI, I’d tell them to zone in one thing to accelerate their agentic commerce ambitions: the collection of offline data.

You’ve likely picked a theme over the past few months in Relentlessly Curious, and that’s the importance of physical, real-world data to support AI development.

When it comes to commerce, it’s no different. See, ~80% of commerce still happens offline, meaning in-person. For AI to revolutionize the shopping experience, it will have to offer a step-function improvement on driving to the mall, walking around and poking your head into a few different stores, trying a couple items on, and then making a purchase. As we chatted about last week, subscription offerings (whether they’re on Amazon or DTC) serve as a reasonable facsimile for low-stakes purchases. If consumers have already automated these purchases through subscriptions, it’s unclear how much additional value AI can create.

My anecdote last week covered my Banana Republic outlet mall success story. To make curated clothing suggestions, AI needs to understand my clothing style based on actual data rather than self-description. Because remember, I don’t even know what my style is, nor did I know what type of long-sleeved button-down shirt I wanted. If I could take a few pictures of my closet so AI has a sense of what I already own, it may be at a better starting point than the gibberish I came up with. It’s a classic GIGO (garbage in, garbage out) problem. For highly customizable, emotional purchases like clothing, more data will likely lead to better recommendations. In one line, preferences are best revealed through behavior, not language.

The elephant in the room is that shopping in person can be more productive than checking out online. It’s not a pleasurable experience to buy a few shirts online and then have to figure out how to return them when you realize the brand’s style doesn’t fit your body type.

Also, there are people who really enjoy the experience of shopping. Ever hear of the term “retail therapy”? Sure, this concept can apply online, but shopping in person and touching and feeling clothes or test driving a car can serve an emotional need.

What OpenAI will need to do is bring the real world to the AI experience.

Share Relentlessly Curious

I recommend OpenAI tackle categories one at a time, instead of tackling every shopping sub-industry. Start with fashion and apparel and build technology to capture and ingest someone’s wardrobe based on a few pictures. Have them take pictures and videos of their living space, to help with interior design for virtual furniture buying.

Investments in AR and VR will be critical to convert user-taken images into high-intent suggestions. Productize the process for one (large) industry and then move on to the next one. AR and VR are really data collection mechanisms, allowing your customers to create a structured data set in their own world.

They’re consumer preference extraction tools; bridging the gap between what you think your style is (in the case of fashion) and what it actually is. If you remember from You Are What You Buy (Part I), what you purchase says a lot about you. I’ll call back to the article for this helpful quote:

“In a world of nearly infinite choice, what we buy says a lot about us. Choice is not only a convenience, but also a signal. It’s an extension of our identity.”

AR and VR can help communicate who you are to AI. Sounds a bit ominous, but if you’re willing to allocate your resources to buying an item, that item becomes another data point in a preference graph that AI can use to understand what other products you are interested in.

With that said, I don’t believe OpenAI should look to create a pair of Google Glasses or Meta RayBans. AI wearables have proven to be a flop from an aesthetic perspective so far.

An acquisition idea: buy Snap for its camera and AR technology, assuming Evan Spiegel would give up voting control. Its stock has been in a constant free fall over the past few years and could be a bargain for what Snap has already normalized when it comes to the collection of highly personal visual data.

Leave a comment

Another facet of the offline data collection conversation is understanding how people interact when they shop in person. What insights can be gained from understanding which clothes people try on but don’t buy? Is it size related? Pattern related? Color related? Or something else? What about how long it takes someone to find just the right dress? How many dresses do they try on?

OpenAI should partner with major retailers (think Walmart or Zara) to gain anonymized data on how shoppers behave in physical retail. Large retailers already collect plenty of data on in-store behavior through cameras, sensors, and transaction systems.

By understanding how people behave while shopping, OpenAI can build more curated recommendation algorithms, and guide the user through a more efficient questionnaire that fits how they shop in person, whether they realize their behavior or not. Maybe the insight is when someone shops for a shirt color like what they have in their closet, they are more likely to convert than taking a risk on another color, unless specified by the shopper. Having this data can create a meaningful advantage for OpenAI throughout the ambiguity of discovery commerce.

What will be critical to the collection of offline data is trust. OpenAI doesn’t have the best public image right now (to say the least), and for people to feel comfortable taking a picture of their belongings or their home, they’ll need to believe OpenAI is a trustworthy steward. Snap has already normalized sharing highly personal visual data with consumers, thus an acquisition could help the cause.

The winner of discovery commerce may be whoever builds the largest dataset of real-world human preferences, not whoever builds the best model.

If you want to learn about a company’s future strategy, look at the roles they’re hiring for.

See below for open product management roles at OpenAI.

I’d like to zone in on the last product manager open role in the screenshot above. The shopping role caught my attention. You see, if a company’s open roles list is any indicator of their plans, then it looks like OpenAI is going to take another crack at being able to check out within ChatGPT.

We’ve talked a lot about agentic commerce in Relentlessly Curious, including OpenAI’s struggles getting Instant Checkout off the ground. So here’s the short version: OpenAI launched their shopping product last October. It was a slow roll-out to several big consumer brands (like Skims) and retailers (like Walmart). Consumer interest was limited and Walmart was vocal that they saw a materially lower conversion rate within ChatGPT than on their website. Simply put, it failed to gain meaningful traction.

Why did it fail? Operationally, agentic commerce requires an enormous amount of infrastructure. And from a conceptual lens, AI is best suited for purchases that you have a general sense of what you want to buy from the get-go. It’s not effective for discovery.

Need a refresher on agentic commerce? Check out AI + Shopping, When AI Buys Your Groceries, or The Case Against Instant Checkout.

In AI + Shopping, we chatted about how AI will bifurcate shopping into two forms: monotonous commerce and discovery commerce. The former focuses on routine purchases like paper towels and toilet paper. Grocery shopping can fall into this category too. Monotonous commerce is a natural fit for agentic commerce, as these types of purchases are low-stakes, frequent, and commoditized.

On the other hand, discovery commerce focuses on higher-ticket, emotional purchases. Think of jewelry, a car, or even a house. You may use AI to help you research which necklace, SUV, or property you would like to buy, but most people probably wouldn’t let AI make the purchase on your behalf. Likewise, the seller may not feel comfortable selling you such a high-priced item either via an agent either.

There are so many considerations when it comes to agentic commerce. Like which forms of payment are accepted and who maintains their security, who handles fulfillment and returns, and who accepts responsibility if an agent makes the wrong purchase. It’s an incredibly tall order to take on and requires coordination across many counterparties (see Universal Commerce Protocol).

If I was advising OpenAI on integrating shopping back into ChatGPT, I’d tell them to focus entirely on routine commerce. Make the user experience as seamless as possible to order groceries, soap, and napkins. Whether that experience is when a user wants to check out for themselves within ChatGPT or set a few guardrails and let an agent buy these products for the user on a recurring basis, the user must go from discovery to transaction quickly and smoothly. The goal is to gain consumer trust and associate ChatGPT with shopping on the “no-brainer” items, so that OpenAI can gain methodocially gain trust on the more ambiguous shopping requests. And associate their brand with all types of shopping.

But here’s the thing. There’s another company that has already mastered low-stakes, routine commerce in the eyes of the consumer. It’s Amazon. They’re arguably the furthest ahead when it comes to agentic commerce too.

I believe Amazon is all over the monotonous commerce element and most likely to succeed in the long run because of how operationally sound they are. They have decades of commerce data, what appears to be the most functional AI shopping chatbot (Alexa for Shopping), and a fulfillment service that no other company can reasonably compete with.

But I want to address the elephant in the room. These types of purchases are routine-oriented, which is best fit for a subscription. Is there really that much value in an AI agent knowing the stock in your refrigerator and auto-ordering when your egg carton has only a few eggs left, when your weekly Instacart order is going to restock you soon anyway? Subscriptions already solve predictability without introducing another layer of decision-making. Is OpenAI trying to pair its incredible technology with a pain point that really doesn’t exist? Don’t subscriptions (either on Amazon or on a DTC website) already solve this problem?

Share Relentlessly Curious

My point is, AI agents today are best suited to shop on behalf of people on low-stakes, routine purchases. If these purchases are so predictable, wouldn’t it make more sense to order a product on a subscription than leave it up to an agent to execute? Then you have certainty of the price and timing, versus providing guardrails and introducing a probabilistic element with AI.

So, if AI is best suited for monotonous commerce and Amazon already has major structural advantages, OpenAI will need to win in discovery commerce.

However, today’s AI struggles with the discovery phase of shopping. Rarely do I know exactly what I want. And if I don’t know what I want, I don’t know what to prompt ChatGPT with either. AI works best when you have a solid idea of what the end state should look like. What happens when you have no idea what you want to buy? How does that work with AI?

A few weeks ago, I was in the market for a few long-sleeve button-down shirts. I needed a shirt that was professional, yet laid-back in the sense that you don’t have to tuck it in.

As someone who doesn’t shop frequently, I didn’t really know what kind of shirt I wanted. It just couldn’t look too much like the shirts I already owned.

I went to an outlet mall and walked around to a few stores. I figured I’d know what I was looking for when I saw it.

And that happened. I walked into Banana Republic and quickly saw a shirt I liked. It wasn’t what I originally thought, nor was it something I would have prompted to AI. I just liked the shirt, tried it on, and I bought it. Simple as that. It looked cool and fit well.

Although a relatively low stakes purchase regarding the amount I spent, I’d classify this transaction as discovery commerce. I had only a general idea of what I wanted to buy, and seeing options in person, being able to touch them and try them on, helped seal the deal for me.

OpenAI will need to create a digital equivalent of walking through a mall or scrolling Pinterest. We will talk about how OpenAI can create the discovery side of commerce for the agentic era in next week’s issue. Stay tuned.

Last week, we touched upon the importance of branding in the physical AI narrative. As robotics companies look to purchase real-world data sets, people will be faced with a new reality: are they willing to delegate traditional human tasks like cleaning and driving to AI? Do they trust AI enough to do so?

Regarding driving, I’d say the answer is yes. In March, TechCrunch reported that Waymo is completing 500,000 paid rides across 10 cities. There is interest, but there is also pushback. Recently New York City Mayor Zohran Mamdani let Waymo’s autonomous vehicle testing permit expire at the end of March due to concerns on how autonomous driving will impact the local economy (read: taxi drivers, who have already taken a critical blow already from Uber/Lyft).

Although having a robot maid is tempting because doing the dishes and vacuuming aren’t the most fun tasks, I believe people will be hesitant to let a robot roam around their home cleaning and interacting with their precious belongings. Sure, iRobot’s Roomba has been sucking up dust from homes for over two decades, but it doesn’t have limbs and stays close to the ground. No real threat there (or not yet).

But then I read an article like this and wonder, maybe we are closer to the future than we think. Pepsi is testing out a fleet of autonomous vehicles to transport its products from distribution centers to retail partners such as Walmart.

Fascinating stuff. This got me thinking: what will the future look like if physical AI gains widespread adoption? Here’s a potential day-in-the-life scenario. Entirely fictitious, but food for thought.

Your alarm clock goes off at 7AM. You immediately regret going to bed late last night and subsequently hit the snooze button.

You finally drag yourself out of bed and step outside your bedroom. Who greets you right away, with a cup of coffee in hand? Your robot maid, of course.

After handing you your double shot of espresso with whole milk (you programmed the maid to make and deliver this), it waltzes into your bedroom and starts making the bed. In a matter of two minutes, the bed is made with military precision and may also be on display at Crate & Barrel.

You stumble down the stairs to make breakfast, and you hear the toaster going “click”. Your robot loaded bread into the toaster before it went to make your bed. You spread butter on the bread, while your refrigerator routes your low butter supply to your Instacart account so that butter is included in your next order. Your refrigerator knows you better than you know yourself.

Time to go to work (robots haven’t replaced all jobs yet). You pull out your iPhone (yes, Apple finally got its AI act together and avoided mobile phone disruption) and tap a few buttons so your autonomous car can drive you to work.

Where did you buy your robot maid? Well, you put down a deposit and pay a monthly fee to Tesla, which owns the entire autonomous “everything” market. The car, the robot, and your spaceship. Tesla and SpaceX merged and bought all car dealerships. They now use them as showrooms for their line of consumer and commercial robots. Ford and GM failed on electric vehicles and just couldn’t cut it when it came to autonomous driving. Elon won out.

Your car drops you off at work and then returns home. Heck, you don’t even need to own a car anymore as cars on demand arrive so quickly. But you bought it to feel some sort of nostalgia for a world where you used to own things, versus opting to lease the car from Big Tech.

During your work break, you log into Instagram and scroll through a feed of entirely AI-generated content. Humans have given up creating videos and taking pictures as AI does a better job of entertaining us than we do ourselves.

The workday wraps up and the moment you walk outside, your Tesla is waiting in line to pick you up because you scheduled it to do so the night before. Reminiscent of your mom waiting to pick you up from preschool.

On your way home, you see drones flying through the sky, hovering close to people’s doorsteps. The drones have the Amazon logo and appear to be dropping packages. A perk only available to Amazon Prime members.

You arrive home and see your robot maid picking up your drone-delivered set of Swiffer dust pad refills. The maid saw dust pads running low and pushed a notification to your Amazon app so you can approve the purchase.

Share Relentlessly Curious

This is a scenario that could plausibly happen if the physical AI narrative really picks up steam. We’ve talked over the past few months about how robotics firms are actively looking to acquire data that allows them to build the robots to handle so many unique edge cases associated with consumer applications.

However, this scenario greatly depends on trust. Trust in the companies building robotics to not compromise your safety or security. Trust that autonomous vehicles can function more effectively than humans driving cars. Trust that the labor market will reinvent itself in an AI-first world.

The robotics company that wins won’t necessarily be the one with the best robot. It may be the one people trust enough to let inside their home. And they’ll focus on building trust long before they look to sell you something.

Well, we talked about more than that. DoorDash had one of the first truly public examples of involving consumers in acquiring training data for physical AI. DoorDash began paying its Taskers to complete various household chores for the sake of data collection in the real world. As a refresher, foundational model companies will eventually run short on internet-scale text data to train LLMs. At the same time, any company building robotics must have access to niche, physical data to even start their endeavor.

If your goal is to build a robot maid, you’re going to need plenty of real-world data of people washing the dishes, folding laundry, and vacuuming.

Turns out, plenty of other companies (besides DoorDash) are serving as an intermediary for real-world data collection too.

I’m rarely on X, but I popped on the app this weekend and saw this announcement from Shift, a NYC-based free cleaning service. I highly recommend you click on the X post.

You read that right: free apartment cleanings for New Yorkers. The catch: the cleaners record themselves doing chores in your apartment. And Shift is super transparent about the transaction: “the value of [the] recording is what funds the service.”

This is pretty wild to think about. I can get someone to clean my apartment for free if I let them record it.

But it’s not really free. If you’re wondering why something is free, it’s because you are the product.

Same deal with most of the internet today. Paywalls and app fees are few and far between because the monetization model has shifted. Advertisers pay marketplaces and social media platforms to influence the decisions you make day in and day out. Why are Meta and Google free for users? Well, it’s because brands spend hundreds of billions of dollars each year to capture a sliver of your attention while you’re using their technology. Whether it’s Meta, Google, TikTok, or even while you wander the aisles of the grocery store (everything is a retail media network), your attention pays for discounts, entertainment, and social connectivity.

Next time you provide your email to a brand for “20% off your first purchase”, you should realize that you’re opting into giving the brand full rein to barrage you with marketing emails. It’s similar with a brand’s ads on Instagram. They are trying to meet you on the channel you are on and capture your attention, so you engage and buy from the brand.

Attention is monetizable. And in 2026, your physical dwelling is monetizable. Or at least, according to Shift.

As a self-proclaimed neat freak, I have no problem cleaning my own apartment. In fact, I find the act of cleaning relaxing at times. But I am tempted to try out Shift as a technology enthusiast.

But if I do, I’ll be careful beforehand. I imagine Shift is mainly looking to capture data points on all different kinds of furniture, kitchen sizes, and bathroom nuances to train how robots interact in a long-tail of environments. However, it’s quite likely they’ll also pick up images of family photos or valuable belongings. This data getting into the wrong hands could have serious consequences.

Shift does state the following in their privacy policy:

“We blur all personally identifiable information from screens and ID cards, to pieces of paper and cell phones to help protect both you and our home.”

But who knows what happens if they’re hacked.

Taking a step back, the AI narrative has created a lot of anxiety for society. U.S. states are considering data center bans, and the founder of OpenAI’s home was recently attacked by an anti-AI protestor. A growing subset of the population believe AI will take over the world in a negative way. Although this is heavily influenced by mainstream media who run sensational headlines to drive clicks and engagement, in return for advertiser’s budgets (you know, monetizing attention).

Things are rarely as good or bad as they seem. I’m pro-AI, but I also believe it is a “net positive”. Emphasis on net. I don’t believe AI will have as much of a long-term impact on the labor market because people find fulfillment in work and will create new jobs for themselves to add value. I also expect it’ll take longer than we think for AI to really push humans out of the loop, regardless of what any model benchmark says. But I’ll admit that AI may have drastic negative consequences on a person’s willingness to be social if they can talk to a chatbot without judgement.

Share Relentlessly Curious

Back to Shift and DoorDash. I believe their services will further exacerbate the negative press around AI. It’ll be branded as destructive to the physical services economy, as well as hurt lower-income workers who typically serve as the delivery drivers or cleaners.

On the flip side, there is a short-term angle here that helps lower-income earners. Gas prices are soaring thanks to the war in Iran, and food supply and materials inflation is likely to tick up materially if the conflict continues for much longer. Expanding the gig economy with opportunities to pick up additional work through task recordings provides additional income streams for those who are struggling to make ends meet.

But it’s dystopian. The cleaners and delivery people are training their replacement. There’s a short-term win because it creates jobs, but over time those same workers may end up training the models that automate parts of their work. And it’s not like these jobs come with health insurance or benefits; they’re contractors.

The tech-enabled marketplaces are best-positioned to serve as the middlemen between gig workers and AI companies looking for real-world training data. I’m watching closely to see which companies are buying this data to help understand what the future will look like. The buyers will do a lot to shape how consumer-facing robotics are perceived going forward. I hope they have strong branding and are seen as trustworthy. The physical AI narrative may ultimately shape AI’s path in society because for the first time the technology will quite literally be inside the home.

Same format as prior Hot Takes in AI editions, but this one is going to be hotter than usual. Let’s get going.

Hot Take #1: SpaceX, OpenAI, and Anthropic IPOs are bad for the public investor.

If you’ve read last week’s Relentlessly Curious, it’s fair to say this take feels a bit redundant. However, I’m going to come at it from a different angle. AI start-ups have seen unprecedented growth, but often at all costs. As these AI labs enter public markets, they will be subject to the same quarterly earnings announcements as everyone else. Higher-than-expected compute costs or slower-than-expected user growth figures are bound to send the market into a tantrum given how much investment has gone into the new frontier from existing Big Tech players. Right now, no one really knows what’s going on under the hood at the AI labs. The numbers they release to the public are at their own discretion, and those figures are chosen deliberately to best position the narrative.

So, when the AI labs are forced to show their comprehensive financial statements every three months, some screws will come loose. In this scenario, more information may actually make things worse. Similar to how the broader market now seems to hinge on NVIDIA earnings announcements, I believe each AI lab earnings call will be a major volatility event for the market.

“Okay, so I won’t buy these IPOs. Thanks for the tip.”

Well, it isn’t that simple, because you’ll probably end up owning these companies whether you like it or not. If you invest in broad-based market-tracking index funds, you’re going to be an owner of each of these companies. SpaceX plans to list their stock on the NASDAQ, which has fast-entry rules that can accelerate inclusion into the Nasdaq-100, within 15 trading days. It’s still to be determined which exchange OpenAI and Anthropic list on, but they’ll likely qualify based on the Fast Entry market capitalization requirement, being in the top 40 of companies. At an $850 billion valuation, OpenAI would immediately rank among the world’s most valuable public companies. Number 16 to be exact as of May 22nd.

As an index investor, why should you be worried? Because of volatility.

Volatility isn’t ideal for the broad-based index investor. The problem is that the index investor is going to own the AI labs soon enough due to index fund tracking rules. The indexes are getting even more concentrated around tech and AI, and once the market knows more about these frontier companies, volatility will become harder to ignore. Thinking you’re fully diversified just because you own a broad-based index may become a fool’s errand.

Fluctuations in SpaceX stock aren’t going to materially move an entire index, but the market’s concentration in technology and AI continues to deepen and AI’s share of market attention keeps growing. Whether it’s SpaceX, OpenAI, or Anthropic, I find it hard to believe that it’ll be smooth sailing in public markets once Wall Street can comb through every line in their financial statements. And that’s not to mention private tech firms like Stripe or Databricks that may entertain public markets at multi-hundred-billion-dollar valuations soon, which would only deepen public market concentration in tech.

Also, side note: if SpaceX were to go public at a $2T valuation, that would be wild given it is estimated to have generated roughly $20B in 2025 revenue. Meanwhile, Amazon is trading at a little over $2T and posted roughly $750B in revenue in 2025. Really makes you think.

Share Relentlessly Curious

Hot Take #2: Salesforce is the software stock to own coming out of the SaaS apocalypse.

Last month, I attended Salesforce’s Agentforce World Tour conference in Manhattan. Work conferences can be hit or miss, but I’m pleased to report that the show was worth my time. With that said, the food was bad. Both from a selection and taste perspective. You’d really expect better from a company worth nearly $150 billion.

Jokes aside, I was impressed with Salesforce’s continued success in integrating AI into their core product. Agentforce 360 turns your CRM into an agentic operating system, acting as the intelligence layer for your company’s data. Meanwhile, Slack has continued expanding their Slackbot, which similarly acts as the intelligence layer and integration capability across all your company’s internal and external communication. Employees can leverage Claude (or their AI model of their choosing) with the Slack interface.

Salesforce has what AI labs need: context-rich, hyper-specific data because customers have been storing their company’s central nervous system inside Salesforce for decades. You’re not going to get rid of Salesforce, especially since they’ve successfully embedded AI into their core product. As a Slack user, I’m impressed by SlackBot and the long list of available MCPs. If I’m scaling a company today, I’m leaning heavily into Slack given its ability to consolidate internal communication and AI into one platform. It’s perfectly situated as the horizontal layer within a company’s tech stack.

I’ve completed some really cool vibe coding projects (one where I replaced internal marketing analytics software that is saving the company I work for five figures per month). But I would never even consider trying to vibe code Slack. That would be crazy. Even Anthropic leans heavily on Slack. If the masterminds behind Claude aren’t trying to build a new workflow communication system, why would any other company think they could pull it off? No one is cancelling their Slack subscription.

The elephant in the room is multiple compression. AI does fundamentally challenge how investors think about a software company’s terminal value, due to its impact on the long-term growth rate. Seat-based software contracts may become less sticky, assuming clients reduce their labor force thanks to AI efficiencies. And AI-native companies will rely more on creating their own internal tools.

Even if I don’t believe Salesforce will be affected in the long term, its association with software stocks could impact its valuation. The narrative checks out for me: Salesforce looks poised for steady growth, but multiple compression could keep its stock price down.

Disclosure: I own Salesforce stock.

Last week, Anthropic announced a change to how they bill Claude usage via X, which goes into effect on June 15^th. Big shoutout to @arckollect on X who broke down this technical announcement that candidly goes over my head. I’ll do my best to further summarize (I read through this post five times).

Basically, there’s two ways people use Claude today: hands-on use and agentic use via API. Hands-on use implies that you are typing prompts into Claude, while programmatic use means you’ve built a process (i.e. agentic workflow) that leverages Claude to get work done via API. Not much will change if you are a hands-on user assuming you are not heavily using the API. But if you are heavily relying on agentic systems, your Claude usage now comes out of a separate bucket with a fixed cap. After you hit your cap, it’s up to you to pay a variable rate based on usage. Or you’re out of luck for the rest of the month.

Historically, many heavy users benefited from generous pooled usage limits. If you’re tinkering around with Claude Code, you’re probably not affected by this change. But if you are building a vertical-layer AI company or the first one-person billion-dollar company, you’re relying on AI as an operating system. Thus, your programmatic use may get a lot more expensive.

This subtle change in pricing is just the start of AI labs starting to think more about their own unit economics. It’s a safe bet to assume that AI will cost more for the user in the coming months and years as AI labs need to demonstrate a long-term viable business without the constant injections of massive capital. With public markets ambitions, these private market goliaths will need show they can tidy up their financial statements. The issue isn’t whether Anthropic and OpenAI are valuable businesses. They are. It’s whether public markets will tolerate the economics required to build these companies.

I’m curious about how this impacts the unit economics of vertical AI companies. You know, the start-ups that are built on top of Claude, OpenAI, and Gemini models. I’d imagine that few AI start-ups are profitable (please correct me if my assumption is wrong). For those that are already deeply unprofitable at the unit economics level, they will really struggle to get to profitability if token costs continue to go up. I’m afraid that entire AI businesses may be getting built on temporarily mispriced intelligence.

But these start-ups have other options in the near term. They can switch from Claude models to OpenAI’s Codex. The new agentic coding solution from OpenAI received a major upgrade thanks to the release of GPT-5.5, and OpenAI is letting developers run wild with token usage (much less restrictive compared to Anthropic). Additionally, OpenAI recently formed a joint venture with several major private equity firms called DeployCo for the purpose of growing adoption of OpenAI models at major enterprises. Despite major brand hits over the past few months, OpenAI is turning on the jets on B2B and creating serious competition for Anthropic while Anthropic takes criticism from power users.

Note: Although I rely almost exclusively on Claude models for work, I have tested out Codex and am pleased with its efficient token usage and higher-level reasoning for complex requests. Though switching my entire AI setup is time-consuming, Anthropic would need to significantly increase prices and further reduce token usage for me to fully switch to Codex.

Share Relentlessly Curious

Here’s the thing though: it was only a matter of time until Anthropic curtailed token usage. They are burning through billions of dollars in compute each month and it’s public information that they may look to IPO by the end of the year. And when they file for an IPO, the world will get to see their S-1 offering (think of this as a prospectus for potential investors), showing what is really going on under the hood from a financial perspective.

Even with significant reductions in token usage offered on fixed rate plans, I have to imagine that unit economics on compute are a disaster. Making up numbers here, imagine a situation where users are paying $1 for $10 worth of tokens. That type of pricing arbitrage doesn’t just close overnight. It would take years and require threading the needle while staving off competition that can be subsidized with many hundreds of billions of dollars of capital. I know it doesn’t seem like it now, but eventually investor appetite will dry up. It remains to be seen how much people are willing to pay for AI if prices go up. Are they willing to pay double? Probably. But are they willing to pay ten times? You’ll need a super clear business case to do so.

Last week on Prof G Markets, guest Ed Zitron argued that Anthropic co-founder Dario Amodei’s repeated statement that Anthropic is profitable on inference (the process of providing responses to users) is misleading. Zitron makes the point that being profitable on inference is a helpful indicator as it takes many billions of dollars to train the model in the first place. A major sign of long-term business viability would be if Anthropic were profitable on both inference and training.

But maybe this time is different. Maybe unit economics doesn’t matter because AI is a winner-take-all market and massive scale is paramount to winning. Take Uber for example, who operated at a loss for well over a decade, including several years as a public company. AI is about narrative and it’s the hottest narrative I’ve seen in my career.

Narratives can stretch economics for a long time, but not indefinitely. And when the S-1 comes out, is when Wall Street gets a peek at how bad the unit economics are and what accounting assumptions are being used to support unit economics figures. I’m not suggesting anything fraudulent is going on, but my guess is that there are aggressive interpretations around profitability metrics.

The thing is, Uber eventually raised prices and is a highly cash-flow generative, profitable company now. The best-case scenario is that Anthropic or OpenAI can change their entire pricing model to variable, and people stay on their current plan.

Because the US stock market is so heavily concentrated in technology (Magnificent Seven make up roughly 35% of the S&P 500 value) and Big Tech plans to spend nearly $700 billion on AI capital expenditures this year, anything that adversely impacts the growth of the AI labs is likely to cause a cascading effect on the rest of the market. As Scott Galloway says, “the market is a giant bet on AI”.

When Anthropic and OpenAI go public (likely later this year), they will be at the whims of public market forces and investor scrutiny. They can’t hide their numbers anymore behind hundreds of billions of dollars of private market money. If the narrative around AI labs’ ability to continue growing at a record-setting clip and ability to inch towards profitability breaks, then the rest of the market will take a major dip. For those of you who invest in only broad-based index funds like the SPY (S&P 500 tracker), you’re exposed big time.

Accessing public markets brings enormous amounts of capital to the AI labs, however I believe it increases their risk profile. They’ll be subject to quarterly earnings requirements, and Wall Street will hold them to revenue, user, and profitability targets. If they stumble, the rest of the market will fall on its face given how dependent the market is on AI capex generating meaningful returns. And that is likely to affect valuations far and wide.

I wish I could invest in Anthropic and OpenAI. They are businesses that I believe will change humanity for the better. But sheesh, my investment portfolio is perfectly okay with them remaining private businesses for a little while longer.

Check out the How to Vibe (Claude) Code series!

Part I

Part II

Part III

Take that valuation with a grain of salt, as reports quickly surfaced following the Times publication alleging that Medvi may have engaged in fraudulent behavior. So maybe we don’t have the first credible one-person billion-dollar company yet.

But the concept is starting to emerge, given the step-function development of AI. Those closest to AI seem to believe this type of company structure is imminent, where one person manages teams of AI agents (in lieu of people). Check out this clip of Sam Altman (CEO of OpenAI) mentioning the near-term possibility of ten-person billion-dollar companies, as well as the glorified one-person billion-dollar company.

The idea is mesmerizing. It turns conventional wisdom about building a company on its head. For most businesses with ambitions of being valued at over $1B, conventional wisdom says you need massive teams, capital, and infrastructure.

I’d love for one of my business ideas to take off and for me to build it by myself all the way to $1B.

But what happens if I want to sell the business and get out? Who would be a buyer of a one-person, AI-native company?

At the billion-dollar level, potential suitors tend to fall into two main buckets: strategic acquirers or financial buyers.

Quick aside: a “strategic” is a company that buys another company for strategic reasons (in its purest form). For instance, Unilever, consumer goods holding company, acquired Gruns in April. Gruns fits in well with Unilever’s portfolio and corporate strategy, thus it was a “strategic acquisition”. On the other hand, a “financial” buyer looks to purchase a company, make operational improvements, and then sell the company. Strategics are typically corporations, and financial buyers are usually private equity firms.

For a strategic buyer, a lean AI-native company is attractive because they don’t have to deal with cultural integration risk. You don’t have to worry about the team you’re acquiring being gun-ho about assimilating into the culture of your company. But you do deal with key-person risk. Sure, AI agents aren’t going to jump to your competitor (or maybe they will if AGI is achieved), but that one person at the ten-person start-up could. Each employee becomes mission critical, so retention packages effectively become embedded into the valuation. Although management retention packages are commonplace in mergers, the stakes are higher if there are only ten people at the target (versus thousands).

Enter, financial buyer. If AI allows companies to operate hyper-efficient workflows and each employee managing teams of AI agents that complete various tasks, how does an outside investor come in and justify a valuation premium if there are limited operational improvements left to be made? Synergies will need to be on the revenue side, as the cost of AI agents will likely be as efficient as possible. What operational efficiencies can private equity drive if there aren’t any more operational efficiencies to create? Read: laying off workers. But wait, those don’t exist!

Financial buyers often rely on margin expansion, and financial engineering to drive returns. Venture-backed technology companies typically rely more heavily on equity financing than debt financing.

Quick tangent: I see both sides of the AI agent efficiency coin on my LinkedIn feed. Some people are bragging about how many thousands of dollars they were charged for Claude API usage the prior month as if this is a signal for AI productivity. I’ve also seen reports that entire companies, including Uber, are blowing through their full-year AI budgets within months. This inherently challenges the thesis that AI agents are cheaper than human employees.

On the flip side, I see others on LinkedIn offering Claude Skills they built to optimize token usage. I imagine a new layer of companies will emerge to help AI agents become as efficient as possible with token usage. In general, as the race to build more data centers continues and the cost of compute decreases, per-token usage costs should decrease. Over time, competition and infrastructure buildout should continue pushing token costs lower, even as total AI usage rises dramatically.

AI-native companies are lean in nature. Highly AI-fluent people are managing teams of AI agents instead of teams of employees to get work done. This means a company’s operating leverage could become extraordinarily high, prior to any potential acquirer coming in. Essentially, AI expands a company’s operating margin to the point where it becomes as lean as realistically possible. If you’re looking to make a purchase, you need a clear revenue-synergy story to justify the acquisition price, as the cost-synergy story is unlikely to be there.

Say you decide to acquire a ten-person AI-native company for $1B. You almost need to do it in all-stock, to keep the operating team incentivized to stay (assuming the target company would even accept the deal). Otherwise, those ten people just became insanely rich, and it’s unlikely they’ll be as motivated as they were pre-acquisition to work at the combined company. Because AI systems are still imperfect, you need someone to stick around and remain accountable for the outputs those agents produce.

Look, if you’re able to build a billion-dollar company by yourself, you are crushing it in life. You’ll eventually figure out how to sell the company. But it does pose a tricky question for an acquirer. Not much meat on the bone in the way of incremental operating efficiencies. Combine that with significant key-person risk, and you’re not looking at a great acquisition story.

If you’ve made it this far, let me know how you think AI will impact valuations and exit potential for the super-lean, AI-native businesses being built today. It may ultimately force potential acquirers to ask themselves: what exactly can we do better? Because the list of answers is likely to be shorter.

Models

If you’re using the Claude desktop app, you’ll see the active model in the bottom right of the prompt bar (just look for the Claude logo). Words like Haiku, Sonnet, and Opus are branded names representing different large language models (LLMs) offered by Anthropic. You can click on the model’s name to open the menu of options within the desktop app or web browser. If you’re in the terminal, type “/model” to see the available models.

It’s important to know a little bit about each one before you begin a vibe coding session, as you’ll either be frustrated by output quality or surprised at how quickly you burn through tokens.

Each model is designed for separate use cases. Haiku is best suited for quick questions and answers. If I need a model to handle basic data retrieval or to answer generic questions, I’d rely on Haiku. It’s the fastest and cheapest model. But I don’t recommend using it to build an app or a website as it lacks the strategic reasoning element of the other Claude models.

Next, Sonnet sits in the second tier. This is my go-to model for vibe coding. For any moderately complex question, I’ll send it over to Sonnet. For most of the agents I create, I’ll have them rely on Sonnet. It’s effective and cost-efficient.

Finally, there’s Opus. View Opus as your sharpest coworker. It’s the one you go to for advice or guidance when the problem you’re trying to solve is too thorny. Opus can handle deep research, strategic reasoning, and complex, multi-step/multi-agent processes better than any other Claude model.

However, be wary of how often you rely on Opus. Regardless of whether you’re on a Pro plan or a Max 20x plan, Opus burns through your token limits noticeably quicker than Haiku or Sonnet.

When I use Claude Code, I always start off in Plan Mode before I request Claude to take an action on my behalf. The more context and guidance you can provide to AI, the closer the output will be to what you originally envisioned. I leverage Opus to write the build plan for the product or system that I’m aiming to build and rely on its strong strategic reasoning to question my logic and push my thinking. I regularly ask Opus to find holes in my own logic, as well as its own. When I’m satisfied with the plan, I push the plan to execute using Sonnet. This separates high-level reasoning from execution, which is more cost-efficient.

Also, you may have noticed the phrase “effort” next to the model’s name. Sonnet and Opus both have a dial that allows you to control the performance of the model. Go low, and you’ll get a fast version, but you may create some errors along the way. Go high, and you’ll burn through tokens quickly, but you’ll be able to handle complex tasks like full reviews of your codebase, synthesizing dozens of files at once, and find even slight irregularities between app versions. If you’re wondering, Sonnet at either Medium or High effort is my default setting.

Higher effort increases the amount of compute the model uses to reason through your prompt. More compute means more tokens used up.

Note: there’s also the rumored Mythos (which we chatted about here), but since it’s not publicly available yet, no need to dive into use cases.

Share Relentlessly Curious

Token Usage and Context Windows

“You keep mentioning, ‘burning through tokens’. What does that even mean?”

Good call out. This is worthy of an explanation.

Tokens in the AI world are like tokens in an arcade. You need tokens to play the game, and some games require more tokens than others.

Each word you type can map to one or more tokens. Additionally, uploading file attachments, requesting pictures, and asking Claude to search the web all require tokens of varying amounts to complete. Essentially, everything you do within Claude chews up tokens, but what you do and which model you use impacts how many tokens you’re using.

Another critical thing to know here is that lengthy conversations eat up tokens. When you press Enter on your next prompt, the full conversation is packaged up and sent to Claude for context interpretation. Thus, the more back-and-forth you have with Claude, the more information that gets continually sent back to the model, and the more tokens are burned in the process.

If you’ve picked up that Claude is starting to forget pieces of information you told it earlier on in your conversation, it’s because you hit your context window. This represents the maximum number of tokens that AI can “hold” within a single chat session. Certain pieces of data or information will start to fall off, and you’ll start wondering why your output looks funky. That’s because new tokens are pushing older ones out of the window on a “first in, first out” basis.

Tip: If you plan to change the topic, start a new session or tab. This resets the context window. Also, you can have Claude remember a summary of your conversation by invoking the “/memory” command within the terminal and then reference this memory in a new session. Leveraging “/compact” helps maintain context in a summarized, token-efficient form. Then, “/clear” erases all prior context, creating a new slate within the same session.

Leave a comment

Model Context Protocol (MCP)

MCP is technology developed by Anthropic that can be thought of as a standardized way for AI to interact with external tools and data sources. It allows users to connect various data sources to their AI setup. In an oversimplified one-liner, MCP is like an API for AI. I’ll hold off on the technical elements of MCPs (reach out if you’re interested, I’d be happy to chat) and instead focus on how you can set them up.

Say you are an e-commerce manager and you’re looking to build a dashboard that depicts real-time data from Google Analytics 4 (GA4). Claude can help you do this, and you can leverage MCP to make it happen.

What does that look like? You write the prompt, “rank site referral traffic sources in GA4 by sessions from last month and analyze bounce rate by source”. Claude can call the GA4 MCP (if configured) and return the data to your desktop app or terminal without having to log into a separate website.

Claude has a growing set of built-in connectors that leverage MCP to pull in data (and push actions too). Check out the “Connectors” section within the desktop app’s Settings area. Or you can ask Claude to help you set up a custom MCP, which will require more technical prowess. Still doable for even novice vibe coders though.

A word of caution: When making an MCP call within a chat, Claude is pulling a lot of extra data from the external data source. This will have a noticeable impact on your tokens and chew up your context window. MCPs are useful but difficult to use at scale given the cost impact and context window consequence. This is especially true if you are asking Claude to pull from many MCP connections within the same request. I recommend separating your tasks into various sub-agents that pull data individually from each MCP and then have an orchestrator agent synthesize the information between each sub-agent. This should eat up fewer tokens than trying to put together an analysis by stuffing each MCP into one session since each MCP call injects a large amount of data into the context window.

Reading between the lines of the model announcement, OpenAI subtly jabbed Anthropic by citing GPT 5.5’s token efficiency for problem solving. For context, Anthropic has faced criticism for token limits and model outages.

When I see news like this, I’m tempted to abandon my current AI setup for the flashy new thing. In some cases, that’s a wise thing to do, particularly if the new model’s upgrades are relevant to your workflow. But I believe you can reasonably expect that every major AI company is going to come up with the latest and greatest every couple of months and frequently changing your AI operating system has material switching costs in time and productivity.

My advice? Pick a model ecosystem and commit to going deep. It matters less which model it is, and more that you are putting consistent effort into learning how to get the most out of it for your workflow. I’ve chosen Anthropic’s Claude models as my focal point, but there’s no harm in spending your time in OpenAI’s Codex (Claude Code rough equivalent).

Although I’m not too familiar with the inner workings and lingo of other agentic coding tools besides Claude Code, I imagine the following tips will be transferrable regardless of solution. Without further ado, let’s get into this week’s Claude Code learnings.

Building in Claude Code

Markdown Files

Candidly, I hadn’t heard of markdown files (.md files) before I downloaded Replit a while back. It looks kind of like a .txt file (i.e. the Notepad app on your computer). A key trait of AI-generated markdown files is that they contain a lot of hashtags, asterisks, backticks, and bold-font section titles.

AI prefers markdown files because of its semantic structure. Hashtags signify hierarchy and structure, while backticks denote code. Since the communication method is clear, markdown files are relatively token efficient for AI models to ingest.

Although the formatting can take some time to get used to, writing instructions and context for agents are best handled in markdown files. I open an IDE (Visual Studio Code) and select a folder on my desktop. Then I’m able to see all the markdown files AI (or myself) has created, and either edit existing instructions or create new ones.

Markdown files look odd, but don’t be intimidated. They are essentially Microsoft Word docs for AI.

Claude Skills

Skills are clutch and are becoming an increasing part of my day-to-day. Oversimplifying it, the difference between skills and agents is that skills operate as singular tasks where context is defined ahead of time. On the other hand, agents handle more complex, sequential workflows and can maintain context across tasks. Agents can complete skills, but a skill is not an agent.

Recently, I developed a skill that transforms video transcriptions into YouTube-optimized titles, descriptions, and tags. I fed Claude Code all the information needed to complete this recurring task and now it executes the task whenever I type “/skill youtube-automation” and provide a bit of context.

You can create skills to consolidate notes, provide weekly recaps on meetings, and complete performance analysis.

How to get started? Ask Claude to help you create a skill. Make sure you’re on “plan mode” so Claude can run you through a Q&A on what you’re looking to build. Claude will proceed to generate a markdown file with your necessary information, goals, and steps to take.

Oh, and check your settings section of the Claude desktop app. There are tons of built-in skills that Claude offers today, that you can call directly from the desktop or terminal.

Share Relentlessly Curious

Pushing to Production

GitHub 101

You are happy with what you built and you want to “save it”. If your usual workflow consists of Microsoft Excel, Word, and PowerPoint, you’re going to need to rethink what saving here means. But that doesn’t mean it is harder to do.

GitHub is a platform that stores and tracks changes to your code over time. It’s the standard way developers manage, back up, and collaborate on software. With integrations across most development tools and AI workflows, engineers and vibe coders can upload their code to GitHub relatively quickly.

You commit your code (save a version locally) and push it (upload to GitHub). If collaborating, you open a pull request (PR) to propose changes to the codebase for your colleagues to review.

First, make an account on their website and then ask Claude to help you connect your local project to a GitHub repository and walk you through the setup. There’s a bit of a learning curve once you’re on GitHub and you look through different branches and navigate the codebase, but rest assured, your code is saved in a secure spot.

The Importance of Engineers in the AI Era

There are plenty of headlines about how AI is eating up the software engineer job market. While there are layoffs in pockets at big companies, the broader story shows that engineering job postings are stabilizing in some areas or even rising (as per Lenny’s Newsletter).

As a vibe coder, I’ve gained a strong appreciation for the value formal engineers bring to building software in the AI era. Vibe coding is excellent for spinning up a wireframe or a working prototype, however, if you plan to store user information or launch the product on the web for others to use, I highly recommend you have a well-trained engineer take a look.

Cybersecurity and scalability are two major concerns when deploying software. Even with repeated prompting around closing any vulnerabilities, there are still holes. Hackers can not only find a way to extract private user information, but they can also find ways to take advantage of features on your app or website.

What do I mean here? If your product involves an API call to an LLM to retrieve data, a hacker could exploit the API integration to run their own API calls and stiff you with the bill. Next thing you know, you have a $10K charge overnight because someone else found a way to run up your tab.

If you plan on accepting and storing payment information or user identification information, reach out to an engineer to take a peek at your codebase. Engineers aren’t going away any time soon; it’s more of a flight to quality. The best ones are in more demand than ever.

If you’re just starting out and have questions about getting started with Claude Code, I’d be happy to troubleshoot and share some pointers. Subscribe and reply to this email with your question!

Author’s Note: Today’s article is a sequel to #72: How to Vibe Code 101. If you haven’t read it yet, I recommend checking it out!

I’ve certainly caught the Claude Code bug. Over the past six weeks, I’ve found myself using Claude Code all day at work, as well as in the evenings for side projects. I’m consistently running up against my token usage limits and drawing on additional credits Anthropic graciously gave me.

Taking a step back, agentic coding agents like Anthropic’s Claude Code and OpenAI’s Codex allow users to build software by writing in plain English (natural language prompting). Since the February release of the Claude Opus 4.6 model, agentic coding agents have seen a step-function improvement in their ability to automate processes, reason strategically, and build legit products from a simple written request. The future truly is agentic. As we chatted about in #79: AI + Future of Work, agentic AI is the AI that is fundamentally changing how knowledge workers complete their job.

Today, I’ll dive into my vibe coding learnings as a daily user of Claude Code. For $20 (plus tax) per month, you can gain access to agentic coding agents and experience the magic of present-day technology firsthand. Not a sales pitch, just a guy who is really in awe of the ease of building today. Unfortunately, vibe coding is rarely free because of all the tokens each conversation burns when AI completes sequential workflows.

Looking back on How to Vibe Code 101, I still agree with my thought framework laid out in the article. Strong systems thinking and “starting with the end in mind” are critical to achieving your objective when working with AI.

Getting Started

Organization is a Top Priority

Before you download an agentic coding tool, you need to get organized. Comb through your folders on your computer or shared drive and make sure each file is placed in the proper folder, as well as named accurately. Your meeting notes shouldn’t be intertwined with your performance recaps.

This matters because you’ll ask Claude to complete tasks that will likely depend on where certain files live, as well as where you’d like to place the outputs of Claude’s work. For instance, if you want Claude to create a weekly performance analysis based on last week’s data, it will need to know where the information lives to complete the task.

Tip: Have a slew of random screenshots and pictures without helpful file names to tell you what the picture is? Put these files into a folder, grab the file path, and prompt Claude with “Look through the picture files in ‘X’ folder and rename each file based on the contents of the picture”. You’ll be amazed at the results.

The Interface: Terminal vs. Desktop app vs. Integrated Development Environment (IDE)

For those of you who are non-technical (like me), you probably haven’t spent much time in a terminal. Type “terminal” into your computer’s search bar and you’ll be greeted by an intimidating black screen. When I first set up Claude Code, I wasn’t too confident of my abilities to tackle using a terminal. After a week, it’s what I use 90% of the time I leverage Claude.

Once you install Claude software onto your computer, you can simply type “claude” into the ominous terminal and then you’re brought to a much friendlier user interface where you can type, and paste to your heart’s content.

Next, Anthropic launched a desktop app that allows you to access Claude Chat (a typical chatbot interface), Claude Code, and Claude Cowork all in one interface. You can view settings, manage integrations, and switch between modes here.

Finally, you can run Claude Code from within an IDE, which is essentially an interface that lets you write code (or type in English), view your file system and actual text and markdown files all in one spot. I find it helpful to review AI output and visualize how my files and folders are set up. Examples of IDEs are Visual Studio Code (what I use), JetBrains IDEs, and Cursor.

Personally, I prefer the terminal view because it gives me the maximum flexibility when using many data sources (via model context protocol) and running multiple agents at once. Desktop is much more user-friendly and intuitive, and I’ll use it for one-off tasks.

Share Relentlessly Curious

Operating within Claude Code

Use Plan Mode

I cannot urge you enough: spend as much time in “plan mode” as you can. It’ll save you time and money by using fewer tokens.

Plan mode allows you to talk back and forth with AI about what you would like to build and have it challenge your thinking. Many times, I’ll want to suggest a feature change to a product I’ve built. I’ll toggle on “plan mode” (hit Shift + Tab if you’re in the terminal or click the drop-down menu in the desktop app), type an idea, and finish the prompt with “ask me questions and challenge my thinking.”

Furthermore, Claude asks me questions about my idea, and will even request additional data points to better support why I want to take “X” action. This is critical because it forces me to be clear on what I’m looking for the AI to complete so I don’t send Claude down a rabbit hole.

The default mode in Claude is to receive a request and then execute (i.e., not plan mode). This works favorably for Anthropic because actions burn more usage tokens than back-and-forth questions. It’s a lot less expensive for Claude to write a plan than it is to take the actions within the plan. At the end of plan mode, Claude provides a formal, written plan sometimes with charts, diagrams, and visualizations. You can make edits to the plan or formally allow Claude to begin working.

Connecting Data Sources: Model Context Protocol (MCP)

One of my favorite parts of Claude Code is that I can connect various data sources to the software and have Claude develop insights from these live sources. For instance, a recent project of mine was automating performance analysis for an e-commerce funnel. I connected paid advertising data (Meta, Google Ads), search trends (via Google Search Console), website traffic information (Google Analytics 4), email performance (email service provider), and website behavior (heatmapping) to Claude through MCP.

From there, I provided context about the types of analysis I’d like to complete within the weekly performance recap. This included a mock-up of the final output, details about each data source, the personas and perspectives that Claude should consider, and benchmarks for past performance and future goals.

Then, Claude analyzed all the data sources I provided, incorporated my instructions, and produced a synthesized performance recap along with a prioritized list of priorities for brand, marketing, and technology teams to improve advertising efficiency, website performance, and conversion rates across the funnel.

In an oversimplified definition, MCPs are like APIs for AI. They allow AI to connect to external data sources and software, local files, and shared drives to stitch together context and execute requested tasks.

You can connect built-in MCPs from within the Claude Code desktop app or install them in the terminal by asking Claude for help on how to set up the MCP connection to the data source of your choice.

Note: Setting up MCPs can be “expensive”. When I say expensive, I’m referring to the context that MCP connections take up within a context window. AI currently has a limit on how much information it can recall within a single session, and that calling large external data sources through MCP takes up a large chunk of the context window as you start adding more sources (like I did in the e-commerce funnel analysis).

Be conscious of how many messages you have in a single session when leveraging MCP. Claude may begin to hallucinate or flat-out not remember context you provided earlier in the session. In this case, I recommend writing the terminal command “/compact” to summarize earlier information for Claude, freeing up context so you can continue prompting within this session.

Clarity of Thought

I’ve found Claude most effective when I provide a visualization of the process I’d like to create. For a dashboard I’m building now, I drew with a pen on paper, a front-end UX mock-up and a back-end data source map.

Now that I’m writing this out to you, I’m realizing that it likely isn’t that Claude works better with visualizations than well-written instructions. It’s that by putting pen to paper, it forced me to really think through what type of dashboard I wanted to build, how the user would interact with it, and the structure of the back end to support the various data sources.

The clearer your thought process, the stronger AI output will be.

Next week, we’ll walk through a 301 session, which will involve GitHub and pull requests (PRs), Claude skills, the importance of engineers in a vibe coding world, and plenty more tips!

Let’s get going.

Hot Take #1: OpenAI’s acquisition of TBPN was about buying engagement, not reach.

For those of you not laser-focused on tech news, OpenAI bought a podcast for a price tag rumored to be in the “low hundreds of millions”. Technology Business Programming Network (TBPN) is a daily, three-hour live show where two hosts discuss the latest happenings in technology and interview executives and other interesting figures from across the industry.

In my opinion, it’s an entertaining listen and their roster of guests rivals that of another top tech podcast, All In. But check out this: TBPN has only 85,000 YouTube subscribers and their live stream averages just 4,000 to 10,000 views daily.

Those statistics don’t scream “major media exit.” A few things to unpack.

Despite having relatively low subscriber numbers (for what it’s worth, I would give an arm and a leg to have these figures for Relentlessly Curious), where TBPN shines is in their clips performance. Today, live streaming is the “hot” form of media, and the red-hot form of media is “clips”. Think of small snippets of conversations or scenes from a longer-format video (i.e. live streaming). Their clips regularly reach hundreds of thousands of views, sometimes breaching millions of views on X.com. They also have advertising sponsors for their clips, which makes the claim that TBPN has a $30 million advertising run-rate business more realistic. So, from a pure media economics perspective, the suspected valuation is quite high, but not truly out of this world.

Taking a step back, it may seem that buying a podcast is contrary to OpenAI’s current directive on increasing their share of the enterprise and coding market. They recently shut down Sora and Instant Checkout in a heavily publicized move to reel in any side bets that were diluting company focus.

Buying TBPN wasn’t about expanding OpenAI’s reach into mass consumer adoption of AI. It was the opposite. They want to win back the power users of AI, those who work in tech or in tech-adjacent careers. TBPN consistently attracts the biggest names in tech and has incredible engagement and view counts on their X.com clips, the epicenter of tech discourse.

Despite claiming TBPN will maintain editorial independence, the brand association with TBPN may warm up AI power users to spend less time in Claude Code and more time in Codex (OpenAI’s agentic coding software). Also, the acquisition is an acquihire of people who truly understand the cultural zeitgeist of technology in 2026. If the TBPN crew can help OpenAI craft the right marketing messaging to capture enterprise clients from Anthropic or Google, the acquisition pays for itself and then some.

One last thing. TBPN has a vibey, but approachable brand. If OpenAI can leverage TBPN to repair their brand image post-Pentagon deal, they may win back retail investors ahead of a likely in the back half of 2026. Given their IPO likely could surpass $1 trillion, a few hundred million is a drop in the bucket for the sake of improving “vibes”.

Share Relentlessly Curious

Hot Take #2: Anthropic’s doomsday messaging around their new model Mythos is actually a good thing for society.

If you’re deep in Claude Code (like me), you understand just how great Anthropic’s top model Claude Opus 4.6 is. What I find most impressive about Opus is its ability to act as a strategic partner when designing a workflow or process. Agentic coding capabilities aside, its ability to synthesize the context I provide and quickly identify edge cases and holes in my logic helps build stronger, cleaner process automation.

Apparently, Anthropic’s new model, Mythos, has coding skills so capable that it has identified cybersecurity vulnerabilities across the web.

As a result, Anthropic quickly assembled a coalition of major enterprises named Project Glasswing because Mythos, “has already found thousands of high-severity vulnerabilities, including some in every major operating system and web browser”. Companies like JPMorgan, CrowdStrike, and Amazon are expected to work closely with Anthropic over the next 90 days to understand and fix any of the vulnerabilities surfaced by Mythos.

What’s notable is that many of these cybersecurity risks have been hiding in plain sight. As AI’s coding capabilities keep improving, it will continue to identify risks we weren’t able to comprehend with existing technology, which is particularly scary if this technology gets into the wrong hands such as a foreign adversary or a malicious domestic actor.

Riddle me this. Anthropic has built a differentiated brand in the AI space around “safety”. In fact, their legal entity is listed as a public benefit corporation, which effectively means they can prioritize stakeholders beyond shareholders.

With a brand focused on “safety”, Anthropic chooses to virtue signal frequently about the risks associated with their AI models and the steps they are taking to manage them. They release what is called a System Card with each model launch, that dives into model performance against benchmarks and potential failure modes (as well as how bad actors can exploit their technology).

Their fearmongering is a form of marketing. They focus on the downside of their models as their competitors often don’t highlight them. It makes Anthropic seem more responsible and trustworthy, particularly critical for enterprises considering using Claude models. Furthermore, it allows them to further embed themselves in enterprise and government infrastructure, which bodes well for long-term adoption of Claude. No wonder why Anthropic is crushing the B2B market.

But in the instance of Mythos, I believe caution is necessary and the 90 day pause is justified. If Anthropic is willing to risk their lead in the AI race for the sake of getting major enterprise players prepared for the risks associated with Mythos, this is a moment worth paying attention to.

Leave a comment

Hot Take #3: Perplexity’s move towards agentic solutions represents the only thing AI companies should be building for: outcomes.

If you’re building an AI company, you should pay attention to this. Perplexity’s annual recurring revenue (ARR) jumped 50% from February to March, on the back of their Perplexity Computer release. Their monthly active user counts topped 100 million as reported by the Financial Times, and a new pricing model, including both subscription and usage charges, has catapulted Perplexity out of the characterization of being a glorified search engine.

Perplexity Computer, a series of AI agents that leverage 19 different models to tackle the project you give it, has provided a new runway for Perplexity to build on. It also signals to the market what users value: agentic solutions.

Anthropic has Claude Code, OpenAI has Codex, and now Perplexity has Perplexity Computer. If you can knock out a complex task for a person or a company, they are willing to pay for it. Better yet, if the software you offer allows users to write in plain English what they want, and your software creates a tech product without them writing a single line of code, you’re golden.

People are willing to pay for outcomes. If your AI can’t automate a customer workflow, freeing them up to focus on higher-value tasks or scaling their business without additional headcount, good luck. Software creates insights, AI creates outcomes

*Crickets*

Well, anyway, it seems like a good time to reflect on some of the predictions I’ve made since the inception of Relentlessly Curious. For this week, I pulled out some of the hottest takes from articles that I wrote at least six months ago and am grading myself on whether I was on the right track. If you’ve been a longtime subscriber, I’m sure some of these pieces will ring a bell. If you’re new here, you get a recap of the content to date.

Let’s get going.

Prediction #01: Apple will acquire Perplexity

Article: AI + Acquisitions

Date: July 25^th, 2025

Score: Wrong 🔴

This one was way off. At the time I made this prediction, Apple did not have a clear AI strategy while their peers were seeing blistering progress. Despite not being particularly acquisitive (is that a word?) over the years (Beats by Dre is their largest acquisition to date), Perplexity was an appealing fit for the tech giant. Perplexity’s strong user growth in July 2025, recent launch of the Comet agentic browser, early adoption of advertising and affiliate channels, and relatively low valuation compared to their peers (i.e. OpenAI, Anthropic) put it in a prime position for acquisition discussions.

Spoiler alert: Apple did not acquire Perplexity. Since I wrote this article, Perplexity tried to acquire Google Chrome, even though Chrome was valued more than Perplexity at the time of their bid. Additionally, they ditched their advertising business and launched Perplexity Computer, their AI agent platform. Simply put, I have no idea what Perplexity’s longer-term strategy is, and I’m going to pause on making predictions about them. However, what I do feel confident about is that they need to choose a niche sooner rather than later, as they do not have the capital to compete with OpenAI, Anthropic, or Google on either mass-market consumer or enterprise AI adoption.

Maybe they reach down and try to own a portion of the vertical layer. I’ve always found Perplexity helpful for deep research. Maybe with their Perplexity Computer platform, they can build de facto software for researching anything. Yes, I know that’s vague, but for a company that has a $20 billion valuation, they still need to tackle vague problems.

Prediction #02: ChatGPT will launch in-prompt purchases

Article: AI + Brand (Part I)

Date: May 2^nd, 2025

Score: Correct 🟢

I was early on this take, but I don’t want to take too much of a victory lap as the writing was on the wall. There were Shopify code leaks referring to ChatGPT as a shopping channel that surfaced as early as April 2025. Thus, it was a matter of when, not if.

OpenAI needs to take on trillion-dollar industries, and commerce is certainly one of them. Given their “spray and pray” approach to new product development in 2025, it was a natural fit. They began rolling out in-prompt shopping experiments by October 2025.

How did that go? See below.

Share Relentlessly Curious

Prediction #03: OpenAI’s Instant Checkout will struggle upon launch.

Article: The Case Against Instant Checkout

Date: October 14^th, 2025

Score: Correct 🟢

As a follow-up to the previous prediction, I bet that shopping within ChatGPT would struggle upon launch. Which it did, so I’m marking this one as green.

Yet, I didn’t predict that OpenAI would shut down Instant Checkout as quickly as they did. Instant Checkout first became available in October 2025 and OpenAI pulled the plug in March 2026. The quick takeaway is that OpenAI spread itself across too many initiatives, Instant Checkout did not drive the level of user engagement and checkout volume they predicted, and the operational challenges with integrating commerce into AI proved to be quite challenging. The AI goliath retreated from commerce, as well as from their video generation app Sora, in a prioritization push to capturing the enterprise segment (although they are still introducing advertising into ChatGPT).

My take was that Instant Checkout would struggle because the sheer amount of sales needed to make their investment worth it was unlikely to be achieved. Just because shopping is available within an AI prompt does not mean consumers suddenly have more purchasing power. The bulk of sales volume was likely to be a share shift from existing e-commerce platforms like Shopify and Walmart. Not net new consumer spending.

But Walmart didn’t need to worry about their website losing volume to ChatGPT. They saw conversion rates roughly three times worse within Instant Checkout than on Walmart.com. No share shift of volume there.

I’m still marking this as green as I did call out that people were unlikely to immediately associate ChatGPT with “shopping”, which was observed in the lack of shopping intent among ChatGPT users, according to OpenAI.

Prediction #04: Strava will IPO in 2026 and become the next major social media platform.

Article: Hot Takes on Consumer Brands

Date: August 19^th, 2025

Score: In progress 🟡

Looking to diversify away from AI in your portfolio? Try out Strava*, potentially heading to public markets soon. In February, Strava submitted an S-1 form with the SEC, a legal precursor to a public stock offering.

I marked this take as yellow, as the groundwork appears to be laid for the take to be at least partially correct. When it comes to “the next major social media platform”, that remains to be seen. As of a few weeks ago, Strava claims to have over 195 million users, with 50 million monthly active users. The key statistic to anchor when comparing social media giants is monthly active users (MAU).

Now, the 50 million MAU figure is a far cry from the likes of mid-tier social media players like Pinterest (619 million) and Snap (946 million). If we draw the line of major platforms at Pinterest, then Strava is unlikely to hit the milestone even in five years.

But Strava certainly has tailwinds in its sails, with running more popular than ever. Also, Strava recently announced a substantial global expansion with localization in ten additional countries, along with new sports for activity tracking. I believe Strava is looking to be the social media platform for exercise.

*Not financial advice

Airports are essentially billboards for credit card companies. And trust me, this is no accident. Credit card companies understand psychology and consumer behavior like the back of their hand. They aim for the first card you pull out of your wallet on your vacation or business trip to be theirs. It’s a concept called “top of wallet”, referring to the card being literally at the top of your wallet, or in the most accessible spot for repeated use.

Credit card companies want to subtly nudge you to associate their card with travel because when you are traveling, you are likely to spend significantly more than you do in your daily life. “Top of wallet” quickly translates into “top of mind”.

Airports are critical to the credit card economics flywheel, helping acquire customers and kick it off through spending. Let’s dive into the three margins critical to credit card companies and how airports help start the flywheel and push it into high gear.

Note: We will focus on examining premium rewards credit cards for major players like American Express, Chase, and Capital One. Also, if you’re new here, I used to work in the credit card industry.

Account Margin

This is where customer acquisition costs sit. But thankfully for credit card companies, they’ve found a way to offset the costs.

For an annual fee*, you gain access to exclusive perks like airport lounge access, rewards points multipliers, and travel, dining, and subscription credits. Not just anyone is allowed into an American Express Platinum airport lounge. It’s an exclusive location reserved for people who are willing to fork over the steep annual fee (nearly $900 at the time of publication for the Platinum card).

In the context of airports, these lounges are an enticing offer if you are a frequent traveler and appreciate a cleaner, secluded waiting area away from the masses. Despite being costly to maintain, they pay for themselves and then some, as a strong nudge for new sign-ups.

When it comes to monthly and annual credits, credit card companies rarely pony up all the cash. They tend to be at least partially merchant-funded, with merchants willing to subsidize the cost of the credit for their product or service with the hope that you’ll become a recurring customer of theirs. This is especially true given that premium card members typically spend more than those with $0 annual fee cards.

Rewards points multipliers come at the cost of the credit card company. You know, the 4x multiplier on dining spend (i.e. if you spend $100 at dinner, you will receive 400 rewards points, usually redeemed at ~$0.01 per point). These typically are geared around dining and travel, which tend to be higher-spend moments for people. This is a play to increase spending volume, with issuers willing to accept lower margins as the rate versus volume trade-off will likely work in the credit card company’s favor.

*For cards that do not charge an annual fee, perks tend to be quite limited. Perhaps a rewards point multiplier (i.e. greater than 1x rewards points) on a single category, but rarely any meaningful travel or dining credits.

Share Relentlessly Curious

Spend Margin

Once you have been convinced to sign up for a fancy credit card so you can unwind at your airport lounge of choice, it’s time to get into the habit of spending. Thankfully, the credit card companies have already thought through how to get the spending flywheel going.

They offer big-time welcome bonuses. You know, the “get 100,000 rewards points”, and then in smaller font, “if you spend $6,000 within the first three months of membership.”

These welcome bonuses sit in the Account Margin but are necessary to create an urgency around using the card. And the required spending is steep for a reason. They don’t want you to be able to reach your spending threshold after a few swipes (remember Bilt?); they want you to form a habit of using your brand-new card.

Every time you use the card, the credit card company receives a small percentage of each transaction. This is an oversimplification, but this fee ranges from ~2-4% of the transaction depending on the size and category of the merchant. After paying out fees to various networks and partners, the “discount rate”, commonly used terminology within the industry regarding the spend fee percentage, nets out lower. Given that overall network volumes reach into the trillions of dollars for the likes of American Express and Chase, the spending revenue adds up quickly.

This is a core revenue driver for credit card companies: people swiping their cards to pay for goods and services.

Leave a comment

Lending Margin

So now that you are convinced to sign up for your grandiose premium rewards credit card at the airport and you’ve spent the amount needed to receive the welcome bonus, you are faced with a predicament. You cannot pay back the amount you owe in full.

They sure won’t admit this publicly (given the amount of legal scrutiny they already receive), but credit card companies are licking their lips when this scenario arises. They offer a minimum payment option, allowing you to pay your outstanding balance over time. But it comes at a steep cost: interest payments.

This is often the most profitable source of revenue for credit card companies. They borrow at relatively low rates from the capital markets and lend it back out to consumers and businesses at a punishingly high rate, otherwise referred to as an annual percentage rate (APR). In the case of American Express, they borrow at ~7% and charge APRs that stretch above 30%, depending on the prime rate.

Sure, the credit card companies take on the risk that you may not pay them back. For their lending business, they keep a certain amount of capital on the books in the case of charge-offs. But the interest income and the late payment fees more than offset the occasional instance that a card member stiffs them.

Subscribe now

See, credit card company economics work in a synchronous flywheel. They offer exclusive rewards and benefits to card members to incentivize them to sign up for the card. Then give a reason to spend immediately and frequently on their new card for the sake of unlocking their welcome bonus, which builds habit and association with the card. And finally, provide a credit line which can lead to overspending and the need to borrow. And there’s no better place to get the flywheel going than where people are primed to begin spending more than they usually do: at the start of their journey, or in other words, at the airport.

Side note: I understand I’m preaching, but whatever you do, avoid credit card debt at nearly all costs. Not only can interest rates creep up into the thirties, but also, what most people miss is that once you start revolving a balance, all additional purchases you make are subject to interest until you pay off the balance in full.

We are living in weird times. DoorDash is paying people to film themselves washing dishes.

Yes, you read that correctly. Last week, Bloomberg reported that DoorDash created an app for their “Taskers” to record themselves completing chores around the house, like dishwashing or folding clothes.

Why? It’s all about offline data collection.

As foundational model companies continually look to upgrade their AI, they’ll need new, niche datasets to maintain their edge. And believe it or not, there’s only so much data they can train on that is available online. Thus, they must look in a place no tech bro would have considered. The real, physical world.

Now, there’s plenty of niche datasets out there. So why is DoorDash offering compensation to those willing to complete everyday tasks on camera? Check out the quote below from the Bloomberg article.

“As an example of how the paid video submissions work, instructions for a dishwashing task ask that the person capture footage with a body-worn camera pointed down toward their hands, scrubbing and rinsing at least five dishes and holding each clean dish steady in frame for a few seconds before moving to the next dish. That camera footage may be valuable as robotics firms hone their humanoids’ ability to recognize objects.”

Companies building physical AI, ranging from autonomous vehicles to at-home robots, need to understand all the intricacies and variations of how people complete these everyday tasks to build a mechanical substitute.

It’s a bit dystopian. The gig economy workers may benefit in the short term by providing training data to DoorDash. But they’re training their robot replacements, as DoorDash may sell this data (they already have a partnership with Waymo) to robotics startups.

I believe 2026 is the year when the physical AI narrative starts to gain steam. Waymo has already completed over 20 million rides and is planning to launch in an additional 20 cities by year’s end. Also, you have Elon Musk shifting the storyline of Tesla from an electric vehicle manufacturer to a robotics company. Oh, and let’s not forget that the main highlight of NVIDIA’s 2026 GTC conference (some refer to it as the “Super Bowl for AI”) was innovation in physical AI that their semiconductor chips will enable.

Robotics outside the factory floor have always seemed far in the future. More so reserved for the plot of a sci-fi thriller. But as I went for a walk today in Manhattan, I strolled by a Waymo car in training. A person was in the car, but still, if an autonomous vehicle can figure out how to navigate traffic in New York City, ride-share drivers will need to find a different way to earn a living.

Share Relentlessly Curious

The world is changing. Candidly, I believe it will become quite normal to see someone filming themselves throughout the day because they are getting compensated by an AI company for doing it. AI valuations are through the roof, particularly on the private market side. Tech companies need to justify their valuations, and physical AI presents another avenue to live up to the hype. Yet, the clearest story is how robotics are likely to reduce the labor force. That’s a topic for another article.

Because tech firms need to live up to their current valuations, the demand for real-world data will continue to rise. The issue is on the supply side. It’s not easy to capture niche tasks at scale.

Which is why I believe that tech-enabled marketplaces employing the gig economy stand to reap significant rewards in the coming years as robotics companies look for training data.

In addition to delivery companies (i.e. DoorDash), marketplaces like Angi (formerly Angie’s List), Uber, and TaskRabbit are sitting on a treasure trove of data waiting to be unlocked. Looking to create a robot that mounts televisions? You’re going to want to strap a camera on a TaskRabbit TV mounting specialist. Or have a contractor on Angi film themself cleaning a pool and checking the chlorine levels.

Private equity is famous for rolling up* service firms like plumbers, electricians, and HVAC companies for the sake of streamlining processes and adding modern technology across the group. It’s become such a popular investment strategy that there are plenty of memes about the thesis. The joke here is that, believe it or not, running a real business isn’t as easy as the math makes it out to be in a spreadsheet. But perhaps this business idea has a new runway. There’s opportunity to roll up a business within a particular trade, record how each employee completes the task at hand, and then sell that data to a robotics startup.

For example, try landscaping companies. They complete both routine (lawnmowing, fertilization) and nuanced tasks like tree cutting, irrigation system installations, and patio building. I imagine within the next five years, we’ll see the first widely adopted autonomous lawnmower. But how could one build an autonomous lawnmower? They first must acquire a ton of specialized data on how people cut lawns today.

*By rolling up, I am referring to purchasing several similar companies. The idea is that by operating, say, 20 electrician businesses, there are operational efficiencies to be had with scale and those will lead to more effective operations.

Leave a comment

Fast forward. Thanks to people willing to film themselves folding laundry, robots are here and ready to complete your chores. But do people really want to delegate their daily tasks to a non-human object?

It’ll be interesting to see how consumer adoption of physical AI plays out. I believe consumer adoption will lag, as people are scared by the narrative that Hollywood has portrayed about robots for decades. A robot living in your home and washing the dishes is likely many years past what any tech guru predicts. But then again, Waymo has done an excellent job with branding around autonomous driving. They claim that their cars are involved in “92% fewer crashes causing serious injuries or worse than human drivers.” Perhaps people will warm up to the robot family member sooner than I think.

But in the short-to-medium term, the opportunity for investing or making some extra cash is in the acquisition of offline data. If you told me I could get paid to do the dishes by DoorDash, if I filmed myself doing it, I would have been happy to clear the table as a kid. Sorry about that mom!

Last week, Google announced a major update to Google Maps where AI is being introduced into the core experience. By leveraging Google Maps data, which includes hundreds of millions of places and user reviews, you can now receive hyperlocal recommendations for your next trip or night out directly within Google Maps.

I believe this paragraph from the announcement sums up the new value proposition quite well:

“Your results are personalized based on things like places you’ve searched for or saved in Maps to help you get the most relevant recommendations. So when you ask, “My friends are coming from Midtown East to meet me after work. Any spots with a cozy aesthetic and a table for 4 at 7 tonight?” Ask Maps already knows you like vegan restaurants and finds convenient midway spots with vegan options.”

Personalized, niche, and informed. The makings of an ideal recommendation engine.

Before we dive into what this update means for consumers and businesses, let’s backtrack to discuss the problem that I believe Google is attempting to solve.

AI works best for deterministic problems, or in other words, outcomes that can be reasonably predicted. Sure, there is a long tail of possible questions a customer could ask an AI chatbot (like this), but most questions have something to do with a company’s assortment, return policy, or shipping delays, based on the data I see in e-commerce customer service reports. AI trained on a company’s brand voice and tone, as well as historical customer service conversations and order history, can act as an adequate facsimile for a human customer service representative.

That’s why billions of dollars have flowed into companies developing AI customer service agents. You know, startups like Fin AI, Decagon, and EliseAI. They provide enterprise firms with an option to rely on AI for their customer touchpoints across email, voice, and chat instead of relying on people.

But what about B2C, or consumer AI customer service? Now, you’re looking to solve a probabilistic scenario, where there could be many different answers based on slight variations of the question. Questions about any business or place have a much longer tail than questions that could be realistically asked of a singular business.

Last year, Google quietly shared a feature that lets Google call businesses on your behalf. Maybe you want to ensure the local shore store has your size before running over. Delegate that task to Google (assuming the store’s phone number is online).

This is inherently an example of consumer-facing AI customer service (and an example of agentic voice AI). I believe that Google is looking for niche, offline data to leverage to improve their foundational models and I suspect they launched this product to acquire phone transcript data.

Google, OpenAI, and Anthropic will continually need more data to enhance their models. And eventually, they may see diminishing returns from internet data: they’ll need to find a way to harness real-world interactions and knowledge that only exists in the physical world.

Google’s quest to acquire offline data will lead them to back into creating consumer-facing AI customer service: a highly personalized (and action-oriented) recommendation engine that allows people to enhance their offline experiences with their online presence. Google Maps is an excellent venue for this goal given the sheer amount of user contributions (over 500 million contributors, according to Google), and omnipresence of the product.

Today, Google Maps is a transportation app with crowd-sourced reviews of businesses and places. There’s a massive, untapped opportunity to turn people’s shared maps into a social feed. However, I think Google is more focused on acquiring as much data as they can, particularly data that their rivals wouldn’t be able to capture. In doing so, they are creating AI customer service for consumers that can act on behalf of people (i.e., phone calls) and provide hyperlocal, personal recommendations regardless of their location in the world.

Google states quite plainly, “combining our Gemini models with our deep understanding of the world unlocks entirely new possibilities”, which I believe nods to Google’s interest in further enhancing its offline data collection.

Think about this scenario: you’re on a road trip and you need to make a pit stop. To fill up your gas tank and use the restroom.

“Hey Gemini, find me a gas station within 15 miles that has a restroom.”

Gemini will quickly learn people’s common driving routes and the reasons why they make the stops they do. They’ll combine user reviews with prompt intent to better understand why people choose to go to one restaurant over another.

Share Relentlessly Curious

So, Google Maps gets AI, and it’s a massive data play to learn more about how and why people travel to certain locations across the globe. What’s the business implication?

Generative Engine Optimization (GEO) matters for all businesses. Regardless of size.

Apparently, Google Maps has over two billion monthly users, which makes it hard to comprehend how much influence Google has with this product . At those numbers, AI isn’t just for people who sit behind a desk all day. It’s for everyone.

If you’re a local pizzeria that relies on foot traffic in a busy part of town, you’ll probably still be fine ignoring GEO. But if you rely on customers looking up non-branded keywords in Google Maps (“pizza near me”) to find your spot, you’re going to want to make sure you show up in AI search results because Google Maps now provides AI-generated output. This leads into my core argument: since Google Maps will leverage Gemini, optimizing your presence in Gemini will translate into more frequent and favorable recommendations within Google Maps.

Optimizing the pizzeria’s web presence for AI search results (particularly for Gemini results) will be critical for small businesses thanks to Google adding AI into Google Maps. And this goes for more than just pizzerias. Car dealerships, mechanics, and gas stations will all need to consider how people find out about their business in the first place. Adding content like a frequently asked question (FAQ) guide, directions to the dealership on major highways, or that the gas station has a public restroom, are examples of content likely to perform well on an AI prompt within Google Maps.

If you’re building in the GEO space, please reach out. I’d like to chat more about the implications this Google Maps update has for small businesses.

Either way, consumers are the real winners of AI making its way to Google Maps. As hyperscalers compete to win the AI race, we benefit from their innovation. I look forward to using the new features on my next road trip.

Hot Take #1: AI may eliminate the apprenticeship aspect of work.

Two weeks ago, Block announced that they were laying off ~40% of their workforce. Not due to poor business performance. Instead, CEO Jack Dorsey shared that they simply did not need as many people because of AI.

His tweet must have sent a chill down the spine of any corporate middle manager pushing paper every day. Executives likely know where they can trim their headcount, and if the market rewards layoffs under the guise of “AI”, their stock price (and their compensation) will benefit tremendously. Note: Block’s stock surged following the announcement, helped by strong quarterly results.

I’ve seen it in my day-to-day: AI makes me more effective and more efficient. Tasks that used to take me hours a few years ago now take minutes. And I’m not the one doing them anymore. I do agree that companies can be more productive with fewer people through well-designed systems of AI agents.

Although I’m bullish about AI’s impact on effectiveness and efficiency in the workforce, I am wary of what happens when recent college graduates enter and are told to do everything with AI.

AI works best when you start with the end in mind and work backwards to describe what you want and how you expect to get there. For instance, if you’ve been working in finance for years, you have a strong understanding of what it takes to build a financial model, slide deck, or research report without AI. You understand the workflows and have a strong sense of “what good looks like.”

Imagine it’s the first day of your career and you’re told to use AI to build a financial model valuing a skincare manufacturing plant. You write out some prompts that a coworker gave you and, boom, you receive a financial model.

How do you check the outputs? How do you know what to prompt back? How do you know the problems that you need to solve? What are you learning and how do you get better at your job if all you know is AI? How do you design processes for automation if you don’t understand the intricacies of the process?

AI allows you to become more efficient once you’ve written an effective prompt or system. And that comes from putting in the time and effort upfront to learn the craft.

Hot Take #2: The core tenet of education will need to be communication. Both with people and AI.

Much of my K-12 education centered around memorizing facts. From history class to science, I spent much of my time memorizing elements on the periodic table, world wars, and mathematical proofs.

With where AI is headed, the value of a memorization-based education comes into question. If we can simply ask AI and receive an answer in seconds, what’s the point of spending all this time memorizing much of anything?

Of course, there’s some baseline education that needs to occur. But perhaps our secondary education needs to center around how to get the most out of AI. Which starts with how to communicate with it.

This comes in the form of first principles thinking, systems design, and sequential workflows. Instead of entering the workforce and being a cog in a machine, people should be prepared to become orchestrators of workflows. This starts with understanding what problems must be solved, breaking the problem into tactical steps, and then building automated point solutions that feed into a larger system. Everyone will have the agency to solve problems at scale thanks to AI. They might as well take advantage of learning how to do so.

Learn how to think. Learn how to speak. Learn how to write. In one phrase: learn how to communicate (which requires forming thoughts and then extracting them from your brain). Whether it’s with AI or people.

In a world where memorization is no longer an advantage, the alpha is in your ability to communicate. Structured thought processes will allow you to build with AI, as well as share your ideas with others to gain buy-in. Strong, cohesive writing cuts through the noise of AI-generated slop. Articulate dialogue and the ability to strike up a conversation with anyone will be critical for differentiation as people naturally rely more heavily on machines and less on their ability to socialize with others.

Share Relentlessly Curious

Hot Take #3: OpenAI’s slow start to agentic commerce is more of a supply problem than a demand issue.

Last week, The Information reported that OpenAI would roll back its ambitions around shopping within ChatGPT.

Need a refresher on OpenAI’s agentic commerce initiative? Check out #61: The Case Against Instant Checkout.

In the fall, OpenAI and Shopify drummed up a lot of press about how their partnership to bring shopping to ChatGPT was the future of commerce. They were building the backend infrastructure to support both humans buying products within AI chatbots and AI agents making purchases on behalf of humans.

Since then, OpenAI has been slow out of the gates in adding brands to the platform, and shopping-related inquiries often end up being glorified Google searches*. They likely bit off more than they could chew with the sheer number of challenges around payment guardrails, fraud, security, brand inventory, and overall commerce awareness. With so many competing priorities, OpenAI appears to be slowing down and focusing more on advertising based on recent news releases.

*Next time you ask ChatGPT a shopping-related question, look closely at how it searches the web to answer your question. This post explains very well how ChatGPT shopping is essentially a wrapper on Google Shopping.

Despite being a massive prize (trillion-dollar industry), commerce is incredibly difficult to crack. Especially if you’re trying to reinvent how people shop on the internet through agentic commerce.

I’ve written about my bearish sentiment on ChatGPT shopping (i.e. Instant Checkout) because people use ChatGPT for a whole host of tasks, in addition to shopping research (check out here for more). This take describes a demand problem, where users are not in a state where they are willing to make a purchase while using ChatGPT.

After reading through The Information’s scoop, my opinion has swayed towards slow adoption because of a supply problem. There were few brands on the platform ready to go.

Promising that any brand using Shopify as its e-commerce platform would be included in agentic commerce created lofty assumptions about the level of brand supply available to shoppers.

But that has not come to fruition (likely due to how complicated agentic commerce is to begin with). Anecdotally, I have not come across any products I could buy directly within ChatGPT during my day-to-day shopping research. Even large brands like Nike and Timberland required you to check out on their website. You’d think OpenAI would have gotten the bigger names on the platform first.

If you don’t have a wide assortment, people won’t associate ChatGPT with shopping. You might as well go to Amazon if you cannot guarantee that a product will be on ChatGPT.

But I believe the market overreacted. Sure, there are fair questions about the go-forward margin profile of a software business in the age of AI. However, the key advantages of deep entrenchment within customer workflows and partnerships remain.

Companies like Adobe and Figma have their products deeply embedded in their customers’ workflows. Job descriptions across the creative industry frequently require proficiency in products like Adobe or Figma because they are the default tools. AI isn’t going to take that away overnight.

Another point: if a Salesforce enterprise sales representative takes their Fortune 500 CTO client golfing each month, the CTO doesn’t suddenly want to stop golfing just because AI is here.

Through their deep client relationships, enterprise software companies have built-in distribution. They can also introduce AI (which they are) into their current offerings to satisfy evolving client needs.

Now, who else has suffered a stock price drawdown, yet has a major distribution advantage? Consulting firms. Check out Accenture’s stock, down 20% year-to-date according to Yahoo Finance. I’m sure OpenAI agrees with this.

Last week, CNBC reported OpenAI announced partnerships with major consulting firms Accenture, Boston Consulting Group (BCG), Capgemini, and McKinsey to help OpenAI’s enterprise clients accelerate AI adoption. The gist is, the consulting firms will help OpenAI’s clients get set up and running with AI in the way OpenAI intends.

If you want to learn about a company’s longer-term strategy, check out their careers page. OpenAI is hiring for hundreds of AI customer success, deployment, and partnerships roles that all appear to be client-facing. In tandem with the news article I linked above, they posted a role for a Global Partner Director, whose key responsibility is to scale their partnership with McKinsey. Although OpenAI is often perceived as a consumer-first AI company while Anthropic is often viewed as more B2B-focused, clearly OpenAI is trying to flip the script. With 40% of revenue coming from enterprise customers, they certainly have a point.

On the surface, this partnership is a win-win situation, primarily about speed. OpenAI is staffing up to support their burgeoning B2B vertical and wants to reduce the time it takes to effectively integrate OpenAI technology into their customers’ day-to-day workflows. They can focus on building the technology, and the consultants can ensure it gets used properly. Custom enterprise AI arrangements require deep knowledge of a customer’s ways of working, guardrails, and regulatory environment. There is likely significant enterprise client overlap between the consultants and OpenAI, meaning that the consultants already know how each company operates. This is a great example of each company leaning into its competitive advantage.

In a world where the return on investment (ROI) of AI is beginning to be questioned (i.e. Google and Amazon’s stock price declines following announcements of major AI capital expenditures for 2026), it’s critical to make sure that OpenAI clients are properly implementing AI and setting themselves up to see a strong ROI on the investment.

Remember the NANDA MIT report that claimed “95% of GenAI pilots are failing” from August? Regardless of how misleading that report was, OpenAI can’t take their chances that clients are suddenly going to become tech-savvy and know where and how to implement their new enterprise plan into the business. So many big company CEOs are relieved when they can announce an enterprise contract with OpenAI because that signals that they “have an AI strategy”. But it’s not going to be a good look for either party if by the second contract year there’s been very little adoption nor impact.

Share Relentlessly Curious

So once again, the consultants win. Same set of customers (i.e. enterprise) but selling a new product: the promise of revolutionary technology.

Still, consultants face similar existential risks to their business model as software does alter how work is priced. AI certainly compresses the number of billable hours needed to complete a competitive landscape research report. And clients know this.

Even so, consultants are doing just fine financially when it comes to AI consulting. Ernst & Young saw 4% revenue growth in 2025, yet their AI-related consulting revenues jumped 30% versus the prior year. According to the New York Times, McKinsey expected its 2025 revenue to be 40% attributed to gen-AI projects. That’s billions of dollars.

The continued dominance by consulting firms to move onto the next thing to sell is no surprise. There’s so much hysteria and anxiety in the job market today about job losses (i.e., the Citrini Research report), what much of the gloomy analysis misses is the relationships forged over decades. If a Bain & Co partner has been a close confidant of the Coca-Cola CEO over the past 20 years, the working relationship isn’t going to evaporate overnight because AI is here. The advice will change, but the relationship isn’t subject to automation. There’s plenty to be worried about in this economy, but if certain areas are slower to change, it’s those where relationships serve as the anchor of distribution. Contracts are a lot stickier than you think in the enterprise client services space.

Hopefully their distribution advantage ends up leading to increased economic productivity. Consulting firms can add a lot of value to their clients on the basis of this partnership by helping clients establish their AI fluency baseline. So many companies don’t have consistent standards on what good looks like for AI adoption. Since the consulting firms will have a front-row seat for the latest AI updates as well as a strong understanding of their clients’ operations, they can craft a set of standards from which that organization can improve. Once the standards are set and adopted, the consultants can work together on establishing ROI baselines for assessing AI-related investments. As a result, a more sustainable environment will be created around AI, helping to find signal through the noise.

Shumer’s post on the future of life and work in the age of AI has reached over 80 million views, and he has since appeared on CNN and CBS News.

He makes some bold, dystopian predictions about how quickly AI is developing and the ramifications of its potential to fundamentally disrupt the job market, regardless of industry. Essentially, humans won’t be needed because AI can do the work for them.

I suspect many people will read this and feel confused. Where I believe Shumer missed is not diving deeper into why AI has become so capable. And that’s what I’d like to unpack today: the concept of agentic AI.

Different Types of AI

If you’re thinking, “I use ChatGPT and it sometimes gives me the wrong answer. How will this displace my job?”, I get it. I frequently receive fishy answers from AI chatbots like ChatGPT, Gemini, or Claude, and must regularly question their logic too.

But chatbots are simply the tip of the AI iceberg. The AI that Shumer is referring to isn’t the free version of ChatGPT (in fact, he specifically urges readers to try out a paid AI subscription). Like most services, the premium version is behind a paywall. The same goes for AI.

Side note: Premium AI tools are not expensive today. For a little more than the cost of a fast-casual lunch in Manhattan, about $20, you can see for yourself why experts feel uncertain about what the future will bring. Whether it’s ChatGPT Plus, Claude Pro, or Google AI Plus, pay $20 per month and see what you’re missing.

What’s the difference between “AI” and agentic AI?

Here’s how I’d explain it. What I’ll refer to as “regular AI” is interacting with an AI chatbot like ChatGPT or Gemini. You ask it questions, request tasks, and go back and forth in conversation. This is what most people’s experience with AI has been to date.

Yet over time, you notice the AI chatbot starts to forget key details from earlier in the conversation. That’s the limitation of a context window in effect. There’s a preset limit on what the AI chatbot can remember. So, you open a new session and remind it about what you had previously told it. AI chatbots are, in a lot of ways, like glorified search engines that can complete singular tasks. As you could have guessed, higher-tier subscriptions generally provide larger context windows and better persistent memory features.

Now, agentic AI executes multi-stage workflows. You can provide agentic AI goals and translate those goals into actions for the AI to take, which it can then run autonomously.

Regular AI is reactive. Agentic AI is proactive. Regular AI handles one task at a time. Agentic AI can handle many tasks, including those that depend upon another’s output.

Here’s what I mean.

AI chatbot: Summarize this article for me.

Agentic AI: You are an expert at evaluating newsletter platforms. Research the top three competitors to Substack, analyze each competitor’s pricing model, draft a pro and con list on each competitor, and then provide a recommendation with 80% confidence in a five-slide presentation deck on which competitor is the best fit for Relentlessly Curious.

In say an hour (if not sooner), you’ll have an output that would normally take a junior team member a week to complete.

I believe agentic AI is the future of work. We already have the capability to create AI “employees” by creating an agent to handle specific tasks. We (humans) will serve in the role of agent orchestrators, managing many agents and helping instruct the bigger picture of the project goal.

Share Relentlessly Curious

Agent Orchestration

Check out the agent orchestration for the newsletter platform research project mentioned above.

Research Agent: Educated on first-principles thinking and decision-making frameworks, this agent searches the web and any separately provided data source to distill a curated viewpoint on the topic you’ve requested.

Price Scraper Agent: Specifically instructed to find the current prices of pre-set competitors at a pre-defined time cadence.

Slide Deck Agent: Provided a presentation structure, brand style, and writer’s voice, this agent can build presentations for you based on the data you provide it.

As the user of AI, you can instruct the Research agent and Price Scraper agent to work in parallel and then once both agents’ work is finished, request the output be used by the Slide Deck agent to create the presentation.

You don’t need to know how to code to create the above agents. By instructing AI in natural language prompting, you are simply writing instructions similarly to how you’d teach a junior employee. And thanks to emerging standards like Model Context Protocol (MCP) which aim to connect models to external tools and data sources, agentic AI can access information across the data sources that you feed it, including your Google Shared Drive, Slack, and even Shopify account. If your organization has a written document or chart on best practices for data analysis, you can tell your AI agent to reference this document, and it will factor that into its process.

This scenario is why there’s growing anxiety in the job market. You don’t need to be an engineer to build an AI agent, which democratizes software development and puts the power in the hands of those willing to be curious.

I’m oversimplifying it, but the key point with agentic AI is that it unlocks multi-dimensional workflows typically executed upon by humans. For those of you that are looking to get ahead of the curve and understand what current AI tools allow you to do, I recommend checking out some of the following tools.

Custom GPTs in ChatGPT Plus

Custom GPTs in ChatGPT Plus are not considered agentic AI, but they are still a strong starting point. A much more robust version of an AI chatbot, Custom GPTs allow users to store information in the way of written context, files, external API integrations, and model preference. No need to re-provide baseline context each time you log in.

I encourage this as your agentic AI launch pad because it teaches you to create a knowledge base for AI to regularly pull from. Examples of Custom GPTs I’ve created include written document editors, analytical thought framework partners, and social media copy generators.

Claude Code in Claude Pro

This is my favorite agentic AI tool, and I’d liken it to magic. I’ve been increasingly relying on Claude Code in my day-to-day. It’s worth calling out that I am by no means an engineer. Despite its name including “Code” you don’t need to know how to code to effectively use the tool. I provide access to my computer files and third-party connections, as well as instructions on tasks I’d like to complete. Claude Code handles the rest. I’ll refine my prompting over time and frequently push Claude to question its own logic. But the magical thing is that this agentic AI tool can reason on its own, pressure test its assumptions, and correct itself when it senses an error, all with relatively little oversight.

I will admit, there is a bit of a learning curve with how you set up Claude Code given that you generally access it through your terminal. I recommend checking out the training program, Claude Code Course for Product Managers, to get started. I know it is meant for Product Managers, but it teaches Claude Code basics and prompting best practices in a first-principles framework. This is important for any knowledge worker. Oh, and it’s free!

I understand this may feel overwhelming, but rest assured, you’re not behind. I’d like to share this chart from a LinkedIn post I stumbled across. Most people haven’t used AI before. And only a microscopic percentage are using AI at a high level. All you need to do is be curious and willing to explore!