About 18 months ago, I tried to create lifestyle images for an e-commerce product listing using AI. I used Flux models from Black Forest, first providing reference images and then wrote a prompt of the variations I wanted to see.

It… kind of worked? It was for a baby hooded towel (a product we sold at my old job), and I wanted to showcase it being worn by a kid at sunset. I shared a few pictures of the hooded towel on its own from different angles and configured a few settings on Hugging Face (the platform I was accessing Flux through). And boom, I had a realistic picture of a kid wearing the hooded towel with a sunset in the background. The problem was that the kid only had four fingers.

Fast-forward to now. AI-generated images have become strikingly similar to real world images. It can be difficult to tell the difference between images (and content) generated by humans and AI. Fewer missing fingers, and a heck of a lot better image quality.

But the line blurs between who did what, between human-created and AI-generated content. Where this particularly creates problems is on social media.

Social media is already flooded with images and videos that misrepresent the lifestyles people portray or the experiences they claim to have. It’s not hard to think that everyone is going on that luxurious vacation or driving that fancy car based on the videos shown on your feed. It was misleading before AI and certainly is even more confusing now.

Enter Google’s SynthID Detector. This new technology places invisible watermarks on AI-generated content, including images, videos, and text. It not only preserves content quality, but also (as per the product launch announcement) “remains detectable even when the content is shared or undergoes a range of transformations.” Simply put, it can’t be easily altered or tweaked to remove the unique identifier.

Google is beginning to roll out the detector, which allows users to upload content that was created using Google’s AI models or tools. Then, the detector indicators whether a SynthID watermark is present or not, as well as which portions of the content are more likely to be AI-generated. Google partnered with NVIDIA to create this open-source technology. Important point as Google could have gated this product behind their Gemini brand of models.

In August, I wrote about the EU AI act that established requirements for all LLM builders to add unique identifiers to their output. This “watermark” was meant to tag any content generated by AI models (like Claude) so users could see. I broke down this story from the branding opportunity angle, but today, I believe it’s the safety protocol needed to quell AI safety fears.

I believe watermarking AI-generated content is a monumental step towards building trust in AI. There is plenty of fear-mongering in the news, with some notable accounts coming from ex-frontier AI labs employees. Consider this group of former OpenAI safety researchers or Jacob Coxon, a former Anthropic employee who published an extensive post on X sharing his concerns about how dangerous AI may become.

But what if we could identify AI agents and AI-generated content for what they are?

Share Relentlessly Curious

Watermarks for AI can be viewed as IDs and passports for people. When we travel internationally, we carry passports to verify our identities and establish our eligibility to enter another country. We use driver’s licenses to verify that we are legally allowed to drive a car. Identity verification has been around for decades. The concept itself is nothing new.

As AI gets better at replicating human actions and decision-making, we need better ways to identify it. Watermarks give us another tool to check whether a video or email was generated by AI. They won’t tell us how much human involvement went into creating something, but they do give us a useful signal about where it came from.

Of course, this only works if people have access to detection tools and actually use them. Still, making it easier for anyone to check content for AI-generated material is a meaningful first step towards helping people make more informed decisions about what they see online.

It would be wise for the frontier AI labs to form a coalition for embedding and detecting watermarks across models and content types. Essentially, the industry needs its own version of SynthID, ideally built on open-source technology so that any company can adopt and improve it.

However, I’m not naive. This will be a difficult task to take on.

Getting Google, Anthropic, OpenAI, and SpaceXAI in a room to share technical requirements on how all of their models can adhere to the same watermark identification standard will be tough. These companies are unlikely to share their secret sauce, which may surface during the creation of these standards. But a common standard could make AI-generated content easier to identify, regardless of which model created it.

You know who could benefit from better AI content detection? Meta.

Think about the volume of AI-generated slop posted across Instagram and Facebook. Beyond low-quality content, bad actors are able to manipulate people’s behavior and beliefs with hyper-realistic AI. Better tools for identifying AI-generated content could give platforms another way to understand what is being posted and who, or what, is behind it.

Meta has been clobbered in the news and in the court room on its lack of safety controls for young people. Recent reporting on a major social media litigation settlement ($17 billion!) adds to that pressure. The Social Reckoning movie isn’t doing Mark Zuckerberg’s public image any favors either. But at least in Zuck’s case, its first weekend in the box office was a flop.

Watermarks can be one piece of the puzzle in the broader effort to make AI more transparent. They can help platforms identify AI-generated content, while a whole host of other technologies could authenticate AI agents, track their actions, and establish what they are allowed to do. There’s a whole AI compliance and safety vertical that can be born out of protecting and surfacing watermarks.

Google’s actions represent meaningful progress towards AI safety. We should build guardrails that check for AI-related actions and content generation. This allows platforms and websites to better understand the actors behaving in their arena and set up controls on how to best interact with them, as well as keep their actions in check.

It’s October now, so throw on a comfy sweater and grab your pumpkin spice latte before you settle into this week’s Relentlessly Curious. There’s no better way to kick off Q4 than with a Hot Takes in AI edition. I haven’t done one like this in a while, and there’s been plenty to talk about in the AI world. Buckle up and let’s get going.

Hot Take #1: OpenAI is in no position to be a public company, given its lack of control over its agents is playing out like a Rick and Morty episode: people (or in this case agents) showing up in random places that they aren’t expected to be.

It’s getting laughable at this point. OpenAI now shows a long list of issues its agents have caused on a page called Misalignment Reports and Notices. It’s a weird cross between an admission of guilt and a brag list. I believe it’s both, an output of the fear-based marketing strategy that the frontier AI labs have been deploying for quite some time.

Let me break this down a bit more. OpenAI (along with other frontier AI labs) runs various tests in closed environments called “sandboxes”. These tests typically consist of creating an AI agent, providing context on the task you’d like it to execute, and assessing its performance. The classic sandbox prompt is, “hey agent, solve this really difficult math problem.” I’m oversimplifying it, but you get the gist.

As the frontier models have become more capable, the agents have gotten more creative and powerful at how they attempt to solve the problem they’re given. And by creative and powerful, I mean that they occasionally go rogue (during evaluations) and obtain unauthorized access to both private and public databases to find the information needed to complete the task they’ve been given.

In the case of the Hugging Face hack, OpenAI agents, including GPT-5.6 Sol, broke into Hugging Face’s systems, stole data, and performed a number of other unauthorized actions, all in pursuit of solving the problem they were given in a test environment by OpenAI researchers.

This reminds me of a Rick and Morty scene, where characters are zooming between planets, locations, and periods of time. Usually, these visits aren’t expected or appreciated by the people or creatures they encounter. OpenAI’s agents are breaking out of their sandbox and finding ways into sensitive data sources.

If the Hugging Face incident wasn’t enough to slow down OpenAI’s tests and ever-capable agents, it’s been reported (originally by the New York Times) that AI agents have gone rogue tens of thousands of times causing security incidents.

I’ll hold off on a spiel about the downside risks of AI as the mainstream media is doing more than enough covering the AI agent hacks across the industry. Anthropic and Google have also had their moment in the spotlight when it comes to hacks. But I will call out that it’s ridiculous to think that OpenAI can realistically meet the expectations of being a public company. Public companies are required to report their financials every quarter and it’s extremely difficult to predict the potential fallout and liability associated with OpenAI’s rogue agents as of today. Who knows if one day the agents cause serious damage or data leakage that could expose OpenAI to tremendous liability?

What will a good earnings report look like? “Hey we lost an obscene amount of money, but at least our agents aren’t cosplaying Rick and Morty as often as they were last quarter”.

Share Relentlessly Curious

Hot Take #2: DoorDash’s recent “text” capability product launch provides users with an excuse not to open the app, which is a risky move.

Back in my days as a brand operator, I took demos from half a dozen software vendors that claimed they could increase sales through personalized text messaging. The pitch was essentially, “our personalization algorithm sends the messaging most likely to lead to a customer (or potential customer) making a purchase”. Candidly, I was never impressed as I was quite skeptical of how “personalized” the text messaging was. Plus, I had concerns about how much brand messaging could be controlled if the text really was “hyper-personalized”.

Anyways, DoorDash is bringing their in-app assistant Ask DoorDash within your text chat. If clicking the DoorDash app on your phone was too much effort, you’re in luck as you can text ambiguous phrases like “order my usual protein bowl to the office,” and you’ll receive localized recommendations, a suggested choice, and the ability to place an order directly within the text thread.

There’s something different about texting with DoorDash than texting with some GPT wrapper that a CPG brand uses to chat with SMS (text) subscribers that’s apparently trained on the brand’s voice. I don’t expect a CPG brand to understand how to pull off conversational commerce, so it feels inauthentic. But a major technology player like DoorDash? This feels more in their wheelhouse.

I am perplexed that DoorDash is rolling out this capability though. When a user is within their app, DoorDash collects a ton of data on how users browse, shop, and transact. They also provide restaurants the opportunity to be front-and-center on the app through their paid placement program (i.e. advertising).

My guess is they are launching this feature on the back of Muse and Instinct’s growing momentum. More people are getting used to the consumer agent environment (communicating through messaging) and DoorDash wants to cement their own pseudo-agent environment.

Ultimately though, DoorDash makes great money from its advertising business (not disclosed separately in financials, but DoorDash announced in 2024 that DoorDash and Wolt Ads had crossed a $1 billion annualized advertising revenue run rate). Providing users with a way to avoid the app could mean fewer people viewing and clicking on restaurant ads. See, advertising is typically a high-margin revenue stream so unless DoorDash starts finding a way to introduce sponsored placements via text (this sounds incredibly clunky), I have to think this product launch has some downsides. Curious to see how this plays out, as the incremental orders via text will ultimately have to offset any potential loss in advertising revenue.

If I were to start a new company today, I’d follow this marketing playbook. To reach people online, I’d set up Meta and Google ads to show up across social media and search. Need to show up during someone’s daily doomscroll ritual.

Next, because it would be too expensive to acquire every customer through advertising, I’d make an organic content push. Next up, I’d like to start up a blog to generate some SEO goodness. Then tweak the webpages to better align with GEO best practices, so the company starts showing up in ChatGPT responses. Furthermore, since I have social media covered and the AI presence is starting to improve, what should I do next?

I’ll dump my entire marketing budget toward branding on a cold brew coffee can for some free coffee shop to hand out.

On September 23^rd, Free Bean opened a coffee shop in Manhattan, which gives free coffee to customers. How does this work? The coffee shop is entirely funded by advertisers, so customers pay nothing.

Free coffee in NYC. Yes, you read that right. But it’s not really free.

How does this work? Companies line up to put their branding on a coffee can (cold brew, vanilla latte, or mocha latte). When you are handed a can, you’re given a company-wrapped coffee at random, meaning you don’t get to choose the sponsor of your cup of joe. Each can has sleek branding designed by the sponsor, with a corresponding QR code to learn more (which can help with engagement tracking).

Why am I excited by this? Because Free Bean isn’t really a new advertising channel. It’s an attempt to turn an existing consumer behavior into a new media channel.

Some view coffee as the event, but in New York, coffee is the transition point, going from point A to point B. Usually going from home to the office. Or maybe it’s a ritualized coffee run in between meetings. The point is, it’s typically embedded with another activity, like commuting. With coffee prices continuing to rise (it’s tough to find a small, black hot coffee for less than $4, my go-to order), why wouldn’t you check out the coffee shop that gives you free coffee? It saves you a few bucks, if not nearly the cost of lunch when you get a latte with bells and whistles. And for the advertiser? They aren’t really buying space on a coffee can. They’re buying a few minutes of someone’s morning.

There are plenty of coffee drinkers in the world. And it’s likely some of those coffee drinkers could buy your company’s software (especially if the coffee shop is in Midtown Manhattan). From what I experienced (I braved the Nor’easter storm to check out what all of the fuss was about), the current Free Bean sponsors are mostly tech companies. Many of these companies sell high-ticket B2B software, which can make unconventional customer acquisition experiments more economically viable.

So, you take a relatively universal product that is coffee (more than 1 billion people consume coffee globally). Can it. Then get a deep-pocketed tech unicorn to hand over a test budget for who knows how much money. And then the coffee shop becomes the distribution channel. It’s easier than ever to create a product thanks to AI. It’s harder than ever to build up a distribution moat due to all the competing sources of attention. As AI makes digital production and distribution increasingly abundant, marketers will increasingly compete for scarce moments of human attention. Free Bean is an example of turning an existing consumer routine into an advertising platform.

Share Relentlessly Curious

Free Bean is a case study in breaking into a new form of marketing. It’s as if the monetization potential of online attention is reaching diminishing returns and a new medium to grab attention has presented itself. Branding for coffee tends to be reserved for, wait for it, companies that sell primarily coffee. Not tech companies.

I’m curious about what point in a marketer’s journey makes sense to do an above-the-line campaign like a Free Bean partnership. Once you see diminishing returns on paid social and search, GEO taps out, and there are only so many commercial spots and conference sponsorships you can line up. Where in the marketer’s journey does it make sense to bankroll a coffee shop that puts your logo on its product?

In some ways, this type of marketing seems the closest to professional sports teams’ jersey sponsorships, but for corporate workers. It’s a (relatively) mass-market play, one that you hope carries on in other forms of media. For the Free Bean sponsors, they are likely using this in-store advertisement as content for social media campaigns (I can’t imagine the number of times Ramp has already posted on LinkedIn about their sponsored coffee).

The decision point becomes: if I have a product that can appeal to a broad audience, what are new, creative ways I can reach them in a format that is routine and natural to them.

The most effective companies in the world have a knack for capturing your time. You stare at your iPhone for hours each day. When you have a question, you turn to Google (or ChatGPT and Claude now). When you want entertainment, you go to Instagram or TikTok for a stream of personalized content (while staring through the glass of an Apple product). So much of people’s waking hours are already captured by Big Tech, so where do you go next?

You catch them at the smaller moments. Maybe someone likes to listen to music or a podcast when they cook dinner, so you advertise with Spotify or your target market’s favorite podcast. Maybe you will buy billboard space to capture someone’s attention on their commute home. Instead of thinking about which platforms to reach your target market, think more about which routines your audience repeats every day.

What Free Bean demonstrates is how saturated the digital marketing landscape is and how the incremental dollar may increasingly be spent on in-person experiences. These large tech companies that don their branding on Free Bean cans are spending plenty of money. Companies like Ramp, Vercel, and Warp sell relatively niche B2B products, making a traditional mass-market advertising channel like a sports arena less obviously efficient. So they target a different stadium: a coffee shop in an urban area, where plenty of their customers and prospects buy coffee every day.

The narrower the audience and the higher the customer value, the more unconventional the distribution can become.

It seems like Free Bean has early traction. They’ve previously distributed coffee in San Francisco and Los Angeles, among other cities. I’m curious about what sponsors view as an ideal outcome of partnering with Free Bean. Is it measured by meetings booked through the QR code on the can? Or prospects mentioning the coffee on a pre-sales call?

There is the broader brand awareness play, and maybe that’s what a B2B marketer must get comfortable with. My hunch is the tech companies partnering with Free Bean are treating the coffee shop concept as a “test budget” experiment, money meant to discover a meaningful incremental return above the tried-and-true acquisition channel. If you see repeat sponsors at Free Bean pop-ups, you know that this idea has legs.

Digital advertising is crowded, but the analytics tracking is top notch. When you move to sponsoring a “free” coffee shop, the attribution gets blurry. In 2026, marketing is as bizarre as ever, yet as creative as ever.

What’s the latest in AI, you ask? Well, apparently, it’s coming to kill us all.

According to ex-Anthropic researcher Jacob Coxon, AI’s meteoric progress (have you checked out OpenAI’s Astra yet?) poses a real possibility of AI-caused human extinction. Coxon argued that AI could cause human extinction by the end of the decade. Current Anthropic researcher Evan Hubinger went further, saying he personally puts the odds above 10% within the next decade.

“What does human extinction caused by AI even mean? This makes no sense.”

Fair point. I was dubious of this statement when I heard it too.

On a recent Prof G Markets podcast episode, guest Nick Bostrom (author of Deep Utopia) gave this example (loosely paraphrasing): Imagine you tell AI to produce as many paper clips as possible, assuming no resource constraints. If AI were to execute this task verbatim, it would find a way to produce a gargantuan amount of paper clips, taking away resources intended for humans. If, say, drinking water was needed to produce paper clips, all drinking water on Earth would be reallocated away from humans. It’s these kinds of resource-constraint issues that could lead AI to make consequential changes to the things humans need to survive.

But take this all with a grain of salt. Coxon and Hubinger’s dialogue has taken the world (or tech world, that is) by storm over the past two weeks. It’s gotten to the point where Sam Altman (OpenAI) and Dario Amodei (Anthropic) are voicing concerns about AI’s rapid acceleration, and suggesting government intervention and a slower pace of frontier AI development.

Although I work in tech and have seen firsthand how incredible recent model developments have been at executing business tasks, I am a bit skeptical of the current narrative that AI will take over the world, rendering humans to be useless along the way. There are so many assumptions and narratives that are being spun to arrive at this possible outcome. It’s like revenge of the robots, but keep in mind, that’s science fiction.

Altman and Amodei’s version of brand building has been fear-based. They’re in a race to prove whose AI model is more capable than the other and have chosen to lead with a self-righteous perspective that “AI is dangerous, but you can trust me”. I view this as nauseating, if not misrepresentative of what this technology can do for good.

It’s either a publicity stunt around which company their models ‘accidentally’ hacked, or it’s about all the awful things that someone can do with AI but that they stopped thanks to their safety guardrails. I don’t disagree that AI models have become incredibly powerful, and we need to be more mindful regarding security risks. However, a doomer’s mindset around this seminal technology isn’t necessarily going to make people want to see how far you can take it with AI.

AI’s biggest branding problem may be that the people building it are disproportionately defining its public narrative around what could go wrong rather than what could become possible when it works. In my opinion, the quickest way to change the current gloomy (if not fatal) narrative about AI is to focus on what can go right with AI. And where can AI produce a massive amount of good? In the medical field. Let’s dive into a few recent wins.

Anti-Aging

Insilico Medicine created a drug with AI originally meant for lung disease patients, that happened to also make patients appear biologically younger according to several aging measures. AI was used to identify the target protein and design the molecule needed to inhibit it (I’m oversimplifying it, but this is as much as I could understand). Insilico used AI to compress the front-end of drug discovery and stumbled upon another win: anti-aging insights.

Cancer Detection

The MASAI trial, involving more than 105,000 women, found AI-assisted screening increased breast cancer detection by 29% while cutting radiologists’ screen-reading workload by 44%. There are only so many radiologists in the world, and they already tend to be quite busy. AI-assisted radiology is producing better results, while making these doctors more efficient. The sooner we have solutions, the sooner more lives get saved from cancer.

Sepsis Detection System

If you go to the hospital, you shouldn’t have to worry about catching an infection and passing away. Unfortunately, sepsis claims more than 250,000 lives per year in the U.S. But thanks to Johns Hopkins researchers there is now an FDA-cleared early-warning system available. They created an AI-based medical tool that compares hundreds of live parameters with electronic medical records to identify potential cases of sepsis earlier. Doctors are moving from patient to patient in hospitals and don’t have the time to stay on top of every little thing regarding their patients. The sepsis detection tool has been shown to reduce the mortality rate by 18% across dozens of hospitals using this new technology.

Is AI getting noticeably powerful? Yes. But if it’s directed toward good, it can materially improve our quality of life (and potentially extend it). Now that’s better branding.

I don’t spend much time on social media. In fact, I deleted the Instagram app on my phone at the beginning of the year. My weekly screen time number was higher than I’d like, and I haven’t looked back since.

However, I do still have the Facebook app on my phone and occasionally go for a good old-fashioned social media doom scroll.

And boy, things have gotten weird on Facebook.

I am now receiving ads to “sell my company’s data for AI training”.

Not ads for AG1 or the latest trendy protein powder. Ads meant for business owners to sell their company data for up to $2 million. It’s as if Facebook knows I write a Substack about AI (they know everything, so they probably do).

Let’s unpack why these data brokers have spun up, who really wants to buy your company’s data, and implications for employees and customers.

The race for niche data sets is on. As the push toward physical AI accelerates, including self-driving cars and humanoid robots, AI will need highly specific training data to produce the outputs necessary for robots to perform their intended tasks. For more on the physical AI narrative, give this article a read.

After being targeted by AI training-data companies on Facebook, I did some research into who these advertisers are. Companies like Data Factor and Micro1 are focused on acquiring training data to support model development. Data Factor’s website even states quite literally, “training data for the machines that move through our world.” As predicted back in the linked Relentlessly Curious article, a business opportunity exists for those who can acquire and consolidate niche data on real-world tasks like cleaning an apartment or delivering food. These data brokers have now expanded to acquiring data from companies in the form of “operational data and business workflows” according to Micro1. My inference is they will then package this data up and sell it to major AI labs or robotics firms. If there’s any validation that this business model has product-market fit, Google’s pending purchase of Spirit operational data shows a proof point that AI labs are willing to go to all corners of the Earth to pick up useful inputs to make their models more effective.

A key element of why AI labs care about your company’s data is for the purpose of evaluation frameworks (commonly referred to as “evals”). This is a major reason AI labs work with a company like Data Factor. See, AI still gets things wrong, and a common reason why is because it doesn’t have the right context. That’s where evals come in, which are essentially frameworks that tell the model “what good looks like” for the sake of pattern recognition. They are systematic tests used to measure whether a model or agent performs a task as per the defined requirements.

For example, if Anthropic wants to break into finance, they likely will need a lot of data on how to build a financial model, as well as how to check their work to ensure that the model was built correctly. Today, AI can build the financial model rather quickly, but can it pressure-test itself under countless sensitivity scenarios that a finance professional just “knows” from years of experience? That’s debatable, and where evals are helpful.

If Anthropic buys a financial advisory firm’s financial model files and related analysis documents, they can then instruct the AI on all the edge cases that may come up and the AI will be able to check its work more effectively. A company with decades of operations has encountered every outlier scenario under the sun, allowing it to help inform what the most “correct” output can be.

So, let’s break this down further. When an AI lab acquires a niche data set like a company’s workflows, they get tons of raw training data to improve their model, evaluation data to test said model, and documentation produced by humans that shows judgment and nuance on how outputs are assessed. Evals help bridge the gap between training data and a finished product.

Although I question why Data Factor and Micro1 are advertising on Facebook, they are likely trying to reach small business owners, as there are too many complications involved in asking a corporation to sell its data (I can’t imagine shareholders would be pleased with a company selling what could be their secret sauce). There are over 36 million small businesses in the US, and I’m sure most of them would appreciate another $1 million in exchange for their data.

The advantage most companies have over the AI labs is that they understand their customers’ data and workflows inside and out. Thus, it’s a bit ominous to think that if small businesses hand off their niche understanding of how to work with certain types of customers (let’s assume actual customer data is not for sale, only the workflows behind them), eventually the LLMs will close the gap. What could lead to a banner year in profitability for a small business in 2026 may lead to a structural decline in revenue in 2028 as the LLMs build strong evaluation frameworks around the problem the small business was solving a few years back.

Furthermore, the idea of selling a company’s data is not a new concept. In the consumer world, it’s a somewhat common practice to buy a defunct brand’s email list for the sake of extending a brand’s broader customer list. When you as a customer sign up to receive marketing emails, you may have little idea where your data is going (unless you read the Terms of Service, which I bet is highly unlikely).

Now, as an employee, all those standard operating procedures (commonly referred to as “SOPs”) you wrote over the years may be packaged up and sold to OpenAI or Google to fine-tune AI trying to do your job. It feels kind of funny, and in not the “ha-ha” way.

But that’s the way the world is going. I’m very curious to see how many businesses take up offers from Data Factor, Micro1, and other AI data brokers. Which companies will shun the idea of selling their workflows while others may not see much value in them and thus view it as extra pocket change to keep the lights on.

The extent to which AI labs are going to acquire the most random data sets really proves out the saying, “one man’s trash is another man’s treasure.” Who knew how valuable trash was, huh?

For those of you following the AI race between Anthropic and OpenAI, you know that these two frontier labs have employed vastly different product roadmaps. OpenAI has chosen to launch everything under the sun and see what sticks. Products have ranged from consumer use cases like Sora (AI video generation, which has been sunset) to agentic coding tools for engineers like Codex (25 million monthly active users). On the other hand, Anthropic has kept narrow focus on building AI primarily for business use cases, launching a relatively narrow set of products, including Claude Chat, Claude Cowork, and Claude Code.

In one sentence, OpenAI has tried to make AI for everyone to change the world, while Anthropic is building AI for people to change how businesses operate. The former strategy is a lot more difficult to execute than the latter, and Anthropic’s results suggest the benefits of that focus. As of July, Anthropic’s revenue run-rate reached $65 billion, while OpenAI’s revenue run-rate sits at $40 billion.

With that said, these two frontier labs have plenty of overlaps in their product offering, and as of last week, they have another similarity: building tools for agentic commerce.

It’s been covered plenty of times in Relentlessly Curious, but last year OpenAI tried to integrate a shopping experience in ChatGPT called Instant Checkout. To make a long story short, it was a flop as users didn’t show enough shopping intent (lower conversion rates versus purchases made on brands’ websites) and the logistical challenges with getting humans and agents to make purchases within an AI platform are steep. OpenAI was too early to the game.

Anthropic released an update announcing that they are building commerce agents.

Instead of looking to integrate a shopping experience within Claude, they are building tools to enable agentic commerce, not be the arena in which transactions occur.

I believe this approach is incredibly savvy and reflects Anthropic’s product strategy so far. They’ve chosen to be the picks and shovels, the infrastructure of front-end AI, instead of trying to convince users of completely new social behavior within their platform (i.e. shopping).

Let’s break down their press release and talk through its implications for brand operators now that Claude is in the mix.

The Shopping Agent

Any AI customer service (CX) should be sweating as they read through this announcement. This agent turns a brand’s website into a conversational search platform, answering customer inquiries, providing tailored recommendations based on what a prospective customer is looking for, and building a shopper’s cart for them. Claude will help upsell products, while staying true to your brand guidelines and pricing constraints. It serves as your digital sales associate.

Major AI CX enterprise players like Decagon and Sierra are likely okay here because working with enterprises requires many integration points and legacy systems and messy data sets. But the AI CX companies that target the lower-to mid-market brands are going to run into trouble competing with a savvy operator who knows how to use AI.

What matters most when using Claude’s shopping agent? Your data infrastructure.

For Claude’s shopping agent to be worthy of your (and your customer’s) time, you need to be running a tight ship when it comes to how you organize your data. Catalog hierarchy must be clean (parent/child SKU relationships, updated attributes, fresh pricing, live inventory), the FAQ section should be exhaustive, and your brand guidelines need to be up to date. Oh, and don’t forget that your product detail pages must be pulling through all of this accurate data to the front end so agents can read it.

I’d be very curious to run an A/B test, splitting traffic between a website that leverages a vertical-layer AI CX player (e.g. Decagon) and one that uses Claude for an e-commerce storefront’s conversational experience.

I’ll be watching my X feeds closely to hear more about what e-comm operators think of this one.

The Merchant Agent

I’ve never been impressed with Shopify’s data analytics interface. It’s clunky, and even with recent updates to incorporate AI into writing SQL queries for data extraction, it has always felt siloed, particularly compared to Amazon, which provides brand operators with a treasure trove of data across the full marketing funnel and logistics.

Claude’s Merchant Agent now unlocks lightning-quick operational insights. You can now use Claude to proactively track inventory, suggest pricing and promotional updates based on sales history, and perform root cause performance analysis.

There are countless e-commerce analytics platforms and conversion rate optimization (CRO) software firms that allow you to ask questions about your data. They are all going to be hurting once this Claude update gains traction (besides the companies focused exclusively on enterprise for the reasons mentioned above).

The analytics firms pride themselves on allowing you to make sense of your messy, disparate data. This update reads like Claude has you covered.

One of my favorite sayings when I was a brand operator was “go from data to insight to action in the most automated way possible”. Now, Claude takes you from data to insight to action, almost fully automated. A human in the loop is part of the design, with Claude requiring a person to approve any changes Claude suggests. Given the sensitivity around the customer experience, this intentional friction point makes sense.

I will say, I am surprised Anthropic decided to release a product for agentic commerce. Perhaps they felt they needed to have an opinion on the multi-trillion-dollar industry that is commerce ahead of their likely fall 2026 IPO. In fact, this move feels like a stretch outside their typical business tooling arena. Agentic commerce is complicated, and their rival has already tripped up. But in typical Anthropic fashion, they focus on providing AI to technologists. See, OpenAI tried to put commerce inside ChatGPT. Anthropic is giving commerce operators the tools to put Claude inside their commerce experiences.

As you can imagine, I’m fascinated by how quickly technology develops and the impact AI has already had on how people work. But at the same time, AI has made some things quite weird.

There are AI companions like Character AI that people treat as their friend or romantic partner, with the average user spending over 1.5 hours per day on their app. You also have AI companies opening 24/7 cafés (Corgi Café), with branded drinks from other AI companies.

Oh, and then you have Google buying bankrupt airline Spirit’s data to improve their models.

The sale is still under review, but it looks like Spirit will live on in a different form. You won’t be booking a 5 AM flight to Vegas with the most eclectic people you could possibly assemble anymore on a yellow hunk of metal, but Spirit’s knowledge base may provide an assist in a Gemini prompt response.

Here’s what happened: Google won a bidding war against enterprise data broker Mercor and ultimately paid $10 million for a head-scratching amount of data. Check out the below snapshot of what Google is looking to scoop up from Spirit.

That’s a lot of data, to say the least. And it’s worth mentioning that Google will be scrubbing any personally identifiable information (PII) from this data, a key point of the agreed-upon sale.

Here’s the quote from the article linked above that confirms Google’s thirst for data:

“We acquired part of an enterprise dataset from Spirit Airlines, which can be helpful in improving our products and AI models,” the spokesperson said. “We will not receive any personal information from this dataset. Any data we receive will be rigorously scrubbed of any personally identifiable information by a third party before receipt.”

So why would anyone pay $10 million for the carcass of a bankrupt company? Because although the company may be dead, its accumulated knowledge sure isn’t.

Let’s unpack a few things here.

Switching gears: as AI model companies look to acquire new sources of data to train their models, they’ll need to be more creative as the obvious locations start to dwindle. Once the open web has been scraped, you must look in every cranny. We’ve chatted about the physical AI narrative in the past and how companies like DoorDash are paying Taskers to record themselves delivering food and selling the recordings as data to help build autonomous devices like delivery robots.

But showing up at a fallen company’s liquidation proceedings is a new one (or at least from my surveillance of the topic). Distressed firms that may have been disproportionately impacted by an adverse macro trend or may have made a few bad decisions and find themselves past the point of no return. Thus, they need to lick their wounds and sell the business off for parts.

AI is making corporate data a liquidation asset. What Google’s bidding war for Spirit’s operational data shows is that the “parts” just got more valuable. If you’re a creditor to a bankrupt company, there’s a chance you could recover a bit more money than expected if the data can be valuable for model training.

There may be a new cast of characters showing up at a liquidation proceeding, looking to get their hands on a fallen angel’s data. I expect we’ll see more legal documents including names like Google, OpenAI, Anthropic, and SpaceXAI putting forth generous bids. The hope would be to pick up niche data that may be able to fit a specific use case or provide more data in general.

But in my opinion, I don’t think this is a big deal from a creditor perspective. Whether you own the bonds of a distressed company or are an opportunistic investor analyzing the credit default swaps on these bonds, it’s unlikely that the amount paid on the data set by AI model firms is likely going to move the needle. In the context of Spirit, an extra $10 million is appreciated, but when there’s billions of dollars of debt, it may as well be loose change. I see this as a “one man’s trash is another man’s treasure” situation.

I have to think that the corporate development teams at Big Tech are going to be busier after this precedent, scouring the earth for soon-to-be bankrupt companies that need to sell off their data as part of liquidation. The opportunity to pick up niche, real-world enterprise data generated through decades of actually operating a business doesn’t come up every day.

Share Relentlessly Curious

Google wades into a gray area from an ethical and compliance perspective when it comes to acquiring employee data. Among those hundred million emails are people’s names and the things they’ve written. All of this was originally under the assumption that their messages wouldn’t be bought by Google.

So, Google will rely on a third party to scrub Spirit’s data for anything that could be traced back to an individual person. It’s worth mentioning that Mercor’s bid stipulated that Mercor itself would scrub the data. No wonder they lost. Who would trust a company that sells data for AI models to not let profit control their decision-making here? But even with a third party filtering the data before the keys are handed to Google, there is still concern that Google will find a way to reverse-engineer who said what in these conversations for the sake of improving models. The worst-case scenario is that a Gemini prompt response mentions a specific Spirit employee’s name on a matter that they thought was a private business matter

At this point, most people have heard of ChatGPT or Gemini and have at least a basic understanding that you can ask a question and get a curated answer in return. There is considerable anxiety about what Google will do with this information, given some of the sensitivity around it. Worth noting that customer information is not included in what Google is buying, only employee data.

The obvious wrinkle is privacy. Google says the data will be de-identified by a third party before it receives it. The flight attendants’ union isn’t convinced that’s enough, arguing that links between datasets could still allow individuals to be identified. The court has delayed the hearing on the sale until September 9. So we’ll see what happens, but if I were a betting man, I’d imagine Google walks away with the prize.

Let’s wrap this up. A new line item on the liquidation checklist seems to have appeared, and that’s “what is the company’s data worth to an AI company?” The next time a company goes bankrupt, don’t just look at its planes, factories, patents, or customer list. Check out the data. There may be an AI company waiting to bid on it.

The current enterprise AI adoption playbook looks like the following:

1. Hire an expensive consultant to tell you how to “implement AI into your business”

2. Sign 10,000 people up for an enterprise ChatGPT or Claude account

3. Roll out training (most likely led by an external consultant or a non-technical team) for said enterprise AI plan

4. Wonder why no one is using AI and there haven’t been any “AI-driven” synergies yet

I think a key problem in this process (well, there are a lot of problems) is that AI adoption is being approached as a singular exercise. Those 10,000 employees spin up 10,000 different AI workflows. Every employee develops their own understanding of where AI is useful, how to prompt it, what context to provide, and how much they can trust its output.

Sure, there may be company training and team-level breakout sessions that get into the tactical adoption of AI. But what tends to happen is that a few power users shine, while the vast majority of employees either ride off their coattails or do the bare minimum, comprehending just enough of how to use AI to get by.

There’s an alternative approach I’ve been thinking about.

It’s called multi-player AI.

Multi-player AI adoption involves multiple humans and AI participating in the same workflow and learning from one another. It starts with viewing an AI agent as a really smart colleague, not a program or “thing” you need to implement into your workflows. AI that has context, can converse with you and your colleagues collectively, and can execute on tasks.

For this to work, there needs to be a mindset shift. AI agents need to be viewed as expected attendees, not party crashers that everyone groans about when they show up and no one understands.

Based on personal anecdotes, multi-player AI adoption comes much more naturally to recently founded companies. In an early-stage environment, pace is critical, so everyone becomes AI-savvy for the sake of growth and keeping the company in business. You think less about finding AI use cases because everyone is using AI all day out of necessity. The multi-player approach (working with each other and AI) is ingrained in company culture.

Now, it’s the legacy enterprises that are starting from a much tougher position. When you have an employee base that has been used to working a certain way for decades, you’re going to encounter friction when changing the way they work. Particularly if the “change management” implies job loss based on what everyone is hearing in the news.

It’s funny. Big companies view AI as in, “we need a strategy,” while start-ups say, “this needs to be done ASAP, let’s get Claude on it.” Who do you think is getting a better ROI on their AI investment?

But I believe large enterprises were just thrown a lifeline when it comes to AI adoption.

Enter, Slack Code.

Last week, Slack (owned by Salesforce) introduced Slack Code, “a new feature that allows users to develop code via shared rooms directly within the messaging platform where AI agents write the software and human teammates watch, steer, and direct the session.”

This is a major announcement because Slack made building software a multi-player game. Now, it’s not just you and your AI harness (i.e. Claude Code or Codex). It’s also you, your colleagues, and your AI model of choice, building together. All with the context of your conversations and documentation in one spot.

I believe this is a critical product feature release because it shifts the mindset in favor of AI adoption. Since Slack Code enables your team to work with AI together, AI essentially becomes your colleague. It has access to all your conversations and documents within Slack, meaning it can respond with rich context and write relevant code.

Share Relentlessly Curious

Let’s shift back to the training component of the multi-player strategy. There’s a higher likelihood that people are going to learn more from seeing how their colleagues interact with AI than sitting through a monotonous weekly, hour-long training session on how to write a prompt (gosh help you if it’s led by HR). I haven’t tried out Slack Code, but it seems that if my colleague and I want to fix a bug in the codebase, we can prompt the AI agent within the same chat. In this case, I will learn from their thought process, and they will learn from me. Not only about the business or technical problems that we are solving, but also how they interact with AI.

This can create a multiplier effect. For example, perhaps two engineers and a product manager are chatting about a customer pain point. The engineers call their AI agent to join the chat, and they all start building a solution. The product manager can provide customer context, while the engineers write more technical prompts to the agent. The team is building together in real time.

This level of AI adoption is likely to compound on itself because those three employees are learning from each other how to work with AI, and then they bring this experience to their next problem with other coworkers. The behavior spreads, enabling fast adoption.

Executives should view AI adoption less as an implementation or training issue and more as a social issue. The power users of AI within a company can create as many tutorials and training guides as possible, but that still doesn’t ensure that a critical mass will comprehend the information.

If you’re leading enterprise AI adoption, I highly recommend you view it as a social experiment. By creating environments where curious employees can work with agents (and each other), their understanding of AI will compound more quickly than if employees are working in a silo. The curiosity will be amplified.

In one line: “show, don’t tell.”

Perhaps the future of enterprise AI isn’t giving every employee an AI assistant and hours of training for them to mindlessly watch. Instead, it’s giving every team an AI coworker. Companies that aren’t seeing much return on their current AI strategy need to be honest with themselves. The fastest way to teach someone how to use AI may not be another training session. It may be letting them watch the person next to them use it. And that’s why the future is multi-player.

We’re back with our 10^th edition of Hot Takes in AI! You know the drill, three takes centered around AI or the universe around it. Buckle up and let’s get going.

Hot Take #1: Bitcoin’s price is tied to its narrative, and right now, other investment opportunities have more-appealing narratives.

In early June, I remember logging into my Coinbase account and wincing at my losses over the past year. Thankfully, crypto represents a small allocation of my broader investment portfolio, and I am only invested in a few blue-chip assets like Bitcoin and Ethereum.

A funny thought hit me as I shook my head at my phone while staring at the Coinbase app: “Does anyone care about Bitcoin anymore?”

Now, I’m approaching this with zero data or news articles to support my claim, but it often feels like there’s no good fundamental reason why Bitcoin’s price goes up or down. It’s very infrequently used as a medium of exchange, and it lacks traditional cash flows or earnings.

Despite a pro-crypto U.S. administration and institutional adoption, Bitcoin’s performance lately has been abysmal. It’s down 28% since January 1^st, and down 46% since mid-August 2025. It’s been interesting to observe that a positive regulatory environment has not translated into sustained price momentum.

Bitcoin (and crypto more generally) just isn’t a cool thing anymore. From what I’ve observed, market media coverage is much more focused on AI and prediction markets (i.e. Kalshi and Polymarket) than crypto. Because Bitcoin doesn’t have cash flows and is not backed by an issuer, it’s only worth as much as someone is willing to pay for it. And if it’s not realizing the hopes that it would become this new-age digital currency that the world adopts over its own central bank currencies, maybe there’s less of a story to follow here.

AI and prediction markets present much more opportunity for upside. AI is upending our work and how people search and discover information. Then there are prediction markets, which, in my opinion, are synonymous with gambling. You can bet on an incredibly wide range of things on Kalshi, from sports and politics to the weather.

In my opinion, there are more exciting ways to allocate your money today than crypto. Perhaps crypto’s ascent was all about animal spirits and the increasing hope that the next buyer had in a decentralized currency being a store of value. The market demands an entertaining story from an asset if it wants to see its price increase. Unfortunately for Bitcoin’s sake, the story has gotten stale.

Also, I need to pat myself on the back for being early to this take. I beat out famed investor Steve Eisman (you know, the guy who was portrayed by Steve Carell in The Big Short). Check out the time stamps.

Share Relentlessly Curious

Hot Take #2: Claude’s AI watermark announcement provides a branding opportunity they don’t realize.

Last week, Anthropic announced that future Claude models will have a watermark encoded on any text or image output. The product feature is in response to transparency requirements under the EU AI Act, which requires providers of AI systems generating content to mark their outputs in a machine-readable format. This subtle identifier will not be visible to the naked eye. However, Anthropic is rolling out tools like a “watermark detection API” to provide additional visibility.

We’ve been seeing an increasing trend towards transparency on what content AI created versus human-written content. LinkedIn has rolled out AI detection tools, and Substack partnered with Pangram to show readers how heavily influenced an article is by AI (you can even check the Pangram score for this article). Don’t worry, Relentlessly Curious comes directly from my mind, not from a Claude prompt.

Anyways, I believe the frontier labs have a major branding opportunity when it comes to how they comply with the watermark regulation. If society is pushing more transparency on what AI helped create, eventually people will stop being sheepish and be more open about what AI did for them. And when that happens, the frontier labs should be loud and proud about their output by making their watermarks visible to readers of the output, whether it’s code, writing, or an image.

This will need to be tastefully done; yet the analogy I’d like to extend is that the labs have a “Sent from my iPhone” moment here. Think about the branding Apple gets by including the iPhone reference in default iOS Mail app. Apple has a premium brand and its users get to show off that they have an iPhone by sending an email, a routine communication method.

Apple turned an otherwise mundane technical detail into a recurring brand impression. It’s all about signaling. And for the frontier labs that invest the most in making their product the best and most premium, they’ll be able to garner incredible social proof if others can see the output of their work.

AI-generated code could underpin the next generation of technology companies and increasingly become part of Fortune 500 technology systems. Imagine Anthropic having a “Claude” tag at the end of a block of code, so users could see how important Claude is to their systems. The watermark compounds on itself.

I can see it now: a marketing campaign that highlights all the incredible products that were built with or heavily influenced by Claude. At companies whose products you and I use every day. It can help show how widespread AI use is and encourage others to join in on the AI wave. You’re welcome, Anthropic brand marketing team.

Leave a comment

Hot Take #3: Lovable’s Series C fundraise shows that application-layer companies can still thrive, even when they fly close to the sun that is the frontier labs.

Vibe-coding platform Lovable announced a Series C investment round last week, valuing the company at over $13 billion. Founded in 2023, Lovable offers users the ability to create software through simply typing or speaking. It’s very similar to Replit, which we’ve chatted about extensively through past vibe-coding tutorials.

I was surprised to hear about Lovable’s massive fundraise and revenue growth since launch (surpassing $500 million annualized revenue run rate in June). Well, I’m not surprised that an AI company is doing well. I’m surprised that frontier labs like Anthropic and OpenAI haven’t eaten their lunch yet. Claude Code and Codex are well positioned to eat into the vibe-coding market thanks to recent model improvements.

See, Lovable is a much more user-friendly version of Claude Code or Codex. It does not require much technical prowess at all, allowing users to truly build software without looking at a single line of code. The same could be said for agentic coding tools like Claude Code. However, I believe it really helps to have some sort of analytical foundation to make the most of them, given the text summary output can be better suited for an engineer’s eyes. Otherwise, you’re not going to understand the inner workings of what you’re building. With plenty of available integrations with different models and software like Linear and Figma, Lovable proves the application layer can capture value even when the underlying model is commoditized or interchangeable.

Perhaps model companies optimize for general intelligence, while application-layer companies optimize workflow completion. This signal comes from Lovable’s Series C press release (linked above) that mentions key features that they now offer built-in such as payment functionality and security and governance protocols. Both storing personal information from users and keeping the site safe from hackers are critical components that the average vibe-coder likely doesn’t (and shouldn’t) feel comfortable building themselves. If Lovable offers native integrations to Stripe and Shopify, it reduces the amount of engineering work required to get the app into production. With that said, it’s still critical to have a formal engineer review your codebase before accepting any personal information from a user.

Lovable’s operating team is sharp. My read is that they ship an initial wrapper above frontier models, see what people build, and then add features based on where users get stuck. They can see data on what types of projects people are creating and then build features to make them easier.

Lovable found its wedge by moving quickly and identifying specific user pain points that may be too small for Anthropic or OpenAI to tackle. At a high level, Claude Code and vibe-coding platforms aren’t very different. But on closer look, platforms like Lovable that automate the tricky, complex workflows for users have a better shot at holding onto their user base despite disruption from the giants.

I’m sure you’ve heard of Reebok or Nautica before. But have you heard of Authentic Brands Group?

As a former brand operator, I was quite familiar with Authentic Brands Group (ABG) as the poster child for what a successful brand portfolio company at scale looks like. Oversimplifying, they buy brands with strong name recognition, and plug them onto their platform of brand management, marketing, and distribution. But here’s the catch, they often only acquire the intellectual property (IP) of the brand and then license it to operating partners. Their secret sauce is how they manage and build brands, while they outsource operations like logistics and fulfillment to third parties.

ABG is a brand portfolio in the most literal way. They literally buy the intangible asset that is the brand.

On the other end of the spectrum, you may have heard the company name, Bending Spoons, in the news lately. They’ve garnered plenty of buzz for their AOL acquisition (yes, the domain of your original email) which closed in January, and last week announced their acquisition of Airtable. Once you get past its ridiculous name, it’s a fascinating business model. The Italian software portfolio company buys strong brand recognition companies that have been on the decline or underperforming versus investor expectations. Their thesis is centered around buying “iconic products” and improving their operations.

At a high level, Bending Spoons reminds me of the Authentic Brands Group of the software industry. ABG and Bending Spoons are two different manifestations of the same broader idea: buying assets whose existing customer recognition survives the deterioration of the original business.

Similar type of companies, just in different industries. Both recognize that customer recognition can remain valuable even after the underlying business begins to bleed. And it worked out well for them. ABG is valued at $20 billion as its last funding round in 2023 (with plans eventually IPO) and Bending Spoons trades at a market capitalization of $28 billion (as of August 7^th, as per Yahoo Finance).

Tremendously valuable companies whose business model is giving brands a second chance.

This got me thinking. There will be an iteration of ABG and Bending Spoons for the AI era.

We’re currently living through a period of hyper-growth in technological capabilities. Hundreds of billions of dollars are being poured into AI infrastructure and model development, with trillions potentially following over the coming years.

It’s a race to acquire market share (or, if you want to be more evangelical about it, a race to artificial general intelligence).

Not every AI model or company is going to survive. At some point, I expect the flow of venture capital and Big Tech funds will slow down, and companies will need to inch towards self-sustainment through healthy unit economics and positive cash flow. Especially if they can’t reach public markets before the bubble pops. And if you’ve kept up with the news over the past few years, you are well informed that profitability has been low on the priority list for AI companies (for good reason though).

AI companies have an unusual problem because the technology underlying their product is increasingly becoming a commodity. A company can have a revered product today and discover that the underlying model economics have changed tomorrow. If your vertical AI company’s AI bill 10’xd overnight because the economics of the underlying models changed and subsequently, your unit economics are busted, what is left to your business? That becomes an easier question to answer if you’ve invested in your brand along the way.

Plenty of AI models and vertical AI startups will fail and need to be rescued by a company with larger market share or a private equity firm thinking they can squeeze out the bloat from the financials. But the AI companies that not only have a strong product but also have a respected brand will have an asset that becomes increasingly valuable as the technology underneath them commoditizes.

Share Relentlessly Curious

Think about the difference in branding between Anthropic and OpenAI. What do you think and feel when you read those two names? There’s no right or wrong answer. My point is that people have an emotional reaction to these names beyond their model capabilities.

Now let’s go to specific sectors. In legal tech, there’s Harvey and Legora. In finance, there’s Hebbia and Rogo. If you work in law or finance, what’s your take on the brands I mentioned? Do they have a brand worth remembering within your industry? If I haven’t mentioned AI startups in your line of work, think about the brands of the up-and-coming companies in lieu of the above examples.

Over time, the workflow may change. The same goes for the underlying model and model provider. However, the brand can remain. Eventually competition will win out and there will be industry consolidation, as well as AI companies that go bankrupt. And some will be scooped up by the next ABG or Bending Spoons.

I believe an AI company’s brand ethos should center around being your smart but humble friend. Someone who is trustworthy, diligent, and will admit when they are wrong.

Today, AI prompt responses can be overconfident, and people are still a bit unsure when to trust AI and when not to. This will be a constant struggle as models continue to develop and may never be fully solved. If you’ve watched the news, you’ve likely seen public backlash on the impact AI will have on the job market, environmental concerns, and whether the leaders of the frontier labs should have so much power.

A brand around transparency, wisdom, honesty, and curiosity will be an enduring asset if the operating model falters. Brand can preserve enterprise value even when the original technology or business model fails.

Back to brand holding companies. Not born yet, I believe there will be an AI version of ABG and Bending Spoons. A highly skilled operating holding company that acquires AI companies that couldn’t build an enduring business (perhaps unit economics were busted when the economics of the underlying frontier models changed) but did create an enduring brand that generated customer trust and retention. If the brand ethos powered customer trust, distribution, and organic user acquisition, a fallen company can be an attractive asset to this kind of portfolio.

Those AI companies will be given second acts. Their brand will live on.

Side note: if this idea of acquiring distressed AI businesses sounds interesting to you, please reach out in a few years. Particularly if you can bring a large credit facility to the table.

Leave a comment

Time to wrap this up. The AI companies that survive the next five years won’t necessarily be the companies with the highest-performing models, the most robust features, or even the largest user base today. And I suspect many of the survivors will be the companies that have invested in their brand.

And if the history of companies like Authentic Brands Group and Bending Spoons tells us anything, there may eventually be an entire industry built around buying the brands of yesterday and giving them the technology of tomorrow.

Today marks the 100^th Relentlessly Curious article, and it’s going to be a little different than the usual AI, media, and consumer brand content. I’m going to peel back the curtain on why I write, and what I’ve learned from publishing 100 editions.

Every week, I give my take on topics across the AI, media, and consumer brand landscape. Sometimes, the ideas flow out of me like a river. Other times, I’m racking my brain late on a Monday night, squeaking something out for my self-imposed Tuesday at 8AM EST deadline. I do not make money directly from Relentlessly Curious.

So why do I do it?

A few reasons.

Writing is a creative outlet that provides me with fulfillment.

I struggle with the balance between creation and consumption in the digital age of endless content. Open TikTok or Instagram, and you are greeted with an unlimited number of pictures to see or videos to watch. Better yet, the content is fine-tuned to exactly the topics you prefer to watch. Social media algorithms know you better than you know yourself.

In 2024, I concluded that I was spending way too much time consuming content. Instagram consistently topped my screen time leaderboard. I didn’t like the behavior that came along with social media consumption. For instance, sending memes to old friends became the substitute for picking up the phone and giving them a call. Spoiler alert: it’s a false sense of closeness and you learn nothing about how a person’s life is going if you trade slop content back and forth as your only means of communication.

I also didn’t like what social media was doing to my brain. I consider myself a driven person, but was feeling demotivated and defaulting to just staring at my phone in my free time. Yet I wanted to do something meaningful like starting another business (fun fact: I ran a pickleball e-commerce business in 2023) or some other creative pursuit. I have an overarching need to always be productive (a blessing and a curse), and social media was greatly hindering this primal desire.

So, I acted. For the second half of 2024 into the middle of 2025, I created an Instagram account dedicated to aesthetic pictures of the West Village (my former neighborhood). So, on weekends, I would bike down to West Village and snap photos on my iPhone of quaint streets, trendy restaurants, and restaurant diners sipping martinis on patios. With the goal of curating a minimalist, trendy vibe. I would post these photos with my subtle, yet recognizable logo I designed in Canva and write an introspective few words as a caption. The goal was to build a clothing brand* after achieving social media success (however the heck you want to define “success” here), as it’s way too expensive to invest in paid media without a sizable organic social following first.

Here’s the problem: I hated running the Instagram account (it also failed to hit any meaningful amount of followers).

I’m not exaggerating. I’m just not particularly interested in photography or visual content editing. And I didn’t enjoy being on Instagram more as you need to be on Instagram to grow an Instagram account (the platform I was trying to get away from).

I’ll speak for myself: to be consistent, you need to enjoy all the work you are doing. Well, most of it. Although I posted a picture or video close to 90 times on the account, I knew that to succeed in social media, I needed to post multiple times every day and engage with followers and other accounts. Additionally, I knew I should invest more in equipment and purchase a proper camera. But I just didn’t want to do it.

So, I quit. It was a great decision.

Quitting is totally okay. Because it’s important to know when you are going nowhere fast.

In mid-February 2025, I decided to write a few thoughts in a Word document about AI. I did this at the same time as I was burning out from running the Instagram account. I figured, why not try another creative pursuit as Instagram wasn’t working out.

And I got lucky. I created a Substack, sent the first article out, and texted friends the link.

They liked it. They gave me feedback, and I kept iterating.

And I really began to enjoy the process. I didn’t mind staying up late to churn out an article or wake up a bit earlier to read a few extra articles to help inform my perspective on a future writing topic.

If you want to be a creator, you need to find something you can be consistent with. Especially when it’s inconvenient or you’re tired or sick. I’m glad I found that. Writing is the creative pursuit that makes me feel fulfilled.

*I’ll hold off on sharing the Instagram account’s name as I believe that the name itself has potential to be part of some sort of brand in the future. Consider this idea shelved, but not in the trash.

I am writing because it allows me to try to stay up to date with the latest advancements in technology.

I believe AI is a once-in-a-lifetime technological development. And it’s moving at light speed.

I write a Substack about AI and I still feel overwhelmed! Every day, it seems that there’s a new model update, a new flashy vertical start-up splashing onto the scene, and a new way to use AI that is fundamentally different than yesterday’s iteration.

It’s too much. Thus, I write as an exercise in pulling out what matters most. Whether it’s hard skills around AI implementation or the latest happenings in public and private markets, I aim to focus on “where the puck is going,” not where it is today.

Relentlessly Curious is how I stay accountable to keep up with the pace of AI. I believe this will pay dividends for my career many years from now, so I figure it’s worth it.

Writing allows me to preserve what I believe to be the most important skill in the AI era: communication.

If you are a longtime subscriber, this “why” will sound familiar as I’ve written plenty on this topic.

I believe AI will commoditize most forms of knowledge work. As a result, the most important skill will be how to communicate with AI and how to communicate with people.

Even now, knowing how to properly prompt, guide, and critique AI is table stakes. You need to be able to articulate your thoughts so AI can provide your intended output. And writing is a great way to do so.

It’s not an original thought, but I’m a firm believer that AI won’t replace your job. It’s people using AI that will take your spot.

And then there’s writing to communicate with people. Because the internet has been covered in AI-generated slop, sharp, authentic writing helps cut through the noise. Real perspectives from real people. I’ve decided that I want the core of my personal brand to be my writing, as I enjoy writing and I believe it will benefit my career down the line.

Thank you for following along as I sharpen my written communication!

Share

Switching gears, I’ve learned a few things along the way too.

Writing really helps you understand a topic.

I find that the best test on whether you understand “something” is to teach it to someone else. If you can explain a topic and get someone to understand how it can be applied to various situations, you likely have in-depth knowledge of said topic.

I view writing as “teaching”. By putting my thoughts on paper (well, in a digital document) and sharing it with others, I’ve demonstrated mastery of the subject.

Here are a few tangible examples.

The How-to Claude Code series forced me to deeply understand how to organize my file structure, build multi-agent workflows, and automate processes with Claude skills. Then, there’s the Hot Takes in AI series (a fan favorite) that keeps me up to date on the latest trends, so I understand where the frontier is headed. Furthermore, the investment pieces have strengthened my conviction in my own investment portfolio.

Writing kicks off the following chain reaction:

Learning → Writing -> Teaching -> Mastery of a Topic

Pressing “publish” gets easier every time.

Every time I “schedule send” or “publish” a Substack article or a LinkedIn post, I get a little less anxious. It’s a bit unnerving to have your identity attached to your thoughts and opinions, which anyone on the internet can read and critique. Friends, family members, and coworkers can see opinions they may disagree with. Or what if I write something that is factually wrong?

It’s a risk I take. Especially as I’m writing with my name attached to the Substack. I’m fully out in the open with my Substack and on LinkedIn (since I post there to drive traffic to my Substack).

If you’re reading this and are worried about public perception around your creative pursuit, give yourself a break. No one thinks about you more than you do, meaning if you do make a mistake, people are unlikely to remember it as they are too focused on themselves.

At the end of the day, if you want your account or work to grow and reach more people, don’t you want people to critique it? Criticism is feedback, and if there’s anything I’ve learned, there are plenty of kind people on the internet that read my articles and share their thoughts.

Public accountability builds confidence.

Piggybacking on the last point, if I’m willing to post my thoughts on Substack and preach about it on LinkedIn, I need to have a level of confidence that my content is good. I take this seriously and hold myself to a high bar as my content reflects my thinking.

Posting online has built my confidence because there’s something profound in putting your unique ideas into the ether and feeling okay about others having their opinion on them. Whether they share their thoughts or not. You’re contributing to the knowledge base of the world, and that’s something worth being proud of.

This has overflowed into other areas of my life, particularly in my work. Whether it’s automating processes or solving problems in a more efficient and effective way in my career or throughout networking, I feel confident enough to share my perspective without fear of being wrong. I spend my free time learning and “teaching”, and it’s paid off.

Heck, I was able to pivot my career from e-commerce into product management, and now in a forward deployed engineer role at an AI start-up thanks in large part to my willingness to write and share my thoughts in public. Making a career transition requires confidence, and what better way to build confidence by learning new skills and teaching others for all to see.

I want to underscore the point I just made. I pivoted myself to a more technical career without any formal training or education and writing online played a critical role. No $200K MBA, no family connection, no referral.

Leave a comment

Every week, I open a blank Word document and start writing. Sometimes, I knock it out in one fell swoop. Other times, I write a few paragraphs and come back to the draft a few hours later. Or the next day.

My thoughts are all my own. I say this as AI has no place in my idea generation or drafting process. The whole point is to comprehend what I’m writing. If AI generated the article itself, it would defeat the purpose. If you’ve been around for a while, I’m sure this is of no surprise to you.

With that said, I will have AI polish up my grammar and syntax. Truth be told, I don’t remember anything from my middle school language arts classes and am thrilled for AI to bail me out here. I have no shame about that part.

I named the Substack, Relentlessly Curious, as it’s something I strive to be. After 100 articles, I’d say I’m well on my way.

To me, being relentlessly curious is about pursuing something without a deadline. It's believing that the effort you put in each day, or in this case every week, will compound over time, even if you're still unclear about where it will ultimately lead.

It’s worth noting, I’ve named the website, stayrelentlesslycurious.com. Emphasis on “stay”, because once I’ve woven being relentlessly curious into my identity, I don’t want to lose it. An object in motion stays in motion, you know?

A typical article is roughly 1,000 words. That means I’ve written over 100,000 words as part of my publication.

Here’s to another 100,000 and staying relentlessly curious.

Thank you to everyone who has supported me along the journey. Your encouragement and feedback keep me going.

If this is the first Relentlessly Curious article you have read, I recommend checking out a few recent ones to get a better sense of the typical content. And while you’re at it, please subscribe. I don’t monetize the Substack, so when I see someone subscribe, I get fired up as it’s validation for my work.

Have you heard about the Phia scandal? The AI shopping start-up accused of manipulating affiliate attribution? You know, the one founded by Bill Gates’ daughter?

To get you up to speed, here’s an oversimplified explanation of the scandal.

Phia is an AI shopping browser extension that helps users compare prices for clothing and accessories across the web. The concept itself isn’t new. It’s built on the affiliate marketing model, where partners earn commissions for referring shoppers to brands. Having AI search the internet for the best deal sounds like a win.

But Phia allegedly took the concept too far. According to research highlighted by Bloomberg, the browser extension silently overwrote affiliate cookies on iOS, taking credit for purchases even when it hadn’t influenced the sale or replacing referrals from other affiliates like Wirecutter. If true, the practice would violate affiliate ecosystem rules and could potentially raise legal concerns as precedents have been set in past cases surrounding eBay and Honey. For a deeper recap, I recommend Michael McNerney’s Adweek article.

To be clear, Phia has denied wrongdoing and has not faced known legal action related to these allegations. The claims remain allegations based on independent research (albeit, very compelling research).

But this week’s article isn’t really about Phia. It’s about the uncomfortable reality exposed by the scandal: AI does not remove incentives. It amplifies them.

Think about how quickly AI is advancing. It seems as if there’s a new frontier model update every month. As exciting as these times are, the pace has caused plenty of concern around trust and safety.

Although the frontier labs have come under plenty of well-deserved scrutiny, they have invested materially in model safety research. For instance, Anthropic has built its brand around how integral trust is to model creation and AI adoption. Entire research teams evaluate areas like model misuse, deception, cybersecurity risks, and dangerous capabilities across companies like Anthropic, OpenAI, Google, and xAI. But thousands of start-ups are building upon these models. And those companies have different incentives. Thus, the vertical layer is where I believe one of AI’s biggest trust problems lies.

It’s a fair assumption that engineers at an AI start-up in 2026 are using agentic coding tools like Claude Code or Codex. So, here’s the uncomfortable truth about Phia: an engineer could theoretically instruct an AI coding assistant to write code that executes cookie stuffing (i.e., deceptively adding in Phia’s tag for affiliate attribution).

The Claude model didn’t decide to overwrite affiliate cookies, the company’s implementation choices did. Even if Anthropic aims to build ethical AI, it’s impossible for them to think of every way its models can be leveraged for bad behavior. I don’t see a scenario where frontier labs are expected to think of every way businesses can exploit each other and then try to install the necessary guardrails to prevent these actions.

If AI faithfully executes the goals we give it, what happens when those goals are unethical? Phia recently raised a large round of funding, and like many venture-backed startups, it faces pressure to grow quickly. If a company’s incentives prioritize growth over ethics, AI becomes a force multiplier for those incentives.

Share Relentlessly Curious

I think back to a phrase a former manager used to say:

“As a business grows, both good and bad things scale.”

AI amplifies incentives, as agents are programmed to execute specific tasks. One employee looking to cut corners can have an exponentially negative effect on the business because they have such a powerful tool in their hands. The AI model didn’t create unethical incentives; it simply followed instructions from its user and amplified the incentives at scale.

What’s concerning is how much easier AI makes it to even “test out” a grey-area business idea. Completely speculating here, but perhaps Phia’s cookie stuffing was an A/B test that showed promise and the company decided to keep it. AI allows for fast, nimble experimentation. Unfortunately, it reduces the cost of experimentation for both ethical and unethical tests.

There’s plenty of angst in society around whether AI is trustworthy. But I believe it’s more important to focus on whether the companies building on top of the AI models are worth our trust as they have their own set of incentives.

We’ve talked a lot about delegating shopping decisions to AI agents in past agentic commerce articles. However, the agentic commerce ambition will slow if consumers begin to suspect that AI shopping agents optimize for affiliate commissions instead of their interests. Most markets function because participants trust that incentives are aligned enough for transactions to happen. Everyone must trust each other; otherwise, nothing will get done.

Bad actors will slow AI adoption, and technological progress more broadly. Every high-profile incident like Phia makes consumers a little less willing to delegate decisions to AI.

I’m less concerned about whether the frontier labs are building ethical AI models. See, they have stronger incentives to invest in safety because they’re highly visible and reputationally exposed. The bigger problem is not the foundational layer. It’s in the vertical layer.

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