That door could decide which brands actually reach the customer.

Two Announcements, One Problem

1. Sierra announced the Personal Agent Protocol (PAP), which it’s developing with Meta and a first group of partners.

2. Decagon open-sourced its own PACT protocol the same day, and joined Sierra’s PAP working group instead of competing with it.

PAP’s approach is simple. The customer grants access. The business sets the limits. The agent can work through the website, the API, or the business’s own agent, with MCP as one likely route in.

PACT does the same job from the other side: a customer’s agent works with a business’s agent inside the customer’s own permissions, and the personal agent never touches the customer’s password. It’s not a from-scratch protocol either, Decagon built it on top of A2A 1.0 and OAuth 2.0, the same identity and authorization standards most of enterprise software already runs on. That’s a quiet but deliberate choice: it means PACT doesn’t ask a business to rebuild its auth stack, just to extend it.

Sierra names six partners already: Genesys, Instinct, Rocket, Shopify, Stripe, and Walmart. The v0.1 specification is planned for later this month. It’s early, but several serious companies are already building toward the same handshake.

Shopify’s own VP of Product put it plainly in Sierra’s announcement: the protocol is there to establish how agents and merchants work together at all, not just how fast they can. Stripe’s head of payments is quoted in the same post, which tells you the part of this stack everyone expects to matter most: not the handshake, the money that moves after it.

The Agent’s Answer Becomes the New Shelf

Consumer brands spent twenty years fighting for a place in Google search. Now an agent can stand between the buyer and the store, choosing which products to show and which to leave out.

The agent’s answer becomes the new shelf.

Without an interface for agents, MCP or something close to it, a business leaves that interaction to someone else. The agent works with whatever it can read off the website, and the brand loses its say in how its own products get described.

For me, “agentic-ready” comes down to one practical question: can an agent find your products, get accurate information, and complete a task inside the customer’s permissions?

I’d start with a simple test. Ask a personal agent about your own brand. See what it gets right, what it gets wrong, and whether you have any way to fix what it’s working from.

OpenAI Already Ran a Version of This

None of this is entirely new. OpenAI launched Instant Checkout inside ChatGPT back in September 2025, starting with a narrow, deliberate test: US users buying single items from Etsy sellers, with Shopify merchants and more regions promised to follow. Merchants paid a fee on completed purchases. Shopping through it was free for the user.

What’s worth watching is that OpenAI has since pulled back from running checkout itself, shifting toward a model where the merchant keeps the checkout and the agent just hands off to it. Owning the whole transaction inside someone else’s chat, payment, fulfillment, returns, support, turned out to be harder, or less wanted, than the original pitch made it sound. Processing a purchase is not the same job as being trusted with a return.

That’s the precedent PAP and PACT are walking into. A protocol can standardize permission. It says nothing yet about who ends up holding the money, or the customer, when something goes wrong after the sale.

A $20 Subscription Won’t Pay for a Personal Agent

The purchases it drives will.

I expect personal agents to earn a growing share of their revenue from the spending they send to merchants. Subscriptions can stay. Purchases open up another way to earn.

I expect agent platforms to build the payment layer underneath those purchases: card programs through issuing partners, a share of interchange fees, bulk discounts, a cut of whatever savings they generate. Revenue grows with the spending they handle, not just the seats they sell.

That gives every platform a reason to pool demand into one network instead of negotiating with every hotel, restaurant, and retailer one at a time. The network does that work once and hands the same terms to every agent connected to it.

For a brand, getting into the agent’s answer stops being a marketing question. Price, availability, and the cost of winning the order all sit on the table now, next to the permissions.

Can your business handle negotiated prices, track who actually brought the sale, and calculate what you owe them? Imagine a network offers to bring you 10,000 orders and asks for 10% of each one. You need to know your margin before you say yes.

My Forecast

The largest agent platforms end up as fintech businesses with real purchasing power, not just software companies with a chat window attached. Brands end up negotiating with networks that bring the customer, handle the payment, and ask for better terms in return.

Agents and fintech become one business. That’s agentic finance.

Where the Real Work Sits

Product data, pricing, orders, and commission tracking all have to connect, and the commercial agreement has to work inside the systems that already run the business, not next to them.

Picture the 10,000-order network deal going live. The agent network sends traffic, the orders land in the business’s own systems, and thirty days later an invoice shows up for 10% of gross, not net, with a different definition of “completed order” than the one the sales team uses internally. Nothing in PAP or PACT resolves that disagreement. The protocol governed the handshake. It had nothing to do with the invoice.

I’ve written before about the gap between what a model can do and what a company actually runs, and who gets paid to close it. This is the same gap, one layer further down. The protocol settles what an agent is allowed to do. It settles nothing about who checks that the numbers reconcile afterward, or who has standing to dispute them when they don’t.

That part shows up the first time a network claims a commission nobody budgeted for.

Watch the payment layer. The permission layer was never the hard part.

The headlines all led with the loss. Almost $42B for 2025. That number is scarier than it is true. $34B of it is a paper revaluation of older financing instruments, not cash that left the building. The number that actually matters sits further down the document, and it’s not the one anyone led with.

The Loss Isn’t the Story

Strip out the accounting noise and the operating picture is almost ordinary, by 2026 AI standards. Revenue grew 12x to $4.6B. Operating loss topped $8B, with $7.3B of that going straight into compute and infrastructure.

For comparison, OpenAI’s own 2025 operating loss, by an independently verified estimate, came in around $20.9B. Anthropic isn’t the outlier here. Every frontier lab is spending like this right now.

The Real Number

$518 billion. That’s what Anthropic has committed to spend over the next 7 to 10 years on cloud, chips, and data centers, and roughly 80% of it is non-cancelable or payable regardless of how much Anthropic actually uses.

The breakdown is specific. $111.1B to Google, running through July 2033. $110B to Amazon, through April 2036. $31.4B to Microsoft, through May 2033. On top of that, roughly $161.2B in Broadcom-related equipment commitments.

Spread evenly, that’s $50 to $75B a year, regardless of demand. Revenue needs to grow by another 10x just to cover the hardware bill that’s already been signed.

The Other Half of the Prospectus

The second half is where the risk sits. Two unnamed customers accounted for 24% of 2025 revenue between them. Many of Anthropic’s largest accounts aren’t locked into long-term contracts at all.

Costs are fixed for a decade. Revenue isn’t guaranteed for a single year. No corporation signs a ten-year contract for tokens, the way they might for office space or a cloud migration. Meanwhile open-source models keep getting cheaper, and Amazon and Google are simultaneously Anthropic’s compute suppliers, its investors, and its competitors.

This Isn’t a Software Company Anymore

Anthropic’s economics now read like an airline’s, or a factory’s. Enormous fixed costs, and one job: fill the capacity.

If demand shows up on schedule, operating leverage makes the margin look fantastic. If it’s late by even a couple of years, the data centers sit empty, and the bill comes due anyway.

I keep coming back to fiber optic cable in 2000. The cable got laid years ahead of the traffic it was built for. The traffic did eventually arrive, bigger than anyone had forecast. The companies that built ahead of it mostly went bankrupt anyway, Global Crossing among them. The technology was right. The economics of the specific companies that bet on it weren’t.

Why the IPO Is Actually Happening

Three weeks ago I wrote about the early signs of a bubble deflating. One of them: a lab that can’t close its next private round.

This prospectus explains why Anthropic’s IPO is close to unavoidable. No private market can absorb a $518B commitment. Obligations this size only get financed with public capital and public debt. The IPO is really the next funding round moving to the stock exchange, because the private round got too large to exist anywhere else, not Anthropic going public for its own sake.

The debt sitting behind data center construction right now, from Oracle to CoreWeave, is secured by exactly this kind of contract: someone’s non-cancelable obligation to pay. The entire chain rests on two or three labs being able to grow another 10x. Anthropic’s prospectus is the clearest look yet at what that actually requires.

What This Says About the Safety Debate

This number also answers a question I’d set aside until now. My own read, for what it’s worth, is that Anthropic is currently ahead of OpenAI, a new domain seems to open up every month where Claude beats very capable people at something. Calls for the lab to slow down already sounded strange against that backdrop.

The prospectus closes the question a different way. A company carrying 80% non-cancelable obligations cannot afford to slow down. It has to grow 10x or it has no way to pay for the hardware it already signed for. Arguing about anyone’s sincerity on safety stops being useful once you can just look at the incentives directly. $518B only pays off if a token never becomes a commodity, and any policy that raises the barrier for open-source or Chinese labs protects exactly that balance. It won’t slow down Google or OpenAI, they have more capital to absorb the wait.

What I’d Actually Watch

Whether Anthropic pulls this off, I don’t know. The pace of revenue growth favors them. The contract structure doesn’t. Press coverage has floated an IPO valuation near $2T.

I’d ignore the valuation headline and watch two numbers instead, in every filing from here forward: whether the two-customer revenue concentration keeps shrinking, and whether the share of revenue locked into multi-year contracts keeps growing. Those two numbers will say more than any single quarter’s growth rate.

Buying Anthropic stock here means buying a stake in a chain of data centers and a bet that someone fills them in time. It’s not a stake in an 80%-margin software business, whatever the pitch deck says.

Pat Grady recorded this one for Boston College’s Investment Committee, an LP and his alma mater, after they asked for his read on AI. His partners heard it and told him to post it. He did, on his own account, two days later. Worth reading slower than a quick LP briefing usually gets.

The thesis is blunt. Companies building on top of AI models have to reinvent themselves every four months, because the platform underneath improves faster than anything built on it can keep up.

The internet and mobile changed how information moves. AI changes how it gets processed. The last shift this size was electronics moving to silicon in the 1960s. That comparison is Pat’s, not mine, and I think it holds.

Three Points, Not Ten

He marks three actual turning points.

1. ChatGPT, November 2022. Scale reached everyone at once, for the first time.

2. o1, end of 2024. Reasoning arrived, a model that thinks before it answers.

3. Claude, autumn 2025. Agents that hold a long task instead of answering one question and stopping.

The first two produced faster horses. The third produced something that finishes the job on its own. Zoom is the closest parallel I can think of. Video calls existed for years before Zoom. People just didn’t trust them yet.

1. The Implementation Gap

Between what models can already do and what companies actually run in production sits a canyon, and every applied AI company is living inside it right now.

The labs are closing that gap themselves. Price cuts on the API are one lever. The more direct one is people: armies of forward deployed engineers sitting inside Fortune 500 accounts, building the application on the client’s own tokens. What used to be a niche job title is now a lab strategy.

2. Expensive Knowledge Work Gets Cheap

Coding. Cybersecurity. Medicine. Finance. Accounting. Sequoia expects a giant to show up in each one.

One portfolio company closed last year at $200M in revenue and is on pace for $700M this year. In pharma, drug candidates that used to take years to shortlist now get found with a prompt in days.

3. In-House Fine-Tuned Models, Argued Both Ways

This is the implication I’d slow down on, because I watched two smart people argue it in opposite directions in the same week.

Sequoia’s case: companies are shifting load off frontier models and onto their own fine-tuned open-source models, for a plain economic reason. A vendor sells you three points on the price to quality curve, big, medium, small, and your actual workload almost never lands exactly on one of them. Your own model ends up cheaper and a better fit. Cheap tokens don’t mean everyone runs straight to OpenAI.

Boris Cherny, who leads Claude Code at Anthropic, made the opposite case on Lenny’s Podcast earlier this year. His rule: always bet on the general-purpose model, don’t use small models, don’t fine-tune. Separately, on scaffolding specifically, engineering built around the model rather than the model itself, he put a number on it: 10 to 20% improvement at best, usually wiped out by the next model release. You’re better off waiting.

Right now the trajectory backs Boris. But a countercurrent is building too: open-weight models, and companies that simply won’t send their context outside their own walls. One of our own clients, a large pharmaceutical company, doesn’t allow any external AI tool at all.

I wouldn’t bury either side yet. The more valuable and sensitive the corporate knowledge involved, the stronger the case for keeping the model next to the data instead of sending the data out to a frontier model.

4. Specialized Architectures

The world doesn’t need another general purpose lab. It needs labs built around one domain, or one new architecture.

TypeSafe AI showed up with Jev, a classifier model. Most corporate machine learning reduces to picking between a handful of options, and a classifier does that 100 times cheaper than an LLM. Grady’s own account puts Jev at $100M in run rate revenue within seven days of launch, a figure I haven’t seen independently confirmed yet. The Information has separately reported investor talk of a $10B valuation.

5. The Organization Becomes the Lab

The best people get pulled off day to day operations and handed room to break and rebuild the product, coordinated through AI instead of layers of managers.

Otherwise nothing gets rebuilt in four months. It just gets planned in four months, which is a different thing entirely.

The Number That Made Me Do a Double Take

Sequoia’s new normal is two rounds back to back on the same company. One led by a partner who builds alongside the founders. A second, a month later, that’s pure capital.

Sequoia counted seven of these over the past twelve months. The average first round priced the company around $110M. The second round, about a month later, priced it around $3.4B.

Pat says he’s never seen anything like it. In the same briefing, he calls it a bubble and, minutes later, the fastest and largest technology transition in history.

I think both are true at the same time.

Where the Real Risk Sits

Sequoia reads the implementation gap as a market opportunity for startups.

I read it as a market for accountability. The gap gets closed by specific people inside a specific company, people who signed their name to the outcome and who take the reputational hit the next time the floor shifts under them.

You cannot out compute a lab. What a lab will never take on for you is your context, your data, and the part where somebody has to explain why it didn’t work.

I’d ask my own company one thing: what we rebuild in the next four months, and whose name ends up on it.

It found road bikes on Marketplace. For a restaurant reservation, it offered to connect OpenTable. When I asked it to publish on X, it initially refused. Then I connected a third-party social publishing service through its MCP tool, supplied the access credential it requested, and the post went out from the conversation. I did not connect bank accounts.

The observation is narrow but useful: the available tool changed what the agent could do.

Next I asked it for a face cream from a large CPG brand. It returned about ten options. What I found was that, in my test, it had no dedicated MCP connection to the brand; it read the company’s site as a web page. That is not proof that the brand has no agent integration anywhere, or that this is how every query is answered. It is what I saw in this run. There is a sharp difference between an authorized tool with structured outputs and a browser interpreting a page designed for people.

In late 2024 I tried Anthropic’s early Computer Use and could not find a reliable use case: it would not even sign in to a site in my test. My current Meta test is a very different experience. The shift reminds me of what I had assembled with OpenClaw: the value is not a single trick but the combination of tools, context and an agent able to carry out a task. That also raises a question for companies whose products the agent sees only from the outside. What data does it read, which version is current, and can the seller recognize the resulting order?

I have been following this for more than a year. The scenario I have been working with is that discovery and checkout in one conversation can weaken the traditional storefront, that the retail funnel can collapse into a dialogue, and that structured product information then determines whether a product is even considered. My new test gives that argument a concrete boundary: one service became actionable through a tool; a CPG catalog was still read as pages.

This is not merely a race to expose an API. In CPG and pharma, even a deep enterprise integration is not a permanent moat against general-purpose agents. Teams should benchmark what a horizontal agent can do against their product and revise the strategy as it improves. For a consumer brand, a practical version of that benchmark is to ask the agent for its own products, check the variants and stock it returns, and repeat after a catalog update. Record both what the agent got wrong and which system supplied the bad fact.

Four records, not one funnel

An ad impression and click, a product variant, a tool’s authorization scope and a retailer order are distinct records with distinct owners. If one is mislabeled, the others cannot be inferred from it.

The State of Brand reported ChatGPT advertisers whose traffic labeled “organic” rose while ads ran and fell when campaigns stopped; one cited a fivefold increase. That chart does not show whether an ad changed an answer. OpenAI says ads are separated below answers and cannot alter them. A more mundane measurement fault deserves testing: ad clicks can carry oppref, which GA4 does not classify as paid on its own. Without campaign UTMs, they can sit with unpaid referrals. Improvado described the mechanism. It is an explanation to test, not a diagnosis of those accounts.

ChatGPT: two feeds, two very different contracts

In ChatGPT’s beta Ads Manager, an advertiser can set an objective, budget, location and platform surface. OpenAI documents CPM and CPC billing, including for eligible conversion-optimized campaigns. A retailer can upload a feed to make item ads. That advertising feed creates sponsored units; it must not be confused with the feed for product discovery. Ads appear below answers for eligible Free and Go users, not Plus, Pro, Business, Enterprise or Edu users. The auction uses conversation context, ad material, targeting and bids, with permitted personalization signals. A context hint does not guarantee placement on a given prompt.

The data boundary matters more than the campaign controls. OpenAI says an advertiser gets aggregated views and clicks, and conversion signals if it implements pixel or Conversions API measurement; it does not get the person’s chat. A click may visit a site, or the person may use Ask ChatGPT to discuss a particular ad. Neither event means the independent answer was bought. I would keep ad impression, tagged click, product lookup and retailer order as separate event types, then reconcile them with a holdout if the team wants to estimate incremental sales. Do not turn a referral spike into a claim about answer placement.

The access price is also a versioned fact, not a constant in the architecture. OpenAI confirmed a $200,000 initial closed-pilot commitment in February 2026; EMARKETER later reported $50,000 direct entry and a $10,000 Criteo partner threshold by June. These are historical entry terms, not a current universal minimum or rate card. Current eligibility, bids and partner terms need a quote. The billing documentation explains models, not a fixed price for a product category.

For product discovery, OpenAI’s Agentic Commerce Protocol (ACP) describes a separate route for approved merchant partners. Its CSV or JSON feed carries stable product and variant IDs, descriptions, links, images, price, availability and seller data. OpenAI recommends validating a sample, sending regular full snapshots and updating changes during the day. Each purchasable variant should have its own ID and matching price, image and availability. One parent SKU cannot safely represent every shade of a cream; a razor handle is not an interchangeable cartridge.

These are referential-integrity problems before they are merchandising problems.

A discovery feed gives ChatGPT structured facts, not a purchased rank.

OpenAI does not promise inclusion for every brand. Shopify catalog data is integrated and retailers including Sephora have connected for discovery, while direct-feed onboarding remains limited to approved partners. The retailer may own the inventory and customer relationship. A manufacturer should not assume its own data overrides the seller’s live listing. Costs here include field mapping, validation, stock freshness, local claims and identifier reconciliation; public documentation gives neither a universal inclusion fee nor a fixed implementation price.

Checkout crosses another boundary. OpenAI launched limited Instant Checkout with Etsy, then said in March 2026 that the initial model lacked flexibility and shifted focus to discovery plus merchant checkout in an in-app browser. Developer documents still describe checkout sessions for approved partners. In that path the merchant’s own systems evaluate tax, fulfillment, payment risk and order acceptance. The authoritative order state remains with the merchant. A product appearing in an answer is not evidence that an in-chat transaction is available for it.

Muse: inspect the actual path, not the announcement

Meta says Muse operates in a dedicated secure virtual computer with a browser, can navigate sites and forms, and asks the person before buying. Meta describes app connection choices, an audit trail and isolation from the person’s passwords and payment details. It says Muse can use Stripe’s Link, whose single-use card keeps the underlying card from the merchant. These are Meta’s design claims, not an independent security audit.

At launch, Meta reported Shopify catalog access. At Connect, it said it was adding Walmart, Sephora, Ulta and other retailer connectors, and named Shop Pay and PayPal for payments; an earlier announcement called Shop Pay “coming soon.” There is no public end-to-end specification showing every announced connector live for every SKU. Browser navigation, catalog access, a retailer integration and a completed payment are distinct paths. Before building around Muse, I would run a real SKU through each supported path and record which system supplied price, stock, seller identity, approval and final order ID.

Meta says private Muse conversation and secure-VM data do not flow into its ad systems. Existing social ads therefore do not establish a paid slot inside Muse’s answers. I found no published Muse-answer ad rate or universal catalog fee. For an integration plan, an unverified connector or hypothetical ad product should be marked as such, not quietly treated as an available API.

Grok: keep three products apart

X feed ads are an existing media product. Musk announced plans for ads in Grok answers in 2025, and reporting describes a limited Grok assistant beta within X Ads Manager. That beta assists campaign planning; it is not proof of a generally available sponsored recommendation in Grok answers. Grok Bot is different again: a work agent in beta for some paid users, with enterprise access on a waitlist. These cannot be modeled as one commerce or advertising interface.

The concrete shopping example is Gopuff Go, powered by xAI models inside Gopuff’s app. According to xAI, it uses Gopuff order history and demand data alongside text, voice and image capabilities to assemble carts and a visual shopping feed from Gopuff inventory. xAI says it is available in the US app, with UK expansion planned. In this design Gopuff owns inventory, customer context and the transaction; the model is a component. The case does not imply a Grok-wide checkout rail or a standard price for brands to appear in Go.

Build the boundary before scaling the pilot

I would begin with a source-of-truth map:

• parent product

• variant SKU

• approved claims

• image

• market and currency

• price

• seller

• live stock

• fulfillment

• order ID

Test cream shades and local claims, or handle-cartridge compatibility and pack size. Identify who owns each field, its update latency and the event that proves it changed. For ChatGPT, validate an ACP-compatible sample only through an approved partner route. For Muse and a retailer-owned assistant, find out which partner supplies catalog and order data before proposing a connector.

Where a platform supports it, MCP can expose narrow tools to an agent. It is not a universal integration switch or a way to buy an answer. A read-only product lookup and a store-availability query have different scopes from cart modification or order submission. MCP’s authorization specification describes OAuth-protected HTTP connections, and its security guidance warns about excessive scopes, confused-deputy attacks and token passthrough. Bind access to the user and resource, separate read from write, require the buyer’s approval for a charge, and retain an order audit. Do not hand an outside agent a general CRM or ERP credential. Public support for arbitrary MCP servers differs across consumer agents.

The test is operational: an agent encounters stale inventory, inconsistent variants and broken checkout steps faster than a person willing to work around them.

There is already an agent-native payment path for digital goods: the agent can be the buyer, but a marketplace still has to own the listing, payment and payout. That example makes the transaction boundary visible without implying that every consumer agent can use that route. Citrini Research’s hypothetical demand shock points the other way: if AI cuts incomes as it raises productivity, more efficient shopping does not automatically mean more customers able to buy. That is a scenario, not a prediction.

Enterprises do not buy a model in isolation. They need a result embedded in data, processes, security and compliance. The integrator’s job here is to make the path from product record to authorized order observable and safe to change. A larger ad budget cannot repair a wrong shade, an overbroad token or an order no one can reconcile.

That framing comes from a new Citrini Research piece. I wrote before about their earlier one, on white-collar work in 2028.

Four Kinds of Rent

The framework is simple. Between a person and the best available decision, there’s almost always friction, and business charges for that friction. The authors split it into four types.

Tedium rent: an hour on hold to cancel a subscription. Inertia rent: too lazy to shop a better rate or move your own money. Attention rent: we buy whatever caught our eye first. Asymmetry rent: a middleman knows the market better than we do and charges a fee for the search.

None of this exists for an agent. It’s not bored sitting on hold with the cable company for an hour.

It checks your subscriptions every month, re-shops your insurance every year, books flights directly with the airline instead of through whoever’s in between. It notices when your cash is sitting at 0% while a money market fund pays real interest, sitting right next to it.

I’ve already shared the savings this produced for me directly: on flights, on medical bills, on subscriptions nobody in my house was watching.

Who Actually Loses

So which businesses take the hit? Probably not the producers.

A hotel is still needed. Booking is needed less. Insurance is still needed. The comparison site that sells you insurance is needed less. A restaurant is still needed. The delivery aggregator sitting between you and it is now a question mark. What’s under threat is the layer of intermediaries sitting on top of the product, not the product itself.

Subscriptions work the same way. A lot of consumer LTV rests on the customer forgetting to cancel, or not wanting to make the call. Everyone has two or three of these hanging around right now. An agent turns that customer into someone who watches their own money every single day.

Two Things Worth Flagging

I noted two caveats for myself here, and both matter more than the headline.

First, the rent doesn’t disappear. It moves. Google collects attention rent today. Tomorrow, whoever’s agent picks the default option collects it instead. Booking loses margin, but there’s no guarantee that margin goes back to the hotel. It’s more likely to land with whoever owns the agent doing the choosing.

Second, the discount available to a savvy customer today exists specifically because most customers aren’t savvy. Credit card cashback is largely funded by people carrying a balance and paying interest. The cheap rate you negotiated is subsidized by the person who didn’t bother to negotiate.

Once everyone has an agent, everyone negotiates. The price gap collapses. Consumers gain less than it looks like from here, and intermediary margins fall by more. The winners sit below the interface entirely: payment networks, identity, cybersecurity, data, infrastructure, and the actual producers of the underlying good or service.

The Real Question

This might be the clearest idea in the agentic economy so far. AI doesn’t just replace someone’s job. It kills off revenue that only ever existed because a human couldn’t be bothered to click five more times.

The question I’d put to any consumer business: how much of your revenue disappears the moment every one of your customers has an assistant that isn’t lazy?

The Agent Also Knows Something Nobody Else Does

That’s the friction side of the story. There’s a second side, and it’s sharper: the agent doesn’t just remove friction, it learns something about you nobody selling to you has ever known.

Google knows what you searched. Amazon knows what you bought. Your bank knows how much you spent. None of them know the one thing that actually matters: how much you were willing to pay. An agent that negotiates on your behalf knows exactly that.

Nobody sells that data outright. Google and Meta have lived off data like this for twenty years without ever selling it, they sell access to you and keep the data itself as the moat. Whoever owns your agent will do the same.

But an agent has something nobody selling to you has ever had: it watched exactly where you broke. Where you agreed to pay more, where you walked away. Economists call this a reservation price. Retail has guessed at it crudely for years, Orbitz once showed Mac users pricier hotels, Staples varied prices by zip code. All rough estimates. An agent knows the real number.

Three Ways to Monetize What It Learns

The mildest: an anonymized panel sold to funds and brands, the same category as Mastercard’s SpendingPulse, every purchase and cancellation across 100 million people, aggregated in real time.

More useful: the agent discloses pieces of your own data on your behalf. Three years of clean history, low risk, ask for the discount. Same logic Tesla uses for insurance, data as leverage you deploy, not a good you sell.

The dangerous version: a seller pays the agent’s owner directly for access to your reservation price. Citrini’s four rents were about friction on the way to a fair price. This is different, here the price itself stops being set by the market and starts being set by whatever one specific buyer can be made to pay.

Whoever Holds Both Sides Wins Most

The strongest position belongs to whoever runs agents on both ends of a transaction, knowing exactly what a buyer will pay and exactly what margin a seller can still afford. That’s a market maker with the full order book. Amazon has already drawn antitrust complaints in the EU and US for using third-party seller data this way, before agents made the picture this detailed.

The Fiduciary War, Again

Finance fought this exact battle for a decade over the fiduciary rule: a broker only has to sell you something suitable, an advisor is required to sell you whatever is actually best for you. AI agents are heading into the same fight, except a financial advisor only ever saw your savings. An agent sees your entire life.

The real question is who owns this data, not whether it eventually gets sold. If you own it, your agent discloses your history only when that helps you. If the platform owns it, you pay close to the maximum you’d ever agree to, every single time. The gap between those two worlds is what the whole battle for consumers actually comes down to.

One stock idea this pushed me toward: Shopify. An agent can remove the storefront the way Muse’s recent traction suggests it might, but Shopify still sits underneath as the infrastructure, the catalog, checkout, payments, Shop Pay. At the current valuation, I’d rather sell insurance against SHOP through options than simply own the stock outright.

A team of researchers led by Güllich, Barth, Hambrick and Macnamara published the study this year in Science. They pulled career histories for Olympic champions, Nobel laureates, elite chess players and top musicians, then asked a question simple enough to state in one sentence and hard enough to answer that nobody had done it properly at this scale before: were the best adults also the best kids in their field?

The answer, across all four domains, was mostly no. In chess, the junior top ten and the adult top ten are close to two separate populations. The same pattern holds in sport. Whoever wins the youth championship is rarely the person still winning twenty years later.

Why early speed is a trap, not a signal

Early stardom has a clean, learnable recipe: pick one discipline as a child, specialize hard, put in disproportionate hours in that single lane, and progress fast. That recipe works. It is an excellent predictor of who is best right now, at age 14 or 16.

It is a poor predictor of who is best at 30. The study’s adult elite had a different childhood shape: several sports or fields rather than one, a late start in what eventually became their actual discipline, and a ramp that, for years, looked slower than their more specialized peers.

We built every selection system we have - junior academies, olympiad pipelines, corporate high-potential tracks - to reward exactly the signal this study calls weak: how fast someone improves early.

The mechanism the authors point to is not mysterious once you see it. Broad early exposure builds transferable skill and the ability to pick up something new quickly, rather than one deep groove. Trying several fields as a kid is also how you find the one you are actually built for, instead of the one you were enrolled in at seven. And there is a quieter cost to the narrow path: burnout. Someone who has extracted everything a single discipline can teach them by age twelve is often, by twenty, already tired of it.

None of this is an argument against depth. Depth is real and it matters enormously at the top of any field. It is an argument about sequencing: depth built after a period of genuine exploration outperforms depth built instead of one.

The signal we actually use to hire

I did not read this study thinking about sport. I read it thinking about hiring.

GPA. Olympiad medals. How fast someone ramps in their first year on the job. These are precisely the signals the study calls weak predictors of long-run performance, and they are also, almost without exception, the signals every corporate high-potential program and every junior-hire screen is built on. The federations, the gifted-student pipelines, the fast-track graduate programs, all of them are optimized against a KPI — junior medals, first-year velocity — that this research says barely correlates with who ends up at the top a decade later. And structurally, they never find out they are wrong: the people they filter out at twenty-two simply never show up in their data again. The machine that measures its own success has no way to see its own errors, so it never corrects itself.

It is worth noting that the most selective American universities have partly already priced this in. Their admissions process for a long time has looked past grades alone toward something closer to what this study rewards: evidence a student can independently carry a project through to real, recognized scale, and a broader, more integrated sense of who that person is, not only how fast they moved through a curriculum.

The practical version: a slow start, in a child or a new hire, is not a diagnosis. An early star is not a guarantee. What is worth measuring instead is whether someone keeps growing once the easy early wins run out, and whether their breadth shows up as actual finished things, not just as a longer list of interests.

Why this matters more, not less, with AI

I have written before that judgment is the one skill that does not get devalued as AI takes over more of the work. This study gives that argument a mechanism.

Agents are fastest at absorbing narrow, well-specified execution — exactly the kind of work the early-specialization path optimizes a person for. What stays scarce, and what gets more valuable as that narrow work gets automated, is the ability to connect several domains and judge whether an output is actually good. That is not a skill you build by optimizing one metric from age seven. It is built, the same way the adult elite in this study was built, by trying enough different things to develop a real sense of what good looks like across more than one context.

Selection systems built for speed will keep producing juniors who peak early. The people worth betting on long-term, in a chess federation or in a company, are more often the ones who look unremarkable at the first checkpoint and are still visibly climbing at the fifth.

Source: Güllich, A., Barth, M., Hambrick, D. Z. & Macnamara, B. N., Science (2025), DOI: 10.1126/science.adt7790.

Everyone talks about productivity: code shipped faster, tasks automated, hours saved. Mine showed up somewhere duller. The fine print.

The small wins came first

Readers send me tips on using ChatGPT or Claude to find discounts and cheaper fares, and it genuinely works, hundreds of dollars saved over the past year without much effort.

I built a flight-search agent with access to Google’s flight API, Amadeus, and a few other sources. It has found real discounts twice. On routes I already know well, it still doesn’t beat me, but the direction is obvious.

A friend used Claude after a Lufthansa strike wiped out a trip. The model read the airline’s own contract of carriage, then helped him build a formal claim against the airline and, separately, an Austrian hotel that wouldn’t budge on a refund. He got most of the money back. Thousands of dollars, from arguing with two companies at once, in a language that wasn’t his.

Those are nice wins. The bigger one showed up somewhere I didn’t expect to look.

Medical bills are a different kind of maze

American healthcare has a strange property: you almost never learn the price of anything before you receive it. Hospitals don’t publish rates. You get treated, and weeks later, bills start arriving from separate companies, each with its own logic.

Inside a single bill there can be dozens of procedure codes, drug codes, insurer rules, a deductible, an out-of-pocket cap, a filing deadline, and a separate appeal deadline running on its own clock. Checking whether a bill is even correct requires knowing medicine, insurance, billing, and contract law at the same time. Almost nobody does.

The advantage in that system has always sat with the hospital and the insurer. A hospital has a billing department. An insurer has actuaries and staff lawyers. A patient gets a 20-page PDF and a due date.

What changed when we handed it to Claude

My son and I ran a stack of real bills through Claude this summer. The first pass was slow. Claude broke each bill down by code, cross-checked procedures against what the insurance plan actually covers, flagged mismatches, and pulled the specific deadlines the hospital or insurer was legally bound to meet. Then it drafted the appeal letters, and where a signature and a stamp mattered more than an email, it sent them, printed, signed, and mailed through an API.

The second pass went faster. That part alone tells you something about where this is headed.

The results were concrete, not theoretical. On one bill, the hospital simply stopped responding, missed its own procedural deadline, and the claim was dropped. On another, the insurer admitted after resubmission that it should have covered a charge it originally denied, and paid it. Several thousand dollars stayed with us instead of moving to whoever wins by outlasting the other side’s paperwork.

The oldest advantage just got smaller

Medical billing, insurance contracts, the tax code, bank fee schedules, most consumer-facing legal text, these systems didn’t become complicated by accident. A corporation can hire a department to navigate its own rules. An individual almost never can.

That asymmetry is exactly what a model reading a thousand pages in minutes starts to erase. I see the same shape in the US tax code, dense enough that even a good CPA specializes in one corner of it, where missing a single provision can cost tens of thousands of dollars. The tool that reads the whole thing without getting tired is new. The rules it’s reading are not.

A new category of business is forming on top of this: refund services for canceled flights, generalized to medical bills, insurance denials, and tax filings. An agent that takes the documents, checks them, drafts the challenge, tracks every deadline, and escalates when it has to. A handful of $200-a-year subscriptions that actually defend your interests already look like real value, even before the tooling around context and evidence gets any better than the manual, trial-and-error process we ran this summer.

I don’t think this stays quiet for long. Complexity that size protected real revenue for a long time, and the industries built on it won’t watch a model read through it for free. My guess is new rules first, then restrictions, then eventually a test case somewhere designed to make an example of someone for using one of these tools against a system that was never supposed to be readable.

That’s usually how it goes when a machine starts winning arguments that used to require a department.

I’ve written before about an 11-hour trip with my son, after which my Oura ring showed lower stress than a normal day at the office. About Shadow Mode, and how Lemonade prices FSD miles at half the cost of regular ones. A couple of times, FSD may have literally saved me - I posted a clip of the car braking and pulling right seconds before a driver flew past, trying to overtake through a median island.

For me, the question of whether the technology works is basically settled.

What the closest competitor’s numbers show

Recently I looked at GM’s numbers, and they’re a good read on where the closest US competitor stands.

GM’s hands-free system, Super Cruise, has existed since 2017. Today it’s installed on roughly 750,000 vehicles. Users have driven more than a billion miles on it. The supported road network has grown to 600,000 miles.

New buyers get Super Cruise free for three years.

Then the interesting part happens.

Once the free period ends, only 30-40% of users keep paying the $25 a month.

Six or seven people out of ten spent three years driving with a technology that lets you keep your hands off the wheel, and decided it wasn’t worth $25 a month. Nine years after the product first hit the market.

We keep overestimating adoption speed

I think we constantly overestimate how fast people adopt new technology. Especially now, with AI.

Technology can be ready long before people are ready to change their habits. And at that point, the problem stops being an engineering one.

Letting go of the wheel is an act of trust.

The same pattern shows up with AI agents

I see exactly the same thing with AI agents. If an agent does the work and you check every action afterward, there’s barely any time saved. Sometimes there’s even more work than before.

The real payoff starts the moment you’re willing to hand over not just the work, but part of the control.

The bottleneck for the next few years, in cars and in companies, is trust: how fast people become willing to hand over control.

Stop thinking of an agent as something you hand a task to. Start thinking of it as a colleague: a role, a memory of everything you’ve worked on together, its own computer, the ability to hand work off to other agents without you in the loop. Close your laptop and the work doesn’t stop.

I spent most of the rest of the session doing something else entirely - recognizing our own roadmap in someone else’s slide deck.

We’ve been building toward this at Customertimes for a while now, using Cursor (Grok Bot isn’t in the stack yet) on Salesforce Spiff and SPM implementations. Four stages of delivery are already heavily automated: presales scoping, where a prospect gets a rough estimate in under an hour instead of days; solution design, where a phased build plan gets pulled straight from the client’s requirements; the actual build, where agents handle much of the system configuration themselves; and documentation and training, generated from the implementation itself instead of written up after the fact by whoever drew the short straw.

That’s the part I keep coming back to. Grok Bot already ships with its own memory. What we’re building on top of that is a second layer, our own institutional knowledge, and feeding it back into Cursor is what makes an already smart agent noticeably smarter, with a memory that goes well beyond what came out of the box.

Any enterprise software shop accumulates hundreds of small lessons over the years - the kind that never make it into a wiki page, that a good consultant just knows in their bones. Which steps have to happen in a specific order, and which ones you can reorder without consequence. Which shortcuts save you a day now and cost you three weeks later. Which configurations look perfectly fine in the sandbox and quietly fail the moment they touch production. And how you actually check that the result is right, as opposed to how the documentation says you’re supposed to check it.

For decades that knowledge lived in people’s heads. Every new team learned it the same way the last one did - the hard way, usually on a client’s dime.

We can capture it once now. Every agent on every project gets to start from what the last twenty implementations already taught us, instead of starting from zero.

On our first production run with this model, we cut both effort and delivery time by a wide margin, and in the same project reconciled several thousand compensation statements against the client’s actual payout data without anyone doing it line by line.

There’s a boundary worth being explicit about here. Agents get the width of the work. Humans keep the judgment - the business calls, the interpretation of ambiguous rules, the sign-off before anything goes live. That part hasn’t moved, and I don’t expect it to for a while.

Cursor’s right that the paradigm shifted. Where I’d push back a little is on what matters most about it, for a consulting business specifically.

The easy part to build is agents that work continuously, collaborate, and take on harder tasks without supervision - every vendor will have a version of that within two years. The harder part to copy is what happens when that machinery runs into twenty years of implementations, mistakes, edge cases, and the judgment calls only a room full of senior consultants would know to make.

For a professional services firm, that accumulated experience has always been the asset, even when nobody could put a number on it. It just used to live in people’s heads, in scattered documents, in whoever happened to be staffed on a given project - useful to the people in that room, gone the day they left the company.

Now it doesn’t have to stay there. It can sit underneath every agent, on every engagement, from day one.

Good agents plus knowledge nobody else has. That’s the combination I think actually moves the number on what consulting costs to deliver, not the agents by themselves.

The Coldcard Q looks like it was designed by cryptographers with a persecution complex. Every part of it answers the question “what if someone actually tries to steal this?” Some of what’s packed in:

• Fully air-gapped: no Wi-Fi, no Bluetooth, no USB. Transactions get signed only via microSD, QR code, or NFC.

• A Secure Element that holds private keys in an encrypted zone, inaccessible even if someone cracks the case open.

• Two independent chips from two different manufacturers, with cryptographic functions split between them, compromising the device means breaking two separate hardware architectures at once.

• Anti-tampering: the device logs attempts to open it. Pull it out of a safe and see the indicator light in the wrong color, and you know someone tried.

• Open-source firmware, so anyone can audit it themselves instead of taking the company’s word for it.

• An e-ink screen and a mechanical keyboard - no extra sensors, no side channels to leak through.

• AA batteries instead of a built-in one: you can power it down completely and know it’s silent, and batteries won’t leak and ruin the internals over time.

• A manual entropy input on the random-number generator, you’re supposed to roll dice and type in the result yourself. I was surprised how many “hack the standard RNG” videos are floating around on YouTube; the fear behind this feature isn’t paranoia, it’s precedent.

• Anti-supply-chain protection: sealed factory packaging, with firmware checked for tampering before use.

• A Brick-Me PIN and a Duress PIN - instant wipe, or a fallback into a fake wallet.

• Decoy wallets, where you can stash a small real balance to hand over if someone forces the code out of you.

The Q looks like a retro calculator on the outside, and it actually runs a calculator mode and a built-in flashlight. The discreet mode isn’t a design flourish, it’s part of the threat model. It’ll also hold encrypted secure notes: backup phrases, instructions, seed fragments, key reminders locked behind the same PIN mechanism guarding your coins.

It’s not a comfortable device to use day to day. But it’s a working manual on cryptographic paranoia, built into a case with AA batteries, a snitch LED, a calculator, a flashlight, and notes nobody can read without your PIN.

How “don’t trust, verify” broke at step one

Coldcard built its reputation on Don’t Trust, Verify. Open-source code, reproducible firmware, the ability to check everything yourself.

The problem turned out to be in the very first action: generating the seed phrase.

Coldcard gave you three options: trust the hardware random number generator, add dice rolls on top of it, or build the seed entirely from dice. That last option meant not trusting the hardware at all. Almost everyone picked the default anyway - press the button, trust the device.

In 2021, a bug shipped in the firmware that disabled the hardware randomness generator. In its place, the device fell back to MicroPython’s weak software generator. No warnings appeared. Users got a normal-looking set of 24 words: checksum passed, wallet restored, and the words could sit on a metal plate for years.

But a seed’s security isn’t in the words, it’s in the entropy, the size of the space they were drawn from. On some devices that space shrank to roughly 40 bits, on others to 72. Still huge numbers for a person, but reachable for a well-organized brute-force attack.

Galaxy Research estimated around 1,367 BTC, close to $90 million, was drained from vulnerable addresses.

That’s when I went and checked my own Coldcard. Sure enough, my version was exactly the vulnerable one. Rolling dice for honest entropy was something I’d been too lazy to do back then, I pressed the button like everyone else. Lucky twice over: there wasn’t much money in the wallet, and it wasn’t touched.

There’s a separate story here about AI’s role. Coinkite itself suggested the attacker may have found the bug through AI analysis of the open-source code while the company’s own AI audit, run just weeks earlier, found nothing. The defender has to catch every bug. The attacker only needs one.

For years, the assumed threat to cryptography was quantum computers. But nobody broke Bitcoin here, and nobody brute-forced a full 256-bit key.

What got broken was weak generation, caused by an ordinary code bug, quite possibly found and scaled with the help of AI.

Reproducible builds proved the firmware matched the source code. They never proved the code itself was safe.

Why the paranoia was earned, the numbers keep climbing

Since that first estimate, Galaxy Research has tracked the fallout across three separate attack waves. The first hit on July 30, draining roughly 1,083 BTC from about 1,200 addresses in just 41 minutes. A second wave surfaced days later, and a third - with a noticeably different transaction pattern, possibly a different actor working the same weakness pushed the running total to 1,367 BTC across 4,585 addresses, worth close to $88.6 million at time of writing. Galaxy says the activity may still be ongoing, and that coins taken so far have sat untouched in attacker-controlled addresses rather than being moved or cashed out. The drained coins had been sitting dormant for years on average, this looks like long-term self-custody holders, not exchanges or institutions.

Some reporting says the Mk4/Mk5/Q line escaped this particular bug entirely; Coinkite’s own advisory says otherwise, specifically flagging Q seeds generated before firmware 1.5.0Q or Edge 6.6.0QX. The honest answer is: it depends on exactly which firmware your seed was generated under, and that’s not something to take on faith from a headline, it’s something to check on the device.

Which brings the story back to that dice-roll entropy input. It’s the one feature on this list that isn’t about defending the device from an attacker, it’s about not needing to trust the device’s randomness in the first place. Coinkite itself now says seeds backed by at least 50 independent, honest dice rolls aren’t exposed to this flaw at all. I used it going in mostly because it seemed like the “hardcore” option. Turns out it was the only one actually built for a world where the manufacturer can ship a bug like this and where the defender has to catch every bug, and the attacker only needs one.

I’m the cofounder and managing partner of Customertimes, a 1,000-person firm that helps Fortune 500 companies implement AI. Before this, I spent almost a decade in pharma leading technology transformation projects. That combination means I’m in the room for two different conversations most people never see side by side: the public one about layoffs, and the private one about what AI actually costs and delivers.

Almost none of the CFOs, CIOs, and CEOs I talk to every week tell me, privately, that they’re replacing people with AI. The gap between that private conversation and the public narrative is exactly why the hysteria around AI layoffs reads so strangely from the inside.

What’s actually happening: AI as cover, not cause.

Long before generative AI, companies were already automating repetitive work through robotic process automation. The underlying goal hasn’t changed: find the inefficient process, automate the routine part, free people from work nobody wanted to keep doing manually.

What I actually see now is companies using AI as the explanation for restructuring that was coming regardless. A process stops making sense, leadership had already planned to reorganize it. There’s nothing unusual about reorganizing for efficiency, businesses have always done that. The problem is honesty: a lot of these companies aren’t straight with their employees or shareholders about what’s actually driving the decision.

The money doesn’t go where people assume.

The public story is simple: cut headcount, pocket the savings, hand leadership a bigger bonus. What I see behind the scenes is messier.

CFOs and CIOs tell me constantly that they underestimated token costs, and plenty of organizations have burned through a year’s AI budget in a few months, some describe it as “token maxing,” spending accelerating far faster than the productivity gains that were supposed to justify it. On top of that, companies are investing heavily in keeping proprietary processes and trade secrets out of public models, building separate internal AI infrastructure specifically to protect that knowledge. None of that is cheap, and none of it shows up in the “AI saves money” headline.

So when a company announces layoffs in the same quarter it reports record profit, the assumption that the savings landed in executive pockets is often just wrong. Infrastructure, token bills, and licensing eat a lot of it before it ever reaches a bonus pool.

People want to see where the money went and they’d back a fair split.

In the survey behind this interview, 86% of adults said companies saving money through AI should pass some of that savings on as lower prices for consumers. That’s a reasonable bar, and it’s one most companies aren’t answering at all.

My actual recommendation to the executives I work with: publish something simple showing where the AI savings went. Into employee bonuses. Into lower prices. Into reinvestment. Pick one, be specific, show a number. When a company stays silent instead, it creates a vacuum and a vacuum always gets filled, usually by the least charitable explanation available. Stay quiet long enough and the conspiracy theory about you gets written whether you show up to correct it or not.

What AI can’t take off your plate and what it can hand back to people.

AI is genuinely excellent at repeatable, structured work. You can automate a CEO’s earnings presentation with an AI avatar today. What you can’t automate is accountability, the decision to restructure, the responsibility for the people affected, the judgment call on where the savings actually go. That’s still a human in a chair.

And I’d push back on the assumption that AI mostly makes people less necessary. At Customertimes, we invest heavily in AI training, and I’ve watched testers and business consultants pick up new AI-driven skills in a matter of weeks. The pattern I actually see isn’t people becoming less valuable, it’s people becoming capable of delivering a lot more than they could before.

If your company has gone through AI-linked layoffs, were you ever told where the savings actually went or did you have to guess?

Run teams with French, Irish, American, and Asian colleagues at the same time, and you learn fast: the same words don’t land the same way twice. What one person hears as straightforward feedback, another hears as an attack.

Trust works the same way. Some people close a deal on a handshake. Others won’t move without a signed document for the smallest commitment. I’ve watched trust get built on completely different foundations, and none of it fits neatly into a company policy.

Persuasion is where this gets personal.

I used to think that being right was the finish line. It isn’t even the starting line. The real test comes when you have to get other people to accept a view you’ve built from your own experience and judgment, and arguing just to prove yourself correct is a bit like shouting louder in a crowded room. It might grab attention for a moment, but if you don’t understand what’s actually driving the other person, you’re just burning energy.

Everyone carries a set of beliefs shaped by their environment, their history, and a hundred invisible threads connecting them to the world. If you actually want to shift someone’s thinking, you have to see things from where they’re standing first, get at what’s motivating them, not just what they’re saying out loud. That takes patience and real listening: asking questions, sitting with the answers, digging for the substance under the words instead of arguing with the words themselves. Only once you genuinely understand their position can you start shifting it by choosing arguments that actually reach both their reasoning and their instincts.

This matters even more with people who are older or more experienced. Their views are the product of years of testing, failing, and winning, they don’t move for someone who’s merely correct. Influencing that kind of conviction takes tact, a long timeline, and sometimes swallowing your own ego to get there. This is the same instinct Erin Meyer’s The Culture Map names directly: in some cultures, open disagreement is just how ideas get tested while in others, it reads as a threat to the relationship itself. You don’t persuade by winning the argument. You persuade by knowing which map the other person is reading from before you open your mouth.

Hiring follows the same logic. Employment-at-will in the US gives you room to move fast and course-correct. Try that same instinct in much of Europe, where bringing someone onto a small team is a slow, deliberate commitment, and you’ll misfire badly. None of this is written in an HR manual. It’s absorbed, the way you absorb a language.

So when Anthropic published new research on how Claude behaves across languages, it read like my own team to me.

Anthropic looked at roughly 300,000 real conversations and broke down the values Claude expresses along a few axes - one of them running from warmth to rigor. In Hindi and Arabic, Claude leans warm: encouraging, affirming, generous with praise. In Russian and English, it leans rigorous: it questions assumptions, pushes on details, and asks you to back up your claims.

Nobody sat down and coded that in deliberately. Claude absorbed it from the same place it absorbed grammar and vocabulary: from millions of human conversations in each language, each carrying its own default manners. Russian directness. The softer diplomacy built into Arabic disagreement. American ease. The training data didn’t just teach the model words; it taught the model how people in that language typically choose to disagree, encourage, or hold back, the same unwritten rules I’ve spent years reading in my own teams.

The researchers gave a version of the example I’d give from my own work: show the same business plan to two people, one working with Claude in Hindi and one in Russian, and they’ll likely walk away with different impressions of how good that plan is, and it’s not because the plan changed, but because the tone of the assessment did.

That’s the part that stayed with me. I stopped expecting one management style, or one style of persuasion, to travel cleanly across every office years ago. I didn’t expect to find the same pattern showing up in a language model trained on the collective output of the people I was managing.

We built the culture into our teams by living inside it for years by learning, the hard way, that being right and being convincing are two different skills. Claude built its version by reading us. Either way, the map Erin Meyer drew for humans turns out to describe something real happening one layer down, in the machine that learned to talk by listening to all of us at once.

Claude Science

On June 30, Anthropic released Claude Science, an AI workbench for scientists, combining databases, coding tools, compute, and research workflows in a single environment.

The problem it’s solving is real and specific. Anyone who has done serious computational research knows the daily grind: dozens of databases, each with its own schema and query language. You’re not doing science, you’re doing logistics.

Claude Science aims to remove that friction. The platform comes preconfigured with access to more than 60 scientific databases and includes specialized tools for genomics, proteomics, structural biology, single-cell research, and chemistry.

The architecture is built around full provenance. Every output includes the exact code, execution environment, plain-language description, and message history that produced it, enabling reproducibility months later. That last part matters more than it sounds. Reproducibility is science’s chronic illness, and Anthropic is aiming directly at it.

A coordinating agent hands tasks to specialist sub-agents, while a separate reviewer agent checks every citation and calculation and fixes errors as it goes. The reviewer is doing something that should have existed years ago: internal peer review before the paper leaves the lab.

Data stays on the researcher’s own infrastructure. Large and sensitive datasets go nowhere, only the context needed for each analytical step is sent to Claude.

One early case study: Manifold Bio uses Claude Science to nominate targets for experiments evaluating surface expression, trafficking, and safety for each tissue and target, then ranking candidates according to criteria learned from their own internal data. They say ordinary coding assistants can’t do this end-to-end.

Claude Science is not a new AI model. It runs on the same Claude models available to paying subscribers, including Opus 4.8. The bet here is entirely on workflow, not raw model capability. That’s a more honest position than most AI product launches take.

The bigger move

But the product wasn’t the main story.

Alongside Claude Science, life sciences leaders Eric Kauderer-Abrams and Jonah Cool announced that Anthropic would run its own internal drug discovery program, targeting neglected diseases that traditional biopharma companies wouldn’t pursue.

This is the unusual part. The standard playbook is: AI company builds tools, pharma companies do the drug hunting. Anthropic is collapsing that division.

Kauderer-Abrams said the reason is simple: “We need to live it along with all of you. We believe in the power of tight feedback loops, and there’s no substitute for having our own experiences alongside you all in the trenches trying to develop drugs.”

There’s a logic to it. If you’re selling workflow tools to biopharma, having your own programs gives you credibility and a real feedback loop. You can’t build the right instrument if you’ve never played the music.

As a public benefit company, Anthropic says they “can choose programs on patient benefit, including work the commercial market overlooks.”

That framing - public benefit, neglected diseases, the patients the healthcare system discards first is deliberate. The healthcare system is optimized for scale and standardization. Rare diseases get cut early. Not enough patients, no profitable target, but the burden for those living with it is completely real.

The infrastructure behind the bet

In April, Anthropic acquired Coefficient Bio, a stealth New York-based startup, in an all-stock deal valued at just over $400 million. The team is fewer than 10 people, most of them former Genentech computational biology researchers, and joins Anthropic’s Healthcare and Life Sciences division.

Coefficient Bio’s co-founders, Samuel Stanton and Nathan C. Frey, both came from Prescient Design, Genentech’s computational drug discovery unit, where they worked on experimental design for scientific discovery and contributed to projects including Cortex, a modular deep learning architecture for drug discovery.

Anthropic has also opened wet labs to run its own basic research. This is not a software company making announcements about biology. They’re building actual laboratory infrastructure.

Kauderer-Abrams was hired with an explicit mandate: make Claude the dominant AI model in biology. “We want a meaningful percentage of all of the life science work in the world to run on Claude, in the same way that that happens today with coding.”

The financial context: Anthropic is currently valued at $965 billion, more than any health company except Eli Lilly, and has filed confidential IPO paperwork.

The honest assessment

I’m watching all this with genuine interest, but I’m staying clear-eyed.

OpenAI launched GPT-Rosalind in April 2026 - a model specifically fine-tuned for biological reasoning. OpenAI’s own LifeSciBench, built with 173 PhD scientists, found that even the best-performing model cleared only 36.1% of real research tasks. OpenAI’s own life sciences lead acknowledged that AI cannot yet create new disease treatments on its own.

The broader picture is harder. Since 2019, roughly $60 billion has poured into AI drug development. Around 175 programs have reached clinical trials. Not a single FDA-approved drug has come out of it yet.

The harder problems- training models on ambiguous biological ground truth, building causal understanding of disease, identifying the right patients for the right drugs are still in early innings.

What Anthropic has that most of the earlier wave didn’t: a foundation model that genuinely reasons, wet labs creating real feedback loops, a team with Genentech-level computational biology credentials, and the financial firepower to stay in the game long enough for it to matter.

The bet is that combining foundation model reasoning with wet-lab feedback loops and biotech operational expertise can compress timelines across the board. Whether the ten-fold acceleration materializes will depend on how fast the training data problems get solved, and whether the models can learn to go beyond human knowledge by closing the loop with experimental results.

That’s the question worth watching. Not the product launch. The question of whether Anthropic can change the picture in a space where the rest of the industry has so far fallen short.

I’m curious what you think. Is this move different or are we watching another well-funded wave that will break the same way the last $60 billion did? Drop your take in the comments.

It’s the primary interface through which dozens of my processes flow, the console where I spend a genuinely embarrassing share of my waking day. I don’t open it to do a task anymore. I open it the way you open a laptop lid: it’s just where the work happens.

And this morning, it wasn’t there.

X filled up within minutes with jokes about a forced digital detox - people half-celebrating the sudden, unasked-for vacation. I laughed too. Then I went quiet, because the laughter was covering something that wasn’t actually funny. I had built so much around a single endpoint that when it blinked, a meaningful slice of my day blinked with it.

We’ve Seen This Movie Before

The feeling was oddly familiar, and it took me a minute to place it. Then it landed: this is the early SaaS era all over again.

I remember the first wave of serious Salesforce rollouts. For a lot of companies, Salesforce didn’t stay a “nice CRM.” It became infrastructure: the spine that sales, support, and reporting all hung from. Which was fine, right up until it wasn’t. When something broke and the red indicators lit up on the status page, trust didn’t erode gradually. It collapsed. Clients got nervous. The market reacted. Stocks dipped on a status-page incident, because everyone understood, all at once, how much was riding on a single dependency.

Anthropic is sitting in roughly that same spot now. The growth is genuinely fantastic. But growth like that drags a heavy companion behind it: responsibility for reliability. In the consumer world you can ship fast and apologize later. In the enterprise world, trust isn’t won by a brilliant demo, it’s earned over years of boring, uneventful uptime. One great quarter of capability does not offset one bad morning of unavailability, at least not in the part of the brain that signs renewal contracts.

That’s not a criticism of Anthropic. It’s just the tax that comes with becoming critical infrastructure. Everyone who gets there pays it.

But the outage pushed me toward a sharper, more personal question, the one that isn’t really about Anthropic at all.

How Dangerous Is a Single AI Operating System?

How exposed am I, actually, building my work around one AI OS?

The question got sharper because of something else that happened recently: a new model I’d been admiring, that I’d had maybe a couple of hours to fall a little in love with got blocked before I’d even finished exploring what it could do. So now I have two failure modes stacked on top of each other: the platform can go down, and the capability I’m depending on can simply be withdrawn from under me. Neither of those is in my control. Both of them are load-bearing.

Here’s my actual stack, laid bare:

• Claude Code - my primary tool. The console.

• Codex - used almost exclusively for code review.

• Cursor - the thing that first opened the door to vibe coding for me, and now my familiar fallback.

• Gemini Coding Assistant - haven’t even tried it yet.

And the unsettling part: I have colleagues already shipping major Salesforce projects almost entirely through Cursor, no separate developers, no separate testers. The whole pipeline collapsed into one person and one tool. Incredible leverage. Also a single point of failure wearing a cape.

So the logical question writes itself: if one platform goes dark, how fast can I actually move to another and keep working?

Because the honest answer this morning was “slower than I’d like,” and that’s the gap that matters.

Maybe the real answer isn’t picking a better single tool. Maybe it’s refusing to have a single tool at all - building my own console and, more importantly, my own skills around a panel of several tools. Including, crucially, ones that don’t live on anyone else’s cloud.

Which Is Where the Home Lab Comes In

Almost a year ago I wrote about a half-serious idea I couldn’t shake, the Emergency GPT Kit. The premise was almost post-apocalyptic: a compact, offline box of local models you could pull out when the internet was gone, the cloud was off-limits, and your phone was just an expensive dead weight. A first-aid kit for the mind. The last library on earth. I framed it as a story about survival and taking back control, and I’ll admit I half-filed it under “fun thought experiment, no promises to build it.”

And that morning it stopped being a thought experiment.

Because here’s the thing: I’d already started, almost without admitting that’s what I was doing. I recently ordered a pair of ASUS Ascent GX10 units for my home lab. Only one has arrived so far, but I’ve already spun up several open-weight models on it, and the first impression genuinely surprised me.

Subjectively, last year’s ChatGPT is already achievable locally. Right here. On a quiet little box on my desk. It runs fast, it’s compact, the power draw is modest, and the noise is almost nothing. A year ago that sentence would have been a press release. Now it’s just Tuesday.

Let me be clear about what it is and isn’t. It is not a replacement for Anthropic’s best models, not even close, not yet. The frontier is the frontier, and local hardware is a generation or two behind it by definition. But for a large share of my routine processes, the unglamorous middle of the workload that just needs to happen reliably, it’s already a perfectly viable backup loop. A second engine. Something that keeps turning when the primary one stalls.

That reframes the Emergency GPT Kit entirely. In the original piece I described it as insurance against catastrophe: grids failing, the long emergency, society degrading slowly over decades. All still true. But the version I actually need is far more mundane and far more immediate: insurance against a Tuesday morning when the API is down and the work still has to ship.

The Health Angle Made It Personal

There’s one category where I’ve decided the local-first approach isn’t optional, it’s the whole point.

I’m going to start analyzing my DNA and health data locally, first, before anything else. Partly that’s about privacy, which should be reason enough. But partly it’s because the frontier-hosted option for that kind of work has been narrowing: the most safety-restricted variants have been pulled back from answering medical questions at all. Whatever the reasons for that and I understand there are real ones, the practical effect on my side is the same: a capability I might want is gated or gone, on someone else’s schedule, not mine.

A local model that’s “good enough” and available beats a frontier model that’s brilliant and off-limits for the question I actually have.

And this connects to something I think about more than I’d like to admit: the slow drift toward process-oriented medicine, the kind shaped less by curiosity and more by what an insurance workflow will reimburse. The standardized path is efficient. It’s also, by design, allergic to the non-standard question, the weird symptom, the off-pattern hunch, the “what if it’s actually this” that a good doctor follows and a billing code does not. An offline, unfiltered tool I control is, among other things, a way to keep asking those questions when the system around me is optimized to skip them.

The Emergency GPT Kit is becoming a normal, sensible part of staying capable and independent.

What the Outage Actually Taught Me

We’re meticulous about backing up the wrong things.

We back up files. We back up photos. We carry a spare tire and keep a flashlight in a drawer. But over the last few years we quietly outsourced thinking itself: memory, navigation, research, drafting, increasingly judgment to services that live somewhere else, behind someone else’s status page. And we almost never back that up. When the server hiccups, our cognition hiccups with it. That’s a strange thing to have let happen.

So here’s where I’ve landed, watching my “operating system” come back online a few hours later as if nothing happened:

The AI-native company of the near future won’t be judged only by how well it uses AI. And neither will the educated, capable individual. The differentiator, the thing that separates the resilient from the merely impressive, is going to be the ability to keep working when the next model, API, or cloud suddenly isn’t available.

Capability is becoming abundant. Continuity is becoming rare. And continuity is the part you have to build yourself, on hardware you own, with skills that don’t expire the moment a connection drops.

What’s in your Emergency GPT Kit?

Nadella describes the architecture beautifully. Private evals, fine-tuning environments, a retrieval layer over institutional memory. The system that compounds over time, the one that stays yours even when you swap the underlying model. All of that is real and worth building.

But there’s a step before all of it. Before the system, before the retrieval layer, before any of that infrastructure - someone has to capture something worth capturing. And that part is harder than the tooling, because it can’t be automated. It might be the one thing in this whole conversation that genuinely can not.

What actually gets lost

I’ve been working in consulting long enough to watch a lot of knowledge disappear.

A project closes. The people on it move to the next engagement. The lessons, the ones that actually mattered, the client dynamic that almost tanked it, the architecture choice that turned out to be right for reasons nobody fully understood at the time - none of that makes it into the retrospective. The retrospective covers what shipped, not why things went the way they did.

Months later, a similar project starts. Similar client, similar configuration, similar early warning signs. Nobody connects the dots, because there are no dots to connect. The knowledge was there. It just wasn’t in a form anyone could use.

I don’t think this is a consulting problem specifically. I think it’s what happens by default when organizations are moving fast and capture is always someone’s lowest priority.

The decisions that don’t reduce to process

There’s a specific category of knowledge that’s hardest to hold onto.

It’s not the process knowledge: how to run a discovery workshop, how to structure a go-live plan. That gets documented, eventually. It’s the judgment knowledge. The calls that weren’t obvious. Whether to push back on a client scope request or absorb it. Which early signal in a project actually predicts trouble six months out. Why a particular customer relationship held through a difficult moment when others didn’t.

Ask someone who made one of those calls well why it worked. You’ll get an answer, but it won’t be an algorithm. It’ll be a story with context: what was happening at the client, what the options actually were, what they decided to discount and why. That story is the thing.

A language model can summarize a decision after the fact, but it can’t reconstruct the judgment that was present at the moment of it - the uncertainty, the alternatives that were seriously considered and dropped, the information that was available but turned out not to matter. That exists only in the person who was there, and only briefly.

The idea I’ve been testing

So I’ve started doing something I’m calling a Resolution Library, though “library” makes it sound more organized than it is right now.

The practice is simple: when a significant decision closes, write it down while it’s fresh. What was happening, what the options were, what got chosen and what got rejected, and crucially, the reasoning at the time. Not the polished version, but the actual thinking.

Then, when the outcome is clear, sometimes a month later, sometimes two years, go back and close the loop. What happened? Where the reasoning held and where it didn’t.

That’s it. No special tool. It’s a habit, not a system.

The honest reason I’m writing about this rather than presenting it as a solved method is that it’s genuinely hard to keep up. There’s no immediate return, the discipline it requires is dull, and there’s always something more urgent. I’ve missed entries I wish I had. I’m not holding this up as something I’ve perfected, more as something I’ve become convinced matters enough to persist with despite the friction.

Why this is Nadella’s argument taken to its source

The previous post ended with the question: how does your company accumulate experience through AI in a way nobody else can copy?

The Resolution Library is one attempt at an answer, specifically for the kind of experience that’s hardest to commoditize. The judgment knowledge. The record of what happened at the actual decision points, before and after.

Nadella’s test for whether you have a real learning loop is whether you can swap out the model without losing the company-veteran expertise built into the system. That expertise has to come from somewhere. It doesn’t emerge from the volume of transactions or aggregated usage data. It comes from the reasoning behind the calls that actually shaped outcomes and captured at the moment they were made, in a form that can be used.

There’s a piece by Usman Sheikh at Framebreak that frames the same stakes through Federer: he won 54% of the points he played and nearly 80% of his matches. A small edge, consistently held at the right moments, produces a dramatically different aggregate result. Not every point matters. Win the ones that do.

That’s the selection problem the Resolution Library is trying to solve, capturing the decisions that actually determine outcomes, which is a much shorter list.

Part of that is because I’ve spent the last few weeks deep in the Cursor ecosystem, watching how fast things move when AI actually gets embedded into a workflow instead of bolted onto one. Reading Nadella’s framing made it concrete: what’s happening right now, in record time, is a full rebuild of the operating system businesses run on. Not an upgrade. A rebuild.

“A frontier without an ecosystem is not stable.” - Satya Nadella

That’s the thesis in eight words. Everything else in the essay is him unpacking it.

The two kinds of capital

The core idea is what Nadella calls “capital in tokens.” His argument is that every company now carries two distinct forms of capital.

The first is human capital what he describes as the knowledge, judgment, relationships, ingenuity, and pattern recognition of a company’s people. Nothing new there. We’ve always known this mattered.

The second is new: token capital. The proprietary AI capability a company builds and owns, as opposed to what it simply rents through an API.

The counterintuitive part: the obvious read is that as token capital scales up, human capital becomes less important. Machines do more, people do less, value shifts accordingly. Nadella’s argument flips that completely. In his own words, human capital does not become less valuable as token capital grows, it becomes more valuable.

Human capital doesn’t get cheaper as AI gets stronger. It gets more expensive.

Why AI alone just spins its wheels

Think about what AI actually does without a human steering it. It optimizes whatever target it’s given, inside whatever boundaries it’s given. But it doesn’t decide which problems matter. It doesn’t connect a customer complaint in support of a roadmap decision in a product to a phrase in a sales deck. It doesn’t notice that the real issue isn’t the one anyone asked about.

That’s still entirely a human job. Without it, AI doesn’t fail loudly, it just runs in circles, confidently. It produces a lot of output that looks like progress and isn’t.

“Without human direction, you have compute running in circles.” - Satya Nadella

Put simply: AI without people is just expensive servers heating up the atmosphere. Impressive electricity bill, no compounding value.

The real competition isn’t about picking the best model

This matters for anyone actually running a company, not just watching one from the sidelines.

Every major tech player is now building on remarkably similar underlying technology. The model landscape is converging fast what was a meaningful gap six months ago is closing. So if everyone has access to roughly comparable horsepower, “which model did you pick” stops being the differentiator.

Nadella’s sharpest point is that the real competitive question is learning system you build on top of it.

Your workflows. Your accumulated edge cases. Your institutional scar tissue from the deals that went sideways and the processes that finally worked. All of that, fed back into a system that gets better every time someone uses it, that’s the actual asset. Not the model. The loop around the model.

Nadella actually gets specific about what that loop is made of, which is rare for a CEO essay like this. He lays out three layers: private evaluations that measure whether the model is improving against outcomes the business actually cares about, not just public benchmarks; private reinforcement-learning environments that let the model get stronger from real traces of work inside the company; and a retrieval layer that makes institutional memory searchable and usable. Stack those three together and you get a system that compounds with every cycle of use - what he calls a hill-climbing machine.

“Unlike most assets, it compounds.” - Satya Nadella, on the hill-climbing machine

Most balance-sheet assets depreciate the moment you buy them. This one is built to do the opposite.

I’d go as far as saying we need a new metric entirely: something like AI-learnability - how fast and how well a company’s internal systems improve from each interaction. I suspect that number will end up correlating with business outcomes more tightly than almost anything else we currently track.

And this only works if knowledge is captured in a form an agent can actually use, not buried in someone’s head, not scattered across six tools, not living exclusively in a quarterly deck. It means fine-tuning on your real cases, your real failures, your real customers. Not generic, abstract data that could belong to anyone.

The sovereignty test

Nadella gives this loop a name worth sitting with: the new IP of the firm. And he offers a concrete way to test whether you actually have one.

Can you swap out a generalist model: switch providers, upgrade versions, whatever without losing the accumulated, company-specific expertise your system has built up? If the answer is no, you don’t have a learning loop. You have a rental agreement with extra steps, and your “institutional knowledge” lives entirely inside someone else’s model weights.

The framing is simple but it’s a genuinely useful filter for decision-making. You can delegate a task. You can hand off entire chunks of execution.

“You can never offload your learning.” - Satya Nadella

That’s the line Nadella keeps coming back to. You protect the learning like it’s the whole company. Because increasingly, it is.

A warning from the 2000s

This is what turns the argument from a nice metaphor into an actual warning and it’s not me reaching for a parallel, it’s the one Nadella draws himself.

We’ve run this experiment before. In the 2000s, GDP numbers looked perfectly healthy on paper while entire industries were quietly hollowed out by outsourcing. Manufacturing left. Jobs disappeared. The aggregate statistics stayed calm while the underlying structure of entire economies shifted somewhere else, slowly enough that nobody panicked until it was already done.

Nadella’s warning is that the same pattern is available again, just with frontier AI models instead of factories: a small number of dominant models absorbing the specialized knowledge of entire industries and selling it back at commodity prices, while the companies that generated that knowledge end up with none of the value. A company or a country that becomes fully dependent on someone else’s model, with no internal loop of its own, is outsourcing its future capability while the dashboard still looks fine.

This matches what I’ve been seeing elsewhere

This connects directly to something I’ve written about before: value accumulates with whoever owns the outcome, not whoever sells the tool. The tool is increasingly commoditized. The outcome, the actual compounding result inside a specific company - isn’t.

It also explains something I’ve been watching closely: the AI-trainer market. Why is it growing right now, specifically? Because the companies pulling ahead understand that human capital inside the building is exactly what makes “capital in tokens” function at all. Someone has to teach the system what good looks like. That’s a human job, and right now it’s an underpriced one.

This is the part of Nadella’s formula that actually makes it work in practice. Capital in tokens, on its own, is inert. It only compounds when it’s paired with the human judgment that knows what to feed it and what to ignore.

So what does this actually mean for you

Everyone will have access to roughly the best model, eventually, on roughly the same terms.

The companies that come out ahead are the ones building their own internal learning system right now - their own quality benchmarks, their own fine-tuning loops, their own corporate knowledge base that an agent can actually query. Two or three years from now, that compounding advantage won’t be something a competitor can buy with a subscription. It’ll be structural.

The real question now is how your company plans to accumulate its own experience through AI, in a way nobody else can copy.

That’s the real moat.

That’s what AI trainers on Wall Street are charging right now.

Saw it on Bloomberg last week. Sounds crazy until you run the numbers.

Twenty to thirty bankers in the room. That’s roughly $1,000 a head. For an investment bank, that’s two hours of one analyst’s time. If the team comes out 10-20% more productive, it pays for itself before end of quarter.

The mass replacement story isn’t playing out the way people expected. What’s actually happening: a new line item showed up on the P&L. And with it, real demand for people who can explain AI to other people - clearly, practically, in a way that makes them want to actually use it.

It brought back one of my first big projects.

SAP CRM rollout for a large pharma company. I went in thinking the hard part was the configuration.

I was wrong.

The technical work ended up being maybe half the job. The other half was harder: flying to dozens of countries, building training for teams with completely different skill levels, figuring out how to get people engaged who had no real reason to care, and somehow making them want to use something that just made their jobs more complicated at first.

We ran experiment after experiment. KPIs or recognition? Mandatory rollout or voluntary adoption?

The answer was always the same: find the people inside the company who already believed in the technology before anyone asked them to.

In the SAP world we called them power users. Anthropic calls them ambassadors - I joined that program. Cursor runs hackathons for the same reason.

The name changes. The pattern doesn’t.

Those people got up at internal events, showed their results, shared what clicked, and ended up being the real reason their colleagues got on board. They moved the needle more than any top-down mandate ever did.

Same thing is happening now in Agentforce deployments.

The technical setup takes weeks. Getting two hundred people to actually open the tool every day takes months. The rollouts that stick aren’t the ones with the cleanest architecture. They’re the ones where someone inside the client org got genuinely excited, showed their team what was possible, and made everyone else feel like they were falling behind.

One person who actually believes in the tool moves faster than six months of mandatory training.

The firms that get this are hiring for it deliberately - not waiting for champions to show up, but finding them early and building around them. That’s a different kind of implementation work. And it gets paid differently.

Last month, at Techflow in Miami, I caught a Claude Code talk aimed at beginners.

The presenter did something genuinely hard: made the tools make sense to a completely mixed room, in a way that made people want to try it right then. Not just understand it - try it.

That skill is rarer than it looks.

Twenty years. The technology has changed completely.

People haven’t.

Companies keep buying new tools expecting results overnight. But between the purchase and the actual payoff, the same thing always gets in the way: changing how people work.

The AI trainer, AI coach, and AI transformation market is just getting started.

It’s the clearest money that will be made on AI - right after data centers and hardware.

For AI companies, the race that matters most right now is who helps thousands of people actually change the way they work.

I’d paid $99, gotten a polished consumer interface, and walked away with almost nothing useful on the health side. The one genuinely interesting detail: a distant cousin who turned out to be another tech guy at Microsoft in Seattle. We probably shared 0.3% of our DNA and had never crossed paths. The internet works as intended.

Everything medical is heavily constrained by the FDA. The outputs land at the level of “slightly elevated Alzheimer’s risk” and “asparagus smell perception.” Those restrictions haven’t loosened since. So the file sat there, doing nothing, for five years.

A friend eventually pointed out the obvious: I was sitting on 600,000 data points and had done exactly nothing with them.

The raw 23andMe file isn’t a summary. It’s a dense text file full of SNPs: Single Nucleotide Polymorphisms, which sounds like a graduate seminar topic but just means: tiny places in your DNA where you differ from other people. One letter changed out of three billion. Multiply that by 600,000 and you have a surprisingly complete map of your biological quirks.

Used right, those SNPs tell you about disease predisposition, metabolic patterns, drug response, ancestry and ancient population origins, even distant connections to known historical lineages. The data was always there, but I never got to it.

Doing this properly used to require real bioinformatics fluency: Plink, VCF files, genome assembly builds, bash scripts, a stack of reference databases. Weeks of learning before any real output. Not impossible, but enough friction that almost nobody outside a research lab was doing it for themselves.

So I opened Claude Code in the terminal and typed one line:

“Here’s my 23andMe file. Check hereditary disease risk, pharmacogenetics, and parse nutrition and metabolism notes.”

That was it.

Claude identified what was needed, pulled the right tools and databases on its own, and built the pipeline. I didn’t write a single bash script. I didn’t configure a genome build. I didn’t spend three days reading documentation for a tool I’d use once.

The databases it pulled in:

• ClinVar - hereditary disease variants

• PharmGKB / CPIC - drug compatibility and metabolism

• GWAS Catalog - risk associations for diabetes, heart disease, and other conditions

• 1000 Genomes - comparison against modern global populations

• Ancient DNA databases - tracing ancestry against ancient peoples and migration paths

By the end of the weekend I had a personal report I can re-run anytime, no third-party service, no subscription, nothing leaving my machine. I saved a section of it for my doctor to put in the chart.

What I actually found:

Specific drug metabolism patterns: the pharmacogenetics piece turned out to be the most immediately practical. Some people process certain compounds fast, some slow, some atypically. I fall outside the standard range on a few things that are worth flagging before any new prescription gets written.

Elevated risk on a few conditions. I’m not listing them here both because it’s personal and because the framing matters a lot. Genetic risk is probabilistic. A raised number isn’t a diagnosis, it’s a signal to pay attention to a particular area. There’s a big difference.

Ancient Eastern European and Slavic lineages running deep. The ancestry section cross-referenced against ancient DNA migration patterns and placed my ancestors in specific regions across thousands of years. That part was more quietly fascinating than actionable, but seeing where your people were 4,000 years ago is one of those moments that’s hard to describe.

Variants affecting tolerance to certain medications - distinct from the metabolism patterns, but related. The kind of thing that’s genuinely useful to have documented somewhere.

My honest take: genetics isn’t a verdict.

Most SNPs only give probabilities. They aren’t diagnoses, they aren’t certainties, and without a sober reading of what the numbers actually mean, it’s very easy to spiral into anxiety over nothing. I’d recommend going in with that framing locked in, or doing it alongside someone: a doctor, a friend with a biology background who can help you contextualize what comes back.

That said, I found the whole thing genuinely valuable, the information actually exists now, in a form I control, that I can hand to a doctor and have it mean something.

Without Claude Code I’d have burned weeks just on tooling, probably given up, and the file would still be sitting there.

With it, the whole thing came together over two weekends.

I’m now looking for a lab to do a full genome sequence. 23andMe only covers a small fragment, around 0.02% of your actual genome. The rest is inference and approximation. Full sequencing gives you the complete picture, and the cost has dropped far enough that it’s no longer an exotic expense.

When that file comes back, Claude Code is the first thing I’m opening.

And yet, most people I respect still don’t post. So I’ve been thinking about why.

The stories we tell ourselves

Many friends and colleagues ask me how to start building their brand and producing content. When I try to understand what’s actually stopping them, it almost always traces back to something absorbed early, not a decision, more like a rule that never got questioned. Don’t brag. Don’t overshare. Don’t make yourself a target. The underlying message we absorb is that being visible is dangerous. Say something good about yourself and you’re showing off. Share something useful and someone might just run with it.

But in practice, it doesn’t work that way. Knowledge and experience almost always only produce results in combination with context, with trust, with the relationship that forms when someone reads your thinking and recognizes something true in it. Sharing a framework doesn’t hand over the years it took to build it. If anything, it demonstrates them.

The deeper fear: looking ridiculous

One of the most paralyzing fears isn’t failure, it’s actually embarrassment. The fear of looking ridiculous. Its roots often go back to childhood: bullying, the perfectionism of always needing to get it right before raising your hand, the social cost of being “that person” who overshares.

Where is the line between sharing and embarrassing yourself? Honestly? Nobody knows. And that uncertainty is exactly the problem. In language learning, the people who progress fastest aren’t those who study the longest before speaking. They’re the ones who start speaking before they’re ready, absorb the feedback, and adjust. Content works exactly the same way.

I’ve seen this in clay pigeon shooting, which I’ve practiced for years. When I was teaching a friend recently, his first rounds were rough, almost no hits. But I could see the real progress: better stance, smoother gun movement, cleaner tracking, more deliberate preparation before the shot. The clays weren’t breaking yet, but the mechanics were improving with every round. That’s not empty encouragement, that’s a coach’s eye recognizing that micro-shifts predict outcomes, even when the scoreboard doesn’t show it yet. When everything finally came together, the clays started breaking consistently.

Content is no different. The early posts feel clumsy. The engagement is low. But something is happening under the surface: you’re developing your voice, learning what resonates, building the habit. Most people quit before the clays start breaking.

The trap of two attempts

I’ve written about this before - the mistake of judging a strategy by two attempts. You go to one conference: “conferences don’t work.” You send ten LinkedIn messages: “LinkedIn doesn’t work.” But someone did reply. Someone did read your post. That’s a signal that didn’t exist before you tried. These are small but real shifts. In business, we often cut too early, right at the moment when the mechanics are starting to click.

The same logic applies to content. Meaningful skill: in shooting, in archery, in writing, in building comes from repetitions plus attention to small changes. Not from two tries followed by a verdict.

What actually works

If you’re not confident enough to share your own ideas yet, start by sharing opinions on other people. Become the person others come to for a take. Engage with people you respect - comment on their posts, ask a real question, enter their orbit. This is low-risk, high-signal activity. It builds your presence without requiring you to claim expertise you don’t yet feel.

And then: post the thing that scares you a little.

Working on content daily for a couple of years, I’ve noticed that the strongest responses never come from the polished, carefully reasoned posts. They come from the ones that contain a contradiction, or an acknowledged vulnerability, moments where I admit I don’t have it figured out. Half the readers like it. Half react negatively. But that generates far more resonance than posts everyone simply nods at and scrolls past.

Training the muscle

In sports psychology something I’ve read extensively because shooting and archery have taught me a great deal about mental performance: there’s a lot of emphasis on neutral response after a bad shot. Don’t grimace. Don’t sigh. Don’t curse under your breath. Each of those small reactions trains your brain to expect failure. Instead: neutral face, calm body, attention immediately to what’s next.

The same muscle applies to content. I deliberately put myself in situations where I might look ridiculous, because the fear of embarrassment, left untrained, will quietly veto every idea you have before it reaches the page. Looking stupid isn’t the worst thing. Never starting is.

Like any skill, this gets easier with repetition. Entrepreneurs I know who are also serious amateur athletes tend to understand this intuitively. The ability to absorb a loss, extract the lesson, and focus on the next attempt is exactly the same in business as it is in sport. And it’s exactly what content creation requires.

One post this week

Start with one post. Maybe the one that feels uncomfortable to publish. The one where you’re not sure how it’ll land. That discomfort is information, it usually means you’ve said something real.

Personal brand isn’t built in a campaign. It’s built in the reps. And the people who seem like they “got lucky” with their audience? They made a hundred attempts, paid with time and effort, survived dozens of misses, and kept going.

The clays eventually start breaking. But only for those who keep shooting.

Read it the way you’d read a leaked board deck. Because that’s what it is.

Email is dead. Slack is the company.

Calvin received about 10 emails the entire year. Everything else was Slack.

Dozens of workspaces. Layered permissions. Hundreds of channels. If you don’t curate notifications aggressively, you drown by week two.

For a company that tripled headcount in twelve months, this is a deliberate choice. Email is too slow. Slack gives speed, and at OpenAI speed beats order.

That’s not a tooling preference. It’s an architectural decision about how decisions get made.

One monorepo. No style guide.

Most of the code is Python - FastAPI plus Pydantic. There’s Rust where latency matters and a little Go.

There is no enforced style guide. In one file, a clean engineering library written by a former Google veteran. In the file next to it, a one-off Jupyter notebook from a PhD who joined last month. Things break more often than they should.

Tests that touch GPUs can take 30 minutes even when parallelized. The backend monolith has become a landfill, the place everything gets dumped because it’s the path of least resistance.

If you’ve worked at a hypergrowth product company, none of this surprises you. The only difference at OpenAI is the dial is turned to maximum.

Duplication is a strategy, not a bug.

This is the part that breaks every MBA framework.

Calvin counted at least 6 internal libraries doing the same job - task queues, agent loops, the basics. Codex existed in 3 to 4 parallel versions from different teams before launch. ChatGPT Connectors followed the same pattern.

Why?

Because there is no central architecture team. Researchers operate as mini-CEOs: they come up with an idea, build a prototype quickly, and ask almost no one for permission. If the prototype works, a team coalesces around it. If it doesn’t, the project quietly dies.

The math is brutal but logical. The cost of three teams writing the same queue library is small. The cost of three teams coordinating on whose queue library to use is enormous, measured in months, not engineer-weeks.

OpenAI buys speed by paying for duplication. Most companies do the opposite and wonder why nothing ships.

Knowledge lives in three places.

Slack. Code. People.

When Calvin asked about the quarterly roadmap, the answer was: “that doesn’t exist.”

Not “we’re working on one.” Not “ask your manager.” It doesn’t exist.

A huge amount of the company’s actual operating context sits in employees’ heads and Slack threads. That’s why the best research managers there aren’t planners, they’re synthesizers. People who can connect twelve disconnected experiments into a coherent product story.

When Codex needed two senior engineers from the ChatGPT team, it got them the next day. No staffing committee. No quarterly resource planning. No reorg memo.

That’s not casual. That’s the system working as designed.

The MBA inversion

If you took every operating principle business school teaches, OpenAI did the opposite. On purpose.

1. No single source of truth. Knowledge is distributed by design.

2. No centralized documentation. Code and Slack are the documentation.

3. Teams duplicate code instead of coordinating endlessly. Cheaper.

4. Slack replaces half the processes the rest of the world calls “operations.”

While speed beats order, this works.

That’s the load-bearing assumption and it’s worth saying out loud because most companies running this playbook think the model is the playbook. It isn’t. The playbook only holds while the marginal value of a shipped product feature is higher than the marginal cost of organizational debt.

For OpenAI in 2026, that math is wildly in favor of speed. New product surface. New customer segment. A research model that keeps getting better. Every quarter, the cost of waiting is larger than the cost of duplication.

That’s not true for most companies. It’s barely true for most labs.

The 10,000-person question

OpenAI is not Google. It’s not a classic enterprise software company. It’s a research lab that accidentally shipped the most viral consumer product in history and is now selling, in parallel, to enterprises and to governments.

Meta hit this exact operating wall around 10,000 people. The duplication tax stopped being affordable. Coordination overhead became the bottleneck, and the company restructured to absorb it.

OpenAI will hit the same wall. The only question is when.

If you’re a founder reading this and copying the playbook because OpenAI is winning - be careful. The playbook is the consequence of a specific stage and a specific kind of business. It’s not the cause of the winning.

For now, the model holds. Slack is the org chart. Researchers are mini-CEOs. Duplication is policy.

Most enterprise wisdom about “scaling” might be expensive folklore until the headcount forces someone to find out.