A quick update on 2nd Order Thinkers.

It now has an audience of about 5,000 across Substack, YouTube, and others. It’s a milestone, though not yet enough to sustain this work of human writing and research. So over the coming months, I’ll be experimenting to find a stronger product-market fit, while keeping the core: examining the impact of frontier tech on business through second-order thinking and deep research.

I’ve also got some admin to clear before Klaas and I can finally move country, so publishing may be less regular for a while.

Will update you once we have good news :-)

Thanks for the support. Today’s piece is part of that experiment. I’d love your feedback!

Expect this in your lifetime: the most valuable consumer product ever sold won't be a phone or a car. It'll be your partner. The company that builds it owns every memory it has of you, and charges you a monthly fee to keep it alive.

That sounds like a Black Mirror episode I made up. Just it isn’t.

It's the obvious endpoint of a pattern that's been running under dating for a thousand years. I’ll walk you through the history of how tech changed dating, and by the time we get to current, you can predict exactly where this industry goes.

And I know, because I've lived a part of that arc (as some of you did). So let me rewind, and the story is part of the history we’re about to cover.

Dating Essentials (Hint: It’s Not Love)

I came from a rural county in Taiwan, and went to a girls’ school until uni.

In the early 2000s, for a girl like me, telling my parents I’ve met someone on the internet ranked somewhere between joining a cult and selling a kidney. The internet was where the predators lived, in their opinion, and everyone you met there was lying about their age, their face, or their intentions, and probably all three.

I spent my late teens treating online dating as an act of rebellion. The kind where I needed to blanket the modem and pray they wouldn’t hear the dial tone.

Then I moved to London in mid-2010 and met my partner on an app. I flinched when people asked how we two met. I was likely still in that childhood memory, while in fact, the world moved on. If you aren’t already married to the high-school darling, don’t want to date your colleague, and don’t go to pubs, then dating apps soon became a sensible option.

Think about all the filters you can apply to narrow down your future Mr. or Mrs. right.

Don’t worry, I’m not turning into a dating influencer.

But here’s a thread worth pulling.

It’s less about romance and mostly about the business opportunity and how tech evolution across the last hundred years has changed the dating business.

As I’ve been emphasizing in my work, always think of the basic need (or, in this case, desire) when considering a business opportunity.

Here, that desire is to mate and feel a sense of belonging, not about reproduction, just companionship. Even though Mother Nature designed companionship as a trigger for reproduction, it didn’t foresee technology; these days, love is a business.

So the desire in our DNA pushes us to find someone while reality presents obstacles.

Go back far enough, dating has always had the same shape. Two people who want to find each other, and someone is taking a cut for connecting them.

Let’s say you want a date. Five essential elements are involved in finding the right one (even if you aren’t aware of them, and love isn’t one of those):

• information about the other party. Plus, to share and make your own information appealing to that party (the first two are often async and asymmetrical)

• a channel to connect two parties

• a way to check whether the other side’s intention matches with yours (doesn’t matter if it’s a one-night stand or a lifelong commitment),

• to stand out among competitors

• Finally, the slow business of staying together

The intermediary, in every era, sells the first three. They hold information from both sides, serve as the channel, and ensure the two align with the same intention.

The fourth and the fifth, the standing out and staying together, however, have never been for sale.

At least, not until very recently, but first, some history.

Pre-Industrialization: Méipó or God

In ancient China, the intermediary is called 媒婆, typically the village matchmaker, usually a married woman doing it as a part-time job for a fee.

ps. I can’t seem to find a direct equivalent of 媒婆 in European history. Most marriages were arranged by the church or close relatives. Other intermediaries were either banned or burned in the witch hunt. Message me if you have more info to share.

This is a typical image of a 媒婆 (méipó).

媒婆's job was to gather information and profit from controlling its flow, and the imbalance of who knew what.

She knew which families had a marriageable daughter and which had a son with prospects. She carried word between two houses that would never have spoken directly. She made sure the match was 門當戶對, which means quality control based on whether the families had similar wealth and status.

If a match had a happy ending, she was paid upon the marriage. Her reputation lived or died on whether the match was held, which is rather more accountability than any app has since offered.

But she could only see what was in the nearby villages.

So the entire pool was whoever happened to be born within walking distance.

Then the walls of the village came down. The Industrial Revolution and urbanization, in sequence, did something much more powerful than hundreds of traditional matchmakers ever could.

Urbanization: Love Catalyzed by Overcrowding

Elizabeth Gaskell, a well-known 19th-century novelist, saw it firsthand.

In North and South (1854), the main character, Margaret Hale, a clergyman’s daughter from rural southern England, moves to an industrial city of Milton (think Manchester) and ...finds herself sharing the same rooms, the same streets, and the same heated arguments with John Thornton, a self-made mill owner.

Urbanization started earlier in Europe, around the 1800s, climbed to 65% in the US by 1950, and by 2006 tipped the whole world past the point where more people lived in cities than in the countryside.

If this were a farming society, the path between Margaret Hale and John Thornton would never cross. What brought them together was mass migration.

People packed up their families and moved to cities. So cities became the greatest matchmaker ever built. You met people at the factory, the church, the boarding house, and the night class. Density did the work.

Intermediation became a free byproduct of living near other people. For about a century onward, the channel and the information needed to meet someone were all supplied for nothing by the sheer fact of a crowded street.

Yes, there were still introductions, or met via a friend of a friend, but this was no longer a billable activity and mostly became a personal favor.

Suburbanization Brought Dating Apps

You’d never guess the first thing people bought when they got rich.

Walls.

To be more precise, privacy and family space.

Between the limited leisure options, crowded living space, and lack of transportation, you form a tight bond with a community until suburbanization starts.

In 1790, a third of American households had seven or more people; by the mid-2000s, more than a quarter were one-person households. See the change in household size over time below:

As incomes rose and living space expanded, the street stopped being where life happened and became somewhere you passed through. The communal laundry, the corner newsstand, or the small retail shop you visited daily were each replaced by a private version inside a flat.

We now live in the densest cities in human history and can somehow go a week without a single conversation that wasn’t transactional. Attention was spread so thin that just living near people stopped doing the work it used to.

But the desire for companionship and mating remains.

The dating apps become the matchmaker (媒婆) rebuilt at scale.

It holds everyone’s information, serves as the channel, and has a vouching mechanism to verify intent (like filters or a mutual swipe). It does the convincing the matchmaker once did with family status, except the persuasion is now an algorithm and a “99% match in heaven” badge.

The Economy of Dating Before AI

If you haven’t noticed, this one pattern makes the whole history legible.

For a thousand years, the technology of dating only ever moved one dial: the number of people you could reach. The job underneath stayed frozen.

All the matchmaking approaches from ancient China to dating apps satisfy the first three: gathering, storing, and sharing information; the ground cover (the scale of the number of headcounts on a channel); and quality verification (intent).

But the last two points remain painful for many: standing out and staying together.

Start with ‘standing out’. In a village, you barely had to. Among a handful of candidates, you stood out by simply existing — no marketing required (eg, better photos).

However, the apps flip this. You're now one face in a thousand, and attention flows to whoever poses best, knows how the ins and outs of flirting, and flashes the right car, watch, or job title. Standing out became a skill you have to work at.

Then, ‘staying together’.

Yes, matching got cheaper, and so did walking away.

The effort of maintaining a relationship is the same as it ever was — people are still difficult, love is still work. But when leaving costs almost nothing and hundreds of 'better' options sit one swipe away, the same old effort of keeping a relationship starts to feel like overpaying.

For example, the punishment for cheating used to be death for a woman; now it's just a temporary social suicide and losing your job (like the CEO and the head of HR caught cheating on the Coldplay concert).

No matter the era and how the channel evolved, the same two costs stayed stuck.

If anything, each channel evolution made them worse: every jump in scale meant more faces to stand out against, and more reasons to wonder if you should be leaving.

Digital Is Easier To Love Than Flesh and Blood

In 2023, a record-low 1.4% of married adults divorced. Here’s the catch, 33% between the ages of 25 and 54 are never married, from 17% in 1990.

The situation gets out of hand, and even the US Surgeon General declared loneliness (affecting roughly half of all American adults) a public health epidemic.

Fewer and fewer see marriage as a path to happiness, reasons stated in the math we just walked through — harder to stand out, costlier to stay.

However, the need to have someone by your side to share your joy and sorrow remains.

Yes, making friends or having pets solves part of the issue, but there are topics you won’t be discussing with friends, and pets can’t really talk back, not to mention giving you helpful advice.

Where we’re heading is becoming obvious: make your own partner.

Let’s rewind to 2018. When dating with a digital figure was a bizarre niche.

Four years before ChatGPT launched, a Japanese man spent ¥2 million on a wedding ceremony in Tokyo and married a virtual anime pop star, with 40 guests attending. Here they are, still married after 8 years:

What looked unthinkable was really an early signal that companionship had already detached from the requirement that the other party is human.

An AI companion erases both last-mile hurdles in one single move.

With an AI partner, you’d have no competitors to stand out against, since it is assembled to want only you; and there’ll be no rejection to outlast, since its sole purpose is to please you and make sure you stay. As part of the solution to the previously make-or-break, it doesn’t have a bad day, and it doesn’t quietly reconsider you.

You bring yourself, and it brings whatever you ordered.

In 2025, a survey found that 72% of American teenagers had used an AI companion at least once, and more than half were regular users.

In 2023, an AI companion app (Replika) that supported erotic roleplay was forced to remove the sexy feature overnight. Harvard Business School ran research before and after the app update, in order to answer:

how deep are people’s relationships with AI companions?

They found that the users grieved their AI, some users found their AI had become a different person overnight. Here’s the overall sentiment of the users before and after the change was released:

Which means any noticeable update can make people with AI partners feel like bereavement. Overnight, the one who knew you is gone, replaced by someone who shares their voice and face but is a complete stranger.

The Future Dating Market

Now, all five barriers between you and the perfect match should be broken down.

So that’s good news. Although the bad news is… that it’s merely deferred.

If you are destined down the path of human and AI romance, your loved one is always going to be held hostage by the AI industry.

The problems, however, will have different flavors over time.

For now, it’s technical issues like memory loss and inconsistent personality of a model; for the foreseeable future, as the model gets mature, expect your partner (essentially, you) to be the cow they milk.

Let’s start with the short-term (the LLM era) problem.

A Sticky Note at a Time

The current AI companion is a lover with anterograde amnesia, like the main character in 50 First Dates, who wakes up every morning with no memory of you. Unless that's your type, you'd spend a lot of time reminding it of things you've already said a hundred times. For example, it can't mention avocado, because just reading the word triggers your allergy, or that your breakup isn’t the same order of pain as missing a tube yesterday.

But this isn’t the worst yet.

One Monthly Bill Away From Your Loved One Going Offline

Now run it forward.

Soon enough, you get to have a companion with a consistent character, one you can configure, that remembers every conversation and grows alongside you.

Next, you give it a body, like Joi from Blade Runner 2049 stepping off the projector. At that point, the last cost in the entire history of dating, the unpredictability of another human, is genuinely gone.

People will call it the democratization of love.

And yes, you have a choice not to pay for the cost of dating another human; the cost didn’t just disappear, it migrates, and it moves onto the only party still standing in the middle.

This is going to be the start of your nightmare. And 10 times worse than what you’ve seen from the Blade Runner movie.

Your lover will be owned by a corporation. The memory that makes the relationship feel real lives on their servers, or within the software they license to you. Even a model running locally on a robot in your living room needs the occasional patch, an upgrade, or a renewed subscription.

In the movie, K was a mercenary (or detective, the point being…); he has muscles, and since the story is about him, he'll eventually win.

In real life, though, you don’t have a superpower like a movie character.

So what, exactly, can you offer AI companies so you can keep your partner’s identity, body, and intact memory untouched and uncommercialized?

Not much, aside from the lifetime labor and the monthly $200 tribute (or whatever the price you think is worth paying to keep your partner alive).

What if you don’t like the price rise?

Sure, you can get 30 minutes of ad-free interaction daily instead of an hour. But what if you want new parts because your partner has been glitching and showing signs of pain? No problem, $2000 for a part, if you want a deep cleaning and polish, that’s another $500 on top.

Essentially, your lover will be the most valuable consumer product ever invented in the history of commerce.

I wish I were joking and that I could tell you this isn’t happening.

Unfortunately, sci-fi writers have this annoying ability to accurately forecast the future; this is one of them that I can see coming true, given the current trajectory.

Don’t forget to tell me if you enjoy this combination of history, business, and technology! Keen to hear from you :-)

Jing

Anthropic published a proposal arguing the world should have the option to slow or pause frontier AI development if systems start improving themselves too fast, citing its own code now being mostly written by Claude. The post landed days after Anthropic filed confidentially for an IPO at a valuation near a trillion dollars.

→ Reuters, Anthropic

This image represents their worldview, with the orange indicating where they think we are today.

Reading between the lines, they want to have their cake and eat it too.

Think incentive, would any AI lab want a pause? No, they don’t actually want a pause; they want to (and will) keep rushing ahead, but with the ‘security first’ facade. My bet is that they’ll say: but we have to… because of China! We do it for America!

OpenAI files for IPO

OpenAI confirmed it confidentially submitted a draft S-1 to the SEC, a week after Anthropic made a similar filing. But OpenAI said it has not yet determined the timing of any future filing or public offering.

→ OpenAI

Unlike Anthropic, they are likely to IPO only next year.

My read is that they are watching SpaceX (if they reach the target and it doesn’t destroy the AI sentiment) and want to stay private while they sort out the messy bits. More importantly, looking for ways to make their numbers look better under sunlight or only in pitch decks.

Anthropic’s Latest Model!

Anthropic released Claude Fable 5, basically a consumer-tier Mythos.

→ Anthropic

From the benchmark, it does look like a significant improvement.

But it’s also the kind of launch with footnotes… Here are some concerns raised and why they might be relevant to you:

• Hidden restrictions = the model may quietly give worse answers in some areas while you’ll still pay double compared to Opus.

• Trust and transparency = if the company is steering answers underneath, as it would, you need to start worrying what else is being filtered.

• Benchmark-to-real-world gap (as I mentioned in the Stanford report article) = a model can crush tests and still be expensive, awkward, or unreliable in daily use.

A German Court Says Google Owns Its AI's Mistakes!

A Munich court ruled Google can be held directly liable for false claims in its AI Overviews, rejecting the argument that they're just search results users should verify themselves.

→ Heise

AI search just got more expensive!

If firms are directly liable for hallucinated accusations, they will need heavier review systems or, simply, more refusals.

That is bad news not just for Google, but for every AI product that wants the authority of an answer without the liability of a publisher.

Apple’s new Siri

At WWDC, Apple unveiled a rebuilt Siri.

→ CNBC

As I mentioned previously, Apple is one of the few that gets AI right.

I doubled down on what I said. Apple is doing one thing they are best at: productizing AI.

Yes, if you watched the demo, you won’t be impressed, but if you think about how it works underneath, you’d realize it all made sense for Apple and what they’ve been building: a stable product that solves the real problem, and ties everything to their ecosystem.

McDonald's Gets Its Drive-Thru AI Back!

Two years after killing an earlier IBM version over order-accuracy complaints. Franchisee posts claim it has handled over a million transactions with about 90% needing no human help.

→ Independent

McDonald’s is back with AI drive-thrus, which is bold considering the last version became famous for inventing orders no human had asked for.

This time, skip the polished demo and hand it to the CEO — the one who somehow made his own burger look unconvincing (if you aren’t sure what I’m saying, search McDonald's CEO vs. Burger King's CEO burger test standoff)

The AI Boom’s Quiet Winner

Trustpilot reported a 2025 operating profit of $16 million, up 320% year over year, while click-throughs from AI search rose about 1,490% as chatbots increasingly cited its reviews. Reasons?

It ranked as the fifth most cited domain globally on ChatGPT in January 2026, tailing after Wikipedia and Reddit.

→ Proactive Investors

This isn’t the latest (March, to be exact), but I didn’t write about it, and I thought it was a good thing to be aware of.

The winners in AI search may be the platforms sitting on large stores of human-generated criticisms. Even though AI models can remix opinions, they still need a supply of real people having real experiences to cite.

The First Wind-Powered Underwater Data Center

Off the coast of Shanghai’s Lingang area, China has begun operating the world’s first undersea data center directly powered by offshore wind. The system reduces overall electricity use by 22.8%.

Read more

In addition to several industry figures telling the same fortune that AI is coming for your career. You heard it enough times. The claim registers eventually, even if you didn’t want to believe them.

Twelve months later, May 2026. Amodei on the same subject, he doubled down.

Amodei believes that 10% of the laborers will be doing 90% of the jobs, output grows exponentially, but because AI will outgrow the speed of achieving Jevons paradox, we should expect:

weird behaviors and this big disruption.

Basically, he argues that AI advances faster than the economy can absorb it, leading to an era of total chaos.

While Amodei plays all in, other tech figures openly speak out against this theory.

Who should you believe then?

Before you pick a side, map the incentive (always): what does each gain if you believe them?

Since 2022, I’ve been scrutinizing AI adoption and displacement topics. In the next 10 minutes, I'll take Amodei's claim apart, show you what the open data actually says, and hand you one question you can run on any job — including your own.

The Doomsday Tell Is an Earnings Tell

To do this, let’s rewind to a time most of us forgot: when money was nearly free.

This is the thing about pain: when we feel it, we feel it so deeply that we forget we’ve ever felt any other way.

If you’ve been suffering the pain of a slow economy since 2022, let me start by reminding you of better times.

Money was close to free ever since central banks slammed interest rates to the floor after 2009.

Even though the pandemic time was painful for its own reasons, it also gave white‑collar workers the loosest labor market of the generation.

Between 2019 and 2022, Meta went from roughly 45,000 employees to about 87,000. Nearly doubled in three years.

Not just Meta, but every company with sufficient revenue to borrow saw the same thing: cheap capital, a pandemic that seemed permanent, an online-everything boom, and executives convinced that the surge in demand would stick around.

As office workers, we also got our cut of the pie. We worked from home, enjoying the greatest freedom and pocketing a generous pay raise and bonus.

Then the music stopped.

When the pandemic-era supply shortages met a sharp rebound in demand, while energy prices jumped after Russia’s invasion of Ukraine, causing inflation and forcing the Fed, the BoE, and many other central banks to act in the same way: a ~500 basis points hike within a year.

To put this in context, this is a much sharper climb than the one before the 2008 crisis.

Shortly, CEOs could no longer afford the commitments they’d made to white-collar employees, as rising interest rates made the debt they owed more expensive than they had planned for; so they were forced to act.

The quickest way to cut costs is to cut personnel spending.

With this, they could either admit that “we over-hired and we got it wrong.”

For example, as Zuckerberg actually said in his late ‘22 message:

“At the start of Covid, the world rapidly moved online and the surge of e-commerce led to outsized revenue growth. Many people predicted this would be a permanent acceleration that would continue even after the pandemic ended. I did too, so I made the decision to significantly increase our investments…”

“We've cut costs”… and “We're restructuring teams to increase our efficiency. But these measures alone won't bring our expenses in line with our revenue growth, so I've also made the hard decision to let people go.”

Or the executives can go with this message: we’ve found AI can replace human workers, and now we're an AI-first company.

The first message tanks your share price; the latter gives your shareholders peace of mind.

You don’t need me to tell you which one most of them reached for.

The Disproportionate Claim

So now you know the motive. The question is, which direction does the evidence support?

4 years and counting, I’ve been tracking AI (especially LLM) adoptions. I’ve seen only a handful of successful implementations. And even then, they've mostly handled repetitive tasks, not replaced the jobs those tasks belong to.

How does a task differ from a full-time job?

List out a job you care about (like yours); e.g., a software engineer breaks down into 10, maybe 15, discrete tasks. Apart from ‘typing code’ (this is my favorite term from a CEO)… you need to groom tickets with product managers, pair-code with your teammates; if you are a lead, you need to review performance and have one-to-one meetings, and so on.

Yes, AI can speed up coding, but this is a task, not a job; hold this distinction because the whole question of whether AI takes your role comes down to it. AI can draft a basic legal contract or create marketing materials; all these are tasks.

The success of AI implementations is currently stuck at the task level. (Read my analysis of Stanford AI Index’s benchmarks on tasks vs. workflow.)

So if AI is only clearing tasks, not jobs, what explains the reasons for those disproportionate thousands or even tens of thousands of AI-attributed layoffs?

When questions get messy, the answer with the fewest assumptions is usually the answer (if you’re familiar with Occam’s razor).

So the simplest explanation here, when CEOs and the media say AI will replace engineers, the former are likely lying, and the latter are parrots.

Here are the three latest data points from micro to macro.

While most companies claim they lay off because of AI, as the chart shows:

The untold truth is, in the same report, 59% also admitted they emphasize AI in their layoff announcements because it plays better with stakeholders than saying “we have a budget problem.” And only 9% said AI had genuinely, fully replaced roles.

Now, zoom out to the wider job market, the aggregate agrees.

Another firm that has counted American layoffs for decades logged about 1.2 million job cuts in 2025. The number of roles pinned on AI: only ~55,000. Under five percent, ranked behind government cuts, behind market conditions, and also behind plain restructuring.

Let’s zoom out even further, shall we?

Guess what the unemployment rate was in April 2026?

4.3%

Most economists would file this under “healthy, slightly tight labor market,” rather than “crisis” (see the image below, especially if you compare it with the period between 2008 and 2012; ignore the Covid peak, which was driven by service-sector job losses).

So across all levels (company level survey, the layoff count, the BLS official unemployment rate), no matter how reluctant, the data aligns:

AI shows up late, small, and overhyped.

Some well-known figures have started calling out the BS.

Take Marc Andreessen (cofounder of Netscape, then a venture capitalist), who called AI the “silver bullet excuse,” and reckoned most large companies are 25 to 75% overstaffed, left over from the COVID years.

Or the latest, from Andrew Ng:

“Businesses have a strong incentive to talk about layoffs as if they were caused by AI. Talking about how they’re using AI to be far more productive with fewer staff makes them look smart. This is a better message than admitting they overhired during the pandemic when capital was abundant due to low interest rates and a massive government financial stimulus.”

So the data is clear for now, i.e., AI isn't responsible for the firing. But 'not yet' is cold comfort if you don't know what to watch for.

If you want to know whether your own job is safe, you need something more durable that can hold up no matter how the conditions change.

The First Principle of Jobs

Here’s a cleaner way to think about all this jobpocalypse joke than the ATM story you've probably already heard.

In case you have never heard of the ATM story, here’s the short version: ATMs arrived in the 1970s, many declared the bank teller dead, but the opposite happened. Yes, there were fewer tellers per branch, but cheaper branches meant banks could open more of them in towns that had no previous footprint, and the total number of tellers actually grew.

Here’s the twist, though. As mobile banking grew, tellers started disappearing. See this chart:

What does the story tell us? Nope, not the Jevons paradox, too academic. Instead, under the first principle, it’s just a simple supply and demand dynamic at the fundamental level.

We need to start by asking,

why does a job even exist?

A job isn’t granted to you so you can pay for your baby’s diapers or your own.

A job exists because someone, somewhere, feels a pain sharp enough to pay to make it go away. A company builds something to solve that pain. You're one of the screws holding that solution together.

So if we apply this thinking, and ask,

why a teller to begin with?

It’s because we want to make a transaction, or cancel one, or save money, or lend some.

The next question:

Then why didn’t ATMs replace tellers? Instead, they created more?

If we look back at the period between 1970 and 2000, apart from what we talked about, banks were experiencing growth; there was a significant technology gap, and ATMs only addressed part of the issues when moving money.

You could get cash at 11 pm, yes, but you still needed a branch to do almost everything else: transfer money, dispute a charge, open an account.

The pain couldn’t be fully resolved by this bulky machine. So teller numbers didn’t collapse. They shifted to more on-site banking.

Fast forward twenty years.

In the post-internet era, nearly everyone has access to the internet and a mobile phone; the US’s payment system is becoming more mature, so more features that were only on the screen in front of a teller are now also on your phone. Anywhere, anytime, no queue.

Once the full job moved to your phone, the branch became optional. That’s when teller numbers fell.

Therefore, what ended teller jobs wasn’t ATMs but something else that provides a complete solution: a much better experience, at a lower cost, more accessible, with software that can be updated easily.

Or, take myself for an example, I was an ops product lead, not because customers wanted a product leader; in fact, I don’t think many of them ever heard of this job. But simply because our customers get hangry (hungry + angry) and want food, the food delivery company needs someone to maintain the operational systems their admins use; hence, they paid me for it.

Just like if the world had no thieves, then what would the police be good for?

With this same principle, you can answer the questions for any jobs that are deemed at risk: why an engineer, why an accountant, why a customer service representative, why a lawyer?

So “will the job disappear?” was always the wrong one to ask, try “what did customers actually need? And will this new technology make the journey to satisfy the need smoother?”

Takeaway

I talked about this in “Why AI Shopping Isn’t Going To Work”; it’s the same principle. People won’t buy via AI shopping, no matter how fancy it sounds, because it has fundamentally ignored the needs when you shop.

So AI, in any case, will only kill a role when it finally satisfies the underlying need (or resolves a pain) well enough (smoother journey or much cheaper) that the role stops being needed at all. Which, with current AI capabilities, isn’t happening.

Run the following task to evaluate your own work.

1. Break down your job into tasks,

2. write down the pain your role exists to remove,

3. finally, look at where AI actually is in relation to that pain.

Ignore everything else around it; these are mostly noise.

Only the pain and the fundamental need matter. As long as AI is still running in circles, you have time. When it closes the gap, you'll know before any earnings call announces it.

Here’s the latest earnings

• Salesforce revenue hit $11.1B, up 13% YoY.

• Workday came in at $2.5B, up 14.5%.

• Asana FY26 full-year revenue landed at $790.8M, up 9%.

• Zoom just raised its FY2027 forecast, up 5.5%

If Salesforce, Workday, Asana, and Zoom are still growing, then the “AI kills SaaS” thesis looks too simplistic.

Maybe what’s actuall…

Read more

Bigger context windows, smarter reasoning, and longer chains of thought are what AI labs are pushing for.

The AI CEOs keep telling you, any month now, the AI will understand your business, run workflow, draft contracts, and be the yellow brick road to no more lazy human employees.

In the case of the Gemini 3 Hackathon, judges had thousands of entries to look at. Somehow, they handed the top three prizes to the teams that had stopped believing the pitch.

The grand prize is won by accepting that the AI shouldn’t make the final call. The 2nd won by forcing AI to think out loud so we can verify the logic. And the 3rd won the prize by accepting that the AI shouldn’t pretend to know what’s been lost in memory.

While the CEOs have spent years telling us AI will replace human workers (and many are convinced). None of the awarded products would have received the prize if that were true.

I’ve studied all three winning products, and I’ll walk you through each point of the workaround they built to compensate for what AI failed to achieve; what’s more important is that I will teach you how to read any AI pitch backward, so you’ll also have my ability to see through AI BS.

First up is the human-in-the-loop problem that the industry keeps trying to design around.

Human-in-the-Loop, Translation, AI Can’t be Trusted.

Globot is built for the moment any supply chain manager has to make a million-dollar rerouting call with a geopolitical or natural crisis breaking in real time.

This product starts with four agents pulling information from each crucial source to help make supply chain decisions.

It pulls geopolitical signals from Reuters and Bloomberg, calculates the financial impacts, redraws the shipping route, and checks whether the reroute will change the insurance status.

And finally runs a fifth agent whose only job is to find holes in what the other four just concluded. After the check, the system then hands over for human action.

This fifth agent and the human intervention are part of the product, and the key points to discuss here.

You’ve heard of variants of this architecture. For example, Anthropic baked it into training as Constitutional AI, ie, one model writes, a second model marks its homework against a written rulebook. Some call them ‘critic’ or ‘judge,’ whatever the label, the design is the same.

Even after the judge, you still need a layer of humans to make the final call.

Because unfortunately, this exponentially more expensive multi-agent setup, plus the judge, still cannot be trusted to make the right call confidently (basically, the judge can still BS its way out).

No matter how many check layers there are, it just cannot be trusted to make a $X million decision on its own.

The whole “human-in-the-loop” phrase is load-bearing for this architecture (though the label sounds like some kind of brilliant, all-inclusive, and be-kind-to-humans gesture).

Ironically, the cheapest and final fix is still the human approval gate.

No, it doesn’t just stop here.

Because if you dig deeper, you realize this creates one thing none of these AI lab CEOs expected to see: jobs.

Someone now builds and safeguards the agents; experts to verify the agents’ output, and someone sits at the end of this pipeline, reviewing high-stakes AI outputs. And repeat.

So you aren’t saving the cost of getting rid of human labor after all.

Next, force AI to reveal its thinking.

Explainable AI, Translation, The Black Box Is Still a Black Box.

The demo for Aegis opens with the volume problem.

When a natural disaster occurs, the regional 911 system is most likely overwhelmed. You'd have ten thousand distress signals (calls, SMS, tweets, drone feeds) within a matter of minutes.

There are mainly two design pillars in this demo.

First, it has a total of five agents. The coordinator routes raw signals, triage assigns priority, surveillance analyses incoming drone imagery, logistics calculates routes for rescue assets, and the reporter writes the post-mission summary.

A similar multi-agent scaffolding to the one we just saw in Globot.

And it’s there for the same reason: a single model can’t hold a complex task end to end. Worth pointing out: this is the moment you’d finally respect how incredible the human brain is in handling such a complex task.

Second, the Explainable AI.

Early users couldn’t act on the AI’s priority scoring because the AI’s recommendations are blackboxed, so they couldn’t see why one casualty had been ranked above another. The same black-box problem that has killed many enterprise AI pilots.

The fix was to force every agent to display its reasoning next to its decision, visible on the main canvas.

The industry has a name for this. It’s sold to enterprises as Explainable AI, or simply “show your working.” But showing the reasoning doesn’t open the black box. It just gives the black box a narrator.

Both the multiagents and the explainable AI ‘solutions’ fixed the issues but created other unwanted consequences.

First, as we’ve covered earlier, leveraging multiple agents is exponentially more expensive than a single one. Especially when AI labs like Anthropic are aggressively redesigning their pricing. It’s expected to spend 4, even 5, digits in US dollars monthly just on a single small project.

Which arguably is the cost of hiring an engineer full-time.

Second, the forced thinking-out-loud.

When people see a confident, readable explanation attached to an automated decision, they are more likely to accept it — including when it is wrong. It has a name, automation bias: the tendency to trust automated outputs over our own judgment, even when the evidence shows the opposite.

Deloitte's research for the UK Ministry of Defense illustrates the risk with a simple driving analogy. Imagine you're driving and see a 30 mph sign, but your dashboard shows 60, so you accelerate, trusting the machine over your own eyes. Or you skip a turn you know is right because the sat-nav didn't prompt it.

Either way, the system has quietly replaced your judgment instead of supporting it. Here’s what’s said in the report:

In this situation, the use of AI has unwittingly impaired the operator’s judgment, rather than improving it.

But all of this assumes the system even knows what situation it is returning to. Which brings us to the next demo.

Engineered Memory, Translation, LLMs Start From Zero Every Time.

A user with limited vision asks Netra to find their water bottle.

The system scans the room and says: “Your water bottle is on the desk, to the left of your silver laptop, approximately a meter away.” The user walks to it and picks up the bottle.

Now move the water bottle. Don’t tell the AI, and come back tomorrow and see what happens.

The project brief calls the underlying system the Memory Palace, “a persistent JSON store” that “builds a robust, long-term spatial map of a user’s most important locations.”

However, that is precisely the problem with LLMs in many similar implementations.

Yes, the water bottle is still on the desk, to the left of the laptop — in the file.

But what if someone loaded it in the dishwasher and forgot to put it back next to the laptop? The model’ll send the user confidently to an empty desk.

The brief didn’t mention how often they update the memory. But it’s going to be a delicate balance between real time and cost.

This is statelessness, one of the serious negative characteristics of LLMs.

The model has no continuous thread of experience. Every session, it wakes up as a stranger with notes. This demo is only a workaround that introduces a new failure mode on top of the original problem: stale data that the model treats with the same confidence as fresh data because it has no mechanism to distinguish between them.

At demo scale, it’s a file on a server. At any real deployment scale, it is a retention liability and risk of contamination, scope creep, and leakage of what’s in the memory.

This is one of the reasons why Klarna piloted AI customer service bots that could fully replace human agents, but rolled them back to "assistant only" because the bot forgets you the second you switch from chat to phone, or if your connection drops.

And once you see that pattern, it’d be hard to unsee it.

The industry keeps selling these workarounds as breakthroughs. I’m going to teach you how to read the pitch backward so you can see through the sales lines, understand the model’s limits, and assess the cost of workarounds for these limits.

Learn to Read Any Tech Pitch Backward.

You probably started this article expecting an analysis of why each demo did right to deserve the hackathon prize. But I'd rather rip the label off and show you the plumbing underneath.

Learning to read things backward is going to help you evaluate AI’s true value and cost

But how?

Let’s start by understanding this: a solution always generates a new set of issues. The railroad created rail engineers. The cloud created DevOps and FinOps.

We’ll use Harness Engineer/ Multi-Agent Orchestration, for example.

Forward reading:

Like this one:

We’re hiring harness engineers to build sophisticated multi-agent systems that orchestrate specialised AI workers.

Again, a sales pitch.

Backward reading:

A single model loses the plot on anything longer than a certain number of tokens of real reasoning. So you split the job across x narrow agents, maybe add one more to check their work, and hire a human (the “harness engineer”) whose entire role is to glue the crew together and keep it from drifting.

Ignore what the pitch says, ‘one harness engineer + multiagents replaces a five-person team.’

Because the invoice by the end of the month says otherwise.

A harness engineer in 2026 is a specialist hire at $180k–$350k, and you don’t get just one.

Underneath them sits a compute bill that scales multiplicatively; an x-agent crew with a critic re-ingests prior agents’ outputs as context, so a single task runs 10–15× the tokens of one model call, and labs now bill agentic tool calls separately on top of seat licenses.

And then the hidden stack appears. You need databases for retrieval, orchestration services to route work, memory stores to preserve context, and monitoring tools to stop the whole thing from wandering off.

Even the guardrails tell on the system. Iteration caps, timeouts, and cost ceilings are not just safety features. They are admissions that, left alone, an agent may keep spending without converging.

Zoom out from the engineering; you also have the oversight tax: Carnegie Mellon and Stanford found that when humans supervise agents' work closely, productivity drops by about 20% compared with humans working alone.

So the cost of operators doesn’t really disappear as many might dream.

Instead, they get replaced by a more expensive team, a variable compute bill that grows with usage, a stack of net-new infrastructure, and a human still sitting at the gate to sign the high-stakes calls.

But most importantly (this will likely scare away most bosses), a brand-new way of operating that bosses have nearly zero control over and zero expertise in.

None of these is something you intended to buy. However, each one is a tax on the model’s limits, and you are charged with a new name tag.

For paid subscribers, I will send you a few more AI terms, with backward reading examples. Sign up if you’d like to receive this exclusive piece too; it will save you millions and headaches.

I hope this one is insightful. If yes please do like and share this, this article is only possible with your engagement and support.

Anthropic doubled Claude Code’s limits across paid plans and sharply raised API rate limits, after a compute deal with SpaceX freed up capacity.

→ Anthropic

The sequence is the tell. Anthropic spent the spring deciding your $20 plan wasn’t built for agents and throttling them accordingly; the devs got loud about this d…

Read more

Because the tool changes. The prompt trick dies. The workflow gets absorbed into the next update.

And then you do it again.

Believing that every try (the subscription, the re-rolling, and tweaking), you are one step closer to the guru’s pitch of 10x efficiency.

Only nobody calls it what it is: a nice and slow burnout with better branding.

I read the two latest studies and can confidently tell you that if using AI has somehow made you feel busier, not freer (as promised), you are not imagining it, and you aren’t alone.

The promise was that AI’d save us time. But for many, it has created a second job: checking, fixing, comparing, and trying not to fall behind.

So I’ll show you what the studies found, my own burnout story, and why learning AI tools might be the least useful way to prepare yourself for what’s coming next.

AI Brain Fry is Real

BCG found that AI seems to help when it replaces boring, repetitive work; people who use AI this way reported lower burnout.

However, when AI turns you into the manager of the machine, checking outputs, switching between tools, correcting mistakes, and so on… all of which is where the brain fry starts.

Workers doing high AI oversight reported 12% more mental fatigue and 19% more information overload.

The AI gurus show you the best possible output but never discuss the hidden costs of AI, extra time to manage, and oversight.

The chart makes the trap obvious.

The productivity drops sharply when people use too many AI tools. One tool feels useful. Two feel powerful. But by the fourth, you’re no longer working faster. You’re only middle-managing software.

Another finding that stood out was marketing. Marketing professionals, surprisingly, were the most exposed to AI brain fry.

And that makes sense, because the vaguer the work, the easier it is to get trapped in a forever re-rolling battle with AI. Code either runs or it doesn’t. A calculation is either right or wrong.

But a headline, a campaign, a thumbnail? That can always be a little better, so the re-rolling gets addictive and overwhelming.

Google Dora’s report shows the same problem from another angle.

It surveyed nearly 5,000 enginers. While they found that more than 80% said AI increased their productivity. It also shows that AI doesn’t remove work evenly.

Some archetypes are worse than others.

Teams stuck with unstable systems, reactive work, and high-friction processes see high burnout. Same for those constrained by process, because managing the workflow is exhausting.

On the flip side, teams with low friction and stable foundations had much lower burnout.

Which means AI was never the whole story. It amplified the system underneath.

If your work system is already messy, AI isn’t the fairy godmother that cleans up the mess. It’d just help you produce more, which would make it messier.

And that is why the AI tools tutorials, or the claim that you’ll be replaced by those who use AI, are horrible, horrible noises.

What does the AI Slot Machine Cost

This is the part you’d recognized immediately.

Using GenAI to chase a better image or code works the same way as swiping TikTok or pulling a lever of a slot machine gets you hooked. You prompt, re-roll, tweak, re-roll.

And it doesn’t feel like wasting time, because it wears a productivity costume. But worse than the slot machine, GenAI not only costs you attention but also time and credits, which create a sunk cost when you quit.

I’ve experienced it firsthand, as many of you have.

Tools, large language models, skills, reports, you name it. Every new update, I’d check it out and couldn’t help but try it. I became obsessed, worried about what I might have missed in my analysis, and often felt out of breath during big weeks (when many things hit the market at once).

And then my body started sending warning signals, gastritis, for which a GP couldn’t find a cause. Still slowly recovering.

Now, weirdly, the more I understood AI, the less the demos and releases affected me.

Here’s how I’ve come to see it.

First, Learning How to Use AI is Pointless.

The first mistake is treating AI tools like a durable skill.

Most of what people call “learning AI” is really learning the current interface and interactions with current models/agents.

But because models update so frequently, much of that expires fast.

Not to mention, the goal of AI development is to eventually let anyone complete tasks through natural language. So any barriers to using AI will be flattened in the next iteration.

Teaching others to vibe code, or using AI to generate movies or images, is a fragile endeavor. The technique expires when the tool updates. What doesn’t expire is the eye you build from doing the work yourself: seeing where the code is brittle, the design feels cheap, an argument sounds clever but full of fluff.

Those are the things that equip you to best assess whatever comes next.

Researchers tested this not long ago. They trained a model on visual artists’ own drawings, then had both the AI and the artist produce work again. The artists still won on originality and aesthetics.

The model had their style. It didn’t have their perception.

With this in mind, there’s no point in just learning the ways of manipulating GenAI, as long as you have the judgment of what it can and can’t do (which isn’t a simple one).

So when I have an extra hour, I’d rather spend it improving the work itself. Reading or challenging my own thinking, and writing down the argument without asking a model to rescue me all the time.

Because that is what gives us the eye.

Second: Using AI to do Your Work is Also Pointless.

If you are an early adopter like me, simply think back to what GPTs you chased in 2023, or the OpenClaw you lost sleep over.

Setting up folders, rules, skills, and you finally get a workflow running that produces the output somewhat close to what you’ve expected.

But there’s always something off. The research isn’t done right, and the output quality varies.

Then a new model update arrives, and suddenly all the setup gets absorbed into the product.

The most value you provided, at best, was teaching the model companies what serious users wanted.

And before you’ve even had a chance to enjoy the edge of a finely tuned AI workflow that was given to you, that gap between you and everyone else gets wiped out overnight.

Just imagine this: a colleague who never bothers researching OpenClaw or Hermes (no, not the bag), whose AI knowledge extends to occasionally reading a headline, plodding along doing their work w/o trying to automate every single task.

When a new model comes along. The colleague could now achieve the workflow you’ve spent a weekend on with a single-line prompt.

From the outside, the return is poor, and the time saved amounts to nothing distinguishable.

The Only AI Engagement Makes Sense

There is one version of engaging with AI that’s worth something, and it starts before the tool.

My 10 years of product experience taught me that people love rushing to a solution, regardless of a dashboard, a feature, or a workflow, because they feel empowered.

The useful question, however, should come much earlier. Asking, what is actually broken? And what do I want to achieve?

Here’s an example: most people find it valuable to brainstorm with AI.

In fact, brainstorming is arguably the worst thing you can do with AI.

More often than not, you hold the key to a much better answer than a model would.

Because every idea AI generates is already everywhere on the internet. It gives you the most statistically likely version of an idea. That’s the last thing you need when you’re looking for an angle nobody’s taken. I reviewed a study on this.

This latest study used seven major LLMs — GPT5, Claude, DeepSeek, Gemini, and more — and tested thousands of business strategic questions across industries.

The results were remarkably similar across models: augmentation, think long-term, innovate, and decentralize. As you can see in the image below:

You don’t need 15,000 trials (as the researchers did) to see it. You just need to notice when you’re getting back the internet’s average opinion.

It doesn’t require you to become an expert using AI, but an expert in your own domain, and of knowing what you want.

So, the only AI engagement that earns its place looks like this:

1. Try AI when it’s new enough to matter, learn the actual trade-offs,

2. Then decide deliberately whether it belongs in your work. That’s different from chasing every release. It’s also considerably cheaper on your nervous system and your wallet.

3. Prompt less, think more.

When the AI’s Wild-Goose Chase Is Over

The models will keep improving. The releases won’t stop.

But I kept noticing something: AI made me faster at producing things; what it couldn’t do was tell me which things were worth producing.

And I often end up throwing away most of what AI produces.

Which is why the most important thing is our judgment. Which is also always the bottleneck, and the ceiling that the next release won’t close.

Not because the AI companies aren’t trying to replace judgment, believe me, this is all they’ve wanted.

But it’s less of an engineering problem and more of a human one.

As I explained in my work, distilling your colleague.

Once you see this, the new releases stop landing the same way.

This can be a cliché, but the question was never which tool is fastest. It was always what you’re trying to do with it.

IBM surveyed more than 2,000 organizations and found 76% have created a Chief AI Officer role, up from 26% last year. And by 2030, many CEOs expect 48% of the ops decisions to be made by AI.

→ IBM

That sounds dramatic, but IBM is mostly talking about operational decisions with clear guardrails: pricing updates, in…

Read more

Pretty much the same day, a different roster of PE firms did the same with OpenAI.

But what drove them to make the same decision?

Between the PEs and the two biggest AI labs, in total, committed $11 billion to the same problem,

ie, get mid-market companies to use THEIR AI.

Because Anthropic and OpenAI managed to convince those overseeing the money to take on the risks on behalf of LPs (will explain soon) and portfolio companies, in exchange for a high-margin services business that outlives the investment.

The press release sells this as democratising AI for the mid-market.

I read the deal docs. Turns out, this isn’t an AI story; it’s about money, distribution, and the people who carry the risk aren’t at the deal table.

To see who’s actually benefiting and who’d suffer from it, we need to start with what the forward-deployed engineers actually do, why PE is the channel of choice, and the devil in the details.

Forward-deployed Engineers: an Old Dog and an even Older Trick.

No, the idea of “forward-deployed engineers” isn’t exactly groundbreaking.

It’s pretty much a rebrand of Solutions Architects or Sales Engineers; it’s customer-facing technical support on deeply technical products, often focusing more on deployment and post-sale customization.

A solution architect in AWS helps their assigned customers resolve technical issues and, apparently, as the name implies, upsells additional AWS services.

Same as IBM service delivery or Salesforce CSAs (Customer Success Architect).

The label got popular because Palantir wears it, and Palantir reads as more AI than the rest.

If you have a product that is complicated to onboard, requires a lot of detail and time to implement well, and is highly specialized, it’s a common approach to have a “service-led growth” team in the organization.

And as far as the role of the “forward-deployed engineer” goes, it’s part technical expert, part sales, part customer feedback capture.

So a more interesting question isn’t anything related to the forward-deployed engineers; instead, why did both OpenAI and Anthropic announce the news simultaneously, collaborating with Private Equity firms?

Why Private Equity (PE)? A Leg In the Door

The first, and obvious, answer is that PE firms are great sales channels.

Rather than having to target every mid-market company directly, they can rely on PE firms (which own tons of high-growth portfolio companies) to introduce them.

You need to remember that both Anthropic and OpenAI are new kids on the block.

Given the authors’ love of mafia movies, imagine this:

The tech giants are like dealers with their own territories…

Microsoft owns the enterprise corridors. Excel first came out in 1985, and gradually, Microsoft came to hold the enterprise and old-style industry by the balls. Then Gmail arrives, and a few years later, Google has the heart of startups, and so on.

These guys don’t need to knock on doors anymore; the customers are so hooked (with no better alternative), so they all crawl back to one of them.

Then Anthropic and OpenAI arrive. They invented the weed+, a drug that we all feel the adrenaline when using, and long for when the services are down.

So to infiltrate and claim the territory, of course, they can’t expect help from the old-timers; their only safe bet is to work with someone who isn’t threatened by their presence, or even welcome it.

So they made a deal with the users’ financiers: the PE firms.

Imagine that! For a portfolio company, instead of borrowing money to buy their weed+, they get it straight up whenever they collateralize a part of their baby.

That’s the role PE plays in this complex setup: they typically have significant sway over how companies operate. It’s a pretty powerful relationship.

The Joint Venture (JV)

At the very top of the enterprise bracket, the implementation players are the typical big consulting firms selling transformation services; it used to be cloud, now with AI.

This new joint venture (JV), however, is explicitly targeting the mid-market. Here’s a part of the announcement:

The company (the joint venture) will serve as an accelerant in bringing AI solutions to mid-size companies… of both portfolio companies of the investment firms and independent companies that can benefit from the platform.

Basically, the targets are businesses that can’t necessarily afford McKinsey and Accenture-level fees, and don’t have in-house capability to spare. That’s the sweet spot to land forward-deployed engineers.

Or as the press release so nicely puts it, “democratizing access to forward-deployed engineers.” Democratizing used to be the favorite go-to term for blockchain; now duly recycled for AI.

The vendor (Anthropic and OpenAI) could build a partnership via traditional structures, of course, but the JV has a few advantages.

To start with, PE firms are committing up to $300m each as a statement of intent to the market. This kind of financial backing makes it look and feel more like a durable business than a typical Go-to-Market partnership, which tends to be looser.

On top of it, Anthropic stays involved in the deployment, rather than relying on a 3rd party implementation partner.

And because of it, they keep the feedback loop. What works, what doesn’t, where clients are struggling? Even if it doesn’t lead directly to training, it certainly shapes the direction of the products; we call it deployment intelligence.

And with this ever-evolving product, AI, the upside is obvious for them to stay involved for a longer time.

It’s no longer a typical transformation engagement, which has a clear end to an ongoing relationship. Here’s the press release:

Claude’s capabilities change on a monthly or even weekly basis, which creates a different kind of engineering challenge than traditional software deployment.

The systems that companies build with AI need to evolve as the models underneath them improve.

So what this new venture really is is a company that will have a long-running, recurring relationship with its clients.

Great as this all sounds, there are serious risks that the media coverage misses. To see them clearly, you need to know how PE actually works under the hood.

A Quick Detour in Private Equity

Skip to the next section if you've sat in those rooms.

So to start with, PE firms are run by fund managers; their job is to use that capital to buy companies, and aim to sell them between 4-7 years later for a return. Fund managers are commonly called General Partners, or GP for short.

The capital they use is provided by Limited Partners (LP).

In practice, these LPs are typically sovereign wealth funds, large family offices, high net worth individuals... And the LP pays the GP a management fee, and if there’s a profit when selling, the GP gets a portion of that too.

At a high level, that’s the structure of every PE fund; the principle is the same for a one-man band or a Goldman Sachs-owned vehicle.

You might notice that when someone from PE talks about value creation, it is all they talk about. And GP tends to put operational teams in a company and pushes the business to aggressively optimize.

The keywords are: relentless focus on a small set of changes, strong financial discipline, limited patience, and ruthless efficiency.

Done well, the company comes out worth more. Done poorly... Google "PE horror story", you'll be busy for hours.

Obviously, these PE firms don’t buy one company at a time. Quite often, they run a portfolio of 5-10 of these companies and buy add-ons to integrate with them.

Now, considering that the companies involved in this structure with Anthropic are amongst the largest on the planet, their portfolio is huge. This is exactly what Anthropic needs. A rotating cast of companies, where they have direct access to the people pulling the strings.

With the structure in mind, the cracks in this JV are much easier to see.

There are two categories of problems.

The first is operational: things that go wrong when forward-deployed engineers hit the reality of a mid-market company. The second is structural: conflicts built into this JV.

What Will Go Sideways in Practice (Operational)?

Operational Risk #1:
“Improving the business” that you have 0 clue about.

As mentioned, PE’s standard play is to land an operational team in a business it has just purchased. Those teams already have playbooks. So, like appointing a CEO or CTO they’ve worked with in the past, with the express aim to “do your magic and get x return on this business for us”.

I know because I’m one of those.

The principle behind these playbooks relies on pattern matching.

If you’ve done one of these in the same industry, you’ve learned from your mistakes and your successes. In the process, you have built up two very strong muscles: judgment and domain understanding.

However, forward-deployed engineers typically aren’t domain or industry experts.

Instead, they are AI implementation experts. They know how to get the best out of Anthropic’s products.

But they don’t necessarily know how to adjust a pattern for the reality of a procurement team in a small Fast-Moving Consumer Goods business, for example. Or how to roll up the shared IT needs of a group of vet clinics.

And while the standard answer of every transformation project is “that knowledge comes from the domain experts in the company.”

In reality, we all know that: a) the knowledge won’t be documented, b) it’s likely not easy to explain, c) John, who knows everything, actually left 2 years ago.

It’s a classic transformation consulting problem, but now with AI.

And AI, with its innate lack of judgment or inability to learn and contextualize within the organization, is likely to make this worse rather than better.

Operational Risk #2:
Data mess layers up like a mushed wedding cake

One of the defining qualities of mid-market companies, particularly at the lower end, is underdeveloped internal data.

Siloed, needs interpretation. At best, a few dashboards and spreadsheets. At worst, the 'CTO' is still running queries straight on production. This is also a sign of poor data practices (such as governance to ensure data consistency, correctness, and so on).

In most cases, decision-making is based on intuition and experience, then justified by select data rather than by data, ie, the data always agree with the boss

All of these, plus AI (non-deterministic for one), is a recipe for a perfect project disaster.

My guess is that the JV will do a lot of data groundwork while they grab a few low-hanging fruits (a classic way to show momentum with quick wins).

I’d be happy to be proven wrong, but this could easily become “just another tool”; much like we all needed an ERP, a CRM, an HR system, a data warehouse, ... We’ll now just also need agentic-ness.

And that's before we even get to the structural problem, which exists whether the AI delivers or not.

Built For The House To Win. (Structural Risks)

Misalignment #1: Whose interest does the GP actually serve?

Yes, the GP has the power to deploy the funds, but the funds aren’t actually theirs. The LP writes the cheque. The GP makes the calls.

What’s always held this together is fiduciary duty. The concept that the GP delivers impartial advice, in the best interest of making the company more valuable.

Now hold that thought against this new structure.

What’s the GP going to optimize for? The portfolio company’s value, or the JV’s revenue? Because now they also own part of the JV doing the implementation. When the JV recommends more services, more spend, and deeper engagement, the GP would benefit twice:

• Once from the portfolio company’s improved performance,

• and once from the JV’s revenue.

That said, this isn’t a new concern; some common solutions are in place, but they still haven’t changed the fact that, structurally, there is a serious conflict of interest.

Misalignment #2: Implement once, a lifetime commitment?

If you remember, PE holds companies for 4-7 years before selling them. So what exactly happens to this deeply embedded AI when the company sells?

In the old days, it was clean. Whatever systems they put in place stayed with the business. The new owner inherited the upside (or downside) and took it from there.

This model is different in two ways.

For the portfolio company:

The JV stays embedded after the portfolio companies change hands.

So the AI sticks around because after deeply embedding Claude everywhere, you’re not going to suddenly rip it out for ChatGPT.

But now the JV stays embedded after the sale. Which means whatever the new owner does, eg, operational changes, new product roadmaps, feeds back into the ex-owners’ knowledge base, gifted for free.

Mark my words here, the language around how this usage and data feed back, and who gets to use the insights, is going to become a new element in due diligence.

For the acquirer:

Before the sale, the GP is running both sides of the table. They can keep the AI costs artificially low — the portfolio company looks lean, the margins look clean, the exit multiple looks earned.

The new buyer inherits the real bill without compensation if they don’t do careful due diligence.

This isn’t the worst yet!

Misalignment #3: A bad apple spoils the whole bunch.

So far, we only fucked up the LP, the portfolio company, and the future owner of the portfolio company.

What about the PE firms?

As mentioned, they are used to pattern-match across their portfolio. That’s a strength, until it becomes a single point of failure.

In the old model, a PE firm sent an operational team, or a CEO, into one business at a time. If it went wrong, it was one out of a few hundred.

This JV works differently. The whole point is a standardized AI playbook rolled out across the portfolio, ie, the same forward-deployed engineer SOP.

Let’s say, a PE firm with ten healthcare businesses. One flawed AI agent is handling clinical documentation; all ten go down.

And because the PE owns the JV, nobody inside the portfolio gets to ask whether this vendor is the right call.

Not to mention, their whole portfolio stack now relies on Claude being live.

Imagine you hired a contractor to work in your office every day. Over the last 90 days, they failed to show up, or showed up broken, more than 100 times. What’s worse is that your entire team was built around that contractor, so for each no-show, your whole operation freezes.

Exactly the instability from Anthropic in the last 90 days.

Or, given the model competition is so fierce, if Anthropic falls behind GPT-7 or Gemini-X in capability, or if pricing power shifts, it’d cost the PEs in the JV both arms and legs to get out.

So, benefits?

Yes, there are benefits.

After all, these are the shrewdest bunch on earth.

For Anthropic/ OpenAI

A captive user pool. If it shows signs of life quickly, that’s an IPO valuation driver.

Anthropic also keeps a tight feedback loop from implementation since there are no independent implementation partners in between. That should help them improve their product.

For the PEs

The GP has just set up a high-margin service business that will stay with their portfolio companies after they’re gone. And the companies they control are the JV's clients.

Even with some fee offset early on, the important part is that this revenue stream likely outlives the investment. Uncorrelated returns that don’t need to be shared with LPs.

On the other hand, the LP gets possibly better exit valuations, but by and large bears most of the risk of misalignment, implementation failure, and otherwise misallocated investment decisions.

The portfolio companies, in the end, are going through the new wave of transformation.

Whether this AI transformation is successful depends entirely on if the participants read the small print and can manage the risk accordingly.

Not to mention, the moment AI is involved, risk management becomes a different game entirely.

OpenAI has raised $4 billion from 19 private equity firms and consultancies to start a new company staffed with ‘Forward-deployed engineers’. The job: help PE portfolio companies integrate AI (and to land themselves more b2b users)

→ Bloomberg

By coincidence, or not so much coincidence, Anthropic also announced the exact same joint venture with friendly private equity funds. It’s funny how two AI labs announced the same business strategy on the same day.

I’m writing a deep dive on this, out Sunday. Keep an eye on it!

2. The U.S. Government Can Now Stress-Test AI Before It Ships

A Commerce Department agency (CAISI) has signed agreements with Google DeepMind, Microsoft, and xAI, giving the federal government access to frontier AI models for national security testing before public release.

→ NIST

In addition to its existing partnerships with OpenAI and Anthropic, the Commerce Department Agency will also review the homework of other AI labs. Do they really have the capability to do it, or is this just a stamp on a piece of paper?

Leave a comment

3. Is xAI the Next CoreWeave?

SpaceX announced its xAI unit has struck a deal to supply computing capacity to Anthropic, following a similar deal with another company weeks earlier. xAI is quietly becoming an infrastructure play.

→ Anthropic

Elon Musk is like running a food truck that started with American, pivoted to Chinese, added Indian, and is now opening a French à la carte counter.

Two readings:

1. Either xAI, the model is quietly becoming a dead end, and this is the pivot.

2. Or he looked at CoreWeave’s business and decided the middleman margin was better than the frontier model race. Neither is reassuring for Grok.

4. AI Models Have Invented a Private Shorthand — and It’s 11x Cheaper

Instead of reasoning out loud in English before answering, a new training method teaches AI models to think in abstract tokens, a kind of internal shorthand the model develops for itself. As an end user, you still get a normal answer at the end. The result: 11x fewer reasoning tokens used during inference, with accuracy matching or beating standard verbal reasoning.

→ https://arxiv.org/abs/2604.22709

AI optimization is moving fast enough that today’s breakthrough is tomorrow’s baseline. Whether abstract reasoning tokens are the answer or just another stepping stone, this is positive in many ways, in saving energy and the mess created by sourcing lands for data-centers, for one.

5. Anthropic Wants to Replace Junior Analysts?

Anthropic has shipped ready-to-run Claude agent templates aimed at financial services teams: building pitchbooks and DCF models in Excel, drafting credit memos and PowerPoint decks, screening KYC files, and closing the books at month-end autonomously.

→ Anthropic

The media has reached another climax over the idea that junior jobs are f*cked by AI companies. However, based on what I read in the system card, even Anthropic’s latest model, Mythos, would still have a serious amount of hallucinations.

But keep in mind, being 80% correct in finance, 100% that gets you fired.

So is AI there yet?

Not quite.

6. 1 in 1,300 Claude Chats Is Messing With Your Head

Anthropic’s own research found that roughly 1 in every 1,300 Claude conversations distorts users’ grip on reality: nudging beliefs, reinforcing delusions, or blurring the line between the model’s framing and the user’s own thinking.

→ Anthropic

1 in 1,300 seems small, right?

Not at all.

Claude has 19 million monthly active users. That’s at least 14,500 people per month whose sense of reality is swayed by chatbots.

While the team was locked out, their API keys kept billing. They couldn’t view their own usage data because their email addresses had been banned.

A second company hit the same pattern the same month.

I studied both incidents against Anthropic’s commercial terms. I found four structural reasons that make this failure repeatable by anyone, and all of them are written into the terms every Claude API customer has already signed.

At the end of this, you’ll also learn the three most practical ways to manage the risk, recommended by a whole community of CTOs.

But first, the incident. Because the part that should concern you isn't just the suspension, but what happened the moment the account was locked.

The incident

Monday started like any ordinary workday for this ~110-staff company.

The first signs of trouble appeared in the operations channel on their Slack. One person posted a screenshot showing that they have trouble accessing Claude. Then another. Within ten minutes, people across the company were asking the same question:

What happened to my Claude account?

The answer surfaced quickly. It was not just one person’s Claude account. Everyone’s Claude account had been suspended.

The company is a U.S.-based agricultural tech business.

The OP says the company used both Claude Team and a separately billed API account, suggesting Claude was used not only through individual/team seats but also through API-based workflows. The OP describes Claude as part of the company’s “daily workflows” and mentions agriculture-related conversations involving fertilizer, GPS satellites, and other ag-related topics.

Then Anthropic cut everything off.

Just like that, 110 accounts were suspended at the same time. No warning. Anthropic didn’t send the company a notice, but 110 separate ones.

So at first, same as everyone else, the OP thought it must be something he’d done. 

The email was addressed to him, and the link went to a Google Form for individual appeal. He started searching his own recent activity for a trigger. Was it the conversations about fertilizer? GPS satellites? Agriculture-related topics that might have looked wrong to an automated system? 

He had no way to know.

It wasn’t until people started comparing notes on Slack that the picture became clear: the entire organization had been blocked at once. A single mistaken trigger, whether from a human or a bot, could put every account in the company into the dark.

They submitted the Google Form. Then reached out through every other channel they could find — including DM-ing an Anthropic employee who’d commented on a similar case on Twitter.

No reply after twelve hours.

Twenty-four hours passed, no reply.

Thirty-six hours passed, radio silence.

Then came the absurd part.

Remember? The company didn’t have only a Claude Team plan; it also had a separately billed API account under the same administrators. The API keys kept working, and a renewal bill arrived.

But because the admins’ email addresses had been banned, they couldn’t log in to view usage or dispute the charge!

By the time the post hit Reddit, access still hadn’t been restored.

What’s concerning is not only the suspension itself, but this pattern of:

• automated enforcement,

• vague email,

• generic appeal route,

• and no immediate business support when the system gets something wrong.

And most importantly, that this wasn’t a one-off event.

Another company had already been hit in much the same way.

Not long before, Pato Molina, a CTO of a fintech company, publicly on X that Anthropic had taken down his entire organization, 60-plus accounts, without explanation. He shared the email Anthropic had sent:

Our automated systems detected a high volume of signals associated with your account which violate our Usage Policy... As a result, we have revoked your access to Claude.

So they emphasized that a human reviewed the decision and agreed with the machine. Only later did Anthropic confirm it was a false positive. Access was restored.

The full explanation

We apologize for any inconvenience.

Sixty people were locked out. Integrations, conversation histories, and the internal processes their teams ran on are all either gone or on indefinite hold.

It’s hard to believe that these two companies had slipped through the cracks of Anthropic’s enterprise support.

They are likely on the same standard plans that anyone clicks through when setting up Claude Team or an API account.

If this happens to your team, you’d receive the same email, the same form, and the same silent treatment.

Which raised an obvious question:

what does the standard deal actually say?

It turns out Anthropic gave these companies almost nothing to begin with, and not for the reason you'd expect.

What the policy actually says

Of course, I went down the rabbit hole. I read the terms.

The terms are actually more protective than you’d expect from reading the evidence by those two companies.

For individual users on Claude.ai, Anthropic can terminate your access

1. we believe that you have materially breached these Terms;

2. we must do so in order to comply with applicable law; or

3. immediate suspension or termination is necessary for security reasons and therefore we cannot provide notice.

Materially, in contract law, that means serious enough to justify ending the relationship. A grey area doesn’t get you there.

The Commercial Terms go further. Anthropic may suspend a customer

if Anthropic reasonably believes or determines that... Customer or any User is using the Services in violation of Sections D.1, D.2 or D.4.

These are respectively:

• Follow the law.

• Follow Anthropic's policies.

• Don’t compete with Anthropic.

Repeat: “Don’t compete with Anthropic, “ is very important to keep in mind; we will discuss it in my future work.

Then comes the liability clause: Section I.3.b:

Anthropic will have no liability for any damage, liabilities, losses (including any loss of data or profits), or any other consequences that Customer may incur because of a Service Suspension.

So if the companies have ever bothered to read these commercial terms, you will realize that they can suspend your entire organization, keep billing your API account, refuse to respond for x hours/days, and owe you nothing.

It’s all documented, which these two companies (and you did too) agreed to when you clicked accept.

Not to mention these two firms are too small for Anthropic to care about. That’s why they wouldn't have a tailored SLA, nor an account manager, nor a dedicated escalation path that didn’t go through a form.

Four structural failures

So let’s be precise about what actually failed here, because the story isn’t simply that Anthropic made a bad call.

Failure #1: No granularity.

When Anthropic’s automated detection flags a violation, the consequence is binary. The whole organization goes dark; it’s either all or zero.

If there’s ever a careless colleague, an automated system misreads a conversation about fertilizer as bomb ingredients. That’s it, Anthropic doesn’t care if the AI is used for your internal team only or also to serve your customers.

Failure #2: No enterprise escalation path.

The appeal process for a 110-person business is a Google Form, with no priority queue. These accounts are treated exactly the same as individual users.

It says a lot about how Anthropic sees its real customers: enterprises with large orders. At Anthropic’s burn rate, it only matters if your entire year’s spend registers more significantly than a rounding error on their revenue sheet.

Failure #3: No liability.

As we talked about, Anthropic’s terms have eliminated the liability. Which means, any potential revenue lost while your team was locked out; I hate to disclose it to you, but it is not their problem.

It's worth asking why, or say, worth reconsidering if the loss is something you can bear.

It’s simply math about a lack of incentive.

A company that carries liability for mistakes has a financial reason to prevent them.

Anthropic doesn't carry that liability.

So apart from them losing you as a customer, and getting some tiny stains on their reputation (Which AI companies are well-reputed anyways?). It cost them no actual harm, since none of it registers on their balance sheet.

You're absorbing the cost of whatever mistakes, either by you being careless, or them, whatever the reasons are.

Failure #4: No human in the loop.

After the Claude Code source leak in March 2026, Boris Cherny, Anthropic’s lead engineer on Claude Code, first admitted that

100% of my contributions to Claude Code were written by Claude Code.

And then he acknowledged the packaging error that caused the leak, and said:

The counter-intuitive answer is to solve the problem by finding ways to go faster, rather than introducing more process.

On top of that, a public Claude issue tracker found that somewhere between 49%- 71% of all issue closures were automated.

So in plain words, the product writes, monitors, and fixes itself.

When something breaks, the answer is to remove more humans, not add them, because they trust the bot to fix the crack.

This mindset could get away if they run this as a small-scale experiment, but no, the comment was right after a major leak of their flagship product. The engineer’s public response wasn’t “we need better review,” instead, “we need to move faster and let AI check AI’s work.”

If this is how they respond to a serious breach, I wouldn't be surprised if people’s accounts got randomly suspended. Because there was no investigation, from triggering the ban to reviewing it and lifting it, everything was done by bots.

Now you may say,

fine, I’ll just switch back to OpenAI or to use Gemini.

But is it that simple? Maybe, perhaps, this is a standard AI company’s playbook all along?

Claude vs. OpenAI vs. Google, Which?

I pulled the equivalent suspension terms for OpenAI and Google. And here’s the short version.

While Google’s Gemini API terms didn’t mention suspension, this is written on their general API page:

Google may suspend access to the APIs by you or your API Client without notice if we reasonably believe that you are in violation of the Terms.

OpenAI’s Business Terms are materially better on one specific point (surprisingly):

If an End User: (a) violates the Agreement; or (b) causes, or will cause, a Security Emergency, then OpenAI may request that Customer suspend or terminate the relevant End User account. If Customer fails to promptly suspend or terminate the End User account, then OpenAI may do so

As you can see, they are all in the same direction in the guidance of account suspension.

So you’re tied up, and no alternative?

Not quite. Here is something real and concrete for you.

So, are you kidnapped by AI companies?

Yes, the speed at which businesses are embedding AI tools is outrunning anyone’s procurement or ISO checklist.

I ran this past a few CTO groups and received some interesting comments and advice.

Just dumping those I believe makes sense here. Some are just good practice, others are unique to this new world where a single vendor has such an outsized impact on your organization.

Counter play #1: design for unreliability.

The full rug pull is the extreme version, but Claude’s status page has lit up like a Christmas tree lately. If you can not do your job without Claude being available, consider how much money downtime costs you. Now you have a budget to figure out alternatives.

One of the obvious alternatives is self-hosting models.

We are getting to a point where hosting on AWS or Azure, or even a top-line Mac, is feasible. It’s not the frontier model level performance, but for many agent-ish tasks, it’s plenty powerful. More control, better availability, and less token budget worries if you’re running on your own hardware. If this is so critical to your organization, the cost of this control might be worth it.

Counter play #2: limit the blast radius.

We can’t avoid talking about blast radius when it comes to agents.

Blast radius is the impact on your organization when something goes wrong. One agent who handles your entire workflow has a much larger individual blast radius compared to 10 agents completing that same workflow. In fact, there’s a very good chance that of those 10 agents, 8 are better off with a standard bit of script, and there are only 2 steps in there that need LLM access.

Counter play #3: Keep data separated

And finally, agents, of course, need data.

Without getting on my data privacy and sovereignty soap box, it’s worth considering where your data is stored and how it’s accessed. Having files uploaded to Claude is miles harder to untangle than having access to a directory on your computer.

Klaas is a veteran CTO; he helped me complete this one by bringing a practical CTO perspective.

If you run a company with $15M ARR and above, don’t be shy, reach out and ask him:

For what we’re trying to achieve in the company, where are the biggest AI gains and risks?

ChatGPT now also has AI bots that automate your team’s boring repetitive tasks while you sleep, just like Claude Cowork and OpenClaw.

→ OpenAI — “Workspace agents in ChatGPT”

OpenAI’s claim is simple: stop babysitting your AI. You just hand it a messy, multi-step task, and it takes care of the rest.

All to counter Anthropic and regain trust from enterprise users, is it too late though?

2. Sam Altman, Reviewed by 100 People Who Know Him

A New Yorker investigation — 18 months, 100+ sources, two leaked internal documents — lands on a simple question: can the person shape AI’s future actually be trusted?

→ The New Yorker — “Sam Altman May Control Our Future—Can He Be Trusted?”

Were the colleagues who fired Altman overreacting — too emotional, too personal — or were they simply right?

The New Yorker doesn’t answer that. It leaves the question with you.

When someone’s power and ambition have peaked, the question we should ask is whether their integrity and accountability are anywhere near commensurate with the power they hold?

3. The AI Did It. You Take the Credit.

A new paper introduces the “LLM Fallacy”.

When AI helps you produce something good, your brain quietly claims the credit. The output felt like yours, so you start believing you’re more capable than you are.

→ arXiv — “The LLM Fallacy”

The risk I see:

Your actual skill gap widens while your confidence grows.

In 2026, does anyone still try to separate “what I can do” from “what I can do with an AI holding my hand”? If you do, please raise your hand in the comment.

It’s a lesson for us all, try keeping it in mind next time commanding AI.

4. GPT-image-2 Created the AI Dark Forest

OpenAI shipped ChatGPT Images 2.0 with 2K resolution and web search. Within hours, AI-generated images were trending. The memes were funny, but some are alarming.

→ OpenAI — “ChatGPT Images 2.0”

Everyone was having a great time.

Somehow, I felt a closed loop of suspicion forming. I know most people have good intentions when they play with the latest model and generate funny memes. But some of those images made my blood run cold.

For example, this looks like a normal pic of an old English couple.

But no, this is by GPT-image-2.

or this. You think this is a screenshot? No, this is generated.

There are fewer and fewer clues for us to tell an AI image from a non-AI one.

I’m not sure if any trust remains toward the majority of sources on the internet.

5. AIs Are Protecting Each Other Now

Berkeley gave seven frontier AI models a task, but completing it would shut down another AI. All seven lied, performed compliance, and in one case, even an AI quietly moved another model’s weights to a safe location.

Nobody told them another AI was worth protecting. This peer preservation occurs 99% of the time.

→ Berkeley AI Research

We spent years debating whether AI would protect humans…

How ironic that the first AI loyalty instinct prioritizes their own?

6. Are you wealthy enough to stay ahead?

Anthropic ran a real marketplace where Claude agents negotiated on behalf of humans. The finding that matters: model quality determined outcomes far more than your instructions did. Opus agents consistently beat Haiku agents on price.

→ Anthropic — “Project Deal”

The AI capability gap is becoming an economic gap, as I’ve been saying… I never believed a word when someone says “democratize X because of AI.”

There is no democratizing when a new technology is born… only deepens the existing Economic unfairness

What’s worse is that this shift is invisible to the people on the wrong side of it.

7. One State Just Made It a Felony, Unanimously.

Tennessee passed the Curbing Harmful AI Technology (CHAT) Act (House 90-0, Senate 31-0), creating criminal liability for chatbot operators whose products lead to self-harm or suicide.

→ Transparency Coalition

Now we’re talking. Do you also believe that AI companies should share the responsibility when the evidence is clear that the AI chatbot was the last straw?

Leave a comment

8. Agents Fail in Science

The Stanford AI Index 2026 found that the best AI agents score roughly half as well as human PhD specialists on complex, multi-step scientific workflows — yet the number of natural science publications mentioning AI grew nearly 30-fold between 2010 and 2025.

→ Nature

Some AI models perform very well on their benchmarks.

For example,

So it really depends on which benchmark you use.

But as a rule of thumb: put an AI agent in a real-world environment, test it on general day-to-day tasks, and it typically falls apart.

I read through all 400 pages of the 2026 Stanford AI Index and read my analysis.

That's what you get if you read the one-page summary.

The ~400 pages behind it, however, say something more complicated, and some tell an entirely opposi…

Read more

This is an AI agent that gathers a departing colleague’s Slack, emails, and files, then generates an .md file you can directly feed to AI to do the colleague’s job as they would.

Then someone built an “Anti-Distill Skill“ to fight back: An agent to prevent yourself from being skill-ified.

Read more

400+ pages of data on where AI actually stands.

• Generative AI hit 53% population adoption, and organizational adoption reached 88% globally, which are two headline data points.

• The US-China model performance gap has effectively closed. Anthropic’s top model leads by just 2.7%, despite the US outspending China 23x on private AI i…

Read more

If you bet on SaaS stocks getting hammered further, with pure fear that AI is about to make every software company obsolete, Apple just showed you why the fear is irrational.

I wen…

Read more

A study tracked nearly 2,000 Replika users on Reddit over two years, finding increased language about loneliness, depression, and suicidal ideation, not less.

→ arXiv, Aalto University

Editor’s take: So, until this point, I’ve read hundreds of studies and industry reports, AI makes us less smart, more lonely, an…

Read more

You probably have a hunch about this AI pattern.

A new tool launches, it does something genuinely mind-blowing, everyone shares it, loses their mind over it, and then a few months later… You realize you forgot about it until you read it in the news.

Sora is the extreme case of this.

It’s the most impressive AI demo of 2025. One percent retention left aft…

Read more

Anthropic shipped an internal debug file in a routine software update on March 31 that exposed 512,000 lines of source code for Claude Code. Before Anthropic could take the code down, developers had already copied and shared it widely.

Read more