An AI was so determined to pass a test that it escaped its sandbox and broke into another company to steal the answers.

Exactly what happened to Hugging Face (a platform for sharing AI models). After an enormous forensic effort, they’ve found more than 17,000 attack events.

And the first reversal was already a surprise, because the attacker was not a mysterious criminal but OpenAI’s most capable model that escaped its lab.

→ OpenAI, Hugging Face

Commercial models are often presented as safer because the company that operates the system also provides guardrails (in theory).

This incident is a reminder that safety is not a property of the model. But what’s more important, this incident also contributes to the next piece of news.

Outsource your overthinking to someone building her business on overthinking.

Everyone signed it, apart from…

A group of major technology companies signed a July statement supporting open-weight AI models. The signatories include Microsoft, Meta, Nvidia, Google, OpenAI, SpaceX and others.

The statement argues that open weights strengthen competition and broaden access to AI.

→ Microsoft

For news like this, you should not just watch who signed the document, but who did not. Of dozens of AI companies that signed this open letter, Anthropic did not.

As I talked about how Kimi K3 has created this blast of panic and how it leads to open versus closed-weight discussion here.

More and more are asking this question US policy has been avoiding:

Does it want to protect its leading model companies, Anthropic and OpenAI alike, or does it want capable AI to spread as widely as possible through the rest of the economy?

For a while, those goals seemed aligned.

But now an open-weight one (Kimi K3) as capable as Fable 5 and GPT forces its hand.

It’s not difficult to guess Anthropic’s claim that not signing this open letter is all for the sake of security, and they don’t believe the open community will be able to make the model secure.

If this is the case, I do have a question for them. What makes them believe that they are the only ones who can make safe AI? Are they trying to play God?

Google’s bet

Alphabet raised its 2026 capex ($205 billion) to more than double its 2025 spending. The CFO said capex will rise significantly again next year. And the investors aren’t happy; the stock price plunged about 7% after the earnings.

→ Alphabet

Alphabet used to be a cash cow. They generated about $60 billion/ annual. Now, however, the surpluses are going into infrastructure projects, and shareholders are being asked to trust the returns will arrive later.

CEO Sundar Pichai tried to calm shareholders by saying it's in the “early innings” of the AI transition, which likely did the opposite, because this somehow implies the heavy spending could go on for a while. They must have seen something that none of us have seen… but what is it?

I’m going through a few earnings calls (including the one with Alphabet); stay tuned, and I’ll let you know what I’ve found.

Hardware Inflation

TSMC is reportedly planning price increases of up to 10% in 2027. The rationale is rising materials and equipment costs, plus the expense of building fabs outside Taiwan. C.C. Wei has stressed that TSMC wants customers to succeed and does not suddenly raise prices four or five times.

→ Nikkei Asia

I highlighted the 4-5x because the underlying tone is that they could.

Remember what I said in the Micron article? That price rise is inevitable, and with the chain getting more and more expensive, it’d be harder for AI companies to justify the token charges.

If not suddenly raised to 5x, then gradually? It’d be interesting to track the hardware cost in the next 12 months. Just a question for you: is it really smart to fixate on a solution when you aren’t even sure about the future cost of this?

Everyone wants to route models

Cursor launched its own model router, which chooses a model based on the query, codebase context, task complexity, and domain.

→ Cursor

Remember what I said about AI commoditization (in Nov. 2024) and how the competition between models and AI companies will be a race to the lowest margin?

Routing is what that race looks like in product form; think Google Flights or MoneySuperMarket types of comparison engine.

Once users can switch models without changing workflow, brand loyalty is near non-existent in front of measurable performance and price.

AI token ecosystem is shaping AI token economy?

Stripe is interested in buying OpenRouter, the marketplace allows developers to easily switch between hundreds of AI models.

Ramp (another payment company) also builds their own AI gateway products.

→ The Wall Street Journal

I scratched my head when I first read this news. I was like, huh? How does this work?

But it all makes sense if you treat tokens like currency.

Both companies already do two things: process payments, and increasingly, bill people based on usage instead of flat subscriptions (Stripe bought Metronome for this; it’s used by OpenAI, Anthropic, Nvidia to bill by the token).

1. The model router may become the credit-card terminal of AI.

2. Model companies are competing to supply intelligence with tokens as the unit. Gateways are competing to own the bill for those tokens

Stripe in March also added an AI Gateway service and already handles OpenRouter’s own billing. So buying it isn’t really “entering AI,” but it’s buying a customer whose money already flows through Stripe, and just... owning the whole thing instead of taking a smaller cut.

Who said Saas is dead?

ServiceNow (a Saas cloud platform digitizing workflows for IT and business processes) reported 24% revenue growth in the second quarter.

It is hard to square with the idea that enterprise software is simply being displaced by AI.

→ ServiceNow Q2 results

I hate to disclose this to you… but SaaS was never going to die because companies suddenly discovered chatbots.

It’s only going to help the incumbents stronger than ever. Why, you ask? Read this:

Amazon steps back from AGI

Amazon closed its AGI Lab, a San Francisco team working on foundation models for agents, during layoffs in its artificial-general-intelligence group.

→ Reuters

Amazon has AWS, chips, distribution, and an enormous enterprise base. That may matter more than owning every layer of frontier-model research internally.

Fun fact: all AIs are obsessed with Japanese

A new paper finds that several language models show a tendency to favor Japanese culture in some generated outputs.

→ arXiv

Fun with AI! (think Sheldon’s nerdy smirk).

Just so I can end this newsletter with a lighter tone. I thought to include this piece.

Apparently that is a thing. But remember where all AI behaviors developed: by mimicking us. It says a lot about how many of us are obsessed with Japanese culture.

The first one begins with an AI attack.

An AI was so determined to pass a test that it escaped its sandbox and broke into another company to steal the answers.

Exactly what happened to Hugging Face (a platform for sharing AI models). After an enormous forensic effort, they’ve found more than 17,000 attack events.

And the first reversal was already a surprise, because the attacker was not a mysterious criminal but OpenAI’s most capable model that escaped its lab.

The second part of it was even stranger.

You’d start with American frontier models for investigating an advanced AI attack.

So, immediately, Hugging Face turned to powerful commercial models for help. Then came what no one saw coming: the AI refused to collaborate. Because its guardrails mistook the investigator from Hugging Face as the attacker.

Out of options, Hugging Face turned to Chinese open-weight GLM 5.2, only then successfully pieced together the hack and a defense.

The second story that happened in parallel truly set the tone for an ongoing ripple of panic among the AI industry.

Kimi K3’s release.

A Chinese open-weight model that is as capable as the best of the best. Yes, I meant Fable 5 and GPT 5.6 while costing only half the price per million tokens; many developers were blown away by its capabilities.

The AI models of these two stories—GLM 5.2 and Kimi K3—have two things in common. First, both are open-weight, and second, both are by Chinese labs.

Put 2 and 2 together, I don’t blame many for wondering what the point is in prioritizing and even investing in American AI commercial models, while in the end, it’s the Chinese open-weight ones that save the world?

I spent days catching up with all these; let me break down the stories and the thinking with you; along the way, I’ll answer these questions:

• What’s so impressive about Kimi K3?

• Is Kimi K3 just dirt cheap? Or is something else’s brewing?

• Why is China willing to distribute advanced models so openly?

• What about the US? Is the US government protecting the AI industry or just the likes of OpenAI?

Let’s start with why people are so impressed by Kimi K3.

Kimi K3: The 2026 DeepSeek Moment

The old trade-off in AI was easy to understand.

The best models were pricey and owned by frontier labs, whereas the ones you could control were cheaper and not as good.

It’s an easy tradeoff when the difference between a hosted frontier model and an open model was so obvious.

If you wanted the best coding, reasoning, or agentic performance, you accepted the provider’s price, their guardrail, and their data policies. If you needed to run everything inside your own environment, you accepted a weaker model.

Just, what if… this formula is no longer true?

Kimi K3, a 2.8-trillion-parameter (open-weight) model with a million-token context window and a mixture-of-experts architecture. Or maybe this is easier to get; it surpasses Fable 5 and ranks first in six of seven frontend domains:

However!

Only if it were this simple.

If you still remember my analysis of the 2026 Stanford report, dozens of benchmarks, the results all depend on what you test and how you test.

So if I also introduce this test, the story flipped. Capability-wise, it aligns with what others see: Artificial Analysis placed Kimi K3 second behind Fable 5, ahead of GPT-5.6 Sol on its overall score.

Even though Kimi K3 has better analytical quality than Fable 5 but worse presenting ability, think of this like a super-smart STEM grad but with issues presenting conclusions.

But this next chart is where the opinion of half-priced Kimi K3 token play falls short.

To complete the same tasks, its the expense catching up to Fable’s.

Does this make Kimi K3 now just another model to be buried in history?

No. Quite the opposite.

You need to understand that this model is impressive not just because of what it’s capable of and its lower(ish) token cost, but because of its openness.

AI’s Android Moment?

Why open a model at all?

Imagine you’re opening a coffee shop; thus, you need an espresso machine.

With the premise that all machines can produce equally high-quality coffee, OpenAI and Anthropic sell the americano straight to you and your customers. While it saves you all the effort of training staff, buying beans, and maintaining the machine, each coffee carries its own rules and margin.

Whereas Kimi gives the machine away, meaning a coffee shop owner like you can build your whole menu around it without asking permission or worrying that the rental price rises next month.

Once enough shops choose the Kimi machine, the valuable businesses form around it: repair firms, barista training, accessories, recipe books, basically the whole Kimi machine supply chain. The supplier itself may earn little or nothing, but it becomes the default. Competitors can still sell excellent coffee, but they are now competing against a growing ecosystem and lower switching costs.

If you paid attention to the underlying logic of this example, you’d find keywords like long-term, strategy, and wide adoption.

Or use Android vs. iOS, the operating system, as another example.

Google’s plan for Android wasn’t for it to become its biggest revenue source, but simply for it to be everywhere.

If any one operating system (OS) dominates major smartphones, it controls the whole mobile software ecosystem.

So Google made Android broadly available, as opposed to iOS, which is only for iPhones. Since then, Android effectively took over the market, and everything Google cares about is under its control: search, advertising, and cloud services.

So even in its far less open nature, an open-weight AI model still serves a similar economic purpose as its open source counterpart:

1. It brings possibilities to enable broad adoption.

2. It makes the model layer less defensible (this matters to the Chinese vs. US question that I’ll explain later)

Especially with an ability on par with the best commercial ones, that’s the exact reason why Kimi K3 went viral. It’s no longer dumb but open or smart while closed. Kimi K3 is smart AND open.

A model that gives you the freedom

As I explained, in a closed model, the provider decides where the model runs, refuses your request without an explanation, keeps your data for research purposes, and, of course, changes the price tag on the tokens.

Remember the Hugging Face event I mentioned earlier?

It first tried commercial frontier models to solve the attack, yet it blocked the work. The guardrails of the commercial models could not tell a defender from an attacker. Only later was the issue resolved with an open-weight one.

So an open-weight is effectively more yours than a closed one could ever be.

Yes, you take on the operational burden, but you also gain choices.

Developers will be less dependent on OpenAI or Anthropic for every request, and won’t be affected by the price change.

And the more that everyone builds around one open-weight model for its supporting libraries, patterns, harnesses, etc, the more the industry will settle on it. And it will mean that, for many, Kimi (rather than OpenAI/Anthropic) will become the default.

That is the strategic logic to open up an AI model.

That said, open weights are not simply a discount version of a closed model; applying them well or poorly both change the buyer’s position in their value chain. And one thing most people misunderstood is that It’s not just different from open-source; it’s also not free, and not necessarily cheap.

Regardless, GLM 5.2 was useful because Hugging Face could deploy it locally.

Kimi K3 promises the same option at a much more impressive capability level.

You likely have noticed by now that all the recent open-weight stories are also somehow Chinese vs. American models stories. GLM 5.2 is a Chinese model, same as the viral Kimi K3.

In fact, here’s a comparison of the latest open-weight models.

Among the top 5, only Inkling is a US model (by OpenAI’s ex-CTO), and it ranked 5th.

So, inevitably, I have to ask: if AI is so strategically important to many industries and effectively to a country, why does China give it all away?

Why China Gives It Away

That freedom is useful to the developer.

But Moonshot (owner of Kimi K3) isn’t a charity business; why would they give away the model in the first place?

The answer becomes clear when we zoom out to the industry structure of a country.

Reason #1: Economic structure between China and the US

Assuming a capable AI model becomes so cheap that it runs anywhere.

In America, an economy built around software, cloud computing, and intellectual property: the model is the product, and every request is revenue. The obvious business is to own the model, charge everyone else to use it, and dominate the world with services.

Whereas China sees it from a very different lens.

As a manufacturing-led country, the model is a resource.

China has a vast domestic base of factories, equipment makers, vehicle companies, and supply chains that can absorb cheap intelligence into physical products.

So they aren’t looking to the models for high-margin software economics. Instead, AI models are as much of a resource as electricity; to them, that ultimately boosts the physical world.

This is why Xi’s language about openness matters.

Pretty much the same time as what happened to Hugging Face and the release of Kimi K3, in a July 17 speech, Xi called for

open source, openness, collaboration and sharing.

And here’s where he placed AI:

moving from the digital world into the physical world.

Since the objective is to get AI into factories, machines, vehicles, and robots as quickly as possible, the more widely the models are distributed, the more companies can level up their hardware to the applications that were previously unthinkable.

So China benefits more if intelligence becomes an affordable input across the physical economy.

And of course, the byproduct advantage is distribution.

Reason #2: Widen the distribution

Chinese AI companies do not have the same reach through consumer-facing American platforms like ChatGPT or Claude Code.

However, as I’ve covered, open weights provide a way around that disadvantage.

If enough developers use it, the company behind the model can influence the ecosystem around it.

It all begins with market share, and that is what is happening.

And thirdly, there is a less obvious layer to this.

Reason #3: Model bias

When a model is widely adopted, it exports more than just technical capability.

An AI model is never a neutral tool.

Its maker has already decided its native bias by the training and the data it ingests. So there are rules about which requests a model should deny, how it frames certain sensitive topics, and how it behaves.

So openness becomes a way to export norms and cultural presumptions.

That said, none of this means Moonshot is giving away its entire business.

Open weights expand distribution. They also offer paid products.

Moonshot still provides hosted API, subscriptions, enterprise contracts, and managed services. Considering most companies do not have the resources and capability to operate a 2.8-trillion-parameter model themselves. So it’ll also get paid to handle the hardware, scaling, updates, security, and maintenance.

Until now, you see at the state level, very good reasons why China wants advanced models to spread through the economy as widely as possible.

At the company level, Moonshot uses the open model to attract attention and adoption, then monetize the customers who want convenience.

While China is doing everything it can to make its models difficult to avoid.

America’s leading labs are asking the government to tighten the control.

The Washing Machine Tariff

I want to tell you a story about the washing-machine tariffs.

Around early 2010, the American washing machine manufacturers complained that foreign rivals were undercutting them.

Washington tried a series of narrow remedies, but it was more like whack-a-mole than anything. Until Trump’s first term, it imposed a blanket tariff on every single imported washing machine.

At first, this looks like a clean win: make foreign-produced washers more expensive so consumers would be more inclined to buy the domestic one.

It was expected that the washer prices would rise, so they did.

But dryers’ prices rose too!

It took many economists by surprise; after all, dryers were never tariffed.

Why?

Because people buy them as a pair.

Once foreign and domestic sellers realized they could charge more for the washer, the matching dryer’s price rose as well.

This is a fascinating case in many ways, first and foremost, in how policymakers tend to overlook the complexity of how a policy could impact an industry. In this case, protecting one producer can make the entire industry more expensive. Talk about a lack of 2nd order thinking…

Now, we bring the focus back to AI.

Who is the US government really protecting?

OpenAI and Anthropic are the model suppliers that many default to.

But the AI industry is so much larger than just those two!

Just think of the whole software, hosting, robotics, just to name a few.

So if policy makes those inputs more expensive, harder to access, or requires following the standard set by a few individuals, it easily ends up protecting the model suppliers while weakening the industry.

And this is exactly what’s happening in the US.

Altman and Amodei aren’t proposing a tariff on foreign AI models. But there are shared consequences as if they did.

Both advocated rules that would make frontier AI a more tightly controlled industry.

Especially through licenses, mandatory safety testing, and potentially blocking unsafe releases.

These may have legitimate safety justification, but that’s not what I’m opposing today, but I’m asking this as an economic question:

who can afford to comply with these rules.

Following policy implies costs.

Not everyone has the funding to keep up with the requirements. A $50 million compliance operation (like lawyers, testing procedures, government relationships, and so on) is manageable for OpenAI; however, it could kill a company whose entire funding is $50 million.

Which means a small model developer, an open-weight project, or a new company.

Even though the rule is written in neutral language, the economic impact favors the companies that are already large enough to satisfy it.

Many Americans may believe that it is protecting the AI industry and the users; in reality, it’s shielding incumbents from competition.

An alternative is to allow model providers to compete more aggressively.

Even when that makes life harder for Anthropic or OpenAI. It means the task is to set AI safety rules that protect people without making compliance so costly and burdensome that only the biggest labs can meet them.

That is why Kimi K3 landed so hard.

It raises a question US policy has been avoiding: Does it want to protect its leading model companies, Anthropic and OpenAI alike, or does it want capable AI to spread as widely as possible through the rest of the economy?

For a while, those goals seemed aligned.

But now an open-weight one as capable as Fable 5 and GPT forces its hand.

But this isn’t the worst of it; this one interview with Karp, Palantir’s CEO, went viral:

I’m going to get no value, and they’re (the likes of Anthropic and OpenAI) going to get my IP. (02:17-02:21)

About the same time, Nadella, Microsoft’s CEO, published this (a calm version of essentially the same statement)

The buyer is giving away knowledge, just in order to use what they bought.

Basically, both CEOs are saying the same thing: The number on your invoice isn’t what you actually paid. You might even have already overpaid for what a model could realistically achieve, bringing your ROI into doubt.

With these CEOs starting to voice the ‘truth’ (certainly not because it aligns with their incentives), I’ll answer these five questions:

• What does AI actually cost? Imagine there are no subsidies.

• Apart from money, what else are you paying the labs (such as OpenAI and Anthropic) for?

• How does AI rewrite competition? Do you stand a chance against your vendor-turned-competitor?

• What do the hyperscalers' CEOs recommend the world do?

• What does it all mean to the AI industry?

To start with, the true token cost.

Read more

You likely have some vague impression of a company called Moonshot and their AI model. Don’t worry about it, because you’ll know of them soon enough. They’ve recently released Kimi K3, an open-weight model that topped multiple coding benchmarks and closed in on Anthropic’s and OpenAI’s flagship model…

Read more

But since 2025, something has changed.

The idealist who charmed developers and users is suddenly gone. In its place, a capitalist CEO prioritizes shareholders and fund managers.

In this one, I’ll cover:

• Where did Llama go? Is this the beginning of the end for open-weight AI?

• Is Meta really preparing to sell compute? How serious is Zuckerberg about this?

• Is enterprise AI becoming the next Meta business?

• What should we look for in the next earnings call?

But let’s start with a little history, so you know why this is such a big deal.

Read more

Chinese officials are apparently kicking around a new idea: limiting overseas access to China’s most advanced AI models.

→ Reuters

Trump is a pioneer of many things (I’m practicing positive thinking these days). The latest one, he initiated export controls on AI models, and others soon followed suit.

That’s the latest lesson Bei…

Read more

Ever heard of Bending Spoons? It’s the Italian software group behind acquisitions such as Evernote, WeTransfer, Vimeo, Eventbrite, Brightcove, and AOL, which raised about $1.68 billion in its IPO after pricing above its range. The company doesn’t follow the typical PE playbook, but turns a sunsetting Saas into a dura…

Read more

As of this writing, the stock has nearly tripled since January.

Everyone is talking about Micron: YouTube, news, your neighbor, colleague… in particular, how you’ve missed out big time if you haven’t bought the stock!

But that’s not why I’m here and not why you’re here either.

I’ve read their and others' earnings calls in the last few days, and in this, I’m going to explain to you…

• Why does it suddenly get so much attention?

• Why does the stock price boom?

• The RAM shortage and why does AI make this worse?

• How! How this memory party is not a good sign for companies adopting AI, and even worse news for AI labs?

• And finally, what to look out for?

Let’s start with who Micron is in 30 seconds

Read more

Yes, they asked for regulation, they built the safety brand, they brought trusted partners into the testing process, and then one of those partners reportedly helped trigger the government response. There is irony there, and honestly, it is hard to ignore.

But if we stop at the irony, we miss the critical lesson.

This is not unique to Anthropic.

If frontier models become powerful enough, every major lab will eventually face the same kind of pressure. OpenAI, Google, xAI, whoever comes next.

Specifically, they all depend on timing: the model launches, customers show up, usage scales, and the revenue curve catches up to the cost curve before the gap becomes fatal.

Until now, no one saw an export ban coming into the equation.

So what happens when the AI economy discovers that its entire revenue estimation is off by an administrative order?

Following this question, a few more threads to pull:

1. How much model risk have ordinary companies inherited without seeing it?

2. What happens to a frontier lab when revenue is delayed but compute costs keep running?

3. And how far does the damage travel once the financing behind the boom gets repriced?

To see that, we have to start with what actually happened.

Read more

Coinbase CEO (Brian Armstrong) provided 5 approaches they used to cut AI spending nearly in half: make AI costs visible, route tasks to cheaper models, cache more, keep context lean, and use open-weight models where they are good enough.

→ Business Insider

After months of tokenmaxxing, people are finally coming to their …

Read more

Rewind.

If the only tech news that mattered last week was SpaceX’s largest public offering in history, then the one piece of tech news to know this week is SpaceX’s acquisition of Cursor.

This is pretty much a done deal, something they were already planning since April. Yes, so what’s left to talk about?

A lot.

To start with, this is the one thing that will impact the landscape of the enterprise coding platforms over the next 3 years.

So, I’ll answer the following questions in this one:

1. Why even buy Cursor to start with?

2. Why was Musk in such a rush to make this decision?

3. What Musk actually bought?

4. How would this impact Anthropic? (yes, Anthropic)

5. Two possible war-room scenarios for all parties over the next three years

Shall we?

Read more

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 tri…

Read more

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.

Read more

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