1. Was this an inadvertent ‘swatting’?

The reports point to Amazon being the source of the ‘tattling’ to the US Govt about potential malicious use of Fable 5. And then the US Govt asking for a specific unprecedented restriction on Mythos and Fable 5: “no access for any foreign national inside or outside the United States, including Anthropic’s own employees”. Swatting depends on a roided-out SWAT team to react peremptorily to a malicious phone call, sans careful investigation. And this is what it seems happened here - Amazon (Anthropic’s largest investor) sent a written report to the US govt, in which it posited dangerous use of Claude - something that needed careful evaluation on its merits.

If you’re experienced in technology and policy, you know the restriction requested from Anthropic can’t be administered at the level of national citizenship. So it’s all or nothing (Anthropic chose ‘all’, restricting it for everyone). We think it’s wrong to frame it as ‘Anthropic was recalcitrant’ as David Sacks did. Usually, the asker (the USG) should have the sophistication to understand the choices imposed by its request. No company can build the infrastructure to filter out non-US citizens’ access to globally deployed software platforms. Only the US government has done it at scale.

And then there is the matter of whether the ‘jailbreak’ was legitimate. Actual details are scant, even with Anthropic’s official statement and Axios’ reporting, and it comes down to trust: whether you trust the flawed but safety-minded AI company or the top tier of the Trump administration. The internet is unequivocal on that one.

2. Unmasking the real import of frontier models: AI is de-facto geo-political ordnance

This is an important moment in the history of the evolution of Artificial Intelligence. On the surface, LLMs have been driven by research labs for the last 5 years. Even though most of them are for-profit, it seems to have worked well; research papers flowed from Google, OpenAI, Anthropic, Meta, Microsoft, Qwen, DeepSeek, and more. But underneath what seems governed by commercial intent, nations have been jockeying and negotiating for what amounts to: ‘who gets to make the best AI and call the shots’. The US has advantages in energy density per capita, data-center build-out potential, capital flows, and intellectual property. It has explicitly blockaded the chip layer (temporary fix) but has not pulled the same for the model layer. Well, it did this week, in a shocking, ham-handed fashion. It cannot be overstated how important this moment is. The United States imperiously ordered that effectively EVERYONE in the world would not get a certain piece of software, which many depended on. No laws were passed, and very few deliberations were held. And it was publicly framed as adversarial to anyone not a United States citizen, even though the policy eventually caught up to citizens, too.

This reveals LLMs as geopolitical ordnance deployed by foreign powers in exclusionary ways against other nations and their citizens. I’m sure the Chinese are not surprised (they offer a different kind of AI diplomacy with their competent, famous, and cheap open-weight models). But the Europeans are likely very pained at this moment. Awkward.

We expect second and third-order consequences, including the following:

1. Public research on LLMs will continue to decline1.

2. Nations will hoard breakthroughs.

3. For-profit labs will yield more and more to the regulation of their nation-state cradles.

4. Some of the new ones will try to be stateless if they can manage it, and still secure venture or other funding.

It was imprudent for the US to tip its hand this early. Oh well. I think we will all regret it together.

3. The future will be both local AI and API-delivered AI. Not one or the other.

In response to this model yank, we see a lot of speculation about local AI being the future of AI proliferation. The general story goes something like this: Moore’s law will make CPU chips do native inference on phones, laptops, and desktops up to a level beyond Mythos. In 5 to 10 years, everyone will have models running locally and problem-solved.

What this ignores is relative advantage and competition between nations and companies. Moore’s Law is actually a relentless upgrade cycle law. When new chips emerge, the most competitive productivity-minded companies HAVE to upgrade to keep up.

Does anyone remember IT departments handing out the beefiest laptops to developers every 2 years? Does anyone remember the big upgrade cycle driven by the emergence of the Apple M1 and M2 chips?

The best models will always run in the data center first, because the best and highest-performing chips always appear there first. It’s a lucrative and easy supply chain target, while phones, laptops, and PCs are incredibly slow to field upgrade.

Frontier models whose capabilities will outstrip your capable PC models will always come to you via API first, in a world economy built around relative performance. And in a world where AI supremacy is a relative acceleration race, local AI will be a second-tier productivity cushion, not the ceiling of performance. And if you think otherwise, you have not contemplated why Azure and AWS exist in the first place.

The future of LLMs is both local and datacenter. It’s not one or the other; it’s complementary. In the very same way, it’s already complementary today. To wit, most businesses use AWS and still issue powerful PCs to their employees. What is yet to emerge is the ‘cooperation layer’ that connects the two more seamlessly.

4. The continued ineptitude of the USA regulatory institutions on AI

One more thing. This ham-handed episode is a prime example of an unforced error. The matter required more delicacy in handling than was exhibited by the Commerce Department and its secretary. This need not have become a domestic and international incident. It has all the hallmarks of an unsophisticated administrator trying to use insufficient evidence to boss a free-willed civilian business around. One that had tried its best to not only submit its IP to oversight but had put in guardrails to prevent circumvention2. It feels like a more consultative approach, looser deadlines, and more assistance would have finessed this last critical engagement. It’s hard not to chalk it up to the ineptness of the named players like Andy Jassy, Treasury Secretary Scott Bessent, and Commerce Secretary Howard Lutnick.

If it’s any comfort, you can be 100% sure Fable 5 will be back; after all, in a year, Mythos class models will be ordinary, even if neutered. But the world of AI will never be the same again.

We believe AI tools will transform knowledge work, and we’ve seen real productivity gains in specific contexts with dozens of teams we’ve worked with. Full stop.

But it’s important that Microsoft’s own research team just published a paper showing LLMs corrupt 25% of documents during delegated workflows

This is the same Microsoft that six months ago rolled out Claude Code to thousands of internal developers with great fanfare, only to quietly cancel those licenses in May 2026 for ONLY “fiscal year” reasons. Now their research division is warning that long-horizon AI delegation remains an open engineering problem to fix slow, insidious data corruption. The two things may not be related, but it does make you wonder.

Overall, we’re starting to see a healthier dose of realism in AI workforce adoption, as the edges of reliability of these new sets of tools and how it’s integrated into the workforce, are explored in more detail.

Let’s talk about the implications of Microsoft Research’s new findings.

Quality In, nice looking corruption out

The results in the paper Microsoft Research published are discouraging. They tested 19 LLMs across 52 professional domains (extremely broad, not just coding) and simulated long-running delegated workflows where you hand off document editing to AI. This BTW is exactly how everyone claims AI will replace most people - long-running iterative work.

They found that even frontier models (Gemini 3.1 Pro, Claude 4.6 Opus, GPT 5.4) corrupt an average of 25% of document content by the end of long workflows. Other models fail more severely. The errors are sparse but severe. They silently corrupt documents and this can compound over long interactions - exactly like humans playing telephone. Even worse, advanced models hallucinated and distorted rather than simply deleting things, making it more difficult for humans to determine that the content was actually wrong.

The degradation gets worse with document size, longer interactions, and the presence of distractor files. Surprisingly, agentic tool use didn’t improve performance i.e. memory and persistence did not help.

This is a “your deliverable is subtly wrong, and nobody will until it costs you money” problem, and until this is addressed, the most common workflow, document management/processing, will need significant human quality steps (Ugh!, even saying it makes us twinge). We’re looking at a seriously documented case of Quality in, Garbage out.

Microsoft’s Claude Code rollback

In December 2025, Microsoft gave thousands of employees across its Experiences and Devices division (Windows, Office, Teams, Surface) access to Claude Code. The goal was to let non-developers prototype with AI and let serious engineers compare Claude Code against GitHub Copilot head-to-head.

Claude Code became wildly popular. According to The Verge, it was “perhaps a little too popular.” Engineers preferred Anthropic’s tool over Microsoft’s own GitHub Copilot CLI. Internal adoption was strong. Usage was high.

On May 14, 2026, Microsoft began rolling back the experiment. By June 30 (end of fiscal year), most Claude Code licenses are being canceled and developers are being moved to GitHub Copilot CLI whether they like it or not.

The official reason is “toolchain unification.” The coterie of emerging reasons are more interesting.

• Token-based billing consumed the annual AI budget far too quickly. The moment Microsoft switched from flat-seat licenses to usage-based pricing, the actual cost became visible and unmanageable. Large-scale adopters like Uber and others are seeing this pattern.

• Also, Microsoft’s own tool was being ignored in favor of the competitor, which is a bit embarrassing when you’re Microsoft, and you own and sell GitHub Co-pilot

So which is it? Is AI so valuable that Microsoft rolled it out to thousands of internal developers, or so expensive and unreliable that they pulled it back six months later?

Real-world productivity research is mixed

The 10x developer productivity projection looks like it will take a lot more model and agentic innovation to materialize. The narrative is definitely cracking.

METR’s February 2026 data indicates a reversal of the initial 19% productivity slowdown for experienced developers, resulting in an 18% net speedup as automation intuition improves. A new cohort of developers also realized a 4% speedup. Overall developers were consistently over estimating their productivity gains but also adapting to the new tools quickly, especially parallel work.

Google’s enterprise study found 21% productivity gains, but with lower quality bars, integrated tools, and corporate tasks. Microsoft’s own 2025 study found that leaning too heavily on AI is associated with weaker critical thinking.

This is all before we talk about cognitive overload and brain fry, which limit humans’ ability to use agents.

Oh, and the Gen Z revolt is happening

Here’s what else is happening: A Gallup survey from April 2026 shows Gen Z anger toward AI jumped from 22% to 31% in one year. Excitement dropped 14 points. Hopefulness dropped 9 points. Weekly adoption growth has stalled at 51%, up only 4 points year over year.

The belief that AI helps them work faster has declined 10 points since 2025.

A Writer and Workplace Intelligence survey of 2,400 knowledge workers found 29% of employees admit to actively sabotaging their company’s AI strategy! Among Gen Z workers specifically, the sabotage rate jumps to 44%. They’re ignoring guidelines, refusing training, and deliberately skewing performance data.

This isn’t irrational technophobia. This is a verdict. Gen Z unemployment for recent college grads hit 5.7% in Q4 2025, above the national rate. Underemployment sits at 42.5%, the highest since 2020. Over half of college students say their school discourages or bans AI use, while employers demand AI literacy from day one. They’re being told AI will take their jobs while simultaneously being told they’re not prepared to use AI at work.

The resentment makes perfect sense, and so does the loud booing at multiple commencements.

The conclusion in the teal leaves

Let’s connect the dots. Microsoft rolled out Claude Code internally in December 2025 with enthusiasm. By May 2026, they’re canceling licenses due to budget overruns and low adoption of their own competing tool. And by the by, Microsoft Research publishes findings that AI delegation corrupts documents at scale. Meanwhile, worker sentiment is collapsing, productivity research is contradictory, and Gen Z is actively sabotaging AI rollouts.

Multiple signals are pointing at the same underlying reality for the current generation of models and agentic harnesses:

They work in narrow, well-defined contexts with high supervision and clear quality bars. They fail in complex, delegated, long-horizon workflows where errors compound silently and where companies don’t provide enough guidance on implementation. They’re expensive at scale when you switch from flat licenses to usage-based pricing. They create dependency faster than capability. And the people being told to use them are increasingly resentful because the promises don’t match the reality they’re experiencing.

Microsoft is the canary because they’re the biggest enterprise AI customer, the biggest AI infrastructure provider, and one of the leading AI research organizations. When they roll something out and pull it back six months later while publishing research showing fundamental reliability problems, that’s not a product decision. That’s a real signal.

What This Means Going Forward

Lots of things jump out at us that the signal implies:

• First, we need an honest accounting of the cost of adoption and weaving AI deep into the enterprise. Token-based billing creates budget exposure that most procurement teams can’t forecast or cap. If Microsoft, with near-unlimited resources, hits this wall, every enterprise will hit it. The pricing models need to change, or adoption will keep hitting fiscal-year boundaries and getting canceled.

• Second, we need to stop treating delegation as solved. The Microsoft Research paper is clear. Current LLMs are unreliable delegates. They introduce sparse but severe errors that silently corrupt work artifacts. You really can’t hand off complex document editing, where correctness matters, to AI and trust the output without human review.

• Third, we need to acknowledge the human capital problem. You can’t tell an entire generation their jobs are being automated while simultaneously failing to prepare them to use AI and expecting enthusiasm. The Gen Z revolt isn’t irrational. It’s a rational response to mixed signals, broken promises, and genuine career risk.

• Fourth, we need better integration and measurement. The gap between benchmarks and field studies suggests we’re measuring the wrong things. AI that works in controlled settings fails in production because production has implicit quality standards, context requirements, and compounding error risks that benchmarks don’t capture.

Microsoft isn’t abandoning AI. They’re adjusting to reality. Claude models are still available through Copilot CLI and Microsoft Foundry. The research continues. The partnership with Anthropic will expand (and maybe they will throw Microsoft a bone), but the easy narrative where you just roll out AI tools and productivity soars is becoming nuanced. Integration will take more work than some assumed.

We believe AI will transform knowledge work. We’ve seen it work in the right contexts with the right support. But transformation requires honest assessment of what works and what doesn’t, not just repeating benchmarks and ignoring field evidence.

🔍 Are you a builder or a technology leader?

Check out our books BUILDING ROCKETSHIPS 🚀 and the new Ageless Peak Performance. Continue this and other conversations in our 💬 ProductMind Slack community and our LinkedIn community.

Aside from the hype, ever wondered why there is so much money going into building data centers, setting an unprecedented pace of investment, and also sparking conflicts in local communities?

The dual silicon-electricity rail

Two rails carry the modern economy: electrons and silicon.

Electricity moves $3 trillion in annual retail sales, and about 95% of global GDP sits downstream of it1. Microprocessors are a $700 billion industry with a $5.5 trillion IT stack riding on top of it (Gartner, 2025), and roughly half of global GDP is materially dependent on the outcomes.

These are not emerging sectors. They are civilizational plumbing. What is new is that both rails are being physically concentrated and rerouted into exactly one kind of ‘building’. The 2025 numbers:

• Data centers consumed about half of all US electricity demand growth (IEA).

• Hyperscalers spent 45 to 57 cents of every revenue dollar on capex (J.P. Morgan; Dell’Oro).

• Total data center buildout will reach $4 to $7 trillion through 2030 (McKinsey), what Jensen Huang (CEO, NVDA) has called the largest infrastructure program in human history

• The rest of the economy grew at 3% (IMF); the rails feeding the data center grew 10 to 57% (WSTS, IEA, Dell’Oro)

That gap is not noise. It is an acceleration and a concentration.

The new factories of the LLM paradigm shift

The data center is becoming the factory of the 21st century, in the same structural role the Manchester mill played in the 19th century. The inputs are different (coal and cotton then, electrons and silicon now) but the pattern is identical: one kind of building, fed by the largest rails in the economy, producing a new intermediate good2.

That good is not compute. Compute was the old product. The new product is inference, delivered as agents, APIs, and embedded intelligence in every surface that used to be static. $7 trillion in capex is not insane once you understand you are not capitalizing more servers; you are capitalizing the birth of a new product category whose demand surface is every industry downstream of intelligence. From here, every large business sits in one of four positions:

1. Supplying silicon

2. Supplying electrons

3. Building the factory

4. Building on top of its output

Everyone else is a downstream customer.

Additional charts

🔍 Are you a builder or a technology leader?

Check out our book BUILDING ROCKETSHIPS 🚀 and continue this and other conversations in our 💬 ProductMind Slack community and our LinkedIn community.

Sources: IEA World Energy Investment 2025; WSTS / Semiconductor Industry Association; Gartner Worldwide IT Spending Forecast 2025; Mordor Intelligence Electricity Retail Market; McKinsey Global Institute; Dell’Oro Group; J.P. Morgan Asset Management; IMF World Economic Outlook; NVIDIA.

After being a leader at places like Bridgewater and Citadel, launching more than a dozen companies and non-profits, serving on boards, and raising a family with special needs children, I found my brain starting to slow down as I neared 50.

I became obsessed with one question:

Why do our best years have to be so short?

I discovered that generative AI provided an answer to this question and the result is Ageless Peak Performance: The Playbook for AI-Powered Excellence.

Your best years can last longer than you think.

For decades, we’ve accepted the traditional performance curve: build expertise in your twenties and thirties, hit your stride for 10 to 15 years, then slowly watch your edge erode as speed fades, cognitive load rises, and the world keeps moving faster.

Thanks to AI this doesn’t have to happen anymore.

Used well, AI can compensate for what tends to decline with age: processing speed, recall, and working memory, while amplifying what experience makes stronger: judgment, pattern recognition, strategic thinking, and perspective.

That is the idea behind Ageless Peak Performance.

This is not a book about using chatbots to write emails faster.

It is about redesigning how you think, learn, decide, and execute, alone and with your team, so AI becomes cognitive leverage rather than just another productivity tool.

Inside the book, I cover:

• How AI can extend peak performance from a short window into a decades-long plateau

• Why experienced professionals may be better positioned for the AI age than they realize

• How younger professionals can compress years of learning into months

• How to build your own AI team of advisors, critics, researchers, coaches, and operators

• How teams can move from individual peak to collective peak

• Why human judgment has to stay at the center

To celebrate the launch, the ebook is available for a limited-time launch price of $9.99.

I’m also sharing two free resources for readers of the ProductMind Substack:

1. The Devil’s Advocate AI Workflow

This is one of the simplest and most useful AI workflows from the book. It helps you use AI to challenge your assumptions, expose blind spots, and strengthen your decisions before the real world gets a vote. Get it here

2. A free book excerpt featuring the first Principle for using AI to perform at your peak. You can read an excerpt here

When you are ready for more Ageless Peak Performance, you can:

Buy the Hardcover Buy the eBook Buy on Amazon

Thank you to everyone who has helped shape this project: my ProductMind partners, readers, reviewers, friends, family - the many people who challenged me to make this book the best it can be.

The AI age does not have to make us less human. Used wisely, it can help us do our best work for longer.

Ted

Together, they talk through the rise of AI agents, why traditional tools like email and PowerPoint may matter less in the near future, and what it means when work shifts from “doing” to directing. The conversation also dives into the stranger side of AI right now: Mythos, facial authentication, government involvement, and the growing feeling that tech companies may already be operating a few steps ahead of the public.

And if you enjoy a little conspiracy theory energy mixed into your AI conversations, this one definitely goes there.

🎧 The End of Clicking

Listen on YouTube, Apple Podcasts and Spotify

👍🏿 Like, subscribe, and share for more conversations on AI, tech, and the future of work.

‘We’ decided the most automatable work in the history of capitalism was the most prestigious. Somehow, that was always going to end badly.

We built an entire class mythology on the wrong assumption: “thinking” was harder to replace than “doing”. It turns out thinking, especially when it happens on a computer in a structured environment, is exactly what machines are built for. ‘Doing’, when it happens on a roof in the rain or under a leaking crawlspace with a flashlight, is not.

The knowledge class is about to find out what the working class has always known: no credential protects you from a cheaper version of yourself.

We want to be precise about what this is and isn’t, because the story is more unsettling than “AI takes your job.” And we’ll tell you exactly what to do about it. But first, you need to understand how we got here.

The Prestige Trap

For a hundred years, white-collar workers lived inside The Prestige Trap, the self-reinforcing belief that cognitive work was categorically different from physical work, and therefore categorically safer. The trap had two jaws: a credential system that made the belief feel earned, and an economy that rewarded it long enough to seem permanent. Both jaws of the trap are now open.

Upton Sinclair, the man who coined the term “white collar” in 1919, saw it first. The white-collar worker, he wrote, was “often the worst exploited of proletarians,” performing middle-class identity on a working-class wage. In 1951, sociologist C. Wright Mills made it the subject of a whole book: a new class of workers with no independent power base (no property, no union, no political identity) whose entire sense of self-worth rested on a status they did not actually control. The credential was a costume. The security was borrowed. Both men saw it coming. Neither got the timing right.

Thanks to AI, the time is now. Just think of the fright the last couple of years has given even the most prestigious of workers - finance and tech workers - and how they have felt super exposed.

Outsourcing was the Proof of Concept

The Prestige Trap held as long as cognitive work was hard to move. The moment it became moveable, leaks appeared.

The rehearsal was outsourcing. Beginning in the 1990s, white-collar work began its first migration: call centers to Manila, software to Bangalore, legal research to Mumbai. What we failed to apprehend was that outsourcing wasn’t a crisis. It was a proof of concept.

The moment knowledge work became transmissible over a wire, it became subject to commodity logic: find the cheapest source, extract the most output, then repeat, repeat, repeat. The offshoring wave climbed the value chain steadily. Junior legal hours at U.S. law firms dropped 60% in the five years leading up to 2014. The jobs that were supposed to be safe because they required intelligence turned out to be unsafe for exactly that reason: intelligence, in structured environments, travels well and has no face.

AI is offshoring at the speed of light, with no visa required.

History Says More Work Will Emerge. But that Misses the Point.

Every major automation wave in history has ended with more work, not less. The spreadsheet didn’t eliminate accountants. It created millions of them. The internet didn’t kill publishing; instead, it detonated the number of independent publishers. Thirty million professional developers were a ceiling until AI coding tools arrived; now, 150 million people may write code who never called themselves developers. Vibe coding is not the death of engineering. It is the democratization of it.

AI will likely follow this pattern. History tells us work expands. People think it might be different this time. Time will tell.

What we do know is that The Great Inversion is not about the quantity of white-collar work. It is about its standing. Its wage premium. Its prestige. The compact that said: get the degree, do the cognitive labor, earn a certain kind of life.

That compact is what AI is breaking. Not the work itself, but instead the deal attached to it. The factory didn’t disappear in the 1970s, but the dignity attached to factory work did. That is the closer analogy, and the one that should make every knowledge worker a bit twitchy with dread.

Geoffrey Hinton won the Nobel Prize for inventing the technology that is about to remake white-collar work. In 2025, a reporter asked him which jobs were safe. He said: plumbers. Not AI researchers. Not lawyers. Not analysts. Plumbers.

The man most responsible for building the AI reshaping the knowledge economy, the man they call the Godfather of AI, looked at what he created and pointed to the guy with the wrench. That tells you everything about where this is going, and it points directly to why.

Blue-collar work has a bigger moat. For now.

Blue-collar work has always had an accidental advantage nobody bothered to pinpoint precisely. We call it The Embodiment Moat: the structural protection that physical, situated, unpredictable work enjoys precisely because it cannot be transmitted over a wire. You cannot email a crawlspace. You cannot Slack a roof in a rainstorm. The moat wasn’t built intentionally. But it turns out to be the only one that matters in the age of AI. In fact, due to a shortage of workers, many tradespeople are seeing annual salary increases of around 3.8%. Construction pay, in particular, has seen substantial gains, with average hourly earnings in the US reaching nearly $40 by August 2025. Specialized roles, such as electricians and HVAC technicians, are now frequently seeing six-figure salaries.

A large language model doesn’t need a body to be effective. It can do in seconds what a junior analyst does in a week: reading documents, synthesizing data, drafting arguments, and running scenarios. Goldman Sachs found that AI systems can now match or outperform up to 47% of industry professionals on economically valuable tasks. McKinsey estimates that today’s existing technology could automate 57% of current U.S. work hours, not in theory, not in ten years, but now. We generally take these self-interested statements with a grain of salt. However, entry-level hiring in AI-exposed jobs has already dropped 13%. Internship offers from Fortune 500 companies fell 22% between 2022 and 2024.

And in case you’re wondering, OpenAI is openly tracking how much of GDP-involved work can be done by the GPT series of models. Go check our GDPVal.

Meanwhile, 500,000 net new trade jobs need to be filled to meet U.S. power demand by 2030. Construction jobs tied to the data center build-out (the physical infrastructure of the AI economy) have increased by 216,000 since 2022. The workers building the systems that will displace knowledge workers are protected by those systems’ own limitations. The Embodiment Moat holds.

There’s a New ‘Collar’ class emerging

History’s optimism about work volume is probably right. But the new prestigious work won’t look like the old prestigious work.

A genuinely new category of worker is forming. We call them The Conductor Class: people whose primary skill is not doing cognitive work directly, but directing the artificial intelligences that do. Not building AI like the labs. Not operating it mechanically. Conducting it: shaping intent, setting constraints, evaluating output, pointing multiple AI systems toward outcomes that require human judgment to define and human accountability to own.

This may be the central professional role of the next thirty years. Every organization will need people who can tell AI what to do, evaluate whether it did it well, and fix the system when it didn’t. That requires domain expertise, taste, and judgment, skills that are hard to automate because they are, by definition, about directing the automator.

The Conductor Class is NOT the old white-collar class. Its power comes not from knowing a domain exhaustively, but from knowing it well enough to direct an intelligence that can know it exhaustively.

In fact, having specific subject-matter knowledge in a single domain makes you brittle and vulnerable. Unless you are literally creating meaningful knowledge on a global level, you will be protected only until the economics make your expertise worth automating. The bar shifts from depth of knowledge to breadth and quality of judgment. The judgment you developed by building specific domain knowledge is what’s important, not the knowledge or the use of the knowledge itself. The value shifts from execution to orchestration and from doing to conducting.

That is a real and valuable job. But it will not command the same premium by default, or simply because you have a degree and showed up. The premium will go to those who combine training with the development of genuine orchestration skills. The credential doesn’t transfer automatically.

What To Do About It

Here are four specific moves, not career advice per se, but decisions you can make in the next 100 days.

Run your exposure audit. Map your actual working week. How much of what you do is transmissible over a wire: document production, analysis, synthesis, research, drafting? That ratio is your AI exposure score. If it is above 60%, you have a structural problem worth solving now, not at your next annual review. Be honest. Most knowledge workers, when they map it clearly, are sitting at 70% or above.

Move one responsibility from execution to evaluation. The work that survives AI is judgment, not production. If you are still the person generating the first draft, running the first analysis, or building the first model, that is your exposure. Pick one of those and flip it: use an agent to generate, and make yourself the one who defines the brief, evaluates the output, and owns the decision. Do that once, deliberately, and learn what it actually feels like to conduct rather than perform. Then do it again.

Develop Conductor Class skills: not prompting, orchestration. Prompting is the floor. The ceiling is workflow design, output evaluation, and the domain expertise to know when AI is wrong in ways a non-expert wouldn’t catch. That last part is the moat. A lawyer who can evaluate AI legal research better than a non-lawyer is valuable. A lawyer who still manually does legal research is expensive. Find the equivalent in your domain and build toward it.

Identify your baton. Every Conductor has a baton: the specific domain expertise that makes their AI supervision worth paying for. What do you know that makes you the right person to direct an AI in your field? Name it explicitly. It is probably not your title or your degree. It is the hard-won judgment about what good looks like in your specific domain, the thing you built from years of doing the work that AI is now doing faster. That judgment is not obsolete. It is the most valuable thing you own right now. Sharpen it.

White-collar work is not disappearing. The mythology about it being better than blue-collar work is. The question is not whether you will have a job. It is whether you will conduct the intelligence or be replaced by the person who does.

Part II, for organizational leaders managing teams through this transition, will be published next week.

🔍 Want to dive deeper?

Check out our book BUILDING ROCKETSHIPS 🚀 and continue this and other conversations in our 💬 ProductMind Slack community and our LinkedIn community.

P.S. AI models just leaked.

We break down what it means and why it matters more than people think.

• 🎥 YouTube → Click Here

• 🎵 Spotify → Click Here

• 🎙️ Apple Podcasts →Click Here

📢 We’re also excited to share that Ted Yang has a new book coming out:

📕 Ageless Peak Performance: The Playbook for AI-Powered Excellence.

In it, Ted explores how AI can amplify human capability and help us reach new levels of performance.

Pre-order now HERE 🎉

We talk about leaks (and whether they’re actually accidental), the growing importance of model personality and loyalty, and how open source is starting to shift the balance of power across the industry.

The bigger takeaway: AI isn’t just improving, it’s becoming more competitive, more accessible, and a lot more unpredictable.

If you’re building, investing, or just trying to keep up, this is a conversation worth listening to.

🎧 So… AI Just Leaked. Inside Claude, Mythos & What’s Next

Listen on YouTube, Apple Podcasts and Spotify

👍🏿 Like, subscribe, and share for more conversations on AI, tech, and the future of work.

If you feel like you can’t keep up with AI news, you’re not alone. Every day brings another model release, another startup announcement, another breathless headline. The firehose is relentless.

Here’s the thing: you’re not supposed to track all of it. You’re supposed to understand what’s actually new.

Every major technology cycle has this phase where noise drowns out signal. We saw it with PCs in the 1980s, the internet in the 1990s, smartphones in the 2000s. Underneath the noise, something genuinely different was happening each time. Not faster or cheaper. Actually new.

That’s what’s happening with LLMs. The way to think clearly about it is simple: focus on the building blocks.

How to Know Something Is Actually New

Most “new” technologies are just improvements on what came before. Faster processors. Cheaper storage. More accessible cloud.

Real paradigm shifts give us capabilities that literally didn’t exist before. The IBM PC launched in August 1981 and moved computing from corporate mainframes to individual desks. New capability: personal access to computing. The web arrived in the 1990s and connected knowledge that had been isolated in libraries and corporate databases. New capability: universal access to information. The iPhone launched June 29, 2007, and computing left the desk. New capability: computing anywhere, anytime.

Each time, we got new Lego blocks. Fundamental capabilities that didn’t exist before.

Counter-example: Blockchain. For all the hype, it was essentially a database with quirks (immutability, non-repudiation). We’d had relational databases for 60 years. Most people don’t actually need those quirks. Your bank can reverse fraudulent transactions; Bitcoin cannot. It was a solution in search of a problem. No genuinely new building blocks emerged.

LLMs are different. Here’s why.

The Five New Capabilities

When ChatGPT launched on November 30, 2022, it reached 100 million users in two months. That’s faster than any consumer application in recorded history. TikTok took nine months to get there. Instagram took two and a half years. UBS analysts, after two decades tracking consumer internet, wrote that they couldn’t recall “a faster ramp in a consumer internet app.” But the adoption curve isn’t the real story. The real story is what LLMs can do that nothing before could.

1. Content Generation: Creating Things That Sound Human

LLMs generate text, code, images, and music that passes for human-created. Not just coherent. Compelling. We’ve casually stepped over the Turing barrier.

Before, content required human creativity, time, and expertise. Now it can be generated at scale, on demand, in any style. GitHub reports that 46% of code written by active developers is now AI-generated, up from 27% when Copilot launched in 2022. Developers using Copilot complete tasks 55% faster. That’s not automation of coding. That’s abundance of code itself.

When content generation becomes abundant, curation becomes the scarce skill. Knowing what content should exist. Having taste. Understanding context. The bottleneck shifts from creation to selection.

Example: A marketing team that once spent weeks on campaign copy can now generate hundreds of variants in minutes. The bottleneck shifts from creation to selection—choosing what resonates, what fits the brand, what works.

2. Natural Language Understanding: Computers Finally Speak Human

LLMs understand imperfect human input: stutters, fragments, ambiguity, context. They respond the way a thoughtful colleague would. The interface between humans and computers has effectively disappeared.

Before, humans learned computer languages. We mastered command lines, filled out rigid forms, clicked through predetermined paths. Now computers adapt to how humans naturally communicate. This is larger than most people realize. It’s why vibe-coding will expand the global developer count from roughly 30 million today to well over 100 million within a decade.

Instead of learning SQL to query a database, you ask: “Show me customers who spent more than $1,000 last quarter but haven’t purchased this quarter.” No syntax cramming required. Intent is enough.

When interfaces become conversational, software becomes accessible to everyone. The distinction between “technical” and “non-technical” workers starts to blur.

3. Data Synthesis: Understand Fuzzy Data, Make Unseen Connections

LLMs make sense of incomplete, unstructured data. They find patterns without pristine inputs. They can predict outcomes before we understand why the prediction works.

The traditional scientific method required theory, then experimentation, then validation. That process can take decades. Now we can feed in unpristine data, get results, and understand the theory later. Or never.

AlphaFold 2 demonstrated this at scale. In November 2020, it solved the 50-year protein folding problem at the CASP14 competition, predicting protein structures with a median accuracy score of 92.4 out of 100. It did so without fully explaining the underlying chemistry. We can now design proteins for drugs, materials, and enzymes without waiting for theory to catch up with practice. R&D cycles compress by orders of magnitude.

4. Personalization: Everyone Gets Their Own Version

Hyper-adaptation to individual preferences at massive scale. Not “recommended for you.” Genuinely customized experiences that previously required armies of humans or millions of lines of code.

Personalization was expensive. It required segmentation, manual configuration, human intervention. True 1:1 experiences were economically unviable for all but the largest companies. Now the cost approaches zero.

When personalization is free, one-size-fits-all becomes a competitive disadvantage. Every product can be individually tailored. The clearest example: educational content that adapts to your learning style in real time. Not “here’s the video for visual learners.” Content that notices you’re struggling with a concept and shifts its approach, pacing, and examples until it clicks.

5. Autonomy: Go Think, Decide, and Report Back

This is the one that changes the nature of work most directly.

Autonomous AI agents don’t just respond to prompts. They receive a goal, break it into steps, use tools, make decisions, and complete the work. A human sets the objective and reviews the output. Everything in between is the agent’s job.

Before, automation required explicit programming for every possible state. Rule-based systems broke the moment conditions changed. Agents reason about and around novel situations, adapt mid-task, and escalate only when genuinely stuck. This is new. Automation that handles ambiguity is categorically different from automation that handles only the expected.

The results are early but striking. Ramp, the corporate spend management company, deployed a finance agent in July 2025 that reads company policy documents, audits expenses, flags violations, and generates reimbursement approvals without manual review. OI Infusion Services deployed a healthcare agent for insurance prior authorization and cut approval times from 30 days to 3. That’s not efficiency improvement. That’s a different category of operation.

When agents handle execution, human judgment shifts to goal-setting and guardrails. What becomes scarce is knowing what to build, not building it.

Why This Is Infrastructure, Not Just Tools

Infrastructure has four characteristics: others build on it, it enables previously impossible capabilities, it becomes foundational to how things work, and it compounds in value as adoption grows.

Electricity wasn’t just better lighting. It let factories reorganize, cities function differently, entire industries emerge. The same pattern holds here. These five building blocks aren’t making existing work faster. They’re enabling new categories of products, services, and business models.

And they compound. Content generation plus personalization equals individualized education at scale. Data synthesis plus agents equals drug discovery that runs overnight. Natural language plus code generation equals anyone can build software. All five together create products that adapt, learn, and evolve.

How to Orient Yourself

Stop trying to keep up with absolutely everything. The news cycle is noise.

Focus on the five building blocks. When a new announcement lands, ask which capability it actually uses. Most “new” things are combinations of these five.

Understand what becomes valuable. When capabilities become abundant, different things become scarce: judgment over execution, taste over production, empathy, ex nihilo creativity, ethics, taste - human elements that AI cannot replicate.

Think in workflows. Don’t ask “Should we use AI?” Ask “Which of the five building blocks applies to which workflows?” The question isn’t whether to adopt. The question is where and how to deploy.

Experiment aggressively. There is a lot of learn. Generals have become Privates overnight. There is no gain in pride. Humility and curiosity are critical values right now.

Once you see the pattern, the noise becomes signal.

Which of these five capabilities could transform your work? Pick one. Experiment with it this week. Then tell us what you learn.

This article is part of ProductMind’s ongoing exploration of AI-era product development. Subscribe for frameworks that help make sense of the transformation.

*LLMs still benefit from nice, clean data. Everything does. But they do not need it. System prompts can take care of a lot of regex and static code.

🔍 Want to dive deeper?

Check out our book BUILDING ROCKETSHIPS 🚀 and continue this and other conversations in our 💬 ProductMind Slack community and our LinkedIn community.

🎧 Prefer to listen? Check our podcast below ↓

🎥 YouTube → Click Here 🎵 Spotify → Click Here 🎙️ Apple Podcasts →Click Here

Listen on YouTube, Apple Podcasts and Spotify

10 Things We Discuss

1. Why AI is becoming a geopolitical and military asset

2. The ethical tension behind Anthropic’s decisions

3. How OpenAI, Meta, and Microsoft are competing

4. The role of leadership (including Sam Altman) in shaping AI

5. What self-improving AI systems actually look like

6. How memory and persistence change AI capabilities

7. The rise of autonomous AI agents

8. How humans and AI will collaborate going forward

9. Why speed and decisive action are competitive advantages

10. What the next wave of AI development could look like

💬 Question for you:
Should AI companies work with the military, or draw a hard line?

👍🏿 Like, subscribe, and share for more conversations on AI, tech, and the future of work.

This one is a little technical. Inside baseball. Everyone can follow but it will mean more to practitioners :) 

Every Product Builder’s Nightmare

A senior PM posted a panicked message on Reddit: their CEO had bypassed the entire product team, coded a solution with the CRO over the weekend, built a landing page, and started pre-selling—all while the product team was “stuck in discovery.” It felt like a unique threat to their value within the company, especially if it worked (spoiler: it didn’t).

This isn’t about one rogue executive. This is the new reality when building software becomes as accessible as writing an email. We have entered the era of infinite code.

The Three-Speed Problem

AI is accelerating different parts of product development at dangerously different rates:

• Discovery/Listening: 2-3x faster (AI feedback analysis)

• Building: 10x faster (AI-powered development)

• Go-to-Market: 1.5x faster (content creation accelerates, attention remains scarce)

It’s like driving a car where the engine goes 200mph, the steering responds at 50mph, and the brakes work at 20mph.

Research validates this mismatch. GitHub studies show massive improvements in developer productivity with AI tools. We have been using Claude Code assiduously to better prepare our advisory and investment clients and founders for the near future, and the productivity gains for non-coders feel like +600%. We can build faster than we can validate. We can ship faster than we can educate markets.

As you can imagine, there are downsides

Low Feature Value and Feature Graveyards: Teams ship everything because they can. One team I know shipped 47 features in a quarter. Customers used three. Not everything that can be built is valuable.

More Organizational Chaos: If everyone can code, who does what? Traditional handoffs—PMs write specs, engineers implement—are dead. Most teams haven’t figured out what replaces them. We’ve long asked product teams to work on the invisible stuff - clear vision/mission, actionable northstar and sub goals, key strategies, and guardrails, clear objectives. Decision rights and decision logs. Rituals that reinforce these things while being loose enough to evolve. if people are involved, some lightweight structure is needed. Infinite code will fracture undisciplined creative organizations.

Frustrated Leadership: CEOs promised 10x gains, but see marginal business outcomes, customer adoption. There is still a competitive market out there, and not everything catches fire. So they take matters into their own hands…if they are that kind.

The Solution is Simple - Accelerate Everything

The instinct could be to slow down building to match other speeds. This is backward thinking that will get you steamrolled.

Instead, accelerate discovery and GTM to match building speed, while completely reimagining how you use that velocity.

Supercharge Customer Listening

The old way: Three months analyzing feedback manually.

The new way: AI tools like Thematic automatically identify themes across massive text volumes without manual tagging, reducing analysis time by 98%.

The winning workflow:

1. Centralize all feedback streams into one AI-powered system

2. Query your data conversationally: “What makes customers cancel?” “Which features drive retention?”

3. Connect insights to action: integrate with development tools so the system can query aggregated feedback, isolate top issues by itself, and write PRDs and first-draft code

As one product manager notes: “We built our internal AI analysis tool, which allows us to automatically collect, analyze, and address all concerns... we can close the feedback loop efficiently, almost in real-time.”

Revolutionize Discovery

Build a multi-channel discovery engine running 24/7:

Community-Powered Discovery: One Discord with 1,000 engaged users validates ideas in hours, not weeks.

AI-Enhanced Research: Use AI research intelligence tools to synthesize social mentions, competitive intelligence, and historical patterns at scale.

Continuous Feedback Farming: Embed discovery everywhere—in-app surveys, AI session analysis, real-time experiments.

The goal: build and test 100 ideas, where you used to test 10.

Master the New Building Paradigm

The reality is: If you can’t build a basic prototype in code, you can’t product manage in 2025.

The era of 20-page PRDs is over. Your new PRD is a lightly annotated (why it matters) working prototype that engineers can run. Use Model Context Protocol (MCP) to formalize the design language and system, even the platform code conventions so rewrites are minimal.

Then use 10x building speed from developers strategically:

1. Polish: Create delightful experiences, not just functional ones

2. Ambition: Solve entire customer workflows, not just slices

3. Tech Debt: Rebuild old non-agentically designed systems with AI at the core

Embed in Go-to-Market

Traditional product-marketing handoffs are too slow. Embed yourself directly in revenue operations:

• Join weekly sales calls for real-time customer feedback

• Define “market-ready” (not just “feature-complete”) criteria

• Use AI to generate sales content and launch campaigns

The New Team Structure: Hyper Creator Shipyards

Traditional: PM owns discovery → Designer creates mockups → Engineers build → Marketing launches. This won’t cut it anymore… soon.

New Shipyard Model:

• Core Hyper Creator Team: 4 product engineers (former PMs and Developers) working with specialized humans and agents. We become highly experienced human + agent conductors. The prior experience is key - AI is not perfect yet.1

• Embedded GTM: Sales/Marketing sits with the team from day one. Things move too fast not to.

• Discovery Ops: New roles managing continuous insight infrastructure.

Success requires: clear goals + autonomy + complete skill sets + customer-facing teams as core partners.

Reclaim Product Leadership

The CEO codes because they see stagnation amid abundance. Your job isn’t to gatekeep—it’s to orchestrate all three speeds into continuous value delivery.

Companies that win won’t just adopt AI tools—they’ll fundamentally restructure how product, engineering, and GTM work together. They’ll build shipyard teams that discover, build, and ship at the new speed standard.

The tools exist. The methodologies work. The question is: will you use them to reclaim product leadership, or watch your CEO ship features while your team debates PRD formatting?

In the three-speed world, surviving product leaders won’t slow down AI—they’ll accelerate everything else to match.

🔍 Want to dive deeper?

Check out our book BUILDING ROCKETSHIPS 🚀 and continue this and other conversations in our 💬 ProductMind Slack community and our LinkedIn community.

🎧 Prefer to listen? Check our podcast below ↓

🎥 YouTube → Click Here

🎵 Spotify → Click Here

🎙️ Apple Podcasts →Click Here

In this chop session, the team explores how AI is democratizing access to knowledge, accelerating skill development, and redefining peak performance in the modern workplace. Ted Yang, a partner in the conversation, shares frameworks from his work on AI-driven productivity and human potential.

Rather than treating AI as an all-knowing oracle, Ted encourages a different approach: use AI as a collaborator, guiding it with judgment, clear goals, and strategic direction.

His principle is simple but powerful: aim first, then accelerate with AI.

When humans provide direction, context, and decision-making, AI becomes a powerful amplifier for creativity, productivity, and learning. Ted also introduces what he calls the “conductor model,” where humans orchestrate multiple AI tools and workflows to unlock new levels of insight and output.

Drawing on ideas from his upcoming book on AI-driven peak performance, Ted explains how individuals and teams can move beyond simply using AI tools to designing systems that help them think better, learn faster, and operate at a higher level.

The conversation also explores how AI is transforming education, enabling faster learning cycles, and expanding access to expertise that once required years to develop.

Join Ezinne, Oji and Ted as they discuss the human potential in the age of AI, how to stay intentional, strategic, and creative as technology accelerates.

Listen on YouTube, Apple Podcasts and Spotify

10 Things We Discuss

1. How AI is democratizing knowledge and expertise

2. The difference between productivity and peak performance

3. Why judgment matters more than ever in the age of AI

4. The principle: “Aim first, then accelerate.”

5. Treating AI like an intern, not an oracle

6. Using AI as a sparring partner for ideas

7. The “conductor model” for working with AI

8. How AI can help create flow states in work

9. What AI could mean for the future of education and learning

10. Frameworks for navigating an AI-powered future

Resources Mentioned in This Episode

Ted references several articles and ideas related to AI, learning, and peak performance. You can explore them here:

Human peaks will be prolonged (pt 1)

Ageless Peak Performance

Many of the ideas Ted shares in this conversation are part of a larger framework he’s been developing on AI and human potential.

About Ted Yang

Right now, Ted Yang is focused on exploring the intersection of AI, peak performance, and the future of work, and how people can use AI not just to automate tasks, but to think better, learn faster, and perform at a higher level.

His upcoming book, Ageless Peak Performance: The Playbook for AI-Powered Excellence, distills many of the ideas discussed in this episode into a practical framework for working with AI to amplify human capability, creativity, and decision-making.

📕 Preorder your copy: HERE

We’ll continue exploring how AI is reshaping how we work, learn, and perform.
Subscribe to follow the conversation and be the first to hear when Ted’s book is released. 📕

We’ve been watching the numbers coming out of GitHub, and something unprecedented is happening.

AI now writes 46% of all code that top developers commit. At some companies, it’s hit 61%. Copilot users complete coding tasks 55% faster—we’re not talking about a marginal improvement here. We’re talking about developer pull request times dropping from 9.6 days to 2.4 days. That’s a 75% reduction in development cycle time.

Over 20 million developers are using AI coding assistants today—a 400% increase in just one year (and a lot of money for Anthropic, the frontrunner with Claude Code). And 90% of Fortune 100 companies have adopted these tools as standard infrastructure. This isn’t experimental anymore. This is the new normal.

And while we’re all celebrating these productivity gains, we’re missing the bigger story. The real story isn’t about code getting faster. It’s about what happens when one part of product development accelerates dramatically while everything else stays roughly the same.

We’re entering the era of infinite code. And most product organizations are completely unprepared for what that means.

The Unprecedented Acceleration

Let’s be clear about how unprecedented this is. The software development market is growing from $532 billion in 2024 to a projected $1.45 trillion by 2033. But that growth is being driven by a fundamental shift in how we produce software.

Consider what’s happening:

• 84% of developers now use AI tools regularly

• Development teams report 10-30% productivity improvements on average, with active Copilot users seeing up to 81% gains

There’s some critical insight buried in these numbers: this acceleration isn’t uniform across the product development lifecycle. We’re seeing a 10x acceleration in building, but only 2-3x in discovery and maybe 1.5-2x in go-to-market.

This creates an equation that doesn’t balance: A three-speed problem.

What “Infinite” Actually Means

When we say “infinite code,” we don’t mean literally infinite, obviously. What we mean is that code capacity has become effectively unlimited relative to historical constraints.

Think about it: for the last 40 years, engineering capacity was the primary constraint on what we could ship. Product teams spent months prioritizing features because we couldn’t build everything. We created elaborate frameworks—RICE scoring, value vs. effort matrices, roadmap planning rituals—all designed to manage scarcity.

That scarcity is disappearing. A CEO can now sketch an idea in the morning and have a junior developer crank out a working feature that is consistent with product design language by lunch. What used to take a sprint team weeks can now be prototyped in hours.

But—and this is the critical point—the rest of the system didn’t speed up proportionally. You still need to:

• Understand if this solves a real customer problem

• Figure out how it fits with your strategy

• Test whether it actually works for users

• Train your sales team on it

• Create documentation

• Ensure it doesn’t confuse or overwhelm customers

These constraints haven’t disappeared. They’ve just become more visible.

WHAT BECOMES ABUNDANT VS. RARE

The New Bottleneck Isn’t Technical, it’s Judgement

Here’s what we’re seeing happen in real organizations:

Scenario 1: The Overnight Feature A CEO “vibe codes” a feature overnight. They show up at the product meeting with a working prototype. The CRO’s team sees it and puts up a landing page because they’re incentivized on the pipeline. The feature gets announced before anyone checks:

• Does this align with our strategy?

• Have we tested this with actual users?

• Will this confuse our current customers?

• Does this create technical debt?

• Can our support team handle this?

The answer is often “no” across the board. But the feature is already live. Oops.

Scenario 2: The Junior Developer A junior developer learns how to use Cursor or GitHub Copilot effectively. They get excited and build three new features over the weekend. Monday morning, they show up asking, “Can we ship these?”

The answer isn’t just “great effort!” The answer is a series of questions:

• Did you think through the edge cases?

• How does this fit the mental model we’re building?

• What happens to users who don’t need this?

• Have you considered the support burden?

• Is this creating parallel paths in the UI?

The pattern is clear: infinite code without infinite judgment creates chaos.

What Product Sense Really Means Now

We don’t particularly like the term “product sense”—it’s often used as a vague code word for something we can’t quite articulate about the primacy of product managers. But in the era of infinite code, it becomes critically important to name what we mean.

Product sense is:

• The ability to say “no” to features that are technically impressive but strategically wrong.

• The judgment to know when “more” is actually “worse”.

• The discipline to maintain simplicity when complexity is free.

• The empathy to understand that customers need time to absorb changes.

• The courage to delete features even after they’re built.

When anyone can code anything, someone needs to decide what should exist.

This isn’t gatekeeping for its own sake. It’s recognizing that customer confusion, support burden, and product complexity are real costs—costs that don’t show up in pull request metrics or code generation statistics.

Think about fashion. When production became cheap (fast fashion), the industry didn’t become less hierarchical—it became more hierarchical. We needed Anna Wintour more, not less. Curation became the entire value proposition.

The same pattern is emerging in product development.

There’s a Larger Pattern

This isn’t just about product development. It’s about what happens when AI makes certain capabilities abundant while leaving others as scarce as they are today.

Everything tied to the five AI building blocks (content generation, natural language interaction, data synthesis, personalization and autonomy) is becoming abundant. These capabilities run on CPUs that never tire, never complain, never need vacation.

But everything that is not easily replaced by these capabilities—politics, chemistry, biology, theater, human psychology, taste-making, judgment—becomes more valuable. The human arts become rarer and more important.

This is one of our bets for the next decade.

We’re not heading toward a world where AI does everything. We’re heading toward a world where AI does a specific set of things extremely well, making human judgment in other domains exponentially more valuable.

The Bottom Line

Here’s what we keep coming back to: In an era where anyone can code anything, the ability to say “no” becomes the scarcest and most valuable skill.

This isn’t the future. This is happening right now. The numbers are clear. The patterns are emerging. The imbalance is real.

The question is: are you preparing for it?

What This Means for You

If you’re a product leader with an AI-tooled eng team:

• Stop optimizing for code velocity. It’s no longer the constraint. Optimize for customer insight, strategic clarity, and judgment quality.

• Build gatekeeping mechanisms now. Before the CEO walks in with an overnight feature, establish clear decision criteria and review processes.

• Invest in product sense. It’s no longer a nice-to-have soft skill. It’s becoming the primary value of product management.

• Think about the absorption rate. Your customers can’t absorb changes as fast as you can ship them. Build that into your planning.

If you’re a founder:

• Don’t confuse building speed with market readiness. The hard part isn’t code anymore—it’s figuring out what people actually want and will pay for.

• Use your infinite code capacity for quality, not quantity. Polish matters more when everyone can ship quickly.

• Hire for judgment, not just execution. The person who knows what not to build is worth more than the person who can build anything.

If you’re an investor:

• Look for teams with strong product judgment, not just engineering velocity. Speed to ship is becoming table stakes. Knowing what to ship is the new edge.

Watch for organizations getting overwhelmed by their own building capacity. It’s a leading indicator of future problems.

🔍 Want to dive deeper?

Check out our book BUILDING ROCKETSHIPS 🚀 and continue this and other conversations in our 💬 ProductMind Slack community and our LinkedIn community.

🎧 Prefer to listen? Check our podcast below ↓

🎥 YouTube → Click Here

🎵 Spotify → Click Here

🎙️ Apple Podcasts →Click Here

Listen to a summary of the article here:

Cogito Ergo BS

Raise your hand if you think human intelligence is something special, something uniquely creative that large language models can never duplicate? I see that most of you did. Cool, cool. Now ask yourself, how do you know that for sure? Did you ever take a course on human intelligence? Or do you just feel the magnificence of the human mind in your bones?

Human intelligence is much closer to GPT 5.2 (or other LLMs) than you think. And this realization has implications.

In 2004, Jeff Hawkins (the founder of the Palm Pilot and neuro thinker and researcher) made a bold prediction in his book, On Intelligence:

Intelligent machines would emerge not from programmed rules, but from systems optimized for prediction. 

Twenty years later, large language models have vindicated half of his thesis while exposing awkward truths about both artificial and human intelligence.

The convergence is substantial but incomplete, and understanding the difference matters even more than the similarity.

The Algorithm We Share with AI

The human neocortex is fundamentally a prediction machine, not a cogitation engine, as we have been told all of our lives. This is measurable neuroscience. A 2023 study in the Journal of Neuroscience discovered dedicated prediction-error neurons1. These are neurons that fire when we notice differences in what our brains predicted vs. what our senses absorb in the auditory cortex. They remain completely silent until expectations are violated.

Yes, your brain and its subsystems have a ‘prediction difference’ engine - we have specialized hardware for prediction. Different neuronal populations fire for different error types: timing violations, volume mismatches, and pitch deviations. The brain constantly generates predictions about what should happen next, compares them to reality, computes the error, and updates its memory and model. This cycle runs at roughly 10Hz across sensory systems.

Large language models do something remarkably similar. During training, GPT-style transformers see a sequence of tokens, predict the next one, compute the prediction error (cross-entropy loss), and update their weights. Repeat this 10^12 times across trillions of tokens, and sophisticated behavior emerges, not from explicit programming of grammar rules or reasoning patterns, but purely from optimizing prediction accuracy.

This empirical alignment between our biology and LLMs is striking.

MIT research in 2024 showed that GPT models trained on as little as 100 million words (roughly what a 10-year-old child has heard) predict human fMRI responses in the language-selective cortex with r=0.65 correlation2. Google’s comparative study found that Whisper’s speech encoder activates superior temporal gyrus patterns, while its language decoder mirrors Broca’s area activity 200-500ms later, matching the brain’s sensory-to-semantic processing sequence. Multiple studies from 2021-2024 consistently show that better next-word prediction correlates with better brain alignment.

This convergence suggests prediction-based learning isn’t just one way to build intelligence. Instead, it may be the fundamental algorithm. Evolution and engineering, under entirely different optimization pressures, independently arrived at hierarchical systems that learn by iteratively predicting and correcting.

De-Mythologizing Human Cognition

If LLMs produce intelligent-looking behavior purely through prediction optimization, what does that say about human creativity and reasoning?

We mythologize human thinking as something fundamentally different from pattern matching. We invoke inspiration, intuition, understanding—as if these transcend statistical learning. But contemporary neuroscience suggests our “creative insights” may be sophisticated prediction errors. When you have an “aha moment,” prediction-error neurons are firing intensely. When you struggle to recall a memory, you’re not accessing a file, you’re reconstructing a prediction based on fragments. Even rote memorization is prediction-based: learning that Paris is the capital of France means learning to predict “Paris” when prompted, with minimal error.

This doesn’t diminish human intelligence. It clarifies it. The brain’s 16 billion neurons and 100 trillion synapses running on 20 watts are doing something computationally extraordinary with prediction as the core algorithm. The question isn’t whether we’re “just” prediction machines, it’s what prediction-based learning can achieve given sufficient architecture and data.

LLMs help us see this clearly, precisely because they lack so much. They have no body, no continuous learning, no lived experience, no self-narrative spanning decades. Strip all that away, and what remains? A prediction engine that can still reason, generalize, and surprise us. This suggests the gap between human and artificial intelligence may be smaller in some dimensions and larger in others than we assumed.

The Gaps That Matter

This is not to say that we are exactly like LLMs. There are real and consequential differences. The human brain operates continuously, learning from single examples through immediate synaptic plasticity. LLMs train in massive batches and then more or less freeze. The brain tests its predictions through embodied action: reach for a cup, feel the weight, update motor models. LLMs never close the sensorimotor loop. At least not yet, since they are disembodied.

Most critically, human cognition includes layers absent from current LLMs: phenomenal consciousness, persistent self-narrative, emotional valence that weights memory formation, learning across vastly different timescales simultaneously. A traumatic event changes you after one exposure. A skill develops over years. LLMs have fixed context windows and no mechanism for the kind of long-term memory consolidation that builds a coherent self over decades.

The brain also runs six orders of magnitude more efficiently: roughly 10^15 operations per watt versus 10^9 for LLM inference. This isn’t just an engineering detail. It suggests that neurons are using computational tricks we haven’t yet discovered.

Prediction Creates Emergent Behavior

This prediction-based convergence carries a serious implication: we may not be able to predict what capabilities emerge when scaled.

If “understanding” and “reasoning” emerge naturally from hierarchical prediction rather than being explicitly programmed, what other capabilities might emerge unexpectedly? The brain’s prediction architecture yielded consciousness, theory of mind, and self-awareness. And these are features nobody would predict from studying individual neurons. We don’t know what the LLM equivalent might be.

This demands both limiters and enhancers. The brain has inhibitory neurons, neuromodulators, and sleep cycles that consolidate some memories and discard others. LLMs currently lack analogous mechanisms for selective forgetting, attention filtering, or dynamic capability modulation.

If we’re going to push prediction-based systems toward greater capability, we should understand which architectural features enable beneficial emergence versus runaway behavior. The brain’s hierarchy isn’t just deeper—it includes feedback loops, lateral inhibition, specialized prediction-error pathways, and mechanisms for detecting uncertainty. These aren’t decorative features. They’re control structures.

Humility and Opportunity

The brain-LLM parallel should inspire humility in both directions. For AI research: we’re not close to replicating human intelligence, and we may not even be asking the right questions yet. For human cognition: our creativity and reasoning, while remarkable, may be less mystical and more algorithmic than we’d like to believe.

The risk is premature certainty in either direction. Claiming LLMs work “just like brains” misses crucial differences in implementation and capability. Dismissing the parallel as superficial ignores that prediction-based learning appears to be a universal algorithm for intelligence, regardless of substrate.

Hawkins was right that prediction-based systems would enable intelligent machines. What he couldn’t foresee was how much those machines would teach us about the prediction engine between our ears and how much both still have to learn from each other.

Understanding the limits of that principle matters as much as celebrating its power.

Bonus 1: Brain techniques that haven’t been tried in LLMs yet

Current LLM research focuses heavily on attention mechanisms and reinforcement learning. But neuroscience offers a much richer toolkit that we’ve barely begun to explore:

Memory attenuation: The brain doesn’t store everything. It strengthens memories through sleep consolidation, weakens irrelevant details, and actively suppresses competing associations. LLMs treat all training data almost equally even with reinforcement learning. What if they had sleep cycles? What if they could forget?

Predictive filtering: The brain filters sensory input before prediction—you don’t consciously process every photon hitting your retina. Visual attention pre-selects what gets predicted in detail. LLMs currently attend to everything in their context window with equal potential weight.

Error-type specificity: The brain has different neurons for different prediction errors—timing versus content, expected versus unexpected novelty, reward versus punishment. LLMs have one monolithic next-token objective. What would specialized error channels enable?

Hierarchical timescales: V1 predicts edges at 10-100ms. Prefrontal cortex predicts events at seconds-to-minutes. The brain natively operates across timescales. LLMs have fixed context windows. Multi-timescale prediction might unlock longer-horizon reasoning.

Active & interventional inference: The brain doesn’t just predict—it acts to make predictions come true or gather information to reduce uncertainty. LLMs are passive predictors although tool calling simulates this. Embodied AI research is beginning to explore this, but we’re still early.

We’re perhaps 5% into understanding what prediction-based architectures can do. The brain represents 4 billion years of evolutionary R&D on prediction optimization. We’ve been doing it intentionally for about five years. The learning curve ahead is steep.

Bonus 2: AGI prediction

Most of the brain’s development arrives not from “thinking” ie learning to solve the hardest math problems, but from managing the body and outside stimulus. The first milestones in a child’s life are from managing vocal cords, managing limbs, and understanding the meaning in faces, etc. AGI arrives much faster as soon as we give AIs a body to sense the environment.
So watch you Waymo cars. They’re gonna come for us first :-)

🔍 Want to dive deeper?

Check out our book BUILDING ROCKETSHIPS 🚀 and continue this and other conversations in our 💬 ProductMind Slack community and our LinkedIn community.

🎧 Prefer to listen? Check our podcast below ↓

🎥 YouTube → Click Here

🎵 Spotify → Click Here

🎙️ Apple Podcasts →Click Here

Over the 2025 holidays, Anthropic opened the floodgates. Higher limits, more Opus 4.5 access — a gift to anyone willing to code through Christmas while everyone else was eating pie. I took the deal. My GitHub commit rate over that break was the best stretch I’ve had in years. Opus 4.5 was sharp. Tool calling was crisp. I was building real things, fast.

Then January hit, and everything got worse. Way worse.

Tool calling got sloppy. I had to keep saying “resume” and “don’t repeat yourself.” The problem-solving ability dipped — not dramatically, but enough that you could feel it in your bones if you’d been deep in it. Reddit confirmed it wasn’t just me. Then Margin Lab started tracking it with actual benchmarks, charting degradation over time. Opus 4.5 performance was sliding — measurably, consistently, in one direction.

The models got dumber. Not metaphorically. Quantifiably.

Model degradation isn’t new, and it isn’t a conspiracy. There are mechanical reasons. These models absorb new data constantly — we demanded that, because nobody wanted to hear “I was trained on January 2023 data, sorry” anymore. But when the data changes dramatically, performance shifts. There’s the synthetic data problem: AI generating training data for AI, the quality of that loop is still an open question. And there’s plain old compute throttling — finite hardware, exploding user base, juice gets spread thinner. When they gave us extra capacity over the holidays, it was partly because half their users were on vacation. January brought everyone back. The math changed.

But none of that is the point. The point is what it means for anyone building products on top of these models.

Intelligence level is Infrastructure. We Don’t Treat It That Way

We’ve been telling our clients: build AI at the core, not at the edge. Still the right call in 2026. But if you’ve done that — if the value your product delivers depends on model intelligence — you’ve made intelligence level a dependency. Not compute. Not storage. IQ.

Unlike every other dependency you’ve ever managed, there is no contractual guarantee that it stays at the level you bought it at.

We have uptime SLAs. Latency SLAs. Throughput guarantees. There is no intelligence SLA. Your enterprise contract covers availability — last I checked, it says nothing about capability. The thing you’re actually paying for, how smart the model is, has zero protection.

If you manufacture physical products and your supplier starts shipping parts that don’t meet spec, you send them back. There’s a tolerance, it’s written down, everyone agrees before a single part ships. With LLMs, there is no spec for intelligence. You just hope today’s model performs like last week’s.

Or try this: imagine you outsourced your entire engineering team to a staffing agency, and every morning, different people show up with different skills, no memory of yesterday. You didn’t contract with the engineers. You contracted with the agency. The agency swaps anyone out anytime. That’s building on LLMs right now. You don’t know who’s going to show up tomorrow.

Even enterprise rates — dedicated clusters, private instances, the whole deal — don’t fully insulate you from drift. You still want model updates, which come with downsides. You’re still exposed.

So if your core functionality depends on intelligence — analysis, generation, reasoning, whatever — and the intelligence drops, what happens? Does your product just get worse? Do your customers notice before you do? Are there legal consequences?

These aren’t hypothetical questions anymore.

The Three-Layer Defense

What do you actually do? Three things, and you need them now.

Model abstraction

Consider a middleware layer that lets you swap providers based on capability and cost in real time. I architected a version of this at Typeform in 2023 — a routing layer that could plug in Anthropic, OpenAI, Google and switch between them. Redundancy and cost optimization. Calculate how much intelligence a feature actually needs, route to whoever supplies it cheapest. Flash for the simple stuff. Claude for complex reasoning. Dynamic switching when something degrades.

This week, we designed a system for a client that uses OpenAI’s lowest model for spam filtering, then escalates borderline cases to Claude. This kind of tiered routing used to be nice-to-have. Now it’s table stakes. If your code is tied to one model and it degrades, you’re one bad week from a production incident.

Portable memory

Context and memory are becoming the real differentiator in agentic systems. But if your vectorized data, your process memory, your accumulated context is locked to one provider — you’re stuck. Model gets dumber, your memory gets less useful because it can’t reason over it as well. Provider goes down, memory goes with it. Intelligence should be portable. Memory should be portable. Two separate problems, both urgent.

Intelligence monitoring

Continuous eval batteries testing your model against the specific thresholds your product requires. Not leaderboard benchmarks — yours. The ones mapped to features your customers pay for. Understand the minimum intelligence threshold to deliver value for each piece of your product, then monitor against it. Catch degradation before your customer catches it for you — because when they catch it, it’s a brand problem, and nobody cares that it was your provider’s fault.

Same discipline you’d apply to any critical dependency. The question used to be “Is the server up?” Now it’s “Is the server still smart enough?”

What’s Next: Monitoring, Contracts, Insurance

Every piece of critical infrastructure eventually gets three things: monitoring, contracts, and insurance. Compute did. Bandwidth did. Uptime did. Intelligence will.

Monitoring is already emerging. Margin Lab tracks degradation in quantified form today. The category will grow. Every company building on LLMs will need something like it.

Contracts follow. Someone will write the first real intelligence SLA into an enterprise deal — not uptime, not throughput, but capability. A battery of evals benchmarked at contract signing, with guarantees that performance stays above a threshold. Doesn’t exist yet. Should. There’s a case for an independent standards body maintaining a shared eval battery — something enterprises run against providers the way you’d run an audit. Sounds early, but so did SOC 2 once.

Insurance comes last, as always. Rented intelligence drops below threshold; someone makes you whole. Sounds like a stretch until you remember every other infrastructure risk got an insurance product eventually.

Monday Morning Implications

Abstract your model. Make your memory portable. Monitor your intelligence thresholds. Start the contract conversation with your providers.

We’ve been saying build AI at the core. We still mean it. But building at the core means this is critical infrastructure now, and critical infrastructure demands rigor. The companies that treat their LLM dependency with the same seriousness they treat their cloud dependency will be fine. The rest will find out the hard way that rented intelligence comes with rented risk.

🔍 Want to dive deeper?

Check out our book BUILDING ROCKETSHIPS 🚀 and continue this and other conversations in our 💬 ProductMind Slack community and our LinkedIn community.

🎧 Prefer to listen? Check our podcast below ↓

🎥 YouTube → Click Here

🎵 Spotify → Click Here

🎙️ Apple Podcasts →Click Here

📣 Live: Escaping the AI Build Trap

After two standout AMAs, we’re excited to welcome Melissa Perri, author of Escaping the Build Trap and CEO of Product Institute, for “Escaping the AI Trap.”
We hope you will join us!

📅 Tuesday, Feb 17

⏰ 9am CST

Register ➡️ HERE ⬅️

Listen on YouTube, Apple Podcasts and Spotify

Ten Things We Discuss

1. We are seeing a huge crunch in the creative core.
2. Hyper creators are people who can do all of these things at once.
3. You have to de-risk by being a hyper creator.
4. The roles are going to start to be combined.
5. We need to look at core trends.
6. The future of work is headed in that direction.
7. There are so many other problems that have yet to be solved.
8. You can't be an AI skeptic sitting on the sidelines.
9. The only question we have is will that cause a shrinking of the workforce?
10. There will be more solopreneurs who are working with just 50,000 people who love their very specific product.

Mark Zuckerberg made a bet on the opposite thing that is actually happening

In October 2021, Mark Zuckerberg stood before the world and declared a new era. ‘Facebook’ was dead. Long live Meta.

The vision was seductive: a billion people inside virtual worlds within a decade, hundreds of billions in digital commerce (hello NFTs!), millions of new creator jobs. “The next platform will be even more immersive,” Zuckerberg promised. “An embodied internet where you’re in the experience, not just looking at it.” We humans would be humans, but in a different dimensionality where we could be whatever and whoever we wanted to be

It was fantastical, but it wasn’t crazy. This is the world as seen through the mind of Ernest Cline, who wrote Ready Player One, and it was adapted into a blockbuster movie by none other than Stephen Speilberg. That story itself was derived from a long line of science fiction, but fully realized in his rendition. For many who were raised in the heyday of Second Life, World of Warcraft, and many MMPORGs; this made a ton of sense. Zuck bought it wholesale, and he had the money to make a bold gamble.

He had been laying the groundwork for years, buying Oculus for $2 billion in 2014, doubling what he paid for Instagram. The metaverse wasn’t a pivot born of desperation. It was the logical extension of everything Facebook had built: connecting people, but more.

And the timing felt right - there HAD to be a sequel to the world wide web as we knew it, it was kind of long in the tooth at that point, since it kicked off hard around 1990. The crypto enthusiasts knew it. The venture crowd sensed it. And Facebook was going to bet on it and build it before anyone could. In a case of confirmation bias, the pandemic forced the physical world of humans online. Remote work became normal. Virtual connections were no longer weird. If there was ever a moment when “teleporting as holograms” to escape physical limitations made sense, this was it.

So Meta went even more all-in. Expanded investments in Reality Labs. Horizon Worlds. Quest headsets… one of us even got one for a birthday present! Billions upon billions in R&D.

The tab? Over $70 billion in losses since 2021, according to Bloomberg. Nearly $5 billion in a single quarter by late 2024.

And then, somewhere in the middle of all that spending, the world changed to what it was always intended to be, in hindsight - enhanced humans bestriding both the physical world and the cyber one.

November 30, 2022: The Day Everything Shifted

ChatGPT launched. Within five days, one million users. Within two months, 100 million: the fastest-growing consumer application in history. Today, 800 million people use it weekly. The explosion caught everyone off guard, including OpenAI. Engineers later admitted they hadn’t expected it to be very successful. But the numbers told a different story. Suddenly, AI wasn’t a research curiosity. It was a product. A tool. Something people wanted, not something they had to be convinced to use.

Google declared “code red.” Microsoft scrambled. And Meta, still bleeding cash on virtual worlds nobody was visiting, faced a choice. The metaverse asked people to leave reality. AI showed up in our reality. One required an expensive headset, a learning curve, social dislocation - an act of faith. The other required nothing but a browser.

The Rise of the Agents - Autonomous AI is finally arriving

The AI story has accelerated fast. Anthropic (founded in 2021 by former OpenAI executives Dario and Daniela Amodei) launched Claude publicly in March 2023. By May 2025, Claude 4 arrived with something different: not just chat, but agentic capabilities. Claude Code and has let people delegate real work (coding, debugging, browser automation) to an AI that showed up in their terminals, their Slack channels, and their actual workflows. OpenAI followed quickly. It also pushed reasoning models. Google released Gemini 2.0 in December 2024 with “agentic” features baked in. The entire industry converged on a new thesis: AI doesn’t need to be smarter. It needs to do things. A fancy thing insiders call “tool calling”.

And the market responded. By 2025, 79% of organizations reported adopting AI agents at some level (Ok ok, let’s maybe call them fancy scripts that fetched things that were hard to fetch, under watchful human eyes.) The global AI agent market hit $7.38 billion, nearly doubling in two years. McKinsey found 23% of enterprises already scaling agentic AI systems, with another 39% experimenting.

The metaverse was supposed to be the future of connection. Agents became the future of life AND work - browsing the web for us, becoming usable software robots that reasoned through new information.

Unscrambling the Metaverse

Meta saw what was happening. In 2023, Zuckerberg declared it the “Year of Efficiency,” code for cutting costs and reallocating resources. The company released Llama, its open-source AI models, positioning itself as a disruptor rather than a follower. Llama reached nearly one million downloads per day by late 2024.

The AI investments accelerated. Capital expenditure guidance for 2025 climbed to $70-72 billion, nearly matching everything Meta had lost on the metaverse. But this time, the money was going to data centers, GPUs, and model training. Meanwhile, Reality Labs kept bleeding. In late 2025, Bloomberg reported Meta was preparing to slash the metaverse budget by 30%, potentially $4-6 billion in cuts. Layoffs followed. The Horizon Worlds team shrank. The Quest hardware unit contracted. In 2026 10% of Reality labs was cut again, drawing down that investment to the lowest level in 5 years.

Meta didn’t and hasn’t abandoned the metaverse entirely. But the pivot has been unmistakable. As one analyst put it: Meta stopped optimizing for quarterly optics and started optimizing for strategic independence. The Ray-Ban Meta Smart Glasses (embedding Llama directly into wearables) became the new bet. Not virtual worlds you escape into. AI that travels with you in the real one.

And let’s not forget Apple bought into this, too. The Apple Vision Pro exists as a testament to how real the taste of the metaverse was, in the mind of Silicon Valley executives. Apple spent over $30 billion on the Vision Pro ecosystem - custom software, custom silicon, custom optics, developer programs, marketing, etc. And it’s estimated that it sold under 1 million units. In comparison, the iPhone sold over 240 million units in 2025 alone.

This Pivot Actually Makes Sense

What Meta and Apple are understanding now is: The metaverse was a platform play. It was a new operating system like Windows, IOS or Facebook, it would need new apps, it would need new app store, it would create lock-in - terms and strategy that had worked well in the past. Build the world, and people will come. But people didn’t come, not at scale, not yet. It was a technology bet looking for a problem to solve (just ask the NFT guys… Oh, too soon?)

The technology was clunky, the value proposition unclear, the adoption curve steep. AI agents are different. They don’t require any great change to customer behavior. They don’t ask you to buy new hardware or learn new interfaces. They meet you where you already are (a chat interface, your email, your code editor, your messaging apps) and they make you faster. Sometimes 100x faster. And this is huge.

The numbers prove it:

• Individuals who get fluent at using Claude or ChatGPT for various workflows, report tremendous productivity increases.

• 66% of companies using AI agents report measurable productivity gains

• Organizations project average ROI of 171% from agentic AI deployments

• 85% of enterprises are expected to implement AI agents by end of 2025

This isn’t hype. It’s adoption at enterprise scale. And Meta, with 4 billion users across its apps, has something no AI-native startup has: distribution. If Meta AI becomes the default assistant for Facebook, Instagram, and WhatsApp, that’s not competing with ChatGPT for premium subscribers. That’s embedding AI into the daily habits of half the planet. “Meta AI is on track to being the most used AI assistant in the world,” Zuckerberg said at Connect 2024. For once, the bravado might be justified. The only real issue to crack for team Meta is monetization. Embedding inside their collection of social apps leaves them at the mercy of the ad market, which AI advances may come along and decimate.

The Inversion Is Complete

So it’s been quite a journey away from the metaverse. In this agentic moment, the inversion is complete. Now it’s the technology learning us, to become more us. More us and faster, because CPUs. No headsets required. The AI doesn’t ask us to come to its reality; it shows up in ours. In our Slack channels, our terminals, our email, doing the work we used to do, in the world we actually live in. When their intelligence and autonomy is paired with autonomous robots, the physical world will be occupied by alien intelligence (that we created, but nevertheless a bit alien), not the reverse.

Ironically, the advent of Transformer-based Artificial Intelligence will accelerate the metaverse because making interesting new virtual worlds will become easier. They are notoriously hard to construct because they have to be handcrafted (detail is important for realism) to be high quality, vs. just procedurally generated. But now we can combine detail and procedural generation as we enter an era of infinite code. The truth is that many of us need the escape of metaverses occasionally. Many MMOs exist and are profitable. But we all need to surface to socialize, work, and play in the real world a lot more than some tech companies imagined.

Meta spent $70 billion trying to get us to leave Earth. Anthropic, OpenAI, and Google are teaching AI to work on Earth. That’s not just a strategy shift; that’s a complete philosophical inversion. And here’s the poetic twist: Meta could keep its rename. “Meta” comes from the Greek for “beyond.” In the metaverse vision, it meant beyond physical reality. In the agentic era, it means beyond human limitations: AI that extends what we can do, where we already are. Same name. Completely different meaning.

But it will also remind everyone of the misfired bet. So the jury is out on whether they will keep that name.

Key takeaways:

1. Meet humans where they are, not where you wish they’d go. Meta asked people to abandon reality. AI agents showed up in their terminals and email. One required faith. The other just worked.

2. AI agents doing real work in the real world is the entire game. The frontier isn’t smarter models—it’s AI that operates autonomously in actual workflows. Embodied intelligence that extends human capability without leaving Earth.

3. Platform plays create lock-in. Tool plays create addiction. The metaverse needed an ecosystem. Agents just need to make you 10x faster. When something works, you don’t need convincing—you need a subscription.

🔍 Want to dive deeper?

Check out our book BUILDING ROCKETSHIPS 🚀 and continue this and other conversations in our 💬 ProductMind Slack community and our LinkedIn community.

🎧 Prefer to listen? Check our podcast below ↓

🎥 YouTube → Click Here

🎵 Spotify → Click Here

🎙️ Apple Podcasts →Click Here

📣 Stay tuned…

After two standout AMAs, we’re excited to welcome Melissa Perri, author of Escaping the Build Trap and CEO of Product Institute, for “Escaping the AI Trap.”
📅 Tuesday, Feb 17 • 9am CST
Details coming soon.

In this CHOP session, Oji, Ezinne, and Ted talk about what’s actually shifting in product management, from the rise of the “hyper creator” to why traditional roles are starting to blur. When one person can now do what used to take an entire team, the way we build (and hire) has to change too.

If you’ve been wondering where you fit in this new landscape, this episode is for you.

Listen on YouTube, Apple Podcasts and Spotify

Ten Things We Discuss

1. We are seeing a huge crunch in the creative core.
2. Hyper creators are people who can do all of these things at once.
3. You have to de-risk by being a hyper creator.
4. The roles are going to start to be combined.
5. We need to look at core trends.
6. The future of work is headed in that direction.
7. There are so many other problems that have yet to be solved.
8. You can't be an AI skeptic sitting on the sidelines.
9. The only question we have is will that cause a shrinking of the workforce?
10. There will be more solopreneurs who are working with just 50,000 people who love their very specific product.

In today’s podcast Oji, Ted and Ezinne talk about the latest developments in AI and technology, focusing on the rise of Claude as a competitor to OpenAI, the hardware landscape at CES, and the implications of AI on the job market.

They discuss the shifting dynamics in the AI space, the importance of adapting workflows to integrate AI, and the future of resources as society transitions from oil to rare metals.

Stick around till the end, Ezinne drops a hot🔥 take on the metaverse you won’t hear anywhere else.

Listen on YouTube, Apple Podcasts and Spotify

Ten Things We Discuss

1. Claude has gained significant traction in the AI space, outperforming OpenAI in some areas.
2. The AI landscape is constantly evolving, with no clear market leader.
3. Hardware advancements from companies like NVIDIA and AMD are crucial for AI development.
4. AI will reshape job roles, emphasizing the need for adaptability and new skills.
5. Companies must focus on integrating AI into workflows rather than just choosing the best model.
6. The hardware market is risky, with potential for significant failures.
7. AI can help streamline and improve codebases, making them more efficient.
8. The future of resources may shift from oil to rare metals and technology.
9. The metaverse concept is evolving, with technology becoming more human-centric.
10. Organizations need to rethink their strategies to leverage AI effectively.

We believe AI coding tools are transformative, 10xing output speed at the same quality, and this is backed up by the dozens of dev teams we’ve spoken to. But a rigorous new study reveals something different.

This study showed experienced developers worked slower, not faster, and all the while these same developers thought that the tools were helping their speed by maintaining quality.

A randomized controlled trial from METR tracked 16 experienced open-source developers working on their own repositories over 246 real tasks. These devs were far from amateurs and we would call them open-source gurus. Each coding task was randomly assigned to either allow or disallow AI tools (primarily Cursor Pro with Claude 3.5/3.7 Sonnet), measure completion time, and compare the results.

When the developers used AI, they took 19% longer to complete tasks. Yet before starting, the developers predicted AI would speed them up by 24%. 

What’s more surprising is even after experiencing the actual slowdown, they still thought AI had sped them up by 20% until they saw the time log data. The perception gap is striking. Developers were systematically and dramatically wrong about their own productivity.

This Matters More Than You Think

This isn’t another “AI hype is overblown” hot take. As we said, we think the tools work, the models are impressive, and Cursor and Claude are powerful enough to recode Slack in 30 hours! But something fundamental is happening that neither developers nor their tools are accounting for.

Let’s start with the basics. These weren’t junior developers playing with new toys. They were veterans with an average 10+ years of software engineering experience and 5 years contributing to their fully mature repositories. Nerd stats: > 1,500+ commits, > 23,000 stars, > 1 million lines of code.

These are the people who should benefit most from AI augmentation as they know the codebase, they know the patterns, and they have the experience to prompt effectively and evaluate outputs critically. If AI can’t speed them up, who can it speed up?

The answer, based on prior research, has always been less experienced developers. Studies on GitHub Copilot consistently showed that junior developers got bigger productivity gains because of skill-compression. AI allows newbies to get peak performance by giving them expert-level scaffolding.

But this study flips that assumption and suggests there’s a ceiling effect, or worse, an inversion. The very expertise that should make AI tools more effective might be making them less effective.

The Five Hidden Costs

The study identifies five contributing factors after analyzing hours and hours of screen recordings that explain why developers feel faster but actually slow down.

Context switching overhead. Moving between the code editor and the AI chat interface introduces friction that compounds over time. Each switch breaks flow and the cognitive load of managing the AI assistant as a separate tool adds up in ways developers don’t consciously track.

Validation tax. Every AI suggestion requires evaluation. Is this correct? Does it fit the codebase style? Will it pass review? Does it handle edge cases? That evaluation work is invisible but expensive, transforming the developer’s role from writing code to reviewing code that looks plausible but might be subtly wrong. In fact it is this near-wrong that is probably so taxing as bugs are harder to catch.

Scope creep from capability. When AI makes it easy to generate more code, developers generate more code without the elegance that might normally go into it. The result is functionally equivalent but unnecessarily complex or brute-force solutions that feel cheap to create but aren’t.

High quality bar friction. Mature open-source projects have implicit requirements around documentation, test coverage, linting, formatting, and code style that take humans years to internalize. AI doesn’t know these requirements, so it generates code that “works” but fails review, and fixing AI output to meet unstated standards takes time.

Task redefinition. With AI, developers attempt more ambitious solutions or explore more options because the cost feels lower. This isn’t necessarily bad, but it means the task itself expands. What would have been a straightforward fix becomes a refactor, and what would have been one approach becomes three prototypes.

Together these factors create a productivity illusion where developers feel like they’re moving faster because they’re generating code faster. But as we’ve said before, a faster rocket engine pointed in the wrong direction doesn’t get you to Alpha Centauri faster.

The Benchmark Paradox

Here’s where it gets genuinely confusing. SWE-Bench Verified has frontier models solving complex programming tasks and agentic coding systems demonstrate remarkable capabilities and anecdotal reports flood socials with developers claiming they shipped features in hours that would have taken days.

So which is real? The answer is both and the gap between them reveals something important about how we measure AI capabilities.

Benchmarks optimize for completion, not production quality. SWE-Bench measures whether code passes author-written tests but doesn’t measure documentation quality, code review standards, maintainability, or adherence to implicit style guides. In mature codebases, those “soft” requirements consume substantial time, and AI generates code that works but doesn’t meet production standards.

Also benchmarks allow unlimited attempts. Frontier models in benchmark settings can sample millions of tokens, try hundreds of approaches, and iterate until something passes tests. Real developers using Cursor don’t do that. They prompt a few times, evaluate the output, and move on. This usage gap alone accounts for a lot of the difference in our opinion.

And of course anecdotes are biased. People who have great experiences with AI tools post about it while people who struggle quietly stop using them. As this study shows, people systematically overestimate their own speedup, meaning the developer who tweets “AI helped me ship a feature in 2 hours” might have actually spent 3 hours but it felt faster because they were generating code quickly. Or might be lying.

Finally quality standards vary dramatically. Prototypes and personal projects have far lower quality bars than production codebases with millions of users like these. Critical infrastructure that powers the internet has a quality bar that is extreme, where every line matters and every edge case counts.

Why Experience Makes It Worse

The study specifically notes that more experienced developers didn’t show adaptation effects. You’d expect that with practice developers would learn to use AI more effectively, figure out better prompting strategies, and develop intuition for when AI helps and when it hinders. That didn’t happen and the slowdown persisted across experience levels with AI tools.

This suggests the problem isn’t just a learning curve but something structural about how these tools integrate into expert workflows. 

Expert developers have internalized patterns, shortcuts, and mental models that make them fast. Remember, Zuckerberg was talking about 10x developers decades ago. They know exactly where to look in the codebase, can predict what will break, and understand the architecture deeply.

AI disrupts that flow by offering suggestions that seem plausible but require validation, generating code in styles that almost match but not quite, and solving problems in ways that technically work but don’t fit the existing architecture. For a junior developer, that’s all upside since they don’t have expert intuition to disrupt. For a senior developer, it’s friction. They’re constantly context-switching between their mental model and the AI’s output, validating instead of creating, and managing a tool instead of solving the problem. Plus that tool isn’t deterministic like coding itself is. As we’ve said before vibe-coding feels more like managing helpful interns with all the pluses and minuses than it does programming.

Design Your AI Coding Environment Carefully

Interestingly, a separate study from Google a year ago found the opposite result where AI tools reduced time on task by approximately 21% in their randomized control trial with 96 engineers. How do we reconcile that with METR’s 19% slowdown?

The Google study used enterprise tasks with presumably lower quality bars than elite open-source projects, and the developers were working in a corporate environment with different standards and constraints. The AI features they used (code completion, smart paste, natural language to code) were integrated directly into their workflow rather than requiring separate tools like Cursor. Most tellingly, the Google study found that developers who spent more hours per day on coding benefited more from AI, which is the opposite of what you’d expect if AI helps beginners and hinders experts.

Context is everything. AI coding tools aren’t universally helpful or harmful. They’re helpful in some settings and harmful in others, and properly designing those settings is key.

The Uncomfortable Truth About AI Augmentation

This study forces us to confront something the AI hype cycle has been avoiding. Augmentation is hard and it’s not enough to build a smart tool. You have to build a tool that integrates into expert workflows without breaking what makes them expert in the first place.

Right now, AI coding assistants are like having a very smart intern who doesn’t know your codebase, doesn’t understand your quality standards, and needs constant supervision. For some tasks, that’s valuable, but for others, it’s slower than just doing it yourself.

The developers in this study kept using AI even though it slowed them down. Some reported they found it more enjoyable, others viewed it as an investment in future skills, and some thought they were faster even though the measurements showed otherwise. This is the same pattern we’ve seen with every productivity tool that feels good but doesn’t measure well. Email feels productive, Slack feels productive, and endless meetings feel productive (ok they don’t but we threw that in there to see if you are still reading), but they all create the sensation of progress without creating actual progress.

What This Means Going Forward

First, we need honest accounting. Stop measuring productivity by “lines of code generated” or “suggestions accepted” and start measuring time to mergeable releases in production-quality codebases, because the gap between those metrics is where the truth lives.

Second, we need better integration. The future of AI coding isn’t chat interfaces and separate tools but deeply integrated systems that understand codebase context, quality standards, and implicit requirements. The tools need to meet developers where they are, not force developers to adapt to the tools.

Third, we need to stop treating AI as universally applicable. Junior developers might benefit enormously while experts working on critical infrastructure might not. Prototyping and exploration might be accelerated while production refinement might be slowed, and that’s fine as long as we’re clear about when to use AI and when to step back.

Fourth, we need to fix the perception problem. Developers systematically overestimate AI’s impact on their productivity, which means organizations making decisions based on developer feedback are getting bad data. The only way to know if AI is helping is to measure objectively, not ask subjectively.

The METR study isn’t the final word on AI coding tools. It’s a snapshot of early 2025 capabilities in one important setting. The tools will get better, the integration will improve, and the slowdown will reverse.

But right now, in mature codebases with high quality standards and experienced developers, AI is making people slower while making them feel faster. That gap between perception and reality is dangerous because it leads to bad decisions, wasted resources, and frustrated teams who can’t figure out why shipping is taking longer despite all these “productivity” tools.

Maybe that’s the most important lesson. In the age of AI, trust data over feelings, even your own.

Want to dive deeper?

We invite you to bring your questions directly to us! To thank you for your support, we are hosting an end-of-year AMA. Join us, Wednesday December 17th at 11 am CST, to connect, reflect, and get practical career and product advice to help you start the new year with confidence.

Please submit your questions when you register using the button below so we can focus on what matters most to you.

We hope to see you there -bring a friend if you think they would enjoy the AMA.

Register Here

Dear CEOs and tech leaders,

As we reflect on the dizzying headlines from this year, there’s one thing we need to get off our chest.

We need to talk about the stories you tell about AI and layoffs because, frankly, they aren’t believable.

But before we continue, our publisher is offering a holiday discount on our book, Building Rocketships: Product Management for High-Growth Companies, throughout December. And with gifting 🎁 season upon us, the book makes a great pick for a colleague, friend, or anyone interested in product leadership.

If you’ve already read the book, we’d be grateful if you could take a moment to leave a review on Amazon. It truly helps us.

Order directly from Damn Gravity to get the discount (Click ‘Order Now’ button above).

We are also hosting our first ProductMind AMA this coming Wednesday. We are so grateful for this community and we’d love to meet with you, so please join us (and bring a friend if you’d like). Details at the bottom of this post.

Thank you for your support, now let’s dive in…

At ProductMind, we work with organizations in real transformation. Business models shift. Technology accelerates. Leaders make choices that shape companies and cultures. We understand how heavy those choices feel because markets are restless and investors want efficiency. And we all know AI is changing how work gets done.

Layoffs can be necessary. Sometimes they are part of responsible, long term strategy and we advise leaders through these transitions every day.

What worries us is the one note claim that AI alone caused this year’s reductions. AI is a factor but it is not the whole story. That oversimplification harms trust inside your company and across the industry.

Call it what it is

When you present AI as the single reason for layoffs, you tell Wall Street a clean efficiency story. You also frighten the very workforce you need to adopt AI well. That is a poor trade.

Amazon fires 30,000 people

The headlines say AI automation but inside accounts point to GPU shortages that constrained AWS revenue and the finance teams needed margin relief so labor expense reductions filled the gap. GPUs needed to go to where they would generate the most revenue. This is not AI replacing people; this is corporate finance under supply pressure.

Meta fires 600 AI engineers

Meta had been bullish on AI until Llama became a base for international models, including DeepSeek. All of the sudden, Meta’s huge expenditures were more about funding its competitors than building the “open source” moat that it lacked. This is not AI replacing people; this is a strategy pivot.

What the studies actually measure

Given how early we are in the AI adoption curve everyone was excited to see actual studies come out about “results.” You saw the MIT study that found most leaders do not yet see positive P+L impact from generative AI. You also saw the Wharton study that found most executives expect or already see positive outcomes.

These apparent contradictions can both be true at once because they measure different things:

• MIT looked for hard P+L movement. Minutes saved only count if they change staffing or throughput.

• Wharton measured perceived outcomes mostly among senior leaders aka confidence in the trajectory.

Neither justifies claiming that AI alone replaced tens of thousands of jobs.

Words matter

Leaders juggle demands from markets, boards, and employees and we respect that. But overstating AI’s role in layoffs wins a headline today but loses trust tomorrow that you will need that trust for real adoption. If there is anything we have learned about technology change it is that cultural barriers overwhelm all other issues.

Employees hear that AI took the jobs and question their future. This will make future candidates hesitate to join. Worse, your customers grow wary. None of that helps you scale AI inside your company.

The honest alternative

“We are reorganizing to align with strategy. Supply constraints or market position or portfolio focus created margin pressure. AI is reshaping workflows and we are capitalizing on that, but our strength remains our people and our culture. Here is how we will help our most valuable resource transition to new roles.”

Clarity strengthens leadership and invites participation instead of fear instead of creating a people vs AI culture.

Three acts of stewardship

1. Honest framing
   Name the broader drivers that drive cuts as your teams deserve clarity and Wall Street can handle complexity.

2. Visible reinvestment
   Pair efficiency gains with investment in learning, internal mobility, and new roles. Show the math and make it concrete.

3. Empowerment over extraction
   Treat AI as an enabler of judgment and creativity and involve employees early to build trust but also to get the best ideas. Invite experiments and reward learning.

A national opportunity

The United States relies on private enterprise to drive AI absorption. That is a challenge and an opportunity. Responsible adoption across manufacturing, healthcare, education, logistics, and government will strengthen both companies and the economy.

Compare that to China which has laid out a government driven plan for AI by 2035. Our strength in the US has never been top down driving change alone, it has been mobilizing our workforce and our people as we did to answer the call in 1941 - government, private enterprise, and our workforce combining to achieve the goal.

Our shared charge

The story we tell about AI will shape its reception for years. If the story is about replacement and efficiency, we will get resistance and distrust. If the story is partnership and empowerment, we will get productivity and innovation.

Our invitation is simple: Communicate with precision. Lead with honesty. Build a narrative that fosters engagement, learning, and shared progress.

AI will not decide the future of work. We will.

With respect and urgency,
Oji, Ted, and Ezinne
ProductMind

If you are wrestling with these questions and need help, let’s talk. We do not have every answer but we are working with teams who are figuring it out with clarity and without the empty headlines.

On that note, we invite you to bring your questions directly to us! To thank you for your support, we are hosting an end-of-year AMA. Join us, Wednesday December 17th at 11 am CST, to connect, reflect, and get practical career and product advice to help you start the new year with confidence.

Please submit your questions when you register using the button below so we can focus on what matters most to you.

We hope to see you there -bring a friend if you think they would enjoy the AMA.

REGISTER HERE