Slow Down in the AI Wave: A Reality Check for Employees and Bosses

中文版 / Chinese Version: 在 AI 浪潮里慢一点:写给打工人,也写给老板

When everyone around you is grinding on AI, your boss demands “AI transformation,” and a new buzzword lands every month — this article’s message is simple: at least half of that anxiety was manufactured by marketing.

Prologue: Everyone Is Being Chased by AI

Over the past few months, I’ve watched “AI anxiety” spread through my circle like a contagion:

• A friend in sales attends a weekly AI class and has spent four months memorizing prompt templates. She sighs: “I thought my job was knowing how to talk to people. Why does this feel like cramming vocabulary?”
• A friend in office admin screenshots her ChatGPT usage daily, because “AI adoption rate” was added to her KPIs.
• A fresh graduate walked into an interview and the first question was: “Do you use AI? Which agent framework?” She froze.
• Someone in my neighborhood group chat asked: “Can anyone install that lobster thing? I’ll pay.” (They meant OpenClaw.)

The tech world is grinding even harder. Two years ago it was Prompt Engineering, last year Vibe Coding, this year Context Engineering, Subagents, Harness Engineering, Skills, MCP… a new term nearly every month.

Everyone wears the same expression — fear.

Fear of learning one step too slowly. Fear the boss will say “we don’t need this role anymore.” Fear that “you won’t be replaced by AI, but by a colleague who uses AI” comes true.

Here’s what I want to tell you: at least half of this anxiety was manufactured.

Let’s set the anxiety aside and look at first-hand data and what the people actually building AI say, on two questions:

1. As an employee, how do you stand firm in the wave without being swept away?
2. As a boss, should your company go all-in on AI transformation — and if so, how do you avoid the traps?

Last Year’s Bold Promises, Unfulfilled

Rewind to January 2025.

OpenAI CEO Sam Altman wrote a line in his blog post “Reflections” that countless companies adopted as strategic gospel:

“In 2025, we may see the first AI agents ‘join the workforce’ and materially change the output of companies.”

Translation: “Starting this year, AI will do your work for you.”

Boardrooms, annual meetings, investor calls — everyone was shouting “AI-Native transformation.” For a while, if you didn’t call yourself an AI company, you couldn’t face your peers.

The result?

In December 2025, The New Yorker published a year-end retrospective whose title says it all:

“Why A.I. Didn’t Transform Our Lives in 2025”

In October 2025, Fortune reported that OpenAI’s own co-founder had walked it back in an interview:

“Real AGI is still a decade away. They don’t currently have enough intelligence.”

The company that first declared the “Year of Agents” now has its own co-founder saying “wait another decade.”

The newer data cuts deeper. On May 19, 2026, PwC released its Global CEO Survey covering 4,454 CEOs worldwide. The conclusion:

56% of CEOs say their AI investments have produced zero financial return.

This isn’t a difference of opinion. It’s an entire industry sobering up from overpromising.

Think about it: OpenAI itself has backpedaled, and PwC’s data is unambiguous. Maybe the accounts recycling “AI will replace 100 million jobs” headlines should cool down first.

The Companies Grinding Themselves Into Absurdity

Story One: Have Employees Train the AI, Then Lay Them Off

The protagonist: Zuckerberg’s Meta.

On May 20, 2026, Meta announced layoffs of 8,000 people (about 10% of staff) while moving 7,000 into its AI division. Zuckerberg’s internal memo read: “In the AI era, success isn’t a given.” (CNBC, 2026/5/20)

Sounds almost normal? What actually blew up was a leaked internal recording from the week before:

Zuckerberg told an internal meeting that the company used internal tracking software to “identify the smartest employees” and use their work processes as AI training data. (WION News, 2026/5/19)

Mother Jones ran a headline on May 21, 2026 that was even more blunt:

“Meta had thousands of engineers train their AI replacements — then fired them.”

WIRED’s follow-up reported “historically low morale” inside Meta. Everyone is wondering: if I teach the AI my job today, is my desk empty tomorrow?

This isn’t science fiction. It’s recent news.

Story Two: Tokens as the New KPI

Silicon Valley coined a new word: Tokenmaxxing — burning tokens for its own sake.

Some large companies started treating “how many tokens an employee burns” as a performance metric.

The resulting absurdities (compiled from Success.com 2026/5, TechCrunch 2026/3, Business Today 2026/5/22):

• Amazon employees reportedly inflated their token usage on purpose to climb internal leaderboards;
• Engineers at Meta and OpenAI compete over “how many tokens I burned this week”;
• One engineer consumed 210 billion tokens in a week — the equivalent of reading Wikipedia 33 times over;
• A single person’s monthly AI bill can hit $150,000;
• An OpenAI executive publicly suggested: “Top engineers should get a $250K annual AI budget — it’s the new recruiting weapon.”

Sounds thrilling? On May 22, 2026, Fortune published a direct rebuttal:

“Microsoft reports are exposing AI’s real cost problem: using AI costs more than hiring people”

The same day, Business Today put it even more plainly:

“AI agents are costing companies more than the employees they replaced.”

Uber’s annual AI budget burned through in weeks.

More sobering still: Microsoft’s own 2026 Work Trend Index shows enterprise AI agent usage grew 18x year over year. Meanwhile PwC’s survey shows 56% of CEOs reporting zero financial return.

18x the usage × zero return = a manufactured boom.

Bluntly: burning tokens is not creating value. If your company starts grading “AI usage rate,” feel free to screenshot this section into the group chat (kidding — you know the risks).

AI Coding Is 10x Faster? The Data Says No

The most common claim: “With AI writing code, we’re 10x faster.”

Sounds great. But product managers slowly noticed something: the programmers got faster, yet release velocity stayed the same.

Why?

Numbers That Silence CEOs

In July 2025, the research group METR ran what the industry now calls the gold-standard randomized controlled trial: 16 experienced open-source developers, 246 real tasks, using Cursor + Claude — top-tier AI coding tools.

Before the experiment, developers estimated AI would make them 24% faster on average.

The measured result: they were actually slower.

They knew their own codebases too well — verifying, correcting, and re-checking the AI’s suggestions cost more than it saved. Fortune’s headline at the time: “Coding is supposed to be genAI’s killer use case. But what if its benefits are a mirage?” (2025/7/15)

In April 2026, DX released a report tracking 4.2 million developers:

Metric	Value
Increase in AI adoption	65%
Median improvement in PR throughput	7.76%
Typical range across companies	5%–15%

One-sentence summary: 93% of programmers use AI, but productivity rose about 10%. (DX Research, 2026)

Why the Gap Between “Flying” and “10x”?

From my vantage point on the front lines, three things:

1. AI writes fast, but humans still review. AI outputs 2,000 lines an hour; nobody dares ship unreviewed code, so the bottleneck moves to the human.
2. AI-written tests have their own bugs. Test cases can be wrong and results non-deterministic; a human still backstops everything.
3. Writing code is one step in the pipeline. Deployment, operations, monitoring, go-to-market — AI is far less mature there. Speeding up one stage doesn’t speed up the whole.

Analogy: you gave the chef a lightning-fast vegetable slicer, but there’s still one waiter and one dishwasher. The meal doesn’t reach the table 10x faster.

Understand this, and you understand why companies spend heavily on AI coding while the boss feels “nothing really changed.”

Enterprise AI Agents: 95% Never Make It

Many bosses’ first instinct for “AI transformation”: “Let’s connect our internal data and build a company AI brain — our own agent!”

Sounds beautiful. But MIT’s NANDA project report “State of AI in Business 2025” (August 2025, via Fortune) threw cold water on every CEO:

Metric	Value
Enterprise GenAI pilot failure rate	95%
Total global investment	$30–40 billion
Success rate of purchased solutions	67%
Success rate of self-built solutions	33%

Self-built succeeds at exactly half the rate of purchased.

The finding isn’t isolated. PwC’s May 2026 CEO survey has 56% of 4,454 CEOs reporting no financial return. Sinch’s “2026 AI Production Paradox” and Deloitte’s “State of AI in the Enterprise 2026” reached the same conclusion: “AI adoption is accelerating, execution is slowing.”

I’ve watched plenty of self-built agent projects. What ships in the end is a slightly smarter chatbot — it can look up data, answer FAQs, write summary reports. Then comes the press conference: “We are now an AI-Native company.”

Why is it so hard?

1. The technology iterates too fast. Spend millions fine-tuning a model this year; next year an open-weights release is smarter out of the box. Yesterday’s moat is today’s ditch.
2. “Data security” is often an excuse. You’re still calling OpenAI’s or Anthropic’s APIs, aren’t you? The data leaves anyway. Full private deployment? Check whether your annual profit can carry that cost.
3. The concepts change quarterly. Last year Prompt Engineering, this year Context Engineering and Harness Engineering — every trend shift writes off the last one’s work.
4. The scarce resource isn’t models, it’s people who understand the full chain. Getting an agent to actually work takes someone who knows the tech + the business + operations + product. Such people are rarer and pricier than GPUs.
5. You can’t out-build Claude Code or Codex. Those products are forged from hundreds of millions of dollars and thousands of PhDs. Thirty hires won’t catch up.

In one line: most enterprise “self-built agents” are a CTO’s vanity project, not business leverage.

What Did the Successful 5% Do Right?

MIT’s report singles them out. Almost none built models from scratch. Their common traits are humble:

• Narrow entry point: no “company brain” — pick one concrete pain point (support scripts, contract review, expense flows) and make one AI tool work end to end;
• Buy, don’t build: use frontier AI (Claude, ChatGPT, Gemini) directly, and spend 80% of the effort on process redesign and employee training, not the model;
• Measure business outcomes: response time, output per person, error rate — not “token consumption” or “AI usage rate”;
• AI augments people, not replaces them: let employees do better work with AI, and invest the saved time in higher-value activity — instead of firing them once the AI is trained.

Plainly: the winners aren’t “AI companies” — they’re companies that know how to use AI.

What the Real Builders Say

When you’re drowning in headlines, listen to the people actually building AI at the frontier. Three of them, different positions, every line lands.

1. Andrej Karpathy (former Tesla AI Director, OpenAI founding member)

On the Dwarkesh Patel podcast in October 2025, he said flatly:

“It’s not the Year of Agents — it’s the Decade of Agents.”

Harsher still, he called current agents “slop” — burdened with serious “cognitive deficits” that need a long climb toward reliability.

The interesting twist: on May 19, 2026, Karpathy announced he was joining Anthropic’s pre-training team. TechTimes captured it precisely:

“The man who called today’s agents ‘slop’ joins Anthropic — using Claude to build the next Claude.”

Why does this matter? Karpathy is signaling with his actions: he believes in AI’s long-term value, and considers the short-term hype a bubble.

In plain words: don’t all-in on the present, and don’t lose faith in the future. Cultivate steadily; don’t chase every buzzword.

2. Yann LeCun (Turing Award laureate, former Chief AI Scientist at Meta)

Note: he left Meta — the very company above that had employees train their AI replacements.

In November 2025, after 12 years, he departed to found AMI Labs with $1 billion raised. His reasoning was hardcore (per Newsweek, The Next Web, and others):

“LLMs are an off-ramp. They are a statistical illusion, not real intelligence.”

“I advise young researchers not to work on LLMs — work on next-generation AI systems.”

Translation: betting everything on the LLM paradigm today — racing to self-build agents — may be betting on a paradigm about to be superseded.

3. Jensen Huang (Nvidia CEO)

Huang sells the shovels, yet what he tells companies and employees is the most restrained (quotes compiled from Fortune, 2025/10 and 2026/4):

“AI agents won’t replace enterprise software — they’ll make it stronger.”

“The future workforce is a combination of humans and digital humans.”

“You won’t lose your job to AI, but you’ll lose it to someone who uses AI.”

“Some CEOs have a god complex — they think they know everything. That overconfidence costs companies their key people.”

“Your job is related to the tools you use to do it — but it is not the same thing.”

That last line belongs on every worker’s monitor.

Translation: AI is the new era’s tool, not its worker. The ones who should worry aren’t people with jobs — they’re bosses who worship tools.

For Employees: You Don’t Need That Much Anxiety

Now for ourselves — this is what the article is really about.

Everyone around you is anxious about learning too slowly, falling behind, being replaced. But look back at what you just read:

• The OpenAI co-founder who declared 2025 the Year of Agents now says “another decade”;
• Meta’s former Chief AI Scientist calls the current path a dead end;
• Jensen Huang keeps repeating “jobs won’t disappear — how you use tools will change”;
• MIT says 95% of enterprise AI projects produced no return.

So: at least half the anxiety was manufactured.

The other half is real — but the answer isn’t panic-learning every new term. It’s these five rules:

1. AI Is an Amplifier, Not a Replacement

What you’re already good at, AI makes better. What you’re bad at, AI can’t save.

If you can’t write copy, AI’s output will seem “pretty good” to you — because you don’t know what good copy looks like. AI’s ceiling is the user’s judgment.

So the priority isn’t learning AI — it’s becoming someone with real insight into your own field first.

2. Strengthen the Fundamentals AI Can’t Touch

The more advanced AI gets, the scarcer these become: taste, judgment, communication, trust, accountability.

• Sales: AI writes the pitch, but the client trusts you;
• HR: AI screens résumés, but reading people is on you;
• Teachers: AI explains knowledge, but kids choose to learn from you;
• Doctors: AI diagnoses, but the patient extends their hand to you.

Human connection is what AI can’t take — for now.

3. Don’t Chase Every Buzzword

Vibe Coding, Context Engineering, Subagents, Harness, Skills, MCP… the words change monthly. You can’t catch them all.

Understand principles, not names. Principles transfer.

Example: you don’t need to know how RAG is implemented — just that “the AI reads your documents before answering.” The rest is engineering detail that isn’t your problem.

4. Pick One or Two Tools in Your Domain and Master Them

Don’t bounce between ChatGPT, Claude, and Gemini weekly.

Pick one. Use it for three months. Integrate it deep into your workflow until it becomes part of how you work — worth 100x more than surface familiarity with ten tools.

5. Stay Curious About Tools, Skeptical About Hype

These two are not in conflict.

The curious don’t fall behind; the skeptical don’t get harvested.

A Few Words for the Bosses

Back to the opening question — should your company go all-in on AI transformation, and how?

My answer: transform, yes — but not the way you’re imagining.

Not layoffs, not building your own LLM, not tokens-as-KPI. Not a press conference declaring yourself “AI-Native.” And definitely not squeezing employee expertise into training data.

Five plain suggestions, all anchored in one idea: don’t panic — return to value itself.

1. Stop Obsessing Over Building Models; Retrofit Your Business Systems

Your moat was never model weights. It’s business data + business processes + domain know-how.

Say you have a 20-year-old internal system (CRM, ERP, ticketing — whatever) that employees know cold, but with messy APIs and poor docs. The right way to “AI-ify” it is not spending $50M training a “sales LLM.” It’s:

• Clean up the API documentation;
• Add an AI-friendly protocol layer (MCP, CLI);
• Write a “business capability catalog” so an AI can call your systems like a new hire reading the handbook.

Then no matter how the model layer churns — Claude 4 to Claude 5, GPT-5 to GPT-6 — your business capability plugs straight in.

One line: let frontier AI use your systems, instead of rebuilding your systems into an AI.

2. Use Frontier AI; Don’t Reinvent the Wheel

OpenAI, Anthropic, and Google each burn on the order of $10 billion a year on model training. Thirty PhD hires won’t close that gap.

Many bosses say “we self-build for data security.” But:

• You’re still calling someone else’s API — the data leaves anyway;
• True private deployment: can your annual profit sustain the cost?
• A multi-million-dollar fine-tune gets obsoleted by next year’s release.

Being a smart user beats being an expensive imitator, 100 to 1.

Put the savings into workflow redesign, training, and customer experience — the returns are far higher.

3. Treat Tokens Like a Utility Bill, Not a KPI

Tokenmaxxing structurally rewards performative work. Employees inflate usage for the leaderboard; companies encourage inflation for the earnings narrative — and the burn is all real money while the business doesn’t move.

Healthy metrics are always business outcomes:

• Is customer response time down?
• Is the defect rate down?
• Are iteration cycles shorter?
• Is customer satisfaction up?

Not the token counter in the logs.

The genuinely smart companies: employees choose freely whether to use AI, but month-end output and quality are visible to everyone.

4. Grow “Full-Chain Understanders” — Worth More Than Poaching AI Experts

The most valuable people ahead aren’t prompt-writers. They’re conductors who understand business + technology + product + operations.

What can they do?

• Decompose a requirement and know which steps go to AI and which need a human backstop;
• Orchestrate multiple AI tools + teams + systems into a flow that actually runs;
• When things break, diagnose whether it’s the model, the prompt, or the data;
• Distill domain know-how into reusable AI assets.

In the 95% failure statistics, this kind of person is what the successful 5% had.

Don’t rush to poach an “ex-OpenAI engineer.” Find the person on your team who knows the business and wants to learn the technology, and grow them into a full-chain understander. Home-grown beats parachuted-in.

5. Don’t Copy the Layoffs — and Never Treat Employees as Training Data

Meta’s play — have employees train the AI, then fire them — looks shrewd and is long-term poison for company culture.

WIRED reports “historically low morale” inside Meta. Consider:

• Will the remaining employees ever teach the company their real skills again?
• Will new hires, having read the news, fully commit?
• Will the genuinely excellent preferentially choose companies that don’t discard people after use?

Strategically, too: the marginal value of AI models is declining, while human judgment, creativity, and trust grow scarcer. A company that feeds its people into the AI hopper will find its talent reservoir dry in three years.

The high-water play: let AI make employees more effective, and invest the freed time in what AI can’t do — customer relationships, product innovation, organizational capability. The company and its people grow together, instead of people being squeezed dry and discarded.

This isn’t just an ethics question. It’s a business question.

Closing

In an era of technological explosion, the easiest thing to forget:

People are the end; technology is the means.

You don’t exist to “know how to use AI” — you exist to live a better life. Companies don’t exist to “become AI companies” — they exist to create value.

Sam Altman says one person plus a swarm of AI agents could someday build a billion-dollar business.

Maybe. But seeing the direction of the tide clearly matters more than chasing every wave.

Less anxiety, more composure. Be the surfer, not the sacrifice.

Slower, steadier — that’s just fine.