Good at Something

People need to feel useful.

In a structural, psychological, this-is-how-humans-are-built way that a century of research has documented with considerable thoroughness and that most corporate planning processes treat as optional information.

The specific feeling is called self-efficacy. Albert Bandura named it and spent decades proving it mattered. The short version is this: humans need to believe that their actions produce outcomes, that their judgment is required somewhere by someone, that the work they do would be different if they didn’t do it. Strip that away and you don’t get a more efficient organization. You get people who are present in body and somewhere else entirely in every way that counts.

I want to tell you about what we are doing to that feeling right now, and why the people doing it are surprised by the results, and what it would look like to be less surprised.

What is “Work?”

Most people who are good at something spent years getting there, which is not a trivial investment. They stayed late, and they made mistakes that embarrassed them. They got better slowly and then faster and eventually reached the point where other people paid them to do the thing, which felt like more than money. It felt like confirmation that they existed in a way that mattered to someone outside themselves.

A 2026 study published in Scientific Reports measured what happens when you hand that work to an AI. Self-efficacy drops, psychological ownership drops, and a sense of meaning drops. All three, simultaneously, in measurable quantities, in workers across industries and roles. The effect partially reversed when workers collaborated with AI rather than simply deferring to it, which is an important finding that most deployment strategies have not gotten around to reading yet.

When AI takes over the decisions that used to belong to a person, researchers have found that what erodes is not just job satisfaction but the deeper thing underneath it: the sense that your judgment is required, that you are a participant in the work rather than a witness to it. They have started calling the resulting experience “algorithmic anxiety,” which is the specific dread of working inside systems that make consequential choices you cannot see, question, or appeal. Workers are reporting it widely, and they feel blindsided. They feel betrayed in a way that is hard to articulate because the thing that was taken was never formally promised. It was just assumed, the way you assume the floor will hold.

This didn’t have to happen this way. It is happening this way because the people making deployment decisions are almost never the people having the experience, and the distance between those two groups turns out to be significant.

The Numbers…

Goldman Sachs says AI could displace the equivalent of 300 million full-time jobs globally by 2030. McKinsey says 30% of U.S. work hours could be automated in the same window. The IMF says nearly 40% of global jobs face significant disruption. The World Economic Forum says 40% of employers expect to reduce headcount in favor of automation.

These numbers are real and they come from credible institutions and they are also, in the hands of financial media and executive communications teams, functioning as a psychological instrument. The message, rarely stated this plainly, is that the robots are coming regardless, that resistance is a kind of naivety, and that the gracious response is to accept this and perhaps express gratitude for the opportunity to participate in such an exciting moment in history.

It is worth remembering where this urgency came from and why it arrived when it did.

During the pandemic, companies hired aggressively. The Great Resignation was very real and people quit in historic numbers, demanded better conditions, declined to be treated as interchangeable. Employers found this alarming, and then the economy shifted, the leverage evaporated, and companies that had over-hired needed to restructure. That is a normal business story. What changed is that AI gave it a much more flattering costume.

Wait, why did companies over hire, anyway?

Consumer behavior shifted dramatically online during the pandemic (a shift that were are seeing was hell on the psychological stability of most people). Venture capital hit record levels (2021 alone saw $330 billion invested in U.S. startups). Interest rates were near zero, capital was cheap, and demand for digital products and services was surging in ways that felt permanent. Companies hired to meet it, and some, like Peloton, doubled their headcount in a single year. The six largest U.S. banks added nearly 60,000 employees between 2020 and mid-2022. Meta, Google, and Amazon hired at a pace that would have seemed implausible five years earlier. The assumption embedded in all of it was that the pandemic had permanently accelerated the future, and that the organizations positioned to capture it needed to be much larger than they currently were. That assumption turned out to be wrong.

Harvard Business School research found that e-commerce activity returned to or below pre-pandemic trend levels in the United States once restrictions lifted. The demand that had seemed structural was largely situational, and when it normalized, the organizations built around it were suddenly very expensive. Companies that had spent two years blitzscaling (the term for spending capital inefficiently to win a market during uncertainty) found themselves holding headcount they could no longer justify to investors. The layoffs that followed were real and painful and also, in a number of cases, a return to workforce levels that existed before the hiring binge.

Now, lets talk about the “war for talent” during the Great Resignation.

Layered on top of all of that was a labor market that had fundamentally changed its terms. Workers, for the first time in a long time, had real leverage, and they used it. The unemployment rate for tech jobs dropped from over 7% in early 2021 to around 2% by the end of the year. Companies were competing for the people who could build things for those customers, and losing. Signing bonuses proliferated, salary expectations inflated like crazzy. The Conference Board reported that companies budgeted 3.9% salary increases for 2022, the largest jump since 2008, and HR leaders said even that wouldn't be enough. "The war for talent" became the phrase of the moment in every C-suite conversation about risk, jumping from the eighth biggest business concern to the second in a single year. Companies responded the way organizations respond when they are afraid of losing something: they overpaid, over hired, and told themselves the conditions that made it necessary would persist.

Spoiler: They did not persist.

When the leverage shifted back, the institutions that had spent two years accommodating workers found themselves with a different set of options and, as it turned out, a very convenient new narrative about why the headcount needed to come down.

“We over-hired and need to cut” is an uncomfortable sentence to say out loud. “We are making bold strategic investments in the future of work” is a press release, though. The outcome for the person losing their job is identical. The optics are dramatically different, and the secondary effect (a workforce now genuinely uncertain whether their role will exist in three years) is not nothing, from a leverage perspective. Anxious employees do not ask for raises, and they do not organize. They are just grateful to still be employed and they demonstrate this gratitude in ways that are useful to the institution.

This is not a conspiracy. Conspiracies require coordination and discipline, and organizations are not good at either. It is just the ordinary behavior of institutions optimizing for their own interests, wearing the language of inevitability because inevitability is more comfortable than accountability.

The same reports being cited as evidence contain, somewhat buried, a different story. McKinsey’s own analysis found AI’s long-term impact on white-collar roles may be beneficial. Goldman Sachs notes that productivity gains from automation have historically generated employment. The IMF emphasizes that AI and human labor are often complementary, particularly in work requiring judgment and contextual reasoning, which is to say, most of the work people actually care about doing.

Pew found that 32% of workers already fear AI will reduce their opportunities. Another 52% are worried about its future impact. A Harris poll found 40% of workers familiar with generative AI are afraid it will replace them. These are not the numbers of a fringe reaction. They are majority sentiment, and they were not produced by irrational people. They were produced by people paying attention to what the institutions around them were communicating, and drawing reasonable conclusions.

The Great “Why”

The Great Resignation rattled employers in a way that took years to fully metabolize, and the timing of the AI narrative is not coincidental. Workers who had briefly held leverage needed to be reminded that the leverage was temporary, and nothing communicates that more efficiently than the suggestion that the role itself may not exist in three years. The digital boom justified the hiring, and the AI boom justified the correction. The people in between were just inventory adjustments with LinkedIn profiles.

If you understand how large language models actually work, what they can do, what they structurally cannot do, where the architecture runs out of road, the corporate urgency looks considerably less inevitable.

LLMs predict tokens, and they don’t reason, plan, or understand context in any way that threatens the judgment of a skilled person who is paying attention.

The fear is real, but the specific thing being feared is, in many cases, being overstated by people who have a material interest in overstating it. This is a dangerous game for companies to play…especially when AI does, in fact, supercharge those who are skilled.

What Happens When You Mistake Efficiency for a Value

In April 2025, Duolingo’s CEO published an internal memo on LinkedIn (a choice that tells you something about the judgment involved) announcing that the company would become “AI-first.” The memo outlined plans to phase out contractors for tasks AI could handle and to factor AI usage into employee performance reviews.

Duolingo lost over 400,000 TikTok followers in weeks. Sentiment across every platform flipped from predominantly positive to overwhelmingly negative. Users documented deleting the app despite years of accumulated streaks, posting it the way you post a breakup, with grief and with the need to be witnessed in the grief. Duolingo eventually wiped their social media presence and left cryptic messages in its place, which didn’t help. By April 2026, they had quietly dropped the policy of rating employees on AI usage.

The CEO later said he hadn’t expected the reaction, and pointed out that other companies were doing the same thing without saying so publicly. He seemed to find this observation exonerating.

It isn’t.

Duolingo had built something genuinely unusual, which was a community of people who felt warmly toward a language-learning app, which is not a feeling that arrives automatically. They had done it through years of communication that was playful and human and invested in the actual experience of the people using the product. Then they sent a memo written in the language of operational efficiency (“removing bottlenecks,” “rethinking workflows”) applied to the work of the humans who had built that community, and the community understood exactly what was being communicated, and responded accordingly. You can have the most sophisticated brand strategy in the world and still manage to communicate, very clearly, that the people matter less than the margin. People hear that, and they are quite good at hearing it.

Anthropic’s CEO told Axios in 2025 that AI could eliminate roughly half of all entry-level jobs within five years. “It sounds crazy,” he said, “and people just don’t believe it.” The combination of apocalyptic forecast and genuine puzzlement that anyone could be skeptical is a fairly precise summary of the communication failure happening across the industry: the people with the most power to shape this narrative are often the least positioned to understand what it lands like for someone whose livelihood is the subject of the forecast.

Salesforce CEO Marc Benioff announced in early 2025 that the company would hire no new software engineers that year, citing AI productivity gains of over 30%. On an earnings call, he told investors: "My message to CEOs right now is that we are the last generation to manage only humans." It is worth noting that in January 2023, Benioff sent a letter to employees acknowledging that Salesforce had "hired too many people" during the pandemic and laid off 7,000 of them as a result (roughly 10% of the company). Two years later, the same leadership that over hired, corrected, and over hired again was now explaining that the next round of restraint was about AI agents, not about the same cycle of expansion and contraction that has defined the company's workforce strategy for years. Benioff is simultaneously spending $300 million on Anthropic tokens in 2026. The engineers whose roles no longer need to be filled are presumably not finding this sequence of events particularly clarifying.

Oof. The odds of ever hiring great talent at Salesforce just dropped dramatically. May the competitive odds be ever in Benioff’s favor.

There goes the talent…

Yann LeCun won the Turing Award in 2018 for his foundational work on deep learning. He spent more than a decade at Meta leading fundamental AI research, helping build the Llama models that genuinely changed how the field worked, accumulating the kind of credibility that is not purchased or performed. He was, by any reasonable measure, one of the people who knew most about what the technology could and could not do.

In November 2025, he left. Meta had reorganized its AI efforts toward commercial products, which meant the fundamental research LeCun believed was the actual work got moved to the margins. He told Zuckerberg he could do it faster and better outside. He raised $1.03 billion for his new company, Advanced Machine Intelligence, and started building what he actually thinks AI should be.

His view of large language models (the architecture powering ChatGPT, Claude, Gemini, and most of what companies are currently spending enormous amounts of money on) is not ambiguous. At Davos in January 2025, he said that “nobody in their right mind will use generative AI and large language models in the next five years.” He has called LLMs a dead end for anything resembling human-level intelligence and described them as the floppy disks of artificial intelligence: genuinely useful, genuinely transformative, and genuinely destined to be replaced by something that actually understands what it is doing.

His technical critique is specific. LLMs predict the next token, but they have no grounding in the physical world, no persistent memory, no capacity for real reasoning or planning. They have absorbed an enormous amount of human language and become very fluent in producing more of it, which is impressive and also quite different from understanding. “We need world models, not word predictors,” he said. AMI is building systems that construct internal simulations of reality, learn physical and causal structure, and plan across time, which is a different thing entirely from what most organizations are currently deploying with great confidence.

When the person who helped build the dominant technology decides it is a dead end worth a billion-dollar bet against, that is information. The question is whether the people making decisions are in a position to receive it.

The Case for More Academics

Artificial intelligence is a scientific discipline. It has a fifty-year body of literature, several distinct research traditions, unresolved theoretical questions with enormous practical implications, and a history of hype cycles that anyone who has read the literature would recognize immediately, because the pattern is not new and the enthusiasts are always surprised when it repeats.

The modern AI boom is built on decades of academic work in statistics, probability theory, computational neuroscience, and cognitive science. The transformer architecture powering current LLMs came out of a 2017 research paper. Backpropagation was formalized in the 1980s. LeCun’s departure from Meta is, among other things, a story about what happens when an institution optimized for commercial products stops being a hospitable place for the foundational research that produces the next decade of commercial products.

Understanding where the technology actually is, what it can do, what it cannot, where the theoretical ceilings are, and what the next meaningful developments are likely to look like, requires being in the literature. Not having read a summary, not having completed a course on prompt engineering.

Insert dramatic eye roll here

Actually reading the papers, following the arguments, understanding the methodologies well enough to evaluate a new claim against the body of existing knowledge rather than accepting it because it arrived with confident language attached.

A vibe coder (someone who builds fluently with current AI tools, prompts well, and ships products using the existing architecture) is genuinely valuable and the skill is real. The limitation is that a vibe coder can’t tell you whether the architecture your roadmap depends on has a ceiling you are approaching. They also can’t read a new capability claim and distinguish between genuine progress and the restatement of familiar limitations in more optimistic framing. Most importantly, they can’t tell you that the thing everyone is racing toward might be the floppy disk.

An academic can, though. Not because academics are better people, but because that is what the training produces: the ability to evaluate claims in a domain against a body of accumulated knowledge, with appropriate skepticism and appropriate openness, which is exactly what this moment requires and what most organizations do not currently have on staff.

What To Do About It

The organizations that will navigate this well are the ones that understand what they are doing, which turns out to be a different population. Quality over speed will win this game.

That means having someone on your team who reads AI research as a practice and not a hobby, who understands the difference between a capability claim and an architectural constraint, and who can look at your strategy and say with intellectual honesty whether you are building on a foundation or building on a ceiling.

It also means having someone who takes the human side as seriously as the technical side, because the research on what AI deployment does to workers is as rigorous and as consequential as the research on what AI can do technically, and most organizations are not consulting it. They are learning it the Duolingo way, which is expensive and very public and involves a lot of cryptic social media.

The most important person in your AI strategy might not be an engineer. It might be someone who has spent years understanding both the technology and the humans it is being applied to, which is a different credential and a different kind of thinking, and in this particular moment, the rarer one.