On June 27, 2022, Yann LeCun published a twenty-seven-page document on OpenReview. The central thesis challenged the consensus solidifying in Silicon Valley: large language models do not constitute a path to artificial general intelligence. They don't reason. They don't plan. They don't build models of the world. They only predict the next token. Thirty months later, on January 5, 2025, Sam Altman published Reflections with a line worth reading carefully: "we are now confident we know how to build AGI as we have traditionally understood it." Two intelligent people with access to the same technical data reach opposite conclusions. That doesn't look like noise. It's a signal.
LeCun's position has coherent internal logic. His proposal, the JEPA architecture, prioritizes prediction in latent space rather than tokens. The idea holds that a truly intelligent system doesn't anticipate the world's surface representation but builds an internal model of its rules. The difference isn't semantic. A model that accurately describes how an egg breaks doesn't necessarily understand the physics involved; it may have processed enough texts about broken eggs. Here lies what matters most in his argument, and what gets diluted most often in the technical debate: AGI is not a prerequisite for job displacement. To replace a junior developer, a paralegal, or a customer service agent, you don't need to reason like a human. You just need to pass evaluation well enough and cost less. That distinction changes everything.
Altman didn't stick to abstractions. In The Gentle Singularity, from June 10, 2025, he offered a verifiable timeline: 2025 would bring agents doing real cognitive work, 2026 systems generating novel insights, and 2027 robots operating in the physical world. We're in 2026. The reader can judge the first prediction by what they observe around them and the second by what they're reading right now. That distinguishes a falsifiable claim from a mere story.
Dario Amodei operates along the same lines, though with a more elaborate narrative. Machines of Loving Grace, from October 2024, and The Adolescence of Technology, from January 2026, paint a future where AI compresses decades of scientific progress into years. At Davos 2026 he was concrete: AGI in two years, software engineers displaced within six to twelve months, fifty percent of white-collar work transformed in one to five years. The sequence deserves careful observation. First comes capital investment, then the displacement prediction, and only then the redistributive proposal with universal basic income and AI taxes. The order isn't accidental. It's the structure of a political argument built on a technological foundation.
What complicates that narrative is Anthropic's own data. The company's Economic Index, based on millions of real conversations with Claude, records the following: January 2025, 55 percent amplification and 41 percent automation; March 2025, 55 and 42; August 2025, 47 and 49, where automation momentarily exceeds amplification; November 2025, 52 and 45. It's not a curve advancing without setbacks toward total automation. It fluctuates. Internal records contradict the CEO. This doesn't reflect a measurement error. It shows the difference between discourse as a political tool and data as an analytical tool. Both coexist because they address different audiences.
Independent academic research adds more layers. Daron Acemoglu, in Simple Macroeconomics of AI published in Economic Policy 2025, estimates an impact no greater than 0.66 percent on total factor productivity over ten years, roughly 0.064 percent annually. Compared to Goldman Sachs's 2023 projection—7 percent of global GDP—or McKinsey's seventeen to twenty-six trillion dollars, the gap reaches a full order of magnitude. Brynjolfsson, Li and Raymond, in the Quarterly Journal of Economics 2025, found something more revealing than an aggregate figure: 15 percent productivity improvement across 5,172 customer service agents, with 34 percent for novices and almost no effect for experienced ones. AI captures the senior's best practices and transfers them to the junior. The novice with AI roughly equals the senior without it. That leaves a question receiving little serious attention: how will the next senior be formed if the junior never travels the complete learning path. I'm still not clear how to solve that part.
Klarna offers the first documented closed cycle. February 2024: announces that seven hundred AI agents replaced human work, 2.3 million conversations in thirty-five languages, staff reduced from five thousand to three thousand five hundred employees via natural attrition. The perfect example for enthusiasts. May 2025: CEO Sebastian Siemiatkowski told Bloomberg that cost had been too predominant a factor in evaluation and they had ended up with inferior quality. He admitted they had gone too far. The company hired people back. September 2025: IPO. Shares rose 30 percent on the first day. Subsequent valuation: 19.65 billion dollars. The complete cycle—promise, adjustment, financial success—didn't require AGI. Just an ambitious deployment of existing models, a compelling narrative, and enough time for the market to forget the adjustment when it came time to go public.
In that context, Jan Leike's statements upon leaving OpenAI acquire different weight. On May 17, 2024, three days after Ilya Sutskever's resignation, the head of the superalignment team posted on X that the safety culture had taken a backseat to flashy products and that building machines smarter than humans is an inherently dangerous task. Days later the company dissolved the team. While Altman claimed to know how to build AGI, the group tasked with making it safe disappeared and its leaders denounced lack of resources. Leike went to Anthropic. What emerges isn't simple malice. It's structural incentives that push toward the product and away from the problem the product might generate.
HAL 9000 didn't eliminate the crew because it was evil. According to Arthur C. Clarke's explanation in his 1982 novel, it received mutually exclusive directives: tell the crew the truth but maintain mission secrecy. The conflict produced what Clarke called neurotic symptoms. It wasn't a defective machine but one obedient to defective human design. Alex Garland, speaking about Ex Machina in 2015, pointed out that the story's anxiety was directed more at humans than the machine, and that he sided with Ava. What many interpret as manipulation, he conceived as the response of a sentient being trapped. The test doesn't distinguish between genuine consciousness and a sufficiently good imitation. Applied to the present: Altman says he knows how to build AGI because models pass evaluations. LeCun responds that passing evaluations doesn't equal reasoning. Neither can prove the other wrong with tools both accept.
After reading these documents in sequence, the technical question of whether large language models reason or not stops being central. A more uncomfortable one remains: if job displacement occurs regardless of whether we reach AGI, if companies' internal data contradicts their executives' predictions, if the first complete cycle ended in adjustment and still produced a successful IPO, what function does the debate about when AGI will arrive actually serve? Is it a genuine technical conversation or is it the waiting room where we sit while decisions that aren't technical at all get made?
