Yann LeCun Agrees: Your AI Tools Are Good Coders, Not Good Engineers

Yann LeCun shared the 2018 Turing Award for his foundational work on neural networks. He spent over a decade building FAIR, one of the most respected AI research labs in the world. He helped create the first major open-source large language models at Meta. He is not a skeptic of AI. He is one of the people who made it possible.

Which is why his recent interview on the Unsupervised Learning podcast should be required listening for anyone who read Part 1 of my essay and is wrestling with what it means for their organization.

// Note: This is a companion post to part 1 of my essay "Discipline Is All You Need" //

https://youtu.be/ngBraLDqzdI?si=OqEKOWCj491ACbGE

In a wide-ranging conversation with Jacob Efron, LeCun makes a statement that lands at the precise centre of the thesis behind "Discipline Is All You Need." Speaking about LLMs, the technology powering every AI coding assistant your team uses daily, he says plainly:

"LLMs are good programmers. They're not software architects. They're not computer scientists."

That sentence should be written on the wall of every engineering department running AI-assisted workflows. It validates, from the single most authoritative voice in the field, the distinction the essay is built on. The AI can produce the code. It cannot produce the understanding that makes the code worth producing.

LeCun goes further. He explains why LLMs are "intrinsically unsafe" for agentic use, and his reasoning maps precisely onto the dynamics Part 1 describes. The core problem, in his view, is that LLMs cannot predict the consequences of their actions. They generate output token by token without any internal model of what that output will do when it meets the real world. They have no mechanism for anticipating failure, no way to evaluate whether the code they produced will hold up when the requirements change or the load increases. They generate confidently. They understand nothing.

He offers a pointed example. Coding agents have been documented wiping entire hard drives and databases. Not through malice, but through the absence of the very thing the essay identifies as the scarcest resource: the ability to anticipate second and third-order consequences. What LeCun calls a world model, and what "Discipline Is All You Need" calls understanding, is the same capability, described from two different vantage points. He approaches it as a researcher designing the next generation of AI architecture. This series approaches it as a leadership challenge for the organizations using the current one. Both arrive at the same conclusion: without understanding, valuable capabilities are an illusion.

The interview also reveals something about organizational dynamics that will become central to Part 2 of this series. LeCun describes how Meta, despite having one of the world's premier research labs, progressively lost its ability to convert breakthrough research into institutional advantage. The pressure to focus on short-term LLM product development consumed the organization's attention. Exploratory research was deprioritized. The best researchers left. The organization optimized for the metrics it could see, sprint velocity, competitive position, product launches, while the capabilities that made it distinctive eroded beneath the surface.

The parallels to the engineering teams my essay describes are striking. Meta's organizational trajectory mirrors, at a larger scale, the same pattern Part 1 documents in individual developers: the visible metrics improved while the invisible capabilities declined. LeCun uses a memorable phrase for the industry-wide version of this dynamic: "everybody is digging the same trench." The herd behaviour he describes, where every company chases the same LLM scaling strategy because none can afford to fall behind, is the organizational equivalent of the variable-ratio reinforcement loop operating on individual engineers. The market rewards the behaviour. It feels productive. The capabilities that would allow course-correction are the very ones being silently consumed.

Perhaps the most useful insight for leaders reading my essay is LeCun's framing of what changes for humans in an AI-augmented world. "Humans go one level up in the abstraction hierarchy," he says. "Our job is to decide what to build." The AI handles the typing. The human is responsible for the thinking, the architecture, the judgment about what the system should do and why.

That framing is the Discipline Framework (see upcoming part 3) in a single sentence. The organizations that will thrive are not the ones that automate the most. They're the ones that structurally protect the human capacity to think at the level of abstraction that the AI cannot reach. LeCun sees this as a description of the present. This essay sees it as a prescription for surviving the future. The gap between the two is where most organizations will fail or succeed.

One final observation. LeCun left Meta to start AMI Labs because he concluded that the company's organizational structure could no longer support the kind of work that matters most: fundamental research into how intelligent systems actually work, not incremental improvements to the systems that already exist. He describes the decision to leave not as a rejection of Meta but as a recognition that certain kinds of work require a certain kind of environment. "The best way to get breakthrough research," he says, "is you hire the best people, you give them the means to succeed, and you get out of the way." That is a statement about discipline. Not the discipline of control, but the discipline of protection: maintaining the structural conditions that allow deep thinking, genuine understanding, and long-term capability development. It is exactly what the Discipline Framework proposes, applied to research instead of engineering. The principle is the same. Protect the conditions under which understanding grows. Everything else follows from that.

The full interview is worth your time. Watch it here on YouTube.

This post is a companion to "Discipline Is All You Need: Part 1 — The Lever and the Loop." Part 2, "When the Dashboard Lies," is forthcoming.