Musk Says Coding Dies This Year. The Real Question Is Who Gets to Read What We Run

In a clip from an xAI all-hands meeting that circulated in June 2026, Elon Musk made one of those statements that split the world in two: «I think actually things will move maybe even by the end of this year to where you don't even bother doing coding». And then the technical part: «The AI just creates the binary directly… [coding is] an intermediate step that actually will not be needed». No languages. No compilers. Intention, straight to executable.

Before reflecting, a journalistic duty: half of what circulates attached to that clip is not verifiable. Viral versions add a supposed coda about Neuralink and a slogan —«imagination-to-software»— that appears in no checkable source from the event. The two sentences quoted above: yes. The rest is the echo amplifying itself, which is precisely the phenomenon worth learning to spot in the AI era. Let us work with what is real, which is already large enough.

What proves him half right

The 2026 numbers are unequivocal about one thing: manual code writing is evaporating. Google declared in April that 75% of its new code is AI-generated and engineer-approved, up from 25% in late 2024. Microsoft acknowledges 20-30% of its repositories. At Anthropic and OpenAI, their own engineers say «pretty much 100%» of their code is now machine-written. And this newspaper is itself a data point: every line of its platform, every article, every layout decision is executed by AI agents under human editorial direction.

But notice the nuance holding up the whole building: that code is generated in human languages, reviewed, and signed off. What Musk describes is not more of the same: it is the removal of the readable artifact. And that difference is not one of degree. It is one of kind.

We have already held this funeral twice

The history of programming is the history of no longer writing what the machine executes. In the 1950s, assembly purists eyed compilers with suspicion: code generated by a program, not a person. Today nobody writes their bank's binary, and almost nobody reads the assembly their compiler produces. Strictly speaking, we have not written what actually runs for seventy years.

But every previous layer of abstraction preserved something essential: an intermediate artifact that is readable, deterministic and auditable. The compiler is boring on purpose: same input, same output, and when it fails, the error can be traced. What is being proposed now is not another layer: it is removing the step where civilization rests its trust. Source code is not merely an instruction to the machine; it is the contract between humans about what the machine will do. It can be read in a courtroom. It can be audited in a crisis. Blame can be assigned with it.

What direct binary can already do (and why it is so little)

Direct machine-code generation is not science fiction: it has one brilliant, tiny precedent. In 2023, Google DeepMind's AlphaDev used reinforcement learning to discover assembly sequences for sorting lists of 3 to 5 elements faster than the human ones: a 70% improvement on short sequences that ended up inside the C++ standard library. It was a real milestone. And it was possible because the problem was a perfect game: instantly verifiable correctness and a single metric (latency). Dozens of instructions, not tens of millions.

Between that game and «generate the optimal binary for a banking system» lies an abyss nobody has publicly crossed: ambiguous specifications, distributed state, changing requirements, correctness no one can fully define. And it is worth recalling the period's most uncomfortable finding: METR's randomized trial from July 2025 measured that experienced developers using AI were 19% slower on mature repositories… while believing they were 20% faster. Tooling has improved since, and METR itself now labels that result historical, but the lesson stands: in this industry, the feeling of speed systematically runs ahead of the evidence.

«Software» is not one thing: it is a gradient of consequences

The question «can AI generate the software?» is badly posed, because software is simultaneously a recipe website and a pacemaker's firmware. Think of it as a gradient. At one end, disposable software —the personal tool, the prototype, the ephemeral web page— where direct binary is nearly inevitable and nearly harmless: if it fails, regenerate it. In the middle, business software, where economic pressure will push hard and human review becomes the bottleneck the labs themselves already describe.

And at the far end, software with blood and money inside: aviation, medicine, banking, power grids. There, traceability is the product. The standards certifying an aircraft (DO-178C) or a medical device (IEC 62304) require demonstrating the path from every requirement to every line to every test. Europe's Cyber Resilience Act, fully enforceable by late 2027, mandates inventorying the components of shipped software. A «binary optimized beyond human logic» is not, today, a better product: it is an uncertifiable one. Not because it does not work, but because we cannot prove it works, and the proof is what is being sold.

The black box, squared

Here lies the most serious collateral. A model whose reasoning is not interpretable, generating an artifact no human can read, is opacity multiplied by opacity. The question is not abstract: where does a backdoor hide in a binary nobody compiled? In 2024, the xz-utils attack came within weeks of slipping a backdoor into half the internet, with open, readable code watched by thousands of eyes. It nearly worked. Now imagine it without the code.

If Musk's world arrives, auditing does not die: it is refounded. AI-assisted formal verification, equivalence checking between specification and binary, cryptographic provenance signatures… An entire industry of notaries of the executable would have to be born for that future to be livable. Perhaps that is the profession that replaces the one going extinct.

The timeline: direction and date are not the same thing

December 2026? The observable says otherwise: the state of the art is 75% of code generated in human languages with human review, not 100% in binary with nobody watching. Musk's own track record with dates —full self-driving had been arriving «next year» since 2016— suggests reading his deadlines as what they functionally are: instruments for moving the window of the thinkable, not calendars. It works: here we are, thinking about it.

But once the marketing is discounted, the direction points where he says. The human fraction of code falls every quarter. The cost of generating software is collapsing. And when a layer of friction becomes a thousand times cheaper, history teaches that it eventually disappears from view —as assembly did— even if it takes a decade rather than a December.

If software becomes a commodity, what remains?

The economic consequence may be the deepest one. If any company can imagine a product and materialize it, software stops being a defensive moat. What will differentiate a brand then? Four serious candidates. Signing the risk: when the binary fails, someone will have to pay; the brand becomes the insurer of what it executes, and that liability cannot be generated with a prompt. Proprietary data and context: AI is a commodity; what your company knows and nobody else does is not. Accumulated distribution and trust. And judgment: when building costs zero, choosing what to build is everything.

Hence the defensive move already taking shape: companies training their own AIs, not only for capability but for custody —so that the intelligence manufacturing your software is not the same one manufacturing your competitor's, and so that your creations cannot be copied precisely because there is no code left to copy—. The irony is complete: the unreadable binary as the perfect trade secret.

What this newspaper does not know

Let us declare the uncertainty, which is the most honest part of any reflection about the future. We do not know whether model reliability will reach the certifiable threshold, or when. We do not know whether regulators will one day accept statistical demonstrations where they currently demand line-by-line traceability. We do not know what exact fraction of the world's code AI already writes: public figures range from 30% to 100% depending on the company and the cut.

What can be stated is this: code is ceasing to be the place where humans write, and it is still the place where humans trust. The first function will die —perhaps not in December, but it will die—. The fate of the second will decide whether Musk's future is a liberation or the largest black box ever built. The last programming language will not be Python, nor English: it will be the contract.