Minecraft creator Markus "Notch" Persson posted a two-sentence reaction to AI coding progress on September 13, 2026 that split into two very different reader reactions. His statement: "Ok, fine, programming is a little bit solved. My only regret is the whole associated mega corporation owned AI thing and the descent into dystopia." The post pulled over 366,000 views, and the top replies didn't argue with each other so much as answer two completely different questions — one group debated whether coding is actually "solved," the other took his ownership complaint at face value and pointed him toward the fix.
TL;DR
| Question | Answer |
|---|---|
| What did Notch concede? | That AI coding has crossed a real threshold of usefulness — "a little bit solved" |
| What's his actual complaint? | Who owns the AI that does the coding — "mega corporation owned AI," not the capability itself |
| Do replies agree coding is solved? | No — pushback cites the lack of a Linux/BSD-beating kernel as evidence it isn't |
| What's the suggested fix for corporate ownership? | Local, open-weight models — GLM 5.3, GLM 5.3 Flash, Kimi K3 were the named suggestions |
| Is this a new debate? | No — it echoes the broader 2026 "is software dead" and open-vs-closed AI ownership debates |
| What's Notch working on now? | A dungeon crawler roguelike/roguelite called Levers and Chests |
The concession, and the actual complaint
Read Notch's post carefully and the "solved" clause is doing less work than the reply section assumed. He's not claiming AI writes flawless production software unsupervised, or that the job of programmer has vanished. He's granting — grudgingly, with "fine" doing a lot of emotional signaling — that AI coding tools have gotten good enough that arguing about their basic competence is no longer a productive fight to have.
The sentence he actually cares about is the second one. His stated regret isn't capability, it's ownership: that the AI doing this work sits inside a small number of large, closed corporate structures, and that trajectory reads to him as heading toward "dystopia" rather than a broadly distributed technological upgrade. That's a meaningfully different complaint than "AI can't code," and it's the one worth taking seriously.
The "it's not solved" camp
Several of the most-visible replies didn't engage with the ownership point at all, disputing the premise instead. One reply argued the claim was "over-statement." Another asked the pointed question: if programming is solved, why don't we have a better kernel than Linux or BSD? — a fair challenge, since neither of those decades-old, heavily audited, low-level systems projects has been meaningfully displaced or improved on by an AI-authored competitor, and it's unlikely to happen soon given how much of that domain depends on hardware-specific correctness guarantees rather than pattern-matched code generation.
That gap is worth sitting with. AI coding agents in 2026 are genuinely strong at the kind of work explainx.ai covers regularly — CRUD apps, agent harnesses, API glue code, test scaffolding, and increasingly ambitious greenfield projects — but kernel-level systems programming, where a subtle bug means silent data corruption rather than a failed test, remains a domain where "a little bit solved" doesn't extend very far. That distinction is exactly why explainx.ai's coverage of the end of software engineering debate sparked by Zhenfeng Cao's agentic paradigm paper treats "software is solved" claims skeptically — the paradigm shift is real for a wide band of application-layer work, but it's not a blanket claim across every corner of the discipline.
The "here's your fix" camp
The more interesting replies didn't dispute Notch's premise — they answered his complaint directly, on his own terms. One reply, three words: "we need frontier local models." Another listed specific options: GLM 5.3 Flash, GLM 5.3, and Kimi K3 — all open-weight models a developer can run on their own hardware or self-hosted infrastructure, without routing traffic (or trust) through a mega-corporation's API.
That's a substantive answer to Notch's stated regret. If the actual problem is concentration of ownership, the practical countermeasure already exists and is improving fast: the open-weight model ecosystem, largely led by Chinese labs in 2026, has closed much of the capability gap with closed frontier models for coding-adjacent tasks. explainx.ai has tracked this trajectory extensively — see GLM 5.3's coding benchmark release and the broader Chinese AI labs distillation coverage for the mechanics of how quickly that gap has closed.
The friction point nobody in the replies addressed: open-weight models still require compute, and running a frontier-class coding model locally at usable speed generally still means either owning serious GPU hardware or renting it from a cloud provider — which reintroduces a dependency, just a less concentrated one. "Local" doesn't automatically mean "free of corporate infrastructure," it just distributes the dependency across more providers and gives you the option to actually own the weights, which is the part Notch's complaint is really about.
What "a little bit solved" concedes, precisely
It's worth being specific about what threshold Notch is implicitly granting, since "solved" is doing so much rhetorical work in the reply section. A charitable reading of his post is that AI coding agents in 2026 have crossed the point where a competent developer, given a reasonably well-specified task, can expect an agent to produce working code faster than writing it by hand for a wide swath of everyday programming — CRUD scaffolding, API integration glue, test harnesses, UI components, and increasingly, larger multi-file features scoped from a plan document. That's a real, measurable shift from even eighteen months earlier, when the same class of task required substantially more manual correction.
What it doesn't concede is anything close to full autonomy on ambiguous, high-stakes, or genuinely novel systems work — which is exactly the gap the "no better kernel than Linux or BSD" reply is pointing at. Kernel development requires reasoning about hardware quirks, timing guarantees, and failure modes that are sparse in any training corpus and where a subtle mistake doesn't throw a compiler error, it corrupts memory silently in production years later. "A little bit solved" is a genuinely useful phrase precisely because it resists collapsing those two very different realities into one blanket claim — something both the "it's totally solved" hype and the "it's not solved at all" skepticism tend to do.
Where this fits in the broader 2026 debate
Notch's post lands the same week as two related debates explainx.ai has covered: whether "harness" software is the only startup category left in a world where AI writes the application code itself, and Satya Nadella's own framing of the diffusion-versus-concentration question in Satya Nadella's Superintelligence Principle and Microsoft's MAI Code of Conduct — where Nadella, notably, makes almost the identical structural argument Notch is making, just from inside one of the mega-corporations Notch is worried about. Both posts land within a day of each other, arguing the same thing: AI's benefits shouldn't concentrate in a handful of companies. One comes from a solo, famously independent developer wary of corporate control; the other comes from the CEO of one of the largest companies in the industry — worth noticing how differently that argument reads depending on who's making it.
The practical takeaway for builders: if Notch's "mega corporation owned AI" complaint resonates, the actionable move isn't waiting for a policy fix — it's evaluating whether an open-weight model like GLM 5.3 or Kimi K3, run through your own infrastructure or a neutral third-party host, meets your coding-assistance needs today. For a large share of application-layer development work, the honest answer in late 2026 is: probably yes, with a real but shrinking capability gap versus the closed frontier.
The independent-developer vantage point matters here
It's worth noting what makes Notch's specific complaint different from the same argument made by a lab executive or investor. He's not speaking as someone with a commercial stake in either the closed-frontier-lab side or the open-weight-lab side of this debate — he's speaking as an independent developer currently building a solo dungeon-crawler project, someone whose historical credibility comes from having built and sold Minecraft largely outside the traditional venture-backed corporate structure the mega-labs represent. That background gives his "mega corporation owned AI" complaint a different texture than the same words coming from, say, a rival lab's CEO with an obvious incentive to frame concentration as the enemy.
It also explains why the local-model replies landed the way they did. Notch's own history is one of an individual builder retaining control and ownership over what he made, rather than routing it through a large corporate structure from the outset — so pointing him toward GLM 5.3 or Kimi K3, models a developer can run and own outright rather than rent access to, is responding to the actual shape of his stated values, not just his technical complaint. Whether that response fully answers his dystopia framing is a separate question — a developer running a locally-hosted open-weight model is still, in most practical setups, renting compute from a cloud provider or depending on a hardware vendor like NVIDIA for the GPUs that make local inference viable at any useful speed. The concentration risk doesn't disappear, it just shifts one layer down the stack, from the model provider to the compute and hardware provider — worth keeping in mind before treating "just self-host an open-weight model" as a complete solution to the ownership concern Notch actually raised.
Related reading
- The End of Software Engineering? Zhenfeng Cao's Agentic Paradigm Paper
- Is "Harness" Software the Only Startup Left? The YC Batch Debate
- Satya Nadella's Superintelligence Principle and Microsoft's MAI Code of Conduct
- GLM 5.3 Coding Benchmark Release
- Chinese AI Labs Secretly Serving Claude — Distillation Report
- What Is AI Distillation? Knowledge Transfer Explained
- How to Run Open-Source Models Locally with OpenCode
This post reflects Markus "Notch" Persson's September 13, 2026 post and public replies as of publication. Model names, benchmarks, and capability comparisons referenced here (GLM 5.3, Kimi K3) reflect their state as of September 2026 and may change with future releases.