A Photoshop clone written from scratch in Rust is the headline. The real story is the cadence behind it, and what that cadence does to the economics of every tool you currently rent.
The number that matters in Latent Space's writeup of ArtCraft is not Photoshop. It is seventeen. According to Latent Space's report from the AI Engineer Code Summit, the young company has already released 17 open source versions of popular software tools, starting with PhotoCraft, a Photoshop alternative built from scratch in Rust. One clone is a demo. Seventeen is a production line.
That distinction is the whole argument of this piece. Most coverage of AI code generation still treats each impressive rewrite as a magic trick: look, the model built a thing. But a repeatable pipeline that turns commercial software into community-owned software changes the question. It stops being "can AI write Photoshop?" and becomes "which tools are cheap enough to rewrite next, and who captures the value once they are?"
Meanwhile, the same week produced three other signals that rhyme with ArtCraft in ways that are easy to miss when you read them separately. Roblox's CEO described a world where anyone can now literally make a game. A robotics startup raised $75 million on the premise that human help becomes training data. And a small open-source plugin quietly made it trivial to run an agent on a local model. Put them side by side and a pattern emerges: the cost of making things is collapsing, and value is migrating to whoever aggregates, maintains, and is trusted with what gets made. For anyone who runs agents day to day, that migration is already reshaping which tools are worth paying for.
Seventeen rewrites is a pipeline, and pipelines change the economics
Start with what the source actually says. ArtCraft used AI to rewrite Photoshop as open source software, and CEO Brandon Thomas framed the ambition as bigger than any single product: in Latent Space's words, "to take back the internet!" The summit's day-two theme, per the same report, was increasing ambition using AI, and ArtCraft was presented as one of the most ambitious projects of the era.
The pack does not tell us how feature-complete PhotoCraft is, how long it took, or how many humans touched it. Those gaps matter, and I will not paper over them. But the count is itself evidence of something. A single heroic rewrite proves a model can be pushed through a hard problem once. Seventeen releases implies a process: a way of scoping a target tool, generating the code, testing it, and shipping it that can be run again.
That is the structural shift. Proprietary software has historically been protected by a simple cost asymmetry. Building a credible image editor took years of specialist engineering, and the incumbent had already paid that cost. Licensing fees were, in effect, rent on accumulated labor that nobody else could cheaply reproduce.
If an agent-assisted team can reproduce a usable core of that labor repeatedly, the asymmetry erodes. Not entirely. Not for every feature. But enough that the rent starts to look negotiable.
A few implications follow, all of them analysis rather than reported fact:
- The target list is long. Any tool whose value sits mostly in well-understood features, rather than proprietary data or network effects, becomes a candidate.
- Speed compounds. Each rewrite likely improves the pipeline for the next one, the way a factory gets faster per unit.
- The choice of Rust is a signal. Building from scratch in a modern systems language suggests the goal is not a quick wrapper but a foundation others can extend.
The instinct to read this as "AI cloned Photoshop" undersells it. The better reading is that ArtCraft is testing whether software licensing can survive a world where reimplementation is a repeatable job.
Rewrites win from the low end, exactly as Disruption Theory predicts
The obvious objection: a from-scratch clone will not match decades of refinement in the original. That is probably true, and it is also beside the point. Disruption Theory says low-end entrants do not need to beat incumbents on their most demanding customers. They need to be good enough for the customers the incumbent overserves.
Think about who actually opens a professional image editor. A large share of users touch a small fraction of the features: crop, layers, masks, color adjustment, export. They pay for the whole cathedral to use the nave. An open-source rewrite that nails the common 20 percent and costs nothing is a credible offer to that group, even if it never touches the long tail of specialist tools.
That pattern resembles every prior wave of open source displacing commercial software, with one difference. Previous open-source alternatives were bottlenecked by volunteer hours. The good ones took a decade to mature because the labor was donated, intermittent, and hard to coordinate. If AI code generation compresses that maturation curve, the low-end foothold could turn into an upmarket climb much faster than incumbents are used to.
Meanwhile, the incumbent's natural defense, adding more advanced features for power users, accelerates the problem. Each premium feature widens the gap between what the median user needs and what the median user pays for. That gap is precisely the space a rewrite factory occupies.
In Wardley Mapping terms, ArtCraft is betting that a category of desktop creative tools has drifted from "product" toward "commodity" on the evolution axis. Once a component is a commodity, competing on its features is a losing game. The value moves to whatever sits adjacent to it.
There are honest limits to this argument:
- File-format compatibility is a moat that code generation does not automatically dissolve. A clone that cannot round-trip industry files reliably is a toy for professionals.
- Ecosystems of plugins, tutorials, and trained staff are sticky in ways that source code is not.
- Quality claims are unverified. Nothing in the reporting establishes how PhotoCraft performs against Photoshop on real work.
So the disruption case is about direction, not arrival. The low end is open. Whether the climb upmarket happens depends on factors the pack cannot yet measure.
When making software is abundant, the aggregator takes the margin
Here is where a seemingly unrelated interview becomes the key to the ArtCraft story. In his conversation with Roblox CEO David Baszucki, Ben Thompson observes that Roblox was a company predicated on the concept that anyone can make a game, "but now, for the first time, literally anyone can make a game." Thompson's conclusion: that means "a world of abundance," and in that world there is an opportunity to be an Aggregator, which Roblox is aggressively pursuing.
Swap "game" for "software tool" and you have the ArtCraft thesis restated from the platform side. If anyone can rewrite an image editor, a video editor, or a note-taking app, then the supply of software stops being scarce. Aggregation Theory is blunt about what happens next: platforms win by aggregating demand and commoditizing supply. The one that owns the user relationship wins.
That reframes Brandon Thomas's "take back the internet" ambition in a more interesting way than the slogan suggests. A company that ships 17 free tools is not primarily selling software. It is assembling a catalog, and catalogs are how aggregators begin. The pack does not say what ArtCraft's business model is, so treat this as a reading of the strategy rather than a description of it. But the shape is recognizable: give away the commodity layer, own the place users come to get it.
This is also classic Commoditize Your Complement logic. If your margin lives in distribution, community, hosting, or support, you benefit when the tools themselves are free. Every proprietary tool ArtCraft rewrites shrinks a competitor's margin and enlarges the pool of things flowing through ArtCraft's front door.
Meanwhile, Roblox shows the incumbent version of the same move. It did not fight the arrival of AI game creation. It leaned into the abundance and positioned itself as the place where that abundance gets discovered and played. Software vendors facing a rewrite factory have the same choice in principle: become the aggregator, or become one of the aggregated.
For readers who pick agent tooling, this is the practical lens. When you weigh openclaw alternatives or any paid tool against a free rewrite, the question is less "which code is better" and more "who owns the relationship, and what do they do with it once the code is free?"
The model layer is commoditizing too, which pushes value into the harness
Meanwhile, a small release from Simon Willison shows the same collapse happening one layer down. The new datasette-llm release adds support for OpenAI-compatible endpoints, so Datasette can use an LLM hosted anywhere that speaks that protocol. Willison notes that if you run local models with LM Studio, Ollama, or llama.cpp (which just hit v1.0.0), you get a local server running a Datasette Agent instance powered by your local model.
This is easy to dismiss as plumbing. It is not. It means the model behind an agent is increasingly swappable, including for a model running on your own machine at no per-call cost. When the model is a commodity, The Harness Hypothesis comes into focus: the value in AI is not in the model, it is in the harness that connects the model to the world.
Put ArtCraft and datasette-llm next to each other and they describe the same squeeze from two directions:
- From above: AI codegen makes the application layer cheap to reproduce, so proprietary apps lose pricing power.
- From below: open protocols and local models make the intelligence layer cheap to swap, so model vendors lose lock-in.
What survives the squeeze is the connective tissue. The workflows, integrations, permissions, and trusted defaults that make a pile of commodity parts behave like a product. That is exactly the layer most agent users actually live in. When you configure an agent, you are rarely choosing a model in isolation. You are choosing a harness and a set of skills that sit between the model and your files, browser, and accounts.
This also reframes the cost question many readers ask, like openclaw cost per month. If the tools are free and the model can run locally, your spend migrates toward hosting, orchestration, and whatever the harness charges for reliability and convenience. The bill does not disappear. It moves.
There is a risk dimension worth flagging here as analysis. A world of 17 freely available rewrites, each wired to a swappable model, is a world with many more components of uncertain provenance. Every free tool a user installs and connects to an agent is a fresh trust decision. The Shadow Agent Problem, individuals installing agents and tools without IT review, gets worse as the supply of free, capable software grows. Abundance is good for users and hard on governance.
Software labor shifts from writing code to tutoring the machine that writes it
Now the labor question, which the ArtCraft story raises but cannot answer by itself. If agents rewrite the code, what do the humans do? An answer comes from an unlikely corner: robotics.
Tutor Intelligence is betting that robots can learn on the job. Remote human tutors help the robots perform useful work, and those interventions become training data, supplemented by a dedicated 100-robot research fleet and a real-to-sim-to-real loop called TRACE. The company announced a $75 million Series B led by Andreessen Horowitz on November 2, 2026, bringing total funding to $117 million.
The parallel to software is direct. In the Tutor Intelligence model, human labor does not vanish; it changes shape. People stop doing the task and start correcting the system that does the task, and every correction makes the system less dependent on them. The pattern resembles what an AI rewrite pipeline likely demands: engineers who specify targets, catch failures, and feed fixes back in, rather than engineers who type every line.
The Sequence names the crucial uncertainty plainly: "The open question is whether assistance falls as useful work grows." That is the single most important question for software labor too. If the human tutoring load per rewrite drops with each of ArtCraft's releases, then the cost of the next clone keeps falling and the moat erodes faster. If the load stays flat, or rises as targets get harder, the factory hits a ceiling.
The robotics piece also supplies a caution that applies to codegen enthusiasm. It asks whether simulated results, such as 99.8% success on a Lego-picking task, hold up across objects, sites, and long operating periods. Translate that to software: a rewrite that looks complete in a demo may fail on the messy long tail of real files, real hardware, and real user habits.
Mapping this onto The Autonomy Spectrum helps. Neither robots nor rewrite pipelines are at full autonomy. Both sit somewhere in the middle, with humans in the loop doing high-value correction. The economic bet in both cases is that the human share shrinks over time. The failure mode, as the framework warns, is deploying as if you are further along the spectrum than you are.
The next molt will be maintenance and security, not more clones
If ArtCraft's pipeline works, what comes next is predictable, because open-source projects tend to follow The Molt Cycle: rapid growth, then a security crisis, then hardening, then enterprise adoption, then commoditization, then the next molt. ArtCraft, with 17 releases and an "rebuild everything" ambition reported by Latent Space, is plainly in the rapid-growth phase.
The growth phase rewards shipping. The phases after it reward something different. Writing a tool once is now cheaper than it has ever been. Maintaining seventeen tools, patching vulnerabilities, tracking file-format changes, and answering user bug reports is a recurring cost that code generation reduces but does not erase. Datasette's own history offers a small illustration: Willison's datasette-vega release was the plugin's first since July 2018, updated for compatibility with newer Datasette versions. Software does not stay working on its own. Someone has to come back.
So the real moat in a post-rewrite world may look like this:
- Maintenance cadence. Whoever reliably keeps free tools current earns trust that one-off clones never will.
- Security posture. AI-generated code at volume is a large attack surface. The first serious vulnerability in a popular rewrite will test whether the community can harden fast.
- Compatibility guarantees. Users will pay, in money or loyalty, for assurance their files and workflows keep working.
- Distribution. As the Roblox interview suggests, the aggregator that owns the user relationship captures the margin.
Notice that none of these are "writing the code." That is the core of the thesis. AI codegen is not abolishing the software business. It is relocating it, away from authorship and toward stewardship.
For incumbents, the uncomfortable implication is that their strongest assets may no longer be their codebases. They are their customer relationships, their file formats, and their reputations for reliability. Those are defensible. Feature lists are increasingly not.
For agent users, the practical advice follows. When a free rewrite of a tool you pay for appears, evaluate it on maintenance and trust, not features. Ask who will patch it, who will keep it compatible, and what it is allowed to touch when your agent calls it. The cheapest tool to install is not always the cheapest tool to own.
/Figures
- 2026-11-02Tutor Intelligence raises $75M Series B
Human tutor interventions become robot training data; total funding $117M.
- 2026-11-12Stratechery interviews Roblox CEO David Baszucki
"Literally anyone can make a game" frames a world of abundance and aggregation.
- 2026-11-13Latent Space reports on ArtCraft
17 open-source rewrites of popular tools, starting with PhotoCraft in Rust.
- 2026-11-14datasette-llm adds OpenAI-compatible endpoints
Local models via LM Studio, Ollama, or llama.cpp can power a Datasette Agent.
/Sources
/Key Takeaways
- ArtCraft has released 17 open-source versions of popular tools, starting with PhotoCraft, a from-scratch Photoshop alternative in Rust. The count signals a repeatable pipeline, not a one-off stunt.
- Low-end rewrites threaten incumbents the way Disruption Theory predicts: by serving users who only need a fraction of the features, for free.
- In a world of software abundance, Aggregation Theory says value moves to whoever owns the user relationship, the same logic Roblox is applying to AI-made games.
- Swappable and local models (as in datasette-llm's new OpenAI-compatible endpoint support) commoditize the intelligence layer too, pushing value into the harness.
- Software labor is shifting from writing code toward correcting the systems that write it; the key unknown, as with Tutor Intelligence's robots, is whether human assistance falls as output grows.
- The durable moats after a rewrite wave are maintenance, security, compatibility, and distribution. Judge free clones on stewardship, not features.

