Not subscribed? Sign up to get it in your inbox every week.

Somebody cloned Adobe this fall.

On September 30, developer Brandon Thomas made the first commit to PhotoCraft, an open-source Photoshop built from scratch in Rust with Claude Opus 5.5. Two weeks later, the project had grown into clones of seven Adobe apps (Photoshop, Illustrator, Premiere, Lightroom, After Effects, InDesign, and Acrobat) consisting of 539,017 lines of code in 2,152 commits, with Claude listed as the author on 277 of them. To be clear, the project says it only studied the Adobe apps to mimic their behavior and look, and otherwise built them from public specs and observation.

The Adobe clones aren't great, yet. According to the project's roadmap, they're about 45% ready for real work, and Thomas himself has said they're "not anywhere close to ready." But he also thinks 99% parity will be "measured in months and not years," and he's far from alone. A public tracker now lists 81 rebuilt projects, including AI-built takes on Word, Excel, PowerPoint, Figma, and AutoCAD. In March, the maintainer of the Python library chardet asked Claude Code to rebuild it from an empty repo and then released it under a new license. By April, a service called MALUS was selling "clean room" clones of open-source projects. None of this has been tested in a court of law yet.

I think this all points to a bigger shift in where software gets its value.

Most software is deterministic: the same input produces the same output every time. That makes software reliable, but it also makes it easy to copy. If every click produces a predictable result, an agent can click on every button, observe the output, and write code until its version produces the same thing. PhotoCraft's apps even run an MCP server so the agent can click around and take screenshots of the clone while it works, and every change gets tested against a corpus of real Photoshop files.

In the past, reverse engineering software took years of specialist work, so it wasn't worth pursuing for most applications. Now it takes agent-hours, and the PhotoCraft roadmap estimates 2,300 to 4,100 of them before the apps reach full parity. While that's a lot of compute, it does put a price tag on it.

Of course, there are exceptions, and they tell you what's still defensible. It's easy to build a Facebook-ish app, but you can't copy the people already on Facebook, so network effects survive. This excellent teardown of the Adobe clones notes that many of the features that sell Adobe subscriptions today, like Generative Fill, run on Adobe's servers, so there's nothing on your computer to copy. The same piece notes that Camera Raw's lens and camera profiles took about twenty years of data collection, and that Pantone's color libraries are licensed. Those moats are made of networks, data, and models, and none of them live in the code you can watch.

So where does the value go? My bet is software that's reliable but non-deterministic. In other words, trained models.

The behavior of a model is encoded in its weights. The weights are produced by data and training runs, and they never ship to the customer, so you can't see them even if you observe every output the model produces.

You can get a rough copy, though. In February, Anthropic accused three Chinese labs (DeepSeek, Moonshot, and MiniMax) of running over 16 million exchanges with Claude through about 24,000 fraudulent accounts, in order to train their own models on Claude's outputs. When Anthropic released a new model in the middle of that campaign, MiniMax "pivoted within 24 hours, redirecting nearly half their traffic" to the new one.

You can copy a model, but the copy is always behind, and you have to start over every time the leader releases a new one. So the companies that train the best models and keep training them are going to win. OpenAI and Anthropic are the obvious examples at the largest scale, but plenty of smaller companies are already winning on a narrower axis.

Win on domain. In August, Harvey released Tenet, its first in-house model for legal work. It's a post-trained version of Moonshot's open-weight Kimi K3, using reinforcement learning graded against expert-designed rubrics. Harvey claims it completes almost twice as many held-out legal agent tasks as the base model. It also says it trained on a mix of synthetic, public, and expert-built data, with no customer data. So its edge comes from the expert rubrics and the simulated legal work environments. Anyone can download Kimi K3, but they don't get those.

Win on cost and speed. In March, Cursor released Composer 2, another Kimi-based model, which beat Claude Opus 4.6 on Terminal-Bench 2.0 (61.7 vs 58.0) at roughly 86% lower cost than Cursor's previous model. Crucially, Cursor can afford to run it across its entire editor, while competitors can't get it at all, since it isn't offered as an API. In a twist, both of these companies trained on a base from one of the labs Anthropic accused of distilling Claude. The copying runs all the way down.

Win on decisions. This is the category I find most interesting. In September, TypeSafe (founded by former OpenAI researcher Diogo Almeida) released Jev, a model that doesn't write anything. Instead, you send it the current state of your system and a set of typed questions (pick one of these options, score this against a rubric, answer yes or no), and it responds with decisions and a calibrated probability for each one. Your code can act when the probability is high enough and escalate to a person when it isn't. TypeSafe claims it runs in 70 to 500 milliseconds at $0.042 per million input tokens, with output free. According to InfoQ, nearly 13% of paid teams on Vercel's AI Gateway were using it within 24 hours of it being added. The speed claims come from TypeSafe so far, and outside testers report smaller but still large gains.

Decision models help answer the question of how you build a defensible product with AI. Anyone can call Jev, so the model alone protects nobody. What a competitor can't copy is the state you feed it: your customer history, your inventory, your policies, the context that makes the decision right for your business. If you also log the decisions and how they turned out, you build up the scored record I wrote about last week. It's the same loop as #153: rent the pen, own the judge.

On the buy side, a big chunk of your software spend is about to get cheaper, because the deterministic parts of it can be rebuilt by a competitor (or by you) faster than ever. Vendors whose products are mostly forms, workflows, and dashboards will feel that pressure first, and your negotiating position improves accordingly. The vendors that hold their pricing will be the ones with a network you can't leave, data you can't get elsewhere, or a model that keeps getting better on work like yours.

On the build side, it's worth asking the same question about your own product. Your features can be watched, so a well-funded competitor with a few thousand agent-hours can probably match them. What they can't watch is your customer data, your decision history, and the models trained on them, so that's where I'd put the next dollar of investment.

Grab the top five items in your software budget and sort them into buckets:

Do the same for your own product. For anything in the first bucket, find the decisions it makes or supports, and check whether you're logging the state and the outcome every time it does.

Code got cheap to copy. The models that keep learning are the expensive part now.

Till next time,
Chris