Adobe clones in an editorial test: what c't 3003 found and what Cisco plans

In short: On October 30, heise reported on a test in which c't 3003 produced a video intro with AI-built rebuilds of Premiere, After Effects and Photoshop. The editors consider the result usable and name the gaps as well. Cisco's reported plan to have AI rewrite legacy code is, by contrast, so far only a target without a result.

In October we already counted how quickly such rebuilds come into being. What is new this week is that a trade editorial team used one to produce a video and published the weaknesses. That is a better source than development-time figures, and the numbers below come from their article.

What c't 3003 built

According to the heise article of October 30, the editors rebuilt a video intro with motion tracking, masks and effects in FilmCraft (a replacement for Premiere) and EffectCraft (a replacement for After Effects), largely one to one. The thumbnail was made in PhotoCraft, the replacement for Photoshop. In the article the programs run under the collective name ArtCraft. The author's verdict: extremely impressive and usable overall, PhotoCraft already fully usable for smaller jobs such as thumbnails, but Premiere still runs a bit more smoothly than FilmCraft.

What was tested matters here: an intro, with no feature film and no teamwork. For production companies the test says nothing.

The measured gaps

The article names concrete weaknesses:

Rendering one minute of 4K footage to H.264 on a MacBook M1 Max, Premiere was roughly a third faster than FilmCraft with hardware encoding and twice as fast with software encoding.
EffectCraft was practically unusable under Zorin OS: the machine froze after one to two minutes every time.
In FilmCraft, in and out points cannot be dragged with the mouse, and effect settings cannot be copied between clips.
In PhotoCraft, masks are lost when layers are copied between documents.
FilmCraft's GitHub page lists, per heise, 86 percent of planned features as implemented and 56 percent as fit for real work (vendor statement, basis not given).

The last figure is not a measurement by c't but by the project itself. That it states at all how much of it is fit for real work makes the state easier to judge.

The glue: no Dynamic Link

What makes Adobe's suites valuable is how the programs work together. The clones have no counterpart to Dynamic Link. The c't author rendered the clips one by one as ProRes 422, forked FilmCraft and used the AI coding tool Codex (about two prompts, by his account) to add an EffectCraft button for editing clips directly.

Legally open

The developer says the tools are a clean-room reimplementation in Rust, built without decompiling; heise notes that Adobe is unlikely to be happy about it. The law scholar Andres Guadamuz (University of Sussex) told Al Jazeera on October 14, in a report on ArtCraft and similar rebuilds, that it is unknown whether the claimed clean-room process would hold up under scrutiny. That article predates the heise test. We are not aware of a lawsuit or a statement from Adobe, and we found no court ruling.

In software, clean room means: one group describes what the original does, and a second group writes the rebuild from that description alone, without having seen the other code.

Cisco: a reported goal, not a result

Fortune reported on October 30 that Cisco president Jeetu Patel gave Padmanabhan Krishnamurthy, a 27-year-old founder and engineer, unlimited AI tokens to rewrite the legacy code and told him he can "go nuts". The figure named is 98 percent of the code in 24 months. Krishnamurthy gave himself four weeks to show that a two-person team can do it. Tom's Hardware reports it too (we read only its headline). Scope, the model used, verification against the behavior of the old code, and cost are unknown. A target is not a measurement, so we present it as a reported announcement. There are no results yet.

What the test does not show

An intro with tracking and masks is a good functional test, but not a stress test. Questions that production companies depend on remain open: How does the program open project files that someone else created? What happens after a crash to the half hour of work before it? How do updates arrive, and who checks them? The article says nothing on this, and we do not assume an answer.

What this means for us

This is our assessment; nobody measured it. A rebuild pays off where an error costs little: a thumbnail, an internal tool with two users. For postings, invoices and stock movements in WinLine or in the TMS, no clone replaces the tested original for us, and uploading that data to hosted models remains ruled out. Each such tool needs a responsible person who maintains the connections to neighboring systems.

Five questions help before a rebuild goes into use: What does an error cost? Can the data be taken out again without loss? Is there a way back to the original? Who maintains the interfaces? And is the rebuild's license stated explicitly in the repository? If one of these answers is bad, the original stays in use. For a pilot we would fix three things up front: a period of two weeks, one responsible person and a fallback rule that applies at the first data incident. Without these three, a test stays an impression, and impressions are exactly what the clones deliver at the moment.