Martin Uecker

Computer Science and Biomedical Engineering
Graz University of Technology

Institute of Biomedical Imaging, TU Graz: https://ibi.tugraz.at/

Last Notes

2026-08-15 18:49:39 CEST

- reply

Your link seems to say just the opposite. (BTW: I do research on diffusion models.) The idea that diffusion models reproduce training data is based on the (incorrect) idea that they would exactly learn the score, which is generally not true. Although by the same argument one could argue that an LLM trained to predict the next token also learns reproducing training data.

2026-08-15 17:31:56 CEST

- reply

Unclear to me. But I am still curious why you think diffusion models are different?

2026-08-15 17:15:01 CEST

- reply

I don't think that this "neither intentional nor wanted" by design is similar to a human being forbidden by some policy. I agree though that LLMs do not (usually) simply reproduce their training data and that they are trained on some data does not necessarily mean their output plagiarizes it. This does not imply though that training them on copyrighted works did not infringe on the rights of the original authors. This is far from clear at this point.

2026-08-15 15:08:36 CEST

- reply

Nobody claims that they record all training data verbatim, just that this is possible, which is a known fact. (BTW: Shannon's theorem would apply to any kind of machine learning model.)

2026-08-15 15:01:28 CEST

- reply

Verbatim memorization is empirically known from LLMs. And why would transformers not be able to do this?

2026-05-10 11:53:16 CEST

Debian will require reproducible builds!This is a huge milestone.
https://lists.debian.org/debian-devel-announce/2026/05/msg00001.html
#Debian #reproduciblebuilds

2026-05-02 15:09:59 CEST

- reply

The Chinese room argument is fairly unconvincing though. It would not be the person mindlessly following a rule book who can speak Chinese, but the system consisting of the rule book (software) and the processor (person in the room). Nobody would argue that a CPU / GPU is conscious either, but this does not logically preclude a computer with suitable software from being conscious. (no, I do not believe modern Chat bots are conscious, but they also do not pass a real Turing test)

2025-03-23 07:51:22 CET

Sie hätten Hitler stoppen können
https://zeitung.faz.net/fas/politik/2025-03-23/d974bf1f6eef25ff6c6c0e6a23436a89/?GEPC=s10

2024-08-11 10:07:40 CEST

- reply

It is because the command line is stable and does not change all the time. GUIs also generally lack consistency and modularity. It could be different in theory. i don't know why GUIs arre so bad and need to be change all time.