Karsten Konrad 🇪🇺
@zabong69.bsky.social
Armchair Data Scientist and nerd. PhD in automated theorem proving just before AI became a subfield of Linear Algebra. Machine Learning, statistics, data visualisation, biology, recreational maths, computation, old pocket calculators, and video games
November 29, 2025 at 4:14 PM
November 27, 2025 at 10:11 PM
How I react when someone shows me a "normal distribution" in data where negative values are impossible.
November 27, 2025 at 3:33 PM
How I react when someone shows me a "normal distribution" in data where negative values are impossible.
Where I come from - Germany -, it's a Bleistiftanspitzer (Lead-Pen-Sharpener). We also have, of course, a manual for how to use it, comparing it to various other techniques available through history.
November 23, 2025 at 3:08 PM
Where I come from - Germany -, it's a Bleistiftanspitzer (Lead-Pen-Sharpener). We also have, of course, a manual for how to use it, comparing it to various other techniques available through history.
Thanks! Tanh hits the spot with only 6 nodes - loss only improves marginally with more nodes. Damn, now I have to experiment with activations whenever I uses NNs ;)
I don't know RBU... not a standard loss in Pytorch... do you have a link for me?
I don't know RBU... not a standard loss in Pytorch... do you have a link for me?
November 18, 2025 at 7:50 AM
Thanks! Tanh hits the spot with only 6 nodes - loss only improves marginally with more nodes. Damn, now I have to experiment with activations whenever I uses NNs ;)
I don't know RBU... not a standard loss in Pytorch... do you have a link for me?
I don't know RBU... not a standard loss in Pytorch... do you have a link for me?
Neural Networks: While the poor little thing below has only six nodes, I did give it 1000 samples and 6000 epochs and a ReLU activation function. I would have expected at least an uptick on the right for that amount of computation.
November 17, 2025 at 4:11 PM
Neural Networks: While the poor little thing below has only six nodes, I did give it 1000 samples and 6000 epochs and a ReLU activation function. I would have expected at least an uptick on the right for that amount of computation.
UMAP is becoming my go-to algorithm for data preparation now that I know that I can use it semi-supervised to separate classes better. Also: LLM embeddings DO look fascinating (below the visualization of 30.000 spare part descriptions).
November 4, 2025 at 9:02 PM
UMAP is becoming my go-to algorithm for data preparation now that I know that I can use it semi-supervised to separate classes better. Also: LLM embeddings DO look fascinating (below the visualization of 30.000 spare part descriptions).
Spooky how the guy looks like me bf. shaving.
October 31, 2025 at 9:50 AM
Spooky how the guy looks like me bf. shaving.
They still sell these, practically a cheat code. You’d need some time to learn the HP Prime though.
October 28, 2025 at 5:48 PM
They still sell these, practically a cheat code. You’d need some time to learn the HP Prime though.
I need to get used to the fact that Neural Networks with Dropout regularization can have a better validation loss than training loss. #machineLearning 🤖🧠
October 27, 2025 at 8:27 AM
I need to get used to the fact that Neural Networks with Dropout regularization can have a better validation loss than training loss. #machineLearning 🤖🧠
The German company whose freight lift was used in the Louvre heist is advertising its lifts now with the power statement „When you are in a hurry again“ and I just can’t stop laughing…
October 24, 2025 at 6:19 AM
The German company whose freight lift was used in the Louvre heist is advertising its lifts now with the power statement „When you are in a hurry again“ and I just can’t stop laughing…
Ich habe ein Meme extra für den Tag, wenn seine Amtszeit endet.
October 23, 2025 at 8:20 PM
Ich habe ein Meme extra für den Tag, wenn seine Amtszeit endet.
Just 4 more years...
October 23, 2025 at 8:15 PM
Just 4 more years...
The USian government accounts here are just trolling us to tell them what we think about them and oh boy am I tempted to do so.
October 18, 2025 at 7:06 AM
The USian government accounts here are just trolling us to tell them what we think about them and oh boy am I tempted to do so.
Today's plans: One VC to go, then a cigar, then meeting with friends. It's day of the octopus and Sigourney's birthday, so a beer or two are AOK. Successfuly finetuned 278 mio. parameter embedding model on a 14" toy laptop yesterday, so general mood is very good.
October 8, 2025 at 12:37 PM
Today's plans: One VC to go, then a cigar, then meeting with friends. It's day of the octopus and Sigourney's birthday, so a beer or two are AOK. Successfuly finetuned 278 mio. parameter embedding model on a 14" toy laptop yesterday, so general mood is very good.
Not to mention recursion. tse3.mm.bing.net/th/id/OIP.58...
August 23, 2025 at 9:02 PM
Not to mention recursion. tse3.mm.bing.net/th/id/OIP.58...
Reading about planned new AI features in software I use.
August 14, 2025 at 2:02 PM
Reading about planned new AI features in software I use.
Using Gradient Boosting in #machinelearning as your go-to-method is incredibly cheesy, but you can't really argue with the results. On tabular data, methods such as CatBoost or LightGBM usually rule supreme, even when using defaults.
August 12, 2025 at 1:24 PM
Using Gradient Boosting in #machinelearning as your go-to-method is incredibly cheesy, but you can't really argue with the results. On tabular data, methods such as CatBoost or LightGBM usually rule supreme, even when using defaults.
Du zeigst gerade sehr schön genau das, was ich hier als Risiko sehe. Du hast das Ergebnis einer Black Box KI ("das überschätzt das Problem") übernommen - ohne es plausibilisieren zu können. Genau so dürfte das dann auch in den Steuerämtern laufen mit den Ergebnissen von KI-Steuerprüfungen.
August 5, 2025 at 4:11 PM
Du zeigst gerade sehr schön genau das, was ich hier als Risiko sehe. Du hast das Ergebnis einer Black Box KI ("das überschätzt das Problem") übernommen - ohne es plausibilisieren zu können. Genau so dürfte das dann auch in den Steuerämtern laufen mit den Ergebnissen von KI-Steuerprüfungen.
Universe, it's so easy to get a machine learning guy to smile.
July 24, 2025 at 8:36 PM
Universe, it's so easy to get a machine learning guy to smile.
July 19, 2025 at 2:08 PM
World coffee production:
upload.wikimedia.org/wikipedia/co...
upload.wikimedia.org/wikipedia/co...
July 19, 2025 at 2:07 PM
World coffee production:
upload.wikimedia.org/wikipedia/co...
upload.wikimedia.org/wikipedia/co...
Austria can be ridiculously beautiful.
July 13, 2025 at 8:50 AM
Austria can be ridiculously beautiful.