Matija Piškorec (matija)

Matija Piškorec

5 Followers

Recent casts

About 7 years ago I started using Arch Linux and documenting all my configuration and usage tips in a set of Markdown files which I refer to daily. Every year or two I push new changes to a Github repo (most recently today!) https://github.com/matijapiskorec/archlinux-usb

  • 3 replies
  • 0 recasts
  • 1 reaction

Five years ago I was looking for a simple and fast English dictionary that could be used in the command line. Could not find it, then I decided to code it myself. It allows you to fuzzy search through 781k English words and their definitions parsed from Wiktionary. https://github.com/matijapiskorec/dict

  • 3 replies
  • 0 recasts
  • 1 reaction

About 8 years ago I prepared hands-on exercises for a summer school in data science on character-wise language modeling with multi-layer LSTMs. This is essentially what modern LLMs do but on a much (much) smaller scale - it was trained on around 20k Donald Trump tweets on a single laptop (without GPU) but was still able to generate believable fake tweets. This was about 5 years before the release of ChatGPT. I had no idea at that time that this would soon become the most dominant way of solving general intelligence. The code and examples of generated fake tweets are available at https://github.com/SSDS-Croatia/SSDS-2017/blob/master/Day-3/3_SSDS_2017_CharLSTMs_Solved.ipynb

  • 2 replies
  • 0 recasts
  • 1 reaction

Top casts

Here is an example of usage - you just run the "dict" command and start fuzzy searching through 781k words of English and their definitions, in real time.

  • 1 reply
  • 0 recasts
  • 2 reactions

I agree with everything except with the "academic research" part - similar to art, science is not a zero-sum game! One person winning does not mean that everyone else is losing. Maybe it seems that way because of the way science is funded (researchers as well as artists compete for limited funding) but in the grand scheme of things, more knowledge sooner is always better for everyone, regardless of who got there first.

  • 1 reply
  • 0 recasts
  • 1 reaction

About 8 years ago I prepared hands-on exercises for a summer school in data science on character-wise language modeling with multi-layer LSTMs. This is essentially what modern LLMs do but on a much (much) smaller scale - it was trained on around 20k Donald Trump tweets on a single laptop (without GPU) but was still able to generate believable fake tweets. This was about 5 years before the release of ChatGPT. I had no idea at that time that this would soon become the most dominant way of solving general intelligence. The code and examples of generated fake tweets are available at https://github.com/SSDS-Croatia/SSDS-2017/blob/master/Day-3/3_SSDS_2017_CharLSTMs_Solved.ipynb

  • 2 replies
  • 0 recasts
  • 1 reaction

Onchain profile

Ethereum addresses