Thomas Winters
LightNews LightNews
banner
thomaswinters.be
Thomas Winters
@thomaswinters.be
990 followers 220 following 16 posts
Postdoctoral researcher in creative AI, large language models & computational humor @ KU Leuven

thomaswinters.be
Posts Replies Media Videos
thomaswinters.be
Ook hun coversysteem is indrukwekkend

Hier een jazz versie van Ochtendgymnastiek van Samson & Gert

Instant kerstshow vibes
September 30, 2025 at 10:18 AM
thomaswinters.be
Supervised a thesis student who extended parts of Weird A.I. Yankovic to newer LLMs. It's still hard to balance adhering to parody constraints with the coherence of the lyrics.

repository.teneo.libis.be/delivery/Del...
Enforcing Creative Constraints In Autoregressive Language Models During Generation For Musical Parodies
repository.teneo.libis.be
December 24, 2024 at 12:48 PM
thomaswinters.be
Interestingly, the closeness of "7" to "11" seem to play a role.

Couldn't get it to mess up over 1.7 and 1.15. But 1.9 and 1.15 almost always causes ChatGPT to respond incorrectly, then contradicts itself immediatly afterwards.
User: "1.7 and 1.15 which one is bigger?" ChatGPT: "1.15 is smaller than 1.7. To compare: 1.7 is equivalent to 1.70. 1.15 remains 1.15. Clearly, 1.70 is greater than 1.15." "1.15 is bigger than 1.9. If we break it down: 1.15 is equivalent to 1 and 15 hundredths. 1.9 is equivalent to 1 and 90 hundredths. When compared, 1.9 > 1.15"
November 28, 2024 at 11:40 AM
thomaswinters.be
A likely culprit? Tokenization. ChatGPT likely sees "11" and "7" as tokens after the decimal. This affects how it compares numbers.
Tokens for "1.11" and "1.7", showing it tokenized as 1 . 11 and 1 . 7
November 28, 2024 at 11:40 AM
thomaswinters.be
Many people argue this answer would be correct in contexts like versioning or section numbering. But that’s not what’s happening here.

If you allow ChatGPT to first use some “draft paper”, it always interprets these as decimal numbers and solves it correctly.
User to ChatGPT 4o: "1.11 and 1.7, which one is bigger? Think first" ChatGPT: "Let's compare the two numbers: 1.11 and 1.7 are both decimal numbers. To directly compare them, let's align their decimal places: 1.11 = 1.11 1.7 = 1.70 Now it's clear that 1.7 (or 1.70) is greater than 1.11 because 70 (the second decimal place of 1.70) is greater than 11 (the second decimal place of 1.11). Final Answer: 1.7 is bigger than 1.11."
November 28, 2024 at 11:40 AM
thomaswinters.be
Lijkt inderdaad goed doenbaar.

Willen we dan oldschool "autocomplete" TorfsBot of de slimmere GPT-versie? Beiden hebben eigen charme
November 25, 2024 at 11:35 PM
thomaswinters.be
Gegenereerd door
> bestand uploaden naar custom ChatGPT die klinkt zoals Rik Torfs zijn tweets (chatgpt.com/g/g-TBDsKAk4...)
> "Professional Voice Clone" met Elevenlabs op anderhalf uur aan Rik Torfs toespraken getraind
> Gegenereerde audio + één foto naar Hydra voor video te genereren
ChatGPT - TorfsGPT
Geautomatiseerde kwakfilosoof aan de KU Leugen
chatgpt.com
November 25, 2024 at 2:26 PM
thomaswinters.be
Het kan eender welke tekst samenvatten naar een volledige Rik Torfs toespraak.

Bijvoorbeeld:
- @lievenscheire.bsky.social's AI boek: www.dropbox.com/scl/fi/j30ao...
- @jbaert.bsky.social's paper over octrees: www.dropbox.com/scl/fi/rf9bn...
November 25, 2024 at 2:24 PM