Bericht (virtuelles) MLUGS Treffen im Januar 2023
17. January 2023
Protokoll
Andreas: ChatGPT: Demo, Einordnung and Discussion
Why?
-
Does chatGPT solve Natural Language Generation? is this game over for AX?
-
current hype (we had hypes the last years for GPT-2 and GPT-3, too)
chatGPT Background
-
summary:
Autoregressive model trained on guessing the next token;
optimized for dialogs by adding a reward model -
successor of InstructGPT, using GPT3.5:
https://beta.openai.com/docs/model-index-for-researchers/models-featured-in-openai-research -
blog post: https://openai.com/blog/chatgpt/
-
latest paper from OpenAI about InstructGPT:
Training language models to follow instructions with human feedback -
"easy" to follow GPT-2 implementation: https://github.com/karpathy/nanoGPT
chatGPT Demo
when still possible: https://chat.openai.com/chat
if not, some examples:
-
Tell me about the Pixel 4a in German:
https://paste.madflex.de/nuWQdMru/+inline (2022-12-18) -
Tell me about the Pixel 4a in German like I am 5:
https://paste.madflex.de/MivgsN24/+inline (2022-12-18) -
Please tell me something about the black Pier One Jumper in Latin:
https://paste.madflex.de/EpFoeczy/+inline (2023-01-12) -
(German product description for a (generated) dataset):
https://paste.madflex.de/AugkjcXf/+inline (2023-01-02) -
How many letters in word "five"?
https://troet.cafe/@MrLaCave/109677109019794666 -
In 1971, Terry Winograd devised the Winograd Schemas, simple sentence understanding tasks that require the use of knowledge and commonsense reasoning. https://hci.social/@Mor/109631432389618519
-
Schreibe mir einen Limerick für Solarzellen:
https://paste.madflex.de/G6MbY22q/+inline (2023-01-17) -
Schreibe mir diesen Text als Rap: https://www.heise.de/meinung/Kommentar-zu-ChatGPT-Keine-Intelligenz-keine-Panik-7392906.html
https://paste.madflex.de/wjnW2WiY/+inline (2023-01-17)
Einordnung / Classification
Dennoch ist Angst fehl am Platz:
Das System ist ein großer Schritt für Sprachmodelle, bleibt aber fern von einer KI-Revolution.
Ein starker Algorithmus, aber nach wie vor keine Intelligenz. Also kein Grund zur Panik.
The idea of an all-knowing computer program comes from science fiction and should stay there.
Despite the seductive fluency of ChatGPT and other language models, they remain unsuitable as sources of knowledge.
We must fight against the instinct to trust a human-sounding machine, argue Emily M. Bender & Chirag Shah.
… generierten Texte sind nur selten völliger Unsinn, oft aber auf merkwürdige Weise nebulös und unverbindlich,
als versuchte jemand, kompetenter zu erscheinen, als er oder sie ist. …
With ChatGPT, we’ve reached the landfill capitalism stage of AI. Throw something on the market that only works if others
clean up. What Bolt, Tier, etc and their scooter are doing with cities, OpenAI are doing with the information landscape.
- Wissing gegen mehr KI-Regulierung (German)
„Mit der kommenden Generation des KI-Sprachmodells GPT-4 werden Dinge möglich, die unser Leben verändern werden“,
sagte Wissing.
Discussion
- Does chatGPT solve Natural Language Generation and all companies in the field are doomed?
No.
The model has no actual concept of the world.
It stitches in a quite good sounding way text chunks
together without actual semantics or pragmatics.
- Is ChatGPT actually helpful?
Yes and No.
Yes, if you (the reader) can distinguish correct parts of the text from wrong parts.
If you cannot, than the text is a trap!
-
How we talk about AI, illustrate AI
-
Meta
next
-
2023-02-21
- mglw. in-person in Stuttgart
- wir suchen Vorträge!