1 of 37

Goas ipmirda ChatGPT sámegiela?

Lars Ailo Bongo�Informatihka instituhtta, �UiT Norgga árktalaš universitehta

​

Sámi allaskuvlla. 18.08.23.

DALL-E: “a bright future in sapmi science fiction style”

2 of 37

Eará AI?

3 of 37

4 of 37

5 of 37

Ipmirda go ChatGPT �sámegiela?

6 of 37

Stable Diffusion: “sami reindeer herders in traditional clothing”

DALL-E: “sami reindeer herders in traditional clothing”

7 of 37

Nugo láve?

Stable Diffusion: “same old story depressive style”

8 of 37

Outline

  1. Technical background
  2. Issues with state-of-the-art
  3. 💡Sami-AI💡

DALL-E: “a bright future in sapmi science fiction style”

9 of 37

Traditional coding - Python

From: https://github.com/uit-hdl/HEImmune

10 of 37

…with the help of ChatGPT

11 of 37

Visual coding - Scratch

Source: https://medium.com/scratchteam-blog/3-things-to-know-about-scratch-3-0-18ee2f564278

12 of 37

Visual coding - Unreal Engine

Source: https://docs.unrealengine.com/5.1/en-US/graphing-in-animation-blueprints-in-unreal-engine/

13 of 37

14 of 37

AI buzzwords

https://medium.com/swlh/artificial-intelligence-machine-learning-and-deep-learning-whats-the-real-difference-94fe7e528097

Source: https://www.spiceworks.com/tech/big-data/articles/what-is-support-vector-machine/

Source: https://www.javatpoint.com/deep-learning

15 of 37

Big data

  • big computer

Source: The Economist

(ChatGPT = 100,000,000,000,000,000,000,000,000 FLOPS)

16 of 37

Training a model

Label = Zebra

Label = antelope

Label = zebra

Repeat a few million times

17 of 37

Using a model (inference)

Label = ?

Label = zebra

18 of 37

Fine tuning a model (transfer learning)

label = gabba

label = muzet

Repeat a few hundred times

19 of 37

💡Boazu �identification💡

20 of 37

From cats and dogs to abnormal lung sounds

21 of 37

From master’s thesis table to medical device

Conflict of interest: LAB own shares in Medsensio.

22 of 37

Gos leat sámi govat?

23 of 37

💡DataverseSME💡

label = gabba

Mas leat govvat!

24 of 37

25 of 37

Diffusion models

Source: https://stable-diffusion-art.com/how-stable-diffusion-work/

26 of 37

💡Fine-tuned sámi image generator model💡

27 of 37

💡Ii leat vattis!💡

28 of 37

29 of 37

Reinforcement learning

30 of 37

ChatGPT

Source: https://openai.com/blog/chatgpt

31 of 37

LLMs (ChatGPT) is the fastest evolving technology on earth!

32 of 37

💡Small is interesting (and beautiful)💡

33 of 37

$10M vs $100 language models

34 of 37

35 of 37

On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? 🦜

Unfathomable training data:

  1. Size doesn’t guarantee diversity
  2. Static data/ changing social views
  3. Encoding bias

36 of 37

💡“take inspiration from movements to decolonize education”💡

​

37 of 37

Oktigeassu

  • ChatGPT ja Stable Diffusion
    • Mii ja movt
    • Manin eai doaimma sámegilli
  • Sámi AI
    • Ráhkadit govaid: hui álki
    • AI modelat: veháš váttis
    • Giella: hui váttis
  • Lea dárbu leat mielde AI ráhkadeames!

Lars Ailo Bongo

lars.ailo.bongo@uit.no

https://uit-hdl.github.io/