The public debate on AI Flashcards
(13 cards)
Rift between Doomers and Boomers
- Doomers
AI is an existential threat for humankind if not strickly regulated
- Proprietary model
- Big investors/companies
- Ethical paradigm → Effective altruism
Rift between Doomers and Boomers
- Boomers
AI is a force capable to turbochange progress
- Open source model
- Venture capital/small companies
- Cultural paradigm → Effective accelerationism
Risk of AI
- Author
- Yoshua Bengio
- Yuval Noah Harary
- Sam Bowman
- Holden Kamosky
- Norbert Wiener
Risk of AI
- Yoshua Bengio
AI is a global political risk
→ human brain is a biological machine
- theoretical possible to build a machine just as intelligent
- once understood the mechanism of human intelligence it will possible to build a superhuman AI (5-20 years)
Risk of AI
- Yoshua Bengio
→ possible risks
→ Terrorism and rogue state could use open-source AI to:
- biological weapons
- cyber-attacks
- manipulate public opinion
→ AI self-perservation (like a human)
- imitate human
- seek power
- intentionally build
Risk of AI
- Yoshua Bengio
→possible measures
→ National and international legislation (UN)
→ Safeguards systems
→ Counter-measures against rogue AI
Risk of AI
- Yoshua Bengio
→ possible measures
- National and international legislation (UN)
→ no AI with dangerous features
such as:
- pharmaceutical
- nuclear
- aviation
→ independent audits
Risk of AI
- Yoshua Bengio
→ possible measures
- Safeguards systems
→ programs
→ protocol
→ governance
Risk of AI
- Yoshua Bengio
→ possible measures
- Counter-measures against rogue AI
→ Reinforcement Learning from Human Feedback (RLHF)
▪ human gives a feedback whether the response to a command was appropriate
Problems:
▪ humans could not concord on “appropriate”
▪ Chat-GPT could learn to imitate human language
▪ AI could learn how to avoid feedbacks
→ Research on AI only on strickly international and supranational basis
▪ only peaceful purposes
▪ never conducted by individual states or for-profit entities
Risk of AI
- Yuval Noah Harary
AI hacked the OS of civilisation → language
→ dangers from AI’s capacity to manipulate or generate language
such as:
- texts
- images
- sounds
the problem could escalate
- religion
- laws
- influence
Risk of AI
- Sam Bowman
→ Pre-launch screening of LLM extremely difficult
→ Use AI to create a Constitutional AI
pro:
- no human involved
- AI better to control
cons:
- who writes the fundamental rights
→ Renounce of Opacity could help
- less performance
- copyright issue
Risk of AI
- Holden Kamosky
AI could leard to circumvent test
Risk of AI
- Norbert Wiener
The understanding should go pari passu with development of performance of the machine
