AI Governance
A complex topic, simply explained
✨ State of AI
- ChatGPT
- Diffusion Model
ChatGPT
Language models are getting really good..
Diffusion Models
“Rick Astley in a hawaiian shirt in Jerusalem”
To be fair, that was not cherry picked.
😨 Risks of AI
- Arms Races: Slaughterbots
- Mind Crime: Kara
- Existential: Paperclip Maximizer
- Evil AI: ChatGPT
Slaughterbots: The Theory
Slaughterbots: The Reality
Kara
Paperclip Maximizer
when the optimization function = max(📎)
the AI will eventually destroy → 🌎
ChatGPT
OpenAI spent a ton of effort to lobotomize ChatGPT from all evil thoughts. And yet…
Facets of AI Governance
- 🏢 Organizational
- 🇺🇸 National
- 🌐 Global
- 💀 Existential
🏢 Organizational
- MLOps
- Fairness
- Adversarial Defense
MLOps
A set of procedures and enabling systems for deploying machine learning models reliably and efficiently. Can be automated…
Fairness
Procedures, science, and enabling systems for preventing AI models from discrimination based on protected classes.
Adversarial Defense
Protecting AI models from intentional attacks intended to degrade their performance.
🇺🇸 National
- Races
- Policy
Races
Two nations develop AI capacity in a national competitive race
The big one:
🇨🇳 vs 🇺🇸
Policy
Can national policy affect change?
🌐 Global
- Autonomous Weapons
- Future of Work
💀 Existential
- AI Alignment
- Instrumental Goals
- The 📎 Maximizer (Again)
- Mind Crimes
- Intelligence Explosion
- Decisive Strategic Advantage
Instrumental Goals
🎺 🥅 are goals that lead to end goals
Remember the paperclip maximizer
when max(📎) →
📎 > 👪
📎 > 🌎
it’s the 🎺 🥅 that get you
Mind Crimes
when 🤖 has 🧠
one can commit crimes against 🤖
Intelligence Explosion
🤖 🦾 self-improving itself
leads to even greater 🧠
🤖 🦾 self-improving itself even faster 🦾 🦾 🦾
and on and on
hence an intelligence 💣💥
Decisive Strategic Advantage
