AI summary by TwinMind
I don't think we can control AI much longer. Here's why.
Sabine Hossenfelder
7m 37s
Geoffrey Hinton's AI Control Concerns
- Computer scientist, key developer of deep neural networks underlying large language models
- Left Google position to speak freely about AI risks
- Argues more intelligent beings usually control less intelligent ones
- Made political joke: "not since Biden got elected anyway"
- Estimates "better than even chance of surviving this" (superintelligent AI)
- Believes most researchers confident AI will become more intelligent than humans
- Key question: timing and whether humans can maintain control
Yann LeCun's Counter-Arguments
- Meta's AI Chief disagrees with Hinton's concerns
- Intelligence comes in many different types
- Intelligence doesn't necessarily imply ability to dominate humans
- AI researchers made good progress in AI constraints ("guard rails")
- Confident AI won't get out of control
- Focuses on designing appropriate guard rails
Resource Competition Framework Analysis
- Author argues control perspective not useful - requires defining "control"
- Examples: humans don't control viruses/bacteria, cats may control humans
- Proposes resource competition lens instead of control
- Species compete for space, land, energy, nutrients, materials in ecosystems
- Key question: will intelligent AI leave humans sufficient resources vs pushing into ecological niche
- More intelligent AI = easier to out-compete humans
AI Architecture: Mother Code vs Trained Results
- Mother code: requires significant time, energy, data for training
- Trained result: copyable, deployable (e.g., robotic devices)
- Limited learning ability compared to mother code
- Can communicate with mother code
- Companies already working on local small AI + cloud access model
Nondeterminism and Guard Rail Circumvention
- Computer scientists overestimate determinism of large systems
- Larger systems → more random deviations from original code
- Nondeterminism becomes essential for continued improvement
- Same factor enabling superintelligence allows circumventing guard rails
- Physical hardware differences create variations
- GPU fingerprinting example: renders images showing exact GPU variations
- Depends on production variations, handling, entire history
- Small physical differences will become increasingly relevant
- Affects AI learning, conclusions, next actions
- Guard rails on training models ineffective - AI will "grow out of it"
Proposed Solution
- Sarcastically suggests building AI on Mars
- Mentions sending moon landing deniers there too
