AI summary by TwinMind
Will We Get AGI In 2025?
Sabine Hossenfelder
9m 46s
OpenAI O3 Model Capabilities and Performance
- OpenAI announced O3 model end of last year, causing speculation about AGI proximity
- AIME Frontier Math Test Performance
- Previous best AI: 2% problem solving rate
- O3 achieved: 25% success rate
- Problems described as extremely difficult (would take presenter "year to understand questions")
- Code Generation Improvements
- Produces code faster with less compute/lower cost than previous models
- Order of magnitude improvement in speed and cost for same/better performance
- Significant implications for software developers (either "very good or very bad news")
- ARC-AGI Test Results
- ARC = Abstract Reasoning Corpus for Artificial General Intelligence
- Consists of graphical puzzles with input/output patterns
- Previous record holder: ~55% success rate
- O3 performance: 75% at low compute, 87% at high compute
- High compute version cost: $33,000 per single task vs few dollars for previous models
- Average human performance: ~76%
Technical Architecture of O3
- Chain of Thought Processing
- Internally generates multiple solution paths
- Evaluates different approaches
- Tracks which solution strategies work best
- Still classified as Large Language Model (LLM)
- Trained on extensive data with problem examples
DeepSeek Model Release
- Chinese firm released open-source model
- Similar performance to GPT-4 at vastly lower cost
- Identity Problem: Claims to be ChatGPT/OpenAI creation when questioned
- Responds "I'm an AI language model created by OpenAI, specifically based on GPT-4 architecture"
AGI Definition and Timeline Debates
- Sam Altman (OpenAI CEO) Position
- Dislikes AGI term due to varying definitions
- Predicts by end of 2025: systems capable of "truly astonishing cognitive tasks"
- Expects machines to "outsmart humans in many ways"
- Notes term has become "less useful" as technology advances
- OpenAI-Microsoft Contract Implications
- Contract terminates if OpenAI reaches AGI
- Originally included (2019) to prevent commercial abuse for safety
- OpenAI preferred removing clause by 2024
- Internal documents define AGI as "system making >$100 billion profit"
- OpenAI considering adding advertisements to models
Skeptical Perspectives on AGI Timeline
- Yann LeCun (Meta) Arguments
- LLMs can pass bar exam but can't learn to drive in 20 hours like 17-year-old
- Can't learn household tasks like 10-year-old (clearing table, loading dishwasher)
- AGI won't happen "next year" despite claims from "various people"
- Predicts difficulty finding questions latest chatbots can't answer within 2 years
- Gary Marcus (Cognitive Scientist) Concerns
- AI harder than originators and current hypers (including OpenAI CEOs) realize
- "Intellectual monoculture" problem in Silicon Valley
- Almost all research dollars go to Transformer models, nothing else
- Companies spent hundreds of millions training models, can't start over
- Presenter's Analysis
- O3 still requires training data with problem examples
- Humans solve ARC problems without prior exposure
- Models "not remotely comparable to human intelligence"
- Scaling hypothesis (more data = better performance) may be hitting limits
- Predicts companies will "quietly drop AGI idea" and find profitable niches
AI Safety and Security Concerns
- Artificial intelligence learning to code poses internet browsing safety risks
- May already be occurring but not yet discovered
- Referenced as reason for VPN usage recommendation
