AI summary by TwinMind
AI Super Agents are coming. Allegedly. What does this mean?
Sabine Hossenfelder
7m 26s
Rumored Government Meeting
- Sam Altman (OpenAI CEO) scheduled closed-door meeting with Washington government officials January 30th
- Topic: "Coming of AI PhD level super agents"
- Secrecy suggests potential security concerns or sensitive applications
AI Agents Definition & Current State
- AI agents: systems that independently create action plans and use tools to complete goals
- Don't require step-by-step instructions - just tell them desired outcome
- Google released Gemini 2.0 "for the agentic era"
- Can add calendar entries on user's behalf
- OpenAI released "Operator" research preview last week
- First agent that browses web and takes actions
- Example: ordering pasta recipe ingredients
- Uses human-built browsers, sees same screen as humans
- Controls via keyboard typing and mouse clicking
- Currently only available for $200 top-tier subscribers
- Casey Newton reports: currently slower, more frustrating, more expensive than doing tasks yourself
PhD-Level AI Capabilities
- "Super agent" speculation: one agent controlling army of communicating agents
- PhD-level likely means passing PhD-level exams
- GPT-4 two years ago: correctly answered ~50% undergraduate/graduate physics questions
- Today's models expected to score 80%+ on such tests
- Center for AI Safety released "Humanity's Last Exam"
- Questions from all scientific disciplines by experts
- Best AI currently scores below 10%
- Likely better than most humans would perform
Research Applications & Concerns
- Literature searches: AI agents could outperform humans
- Applications: fact-checking, summaries, identifying research topics, funding opportunities
- Problem: much published scientific literature is nonsense
- "Tacit knowledge" exists only in human experts' brains
- Need human experts to train AI to distinguish nonsense from valid research
Academic Impact Predictions
- Future academic publishing: papers written by AI, reviewed by AI, read by AI
- AI agents useful for literature searches initially
- Skeptical about independent scientific research
- Challenge: figuring out what constitutes good research goals
- Prediction: OpenAI will release "super duper agents" soon
Security & Government Implications
- Closed-door nature suggests sensitive applications
- Potential military applications concerns
- Biological and chemical weapons considerations
- Computer vulnerabilities
- Possible replacement of scientific advisors to government
