2024-08-15

Toronto AI Safety Meetup
- News
  - https://www.arena.education/
    - hard
  - Hermes 3 model
    - less censored
  - Paper on LLM predicting treatment effect: https://www.treatmenteffect.app/
  - SWE-bench verified
    - Agentless scaffhold
      - https://arxiv.org/pdf/2407.01489
  - office website: https://trajectoryai.org
- Emergent capabilities
  - Paper about this: https://openreview.net/pdf?id=yzkSU5zdwD
  - going from can't do at all -> can do at some size
    - artifact of benchmark?
      - might be because of measurement problems
        
        e.g. pass/fail grading of model responses
  - Prompting strategies are emergent too
    - Explanations
      - Multi-step reasoning requires more layers
        
        not with chain of though tho
      - More parameters for memorization
      - Learning to do in-context learning at some scale makes everything better
        
        once ICL is gets to a certain goodness, the prompting techniques are viable
  - Happens in humans, nature too
  - Papers
    - real - https://arxiv.org/pdf/2201.11903
    - not real - https://arxiv.org/pdf/2304.15004
    - actually real - https://arxiv.org/pdf/2403.15796
    - observational scaling laws - https://arxiv.org/pdf/2405.10938
  - more research needed!