Evaluators and safety organisations · Blog post · Redwood Research
A basic case for accelerating conceptual reasoning capabilities
Read at blog.redwoodresearch.org ↗Read at sourceAll rights reserved
Outline · 170 paragraphs
- Introduction¶1
- Overview¶6
- It’s important how soon AI is good at conceptual reasoning¶26
- Important challenges need conceptual reasoning¶27
- AI R&D and takeover rely much less on conceptual reasoning¶62
- AI being better at conceptual reasoning earlier would be very valuable¶85
- We can probably differentially accelerate conceptual reasoning¶94
- Work on conceptual reasoning is likely counterfactual and tractable¶95
- Effects on generic capabilities are likely small¶102
- Counterpoint: Using our datasets to develop generic techniques¶107
- Our work¶111
- Acknowledgements¶118
- Appendix: Examples and non-examples of conceptual domains¶120
- Appendix: is ML bottlenecked by conceptual insights?¶137