Method: The OpenAI team ran a suite of 63 different benchmarks (including sub-types), each for zero/one/few shot. In each scenario, there are 8 model sizes. I looked at how results scale with model size. With only 8 measurements, there is a large associated uncertainty for predictions. Formally, one would test the trend function using a Bayesian model selection between a linear and (e.g.,) a polynomial. I did this for a few and then eye-balled the rest. So, please take the following as an indication only.
Disclaimer: The smallest model for GPT-3 has 108 parameters, the largest 1011. That's a span of 3 orders of magnitude. Scaling this out to many more orders of magnitude is dangerous. Thus, take these numbers only as an indication.
Results. For the following tests, I find an asymptotic trend. Scaling the model will apparently not yield fantastic results for:
TriviaQA (1013 parameter estimate to achieve 100%)
BoolQ (1015)
MultiRC (1016)
ARC (1016)
SuperGLUE (1018)
WSC (1020)
WebQs (1021)
Cycled (1023)
Some tests scale neither linear nor asymptotic:
Symbol: Near exponential (1012)
Arithmetic: Exponential; one-digit composite may achieve 100% at 1014
Reversed: Near exponential (1016)
Anagrams: Polynomial (1019)
ANLI: stepped, unclear
RTE: stepped, unclear
Summary: About half of the tested skills will likely not scale much with larger models. The other half will (e.g., TriviaQA, SuperGLUE, arithmetic, anagrams). Going to e.g., 1016 parameters - would that make an Oracle AI? Probably it's not sufficient, but I'm interested in hearing your opinion!
I think GPT-3 should be viewed as roughly as aligned as IDA would be if we pursued it using our current understanding. GPT-3 is trained via self-supervised learning (which is, on the face of it, myopic), so the only obvious x-safety concerns are something like mesa-optimization.
In my mind, the main argument for IDA being safe is still myopia.
I think GPT-3 seems safer than (recursive) reward modelling, CIRL, or any other alignment proposals based on deliberately building agent-y AI systems.
--------------------
In the above, I'm ignoring the ways in which any of these systems increase x-risk via their (e.g. destabilizing) social impact and/or contribution towards accelerating timelines.
Regarding GPT-3, there is some discussion whether growing the model would transform it into an Oracle AI. I looked into the actual benchmark results (Appendix H in the paper) to see if we can predict something useful from the actual measurements.
Method: The OpenAI team ran a suite of 63 different benchmarks (including sub-types), each for zero/one/few shot. In each scenario, there are 8 model sizes. I looked at how results scale with model size. With only 8 measurements, there is a large associated uncertainty for predictions. Formally, one would test the trend function using a
Bayesian model selection between a linear and (e.g.,) a polynomial. I did this for a few and then eye-balled the rest. So, please take the following as an indication only.
Disclaimer: The smallest model for GPT-3 has 108 parameters, the largest 1011. That's a span of 3 orders of magnitude. Scaling this out to many more orders of magnitude is dangerous. Thus, take these numbers only as an indication.
Results. For the following tests, I find an asymptotic trend. Scaling the model will apparently not yield fantastic results for:
In the following tests, it is unclear if the trend is asymptotic or better than that:
These tests show a linear scaling:
Some tests scale neither linear nor asymptotic:
Summary: About half of the tested skills will likely not scale much with larger models. The other half will (e.g., TriviaQA, SuperGLUE, arithmetic, anagrams). Going to e.g., 1016 parameters - would that make an Oracle AI? Probably it's not sufficient, but I'm interested in hearing your opinion!