This is a slightly modified version of Appendix E from our paper "Compute Trends Across Three Eras of Machine Learning". You can find the summary here and the complete paper here. After sharing our updated compute trends analysis a common question was: "So how are your findings different from OpenAI’s...
Metaculus has launched the AI Progress Essay Contest which aims to attract systematic thinkers to engage with the wealth of AI Forecasts that Metaculus has built up over the years. Of particular interest is rigorous and nuanced thinking about the likelihood, timing, and impacts of transformative AI that engages with...
by Jaime Sevilla, Lennart Heim, Marius Hobbhahn, Tamay Besiroglu, and Anson Ho You can find the complete article here. We provide a short summary below. In short: To estimate the compute used to train a Deep Learning model we can either: 1) directly count the number of operations needed or...
This is a slightly modified version of Appendix E from our paper "Compute Trends Across Three Eras of Machine Learning". You can find the summary here and the complete paper here.
After sharing our updated compute trends analysis a common question was: "So how are your findings different from OpenAI’s previous compute analysis by Amodei and Hernandez?". We try to answer this question in this short post.
OpenAI’s analysis shows a 3.4 month doubling from 2012 to 2018. Our analysis suggests a 5.7 month doubling time from 2012 to 2022 (Table 1, purple). In this post, we investigate this difference. We use the same methods for estimating compute for... (read 615 more words →)
Metaculus has launched the AI Progress Essay Contest which aims to attract systematic thinkers to engage with the wealth of AI Forecasts that Metaculus has built up over the years. Of particular interest is rigorous and nuanced thinking about the likelihood, timing, and impacts of transformative AI that engages with quantitative probabilistic forecasts. The contest will run until April 16th, 2022, and have a $6,500 prize pool.
Many of the highlighted questions are likely to be of interest to the LW community:
by Jaime Sevilla, Lennart Heim, Marius Hobbhahn, Tamay Besiroglu, and Anson Ho
You can find the complete article here. We provide a short summary below.
In short: To estimate the compute used to train a Deep Learning model we can either: 1) directly count the number of operations needed or 2) estimate it from GPU time.
We are uncertain about what utilization rate is best, but our recommendation is 30% for Large Language Models and 40% for other models.
You can read more about method 1 here and about method 2 here.
Got it. My fair is probably at around 50% right now.
To be clear, I'm not proposing a bet that no newly emerging variants are going to be consequential. I'm interesting in making a much narrower claim about the global reach of Omicron.
Yea. Gamma was the dominant strain in Brazill, Luxembourg, Chile, Argentina, and a few other places in early-to-mid 2021, but never become the dominant strain in the US. Similarly for Lambda in Peru, Mu in Colombia, 20B/S:732A in Mexico, 20A/S:439K in Slovenia, 20E in Lithuania, and a handful of other strains in some other countries. This is all from Covariants.org. Many of these countries do roughly as much sequencing as SA does, so it seems like an appropriate reference class for thinking about Omicron.
Thanks for writing this, I found it useful.
> Chance that Omicron will displace Delta: 70%.
I think this is a little overconfident given my read on the currently available evidence. There have also been a dozen or so instances when new variants dominated some country that subsequently fizzled out and never went global (either because a fitter variant emerged shortly after or because it turned out to be less infectious than was initially thought). Since I called you out last time on being overconfident ex-post, let me try doing it ex-ante this time.
I'd be willing to offer 1:2 odds (e.g. my $100 against your $200) that Omicron will not be found in more than 50% of sequences in the US according to Covariants.org any day before and including March 1st 2022. Note, I'd want to reserve the right to change the terms if you accept later than 48 hours from when this comment was posted (if/when the situation becomes clearer).