Ethan Caballero
Nov 18, 2022
Approver
See Section 5.6 titled "The Limit of the Predictability of Scaling Behavior" of "Broken Neural Scaling Laws" paper:
https://arxiv.org/abs/2210.14891v4
https://arxiv.org/abs/2210.14891v4
Suggestion was deleted
Show more
Show less
Comment details cannot be verified
0 replies
New
Zach Stein-Perlman
Nov 17, 2022
Approver
(I think this model is nonsense: using an uninformative prior on the number of insights required throws away almost all information about AI progress. And for semi-informative priors, Davidson's report seems like a better starting place.)
Suggestion was deleted
Show more
Show less
Comment details cannot be verified
0 replies
New
You're suggesting
Gemini created these notes. They can contain errors so should be double-checked. How Gemini takes notes
Drag image to reposition
Outline
Outline
Document tabs
AI forecasting research ideas from Epoch
Headings you add to the document will appear here.
AI Forecasting Research Ideas
Overview
Projects
Extrapolating GPT-N performance Difficulty: Medium
Qualitatively analysing language model / image generation improvements since ~2000 Difficulty: Easy
Do AI researchers train models using scaling laws? Difficulty: Medium
Revisiting ‘Is AI Progress Impossible To Predict?’ Difficulty: Hard
Algorithmic breakthroughs in machine learning history Difficulty: Medium
Paradigm changes in AI Difficulty: Medium
Study training run lengths Difficulty: Easy
AI development vignettes Difficulty: Hard
Profiler to measure compute Difficulty: Hard
Insights-based models of AI timelines Difficulty: Medium
Brain emulation development Difficulty: Medium
Rethinking the evolutionary anchor Difficulty: Hard
Investigate trends in memory bandwidth, latency and price of memory Difficulty: Easy
Investigating the parameter gap Difficulty: Medium
Investigate possible paradigm shifts in hardware Difficulty: Medium
How much can we scale up production in the compute supply chain? Difficulty: Hard
Will the price of compute go down if hardware performance stops improving? Difficulty: Hard
What has been the share of any chip in a given year of total available compute performance? Difficulty: Easy
Changes by
AI forecasting research ideas from Epoch
