Results
(Update 2020-03-17: I've quit AI alignment research, so this page is not relevant anymore. I might update it when I find new work. Until then I'll leave it here for reference.)
I will measure my results, starting from October 2019, in order to make sure that I fulfill the responsibilities set down in the job description.
Measures and standards
- Number of final articles etc. published.
- Blog entries that could be articles suitable for a conference or journal. Conference (incl. workshop) and journal submissions, even if not accepted. Major software or dataset releases.
- 1 per 6 months
- Number of pieces of feedback received about intermediate products.
- Plans, analyses, summaries, critiques. Research-related comments, blog and forum entries.
- 1 per 2 weeks
- Number of positive responses to research-related help, recommendations, advice, feedback given.
- 1 per 2 months
- Funding received.
- 60000 USD per year
- Number of interactions with (potential) junior AI alignment contributors.
- 1 per 3 months
Reports
2020-02-22
- Publishing – three minor pieces:
- GitHub: rmoehn/amplification – Almost all of this code was written by Paul Christiano, Buck Shlegeris and William Saunders.
- Twenty-three AI alignment research project definitions
- Training a tiny SupAmp model on easy tasks: The influence of failure rate on learning curves
- Positive responses to feedback given: Almost no feedback given.
- Feedback received: Barely anything.
- Funding: Zero. Application for funding from LTFF pending.
- Interactions:
- Still weekly RAAWG meetings.
- Online conversation with Victor Tao about IDA and a research idea of his.
- AISC Toronto research proposal discussion.
2020-01-18
I'm producing again for a while. BLUF: I'm not meeting the standards. Still, I know that I'm on a good path, except that I never know whether my results are good enough to earn me the next chunk of funding. Details:
- Publishing: I don't expect to publish anything by the end of March. Provided that funding doesn't run out, I will have implemented overseer failures in SupAmp and run experiments. And I hope to get RelAmp about half done. (See the Farlamp README for some explanation.) Probably I won't be able to make this software public.
- Feedback received: Barely anything.
- Positive responses to feedback given: No feedback given in Oct/Dec. One feedback given and positive response received in January.
- Funding: 15 kUSD received in October. Funding for February and beyond uncertain.
- Interactions: Weekly since I started the Remote AI Alignment Writing Group (RAAWG) in October. See also the latest announcement, Remote AI alignment writing group seeking new members.
2019-11-10
Shortly after I set down how I will measure my results, I had to change direction. Instead of producing I'm now studying how to implement deep learning. To this the above measures and standards don't apply. I will let everything stand until I resume production.