I’m the co-founder and operations director of WhiteBox Research. WhiteBox aims to develop more AI interpretability and safety researchers in Asia. I’m also a co-founder of EA Philippines.
I previously was a Group Support Contractor for the Centre for Effective Altruism (CEA) for two years, where I helped support EA groups around the world.
You can reach out to me at brian@whiteboxresearch.org or find me on LinkedIn.
My typo reaction may have glitched, but I think you meant “Don’t push the frontier of capabilities” in the last bullet?
I’ve only read the blog post and a bit of the paper so far, but do you plan to investigate how to remove alignment faking in these situations? I wonder if there are simple methods to do so without negatively affecting the model’s capabilities and safety.
Thanks for doing this important research! I may have found 2 minor typos:
The abstract says “We find the model complies with harmful queries from free users 14% of the time”, but in other places it says 12% - should it be 12%?
In the blog post, “sabotage evaluations” seems to link to a private link
Thanks for this analysis! A minor note: you’re probably aware of this, but OpenPhil funds a lot of technical AI safety field-building work as part of their “Global Catastrophic Risks Capacity Building” grants. So the proportion of field-building / talent-development grants would be significantly higher if those were included.
Thanks for making this! This is minor, but I think the total should be $189M and not $169M?
Your last sentence in the first paragraph seems to be cut off at “gets a lot more than”!
I’m following up on Leon’s question—have the results already been posted? If not, when will they be posted (if they will be)? I’m curious to know. Thanks!
And this thread from Dr. Eric Feigl-Ding is worrying too.
Thanks for this. This tweet from Dr. Jacob Glanville, founder and CEO of Centivax, makes me worried about this variant too:
The new B.1.1.529 strain out of South Africa has 15 mutations in the RBD where majority of neutralizing antibodies bind. The current vaccines and even Delta-based vaccines probably won’t work against this new strain. Swift, vigorous containment is needed.
There’s Otter.ai which costs $8-30/month depending on which plan you get. You can try their free plan too to get a feel of how good their transcription is.
I haven’t used rev.com compared to Otter, but I think it also takes ~1x the time of the audio to fix the mistakes of Otter.ai, which would make it similar in time-cost to fixing Rev.com transcripts. So Otter.ai might be a way cheaper option than Rev.com. And the transcripts should be ready within 30-60 minutes of you upload it, given that it’s AI-based, versus Rev, which I think is actual people typing your transcript.
Thanks for linking both of those resources! I hadn’t heard of CETF before. I’m not sure how much to trust CETF, but that’s an interesting resource. Their website led me to the New York Times’ treatment tracker though, and generally I find the NYT pretty reputable. I wonder why fluvoxamine, and to a smaller extent remdesivir, aren’t talked about a lot yet in the Philippines as having promising evidence as a treatment for COVID.
Thanks also for linking Scott’s article. I had heard of it but hadn’t read it much until today. It’s interesting that he only thinks Vitamin D has a 25% chance of being effective. I would defer to him on that, but yeah I agree with him that the benefits of taking it likely outweigh the costs.
This was really helpful Jim!
When I said public advocacy though, I mainly meant making a Facebook post communicating my beliefs about Ivermectin and whether people should take it or not, which 50-500 people in my network might see. So I could convey my beliefs honestly about Ivermectin in that post and at length, and I’m open to conveying I am uncertain about it.
That being said, from your comment, I realize that there’s a large chance that whatever beliefs I communicate about Ivermectin will be wrong, and being wrong about it publicly can tarnish my reputation. It’s also hard to communicate nuance about it, and I’m no expert here. So based on your comment, I don’t plan on making any public post about Ivermectin to my network.
Thanks for linking about UpToDate, PubMed, and medrxiv. I hadn’t known of UpToDate and PubMed, and didn’t really take note of medrxiv. I’ve also heard about and used Google Scholar before, but I forgot to use it for my recent research on COVID. I was mainly searching on Google and adding “NCBI” to the end, like “Vitamin D COVID-19 studies NCBI”.
Also, do you have any resources, articles, or books to recommend on how to get better at doing literature review / distillation work, whether on politicized/fast-moving or non-fast moving topics? I want to get better at this skill and am not sure where to start. Thanks!
Hi Raemon, the link above is broken, could you fix it?
This is my first time seeing a UI/UX design related post here on LessWrong.com. And as a UI/UX designer myself, I’m glad to see this post and that it got curated!
Thanks for linking these! I also want to highlight that Sam shared his AGI timeline in the Bloomberg interview: “I think AGI will probably get developed during this president’s term, and getting that right seems really important.”