Produced as part of the SERI ML Alignment Theory Scholars Program - Summer 2023 Cohort Thanks to @NicholasKees for guiding this and reading drafts, to @Egg Syntax @Oskar Hollinsworth, @Quentin FEUILLADE--MONTIXI and others for comments and helpful guidance, to @Viktor Rehnberg, @Daniel Paleka, @Sami Petersen, @Leon Lang, Leo Dana, Jacek...
Produced as part of the SERI ML Alignment Theory Scholars Program - Summer 2023 Cohort Thanks to ev_ and Kei for suggestions on this post. LLMs can do many incredible things. They can generate unique creative content, carry on long conversations in any number of subjects, complete complex cognitive tasks,...
Produced as part of the SERI ML Alignment Theory Scholars Program - Summer 2023 Cohort
Thanks to @NicholasKees for guiding this and reading drafts, to @Egg Syntax @Oskar Hollinsworth, @Quentin FEUILLADE--MONTIXI and others for comments and helpful guidance, to @Viktor Rehnberg, @Daniel Paleka, @Sami Petersen, @Leon Lang, Leo Dana, Jacek Karwowski and many others for translation help, and to @ophira for inspiring the original idea.
I wanted to test whether GPT-4’s capabilities were dependent on particular idioms, such as the English language and Arabic numerals, or if it could easily translate between languages and numeral systems as a background process. To do this I tested GPT-4’s ability to do 3-digit multiplication problems without Chain-of-Thought... (read 10547 more words →)
I think you're probably right, my feeling is that organic pro-regime internet campaigns are possibly more important than traditional censorship. The PRC has been good at this and I've also been worried about how vocal Hindutva elements are becoming.
I don't know that we've yet found the optimal formula for information control (which is a good thing) and I remain a little agnostic on the balance between censorship and propaganda. This post focused on old-style censorship because it's better documented, but a contemporary information control strategy necessarily involves a lot more.
I've so far been skeptical of a lot of misinformation narratives because I don't think fake news articles for example are labor constrained,... (read more)
I think these are great points. Entirely possible that a really good appropriately censored LLM becomes a big part of China's public-facing internet.
On the article about Iowa schools, I looked into this a little bit while writing this and as far as I could see rather than running GPT over the full text and asking about the content like what I was approximating, they are instead literally just prompting it with "Does [book X] contain a sex scene?" and taking the first completion as the truth. This to me seems like not a very good way of determining whether books contain objectionable content, but is evidence that bureaucratic organs like outsourcing decisions to opaque knowledge-producers like LLMs whether or not they are effective.
Feel free to DM. I think you're absolutely correct these systems will eventually be used by intelligence agencies and other parts of the security apparatus for fine-grained targeting and espionage, as well as larger scale control mechanisms if they have the right data. This was just the simplest use of the current technology, and it seems interesting that mass monitoring has still been somewhat labor-constrained but may not remain so. These sorts of immediate concerns may also be useful for better outreach in governance/policy discussions.
This was a post I wrote during SERI MATS and not my main research. Some of the folks working on hacking and security are more explicitly investigating the potential of targeted operations with LLMs.
Produced as part of the SERI ML Alignment Theory Scholars Program - Summer 2023 Cohort
Thanks to ev_ and Kei for suggestions on this post.
LLMs can do many incredible things. They can generate unique creative content, carry on long conversations in any number of subjects, complete complex cognitive tasks, and write nearly any argument. More mundanely, they are now the state of the art for boring classification tasks and therefore have the capability to radically upgrade the censorship capacities of authoritarian regimes throughout the world.
In totalitarian government states with wide censorship - Tsarist Russia, Eastern Bloc Communist states, the People’s Republic of China, Apartheid South Africa, etc - all public materials... (read 2382 more words →)