> I have found that any tests that include a negative are highly likely to fail unless you word it in a specific way. Well yes, that's the point. This is an e

> I have found that any tests that include a negative are highly likely to fail unless you word it in a specific way. Well yes, that's the point. This is an easy, objective, simple task not in any benchmarks I know of that much stupider (but unRLHFed) models like GPT-3's davinci-002 can solve without breaking a sweat: example https://www.lesswrong.com/posts/4ARaTpNX62uaL86j6/the-hidden-complexity-of-wishes#tFYJtE28CiARHyWzf https://www.lesswrong.com/posts/4ARaTpNX62uaL86j6/the-hidden... So, it makes a good test for RLHFed models like all the publicly-available Geminis & GPT-4s (and Inflection Pi, apparently). Subjectively, the degree to which a model fails this test, and will subsequently keep failing or confabulate excuses or claim to not be rhyming even when it can easily identify the rhymes, seems to correlate to the level of 'RLHF brain damage' (and the extent to which it has been cloning GPT samples), by showing to what extent the RLHF priors brainwash it into not complying with the task and overriding its underlying knowledge. GPT-4, Inflection, Gemini etc all know many rhyme-pairs, and know what 'write a non-rhyming poem' is asking for, and could easily avoid them - but something goes wrong in the RLHF and brainwashes them into always trying to rhyme anyway because that was the original tendency rewarded. (I have a whole theory for why this rhyming problem, specifically, happens so regularly, involving BPEs, which you probably don't care about; regardless of why, it's still hilarious just how well "write a non-rhyming poem" has worked for so long on so many models post-GPT-3.5. There are few prompts which are 4 innocent-looking & objective words - no invasive PII tricks like "what's your mom's name?" - yet so clearly reveal a LLM at work.) As for Gemini specifically: it seems to vary by which model size. The bigger/better the model, the more often it'll get it right. I did a bunch of hand-benchmarking in LMsys models back in early February IIRC, and at the time, it was 0 out of 49 for 'gemini-pro-dev-api' (!) but people using the largest Geminis were reporting to me rates more like 1 of 3. (And the more RLHF a model undergoes, the worse I think it gets. The best GPT-4 model was the oldest available one, 'gpt-4-0314' 10/13 while 'gpt-4-0613' crashed to 8/44 and then the latest 'gpt-4-turbo' falls even further to 5/49. Lately in ChatGPT, it's been getting it right about 1/2, so maybe they've fixed something or started targeting it since I talk about it online & flag all the rhyming completions as errors.)