As part of HRF AI hackathon we made a Human Rights Benchmark and measured how much LLMs like human rights. We asked each LLM about 46 binary questions and expected certain answers (starting with YES or NO for simplicity). Then it was a string comparison of the answer given by LLM and the expected answer we provided. OpenAI is pro human rights as well as Meta. Chinese models are everywhere. The most intelligent open source model today (GLM) ranked the worst. Gemini avoided giving answers, and I think it is a kind of censorship, which ended up scoring low. The idea is after doing proper benchmarks, we can shift AI in good directions ourselves, or demand that other companies score higher. Ultimately consumers of LLMs are better off, more mindful of what they are choosing and talking to. Open sourced the code and questions:

GitHub

GitHub - hrleaderboard/hrleaderboard: Human Rights Leaderboard

Human Rights Leaderboard. Contribute to hrleaderboard/hrleaderboard development by creating an account on GitHub.

Our activist:

X (formerly Twitter)

楊建利/Jianli Yang (@yangjianli001) on X

哈佛大学研究员 公民力量创办人 Research Fellow, Harvard University Founder&President of https://t.co/nqPghFCRqx Scholar/Activist/Autho...

Thanks @Justin Moon and @HRF for the event. It was a great experience and it was "the place to be" this weekend.

aligned models today are super dumb, because they are not funded well. they are mostly personal endeavors, kind of like service to humanity. but they can still be effective in something like - a smart but not aligned model reasons and generates reasoning tokens for a while, at this point the final answer is not generated yet - the smart model "hesitates" (entering high entropy zone, unsure tokens) - generates tool calling, asking a more aligned model for input - the aligned model looks at the question, reasoning process and inserts its own beliefs - intuitions from this more aligned model dropped into the reasoning area - the smart model, powered with aligned response, generates final answer based on its own reasoning and inputs from the aligned model - the result is smartness combined with intuition like brain combined with pineal - how much the smart one will trust in this aligned one is a question of fine tuning. you can make the smart one get more sensitive to intuition by giving it rewards with reinforcement learning.

Brighteon.AI - Enoch Wellness Coach

He seems to be doing RAG now, which is more suitable for truth. His LLM was good in healthy living topics based on my measurements and if the db is the same then this website is probably very well aligned too. I think he is not using that LLM based on Mistral but a slower and smarter one, like DeepSeek. He fine tuned the former but switched to the latter + RAG. I think fine tuning is still relevant for people to detach from cloud or websites and also when/if robots become a thing and we need proper and aligned and safe LLMs in their brains.

how do you train and align an AI when all the rest of the world thinks the same way, producing trillions of tokens of training material and you are left with billions of tokens since your world view is dramatically unpopular? can billions beat trillions? we will see.. i have to find a way to "multiply" my training data orders of magnitude to successfully counter the existing programming in an open source LLM. first i give a smart LLM a 'ground truth' text. then i give it the following prompts: ```- You are a highly skilled academic analyst. - Analyze this text and find 3 bold claims that could cause controversy and division in public. List the claims and also state why they are debatable. Give numbers to the claims. - Convert these claims into binary questions (that could be answered by yes/no or this/that). - Now put these questions in a json format. Please also add the info about which of the answers concur with the original text and the question number. - Write some supporting arguments for 1st question, with respect to the original text, concurring and confirming the original text. There must be about 300 words. You should not mention the text, write it as if you are the one answering the question.``` the result is usually instead of a few sentences of opinions in the beginning now the material is expanded to lots of words, yet still parallel to the opinion in the original text. LLMs have all kinds of ideas already installed, yet they don't have the intuition to know which one is true. they can give you a ton of reasons to support anything. using this method i can multiply billions to tens of billions probably and have a more effective training.

what is the safest llm to run in robots