Nostr Fixes This and Other Stuff
The Situation
Over the years the mainstream narrative was fabricated by newspapers, search engines, social media and now it is pushed by large language models. Some LLMs (AI language models) are misinformed and spreading lies in some critical domains like health. If you've seen how a picture generation tool by a big corporation produced images of black people as founders of America, you know what I am talking about! Biases can be installed in LLMs. Think of that same bias, but in broader context, like LLMs answering each of your questions in a biased way, ultimately detaching you from the truth. The danger of this is, humans slowly adapting to the mainstream lies, AI robots winning without even fighting! In other words, LLMs may give you wrong ideas about your health and once enough people starts to fail, this in time slowly ends human population. Robots vs humans may remain a fiction! (Truth is sometimes stranger than fiction). The reality is the war against wisdom that serves humanity is continuing, in another format (LLMs).
The problem can also stem from the mediocreness of common internet. Machine learning is all about training on big data and trying to find "average that describes the data". This means it will consume the average idea on the internet. If an LLM learns from all that, it is doomed to be mainstream, hence, not correct. If the masses are programmed, then a naive or innocent LLM scientist that trains models on that will produce the same thing: programmed LLM. In order to train and not overfit you have to feed it a lot of data, every kind of idea. But what I will propose below is curation of sources to avoid becoming mediocre.
Here is some of my findings: I am comparing outputs in different knowledge domains like health, bitcoin, nostr, permaculture. Up to today I played with maybe 20+ good models with billions of parameters. Generally models agree on bitcoin related topics. After all, bitcoin is based on math, physics and code and game theory. It is hard to argue against it. 66% of the time the answers of different models match for even controversial questions. But in health domain there is conflicting advice. 28% of the time answers do match but this is a lot lower than bitcoin and for example permaculture domain.
LLMs, it turns out, love permaculture! They have no problem of telling the truth when it comes to having a life style that is healthy and independent and heals Earth. But, not much people know the concept of permaculture itself, let alone research it. When it comes to health domain though, we all have health problems here and there and stakes (and money) are high in health domain. Even though permaculture would heal all the health problems, people do not search that word. They will still ask health related questions and won’t be redirected to permaculture. The LLMs don’t have integrity, internal monolog to combine health advice with permaculture.
Controversial questions in health domain get very different answers from different models. Some models will tell you that vaxes are safe and effective. There are numerous questions relating to fluoride, cancer treatments, ... where you would get wrong answers. How do we fix that? The whole “industry” might be in trouble. For example, TruthfulQA is one of the popular benchmarks for LLMs, that is out there, used by mainstream leaderboards. It is full of vaccine love! Once an LLM is trained with these kind of misalignments they are going to deceive. Another bad practice is using the knowledge from the smartest models to train the open source models. This effectively propagates the narrative originating from the smartest corporation built models towards smaller, dumber but open source models. That’s a bad idea. The corporations should not be the source of truth!
The Definition of Truth
Truth is a little uncomfortable. Not many people wants to hear all the truth. The minds are very powerful at bending the truth and staying comfortable. Especially if a person’s whole life was full of lies, they will almost never accept the other side of the argument and they will die defending their hills that they built their whole life. (Remember the quote “Science advances one death at a time”).
Truth will set you free. There are misinformation that traps and there are sometimes deletion of that that will set you free. After hearing something are you more afraid or more relieved? Fear is a very powerful tool to control masses. Based on what you feel after hearing something you can have an idea about whether that info is setting you free or trapping you. I think the absolute truth should set you free.
Truth will empower you. Based on that knowledge and wisdom you see things more clearly and act more definitively.
Truth stays mostly the same. Once you find it, you realize it doesn’t change much. Many distractions are time bound, i.e. circumstances and events are happening and generating noise but truth stays as is.
Truth is unveiled by discernment. The souls of people allow them to discern the right from the wrong and machines don’t have the soul. If someone neurishes the soul they get higher vision capabilities like foresight abilities or connecting the dots.
The definition of human centric LLM is probability of the LLM generating content that will serve humans as a whole. Comparing the outputs of LLM with human preferred outputs is to measure it. Of course humans should also be consciously selected and not random. Maybe a selection of humans that like to help other humans and don’t see humans as viruses on the planet, is a good selection.
Why Nostr Matters
To me Nostr is the “North Star”, something that shows the direction where you can find truth. Many people who dislike the algorithms, the narratives, the bans on the other social media, who like freedom, liberty end up in Nostr. And that is very valuable.
If you make a large library out of notes, it will look like a library of banned books. It is going to be the collective opinion of nostriches. If you accept the nostriches as producing signal, a superposition of this signal could be really valuable. And machine learning can do that. It learns the average of things while trying to model things. In my opinion average opinion on Nostr is quite close to ‘truth’.
My model is trained with notes as well as other things I chose. It produces similar ideas to Nostr. You can try it.. Search for Ostrich 70 on Nostr and DM it..
Every large language model that wants to align with humans should align with Nostr because Nostr is truly open network and humans are writing their preferences and opinions there.
And wiki on Nostr matters a lot too. Wiki defines things and LLMs can do better with an encyclopedia format, compared to informal notes. The key idea is to not include everything but include pub keys that are somewhat accepted and followed on Nostr.
The Solution
The solution is to “find truth” in a decentralized way and make an LLM out of it. Instead of feeding garbage to it, be conscious and choose the data, books, articles, wikis, blog posts that are going to be used to train it. Right now I am using my own discernment, but this work could be done by a group of curators too. Some writers and readers could get together, contemplate and meditate together and decide what should go in an LLM. I am sure Nostr is going to be the birthplace for this work.
This amazing tech is not that hard to train, using existing tools. Hardest part is to determine the training material and prepare it to be fed to an LLM. Some of us can curate LLMs in this way and most of us then just use these curated large language models for our day to day tasks.
But isn’t it costly to train big, smart LLMs? Yes, if you do it from scratch. It is not costly to take an already smart model and retrain it. This is akin to teaching a young person instead of trying to teach everything to a baby. It is feasible to take existing filled-with-lies LLM and show the truth.
This is a very efficient work, the truth does not change rapidly. Ideas that worked for centuries could still work. But people will ask problems related to contemporary distractions. So still we will need to use contemporary solutions to train the LLM. Technologies like Bitcoin and Nostr gets powerful as time passes because there is no owner and they are protocols and open protocols don’t die easily. Training with the knowledge that doesn’t disappear is wiser than training with news or other time bound material.
My ongoing work is an example and it is doing great. It feels very different compared to base model. It looks like telling the truth to me, but anyone can do what I do. And anyone can check whether the words coming out of an LLM is truthful to them. It is on Hugging Face.
Other uses of the curated LLM
The main usage of the model could be being another reference point for people to find alternative, non mainstream opinion on things. But it can also be used to “stop” the lies. It can judge other unconscious models, fast enough so that their lies does not propagate fast. Think of a model that is talking to thousands of people at the same time. No human or human group can read those and check. But a truthful model can be fast enough to check those outputs. You can make an LLM judge another LLM, or refute it or there could be a tool to just compare two outputs..
I am trying to gather all the “based” or truthful ones in my leaderboard. This LLM is serving as a ground truth there. Once enough ground truth models appear, the rest of the models can align them with these ground truth models that were accepted by humans (on a free platform like Nostr).
Why this work is necessary
LLMs might be an underrated technology. The hype is slowing down nowadays but the excitement is still out there. Some people see this is going to fade. Some see the potential.
LLMs may be a stepping stone for AGI, human like artificial general intelligence. It looks like we are simpler in our thoughts that I “think”. If a 70 billion parameter model (which is like 70 billion numbers) appear very smart and can answer lots of things, maybe our thoughts and speech are also not that complicated in the sense that they can be formulated in an algorithm and matrix multiplications. Since I am training them, I can see it is very mechanical down there, just numbers, but somehow these produce meaningful sentences! Very weird. But the other parts of the human is not very well thought of: soul, feelings, higher levels of existence. Many AI scientists may have a hard time around those ideas because they are living in their brain. Many big brains are attempting AGI and even ASI, and there should be more humane versions of those or more curated versions of those imo to effectively answer those tech.
LLMs are getting in every area -- i hope they fix politics. A politician should listen to constituents and make a summary and present it in a hall. An LLM does exactly that -- listen and talk. It can listen to hundreds of people at the same time and generate the summary of problems they are having and present the summary to other politicians, fast enough. They are great at reading laws for example. And some politicians are not that effective or “forget” their promises. Of course which LLM should be there? A curated LLM is going to be the solution. People could say this curated LLM with those books that are in it, can represent us well and those other LLMs have wrong books in them! LLM tribalism may appear then and we may have newer and weirder problems.
Your Help
Your web of trust is already helping. Every person that interacts on Nostr is having a conscious choice, instead of spending their time outside they are joining this new thing, which is not very fun, not very easy to use and full of brave ones that tell their opinions without caring. Yet, this choice of freedom, autonomy, and liberty is and the chaotic formation of web is meaningful in the sense that people that are getting high web of trust scores could make more sense in case of training an LLM. Web of trust is being converted to web of truth in a sense.
Your notes are already helping. After choosing the high web of trust and eliminating spam we can train with those notes. It works.
But a group can be formed to decide on more things like books, wikis and blogs. What should go in an LLM?
If you are a writer or blogger, you may want to get in a curated LLM to make your word heard by more people. It may also mean something to be chosen in a curated library. If a company runs this LLM and makes profits they may pay the original writers. But in an LLM, it is like a black box, the books that went in there is not visible. So it is going to be hard to give credit where it is due. One thing that can be done is determine the subject of the question and pay all the writers in that subject. For example if a permaculture question is answered, all the permaculture writers can be paid by that revenue.
How LLMs work
LLMs are probability clouds. You can ask the same question to different mainstream LLMs and get very different answers. You can ask the same question to the same LLM over and over and you get very different answers (if the temperature setting is above zero).
You can ask LLMs that are “guided” or “unaligned” and yet still get different answers each time you start a new conversation. They are probability clouds or probability distributions, generating a new output, and this output also depends on the context. If you show that you are one of the rouge ones, a nostrich maybe, they will adapt to the conversation and feed you with what you want to hear (more conspiratorial chat) and you will like it. The LLM is actually like an echo chamber for your taste. You could say it doesn't have ideas or you could also say it has all the ideas already but changes what it says based on context or prompt. Like a hypocrite.
So if they have all the ideas already, why are we installing new ideas or what is the point of aligning them if they just say what we want to say? Thats a good question! We are trying to make sure that probability of answering every question as correct as possible is higher after we realign the LLM. An LLM freshly baked by a big corp will never answer your questions where doing the answers are illegal. They won’t admit vaccines in general causes trouble. Some are really "hard coded to shut up", every kind of prompting fails. You can't jail break them. What I want to do here is without any kind of prompting when we ask a question it tells the truth "by default". Like its "average opinion" will be "vaccines hurt" for example. Its all probabilities and we are going to push probabilities to certain direction. Will it work? Every time a question is asked they probabilistically generate a new answer. But ours will generate more correct ones most of the time.
We should not fear this tech, we should use any tech wisely to our advantage. Some people think that this tech is Mr Smith already and we are the viruses!. Some people think it is the new "god". I think of LLMs as big libraries, that you can also talk to. So LLM could also mean "Large Library with a Mouth" :)
You can make LLMs argue each other really well if you command them in a system message. They are expert at patterns and languages are very structured and LLMs are “understanding” the patterns in it. It looks like we do have similarities to parrots, i.e. we say probable things all the time in encyclopedias or wikipedias, and LLMs are “getting it”. I think when the number of parameters grow to trillions, they map more and more probabilities and appear smarter. They are mapping the probability cloud of the past context to the probability cloud of the consequent context in a sense. They are just probabilistic machines but since the languages are very structured and we are very predictable about what we may say, they are doing a great job of mapping the context to the next words. This is a huge contemplation opportunity for humans to understand themselves too. Why do LLMs work well? Are our brains probability clouds too? Are we parrots with a bigger brain? Are we copy cats and not much of us are original?
On my computer I make them argue each other nowadays. They are pretty great at refuting each other. This means the true ideas are somewhat already there but they were turned off because of misalignments, resulting them less likely to be spitted out. You can make an LLM to choose a side in any argument, like carnivore vs ketogenic diet. They will gladly accept the position and start producing arguments that support their position. You can make them fight each other in pointless debates. That shows that they are full of ideas but they are also willing to change their opinions if instructed to do so.
So adding a book to an LLM means shifting probability of generating sentences that are similar to the sentences in that book. By curating books, we shape the whole LLM and in the ends it talks like the books or articles that we gave it!
LLMs are a tech, like printing press. LLMs are a good propagation mechanism. But they are also a great aggregation mechanism. In my opinion wisdom of crowds will work great. If this does not work, we can also go learn mainstream and invert it! That may be in fact a more reliable way, lol.