Hi all, I haven’t been around much in the last year or partly because I have been finishing my masters in data analytics. At some point I started specializing into building ChatGPT apps to help businesses, and I decided to create a actualism chatbot for fun.
This is a ‘retrieval augmented generation’ chatbot, that takes the users query, finds relevant snippets from a knowledge base, and stitches together an answer using those snippets as context/knowledge. This chatbot uses all the .html files in the actualfreedomlite.zip download from this page as its knowledge base:
It is powered by chat gpt 3.5 and makes a lot of mistakes, the version powered by gpt 4 is better but way more expensive to run. Feel free to check it out here: https://actualismchat.streamlit.app/
Oh, interesting. I was thinking about setting up a RAG myself, but I wasn’t sure the zip was enough and I didn’t want to scrape the entire website (bandwidth, etc) to get better context. I also wasn’t sure how useful it would be in the end, but looking for quotes might not be a bad case. GPT still hallucinated too much for my liking for general conversations.
Have you tried it with any of the Mistrals instead? I’m thinking using another model without the GPT safeguards/mixups might yield a better result, but if only using it for lookups it might not matter.
This is my main concern with this — actualism really demands exacting precision. The human condition is rather murky, particularly whilst being it, and it’s important to be discerning. I think the chatbot is likely to rephrase or reword things and leave one with the wrong impression.
For example it already referred to feeling good as ‘positive feelings’, while that refers to ‘good’ feelings not feeling good. And it talked about the “present moment” which isn’t quite right — there’s no “present” moment as distinct from a “past” or “future” moment, there’s only one moment, this one!
Decided to try it, the first response seemed decent although it is a bit confusing that it makes it sound like the state of actual freedom entails consistent altruistic self-sacrifice, which it doesn’t (Richard has said he is not altruistic):
The second response seems ok-ish. It brings up PCEs as being essential, but doesn’t mention the actualism method at all, which would be the most relevant answer here:
Unfortunately it appears to be more of an Actualism Misinformation Bot than an Actualism Chatbot!
Someone uncritically using it as a source of information would be led completely astray.
I will say that an actually functioning and precise chat-bot, that doesn’t give wrong answers, seems like it would be really useful. Basically a much better version of a search engine. But I would recommend against using it in the state it’s in now.
I also wonder if an LLM framework like ChatGPT can ever really succeed at this task. Every implementation of them has the problem of hallucinating (confidently making up wrong answers). They are trained on large swaths of data, all of which derives from feeling-beings well within the human condition. Although clearly training on the AFT site helped, its basis is still the probably billions or even trillions of words that it trained on prior to being fed the AFT site, and it will pick among that data as its most likely responses.
I haven’t messed with open source models much in general but I do think you are right that the GPT guardrails interfere with Actualism stuff, Kuba’s is a good example lol. For most use cases I work with the guardrails are a good thing but here definitely a problem. The OpenAI assistant API v2 also implements a ton of RAG best practices out of the box which is why I used it.
I agree, it makes lots of mistakes, and as I said I think its best use right now might be just looking for obscure pages/quotes that keyword search struggles to find.
I also agree with your point that it is trained on trillions of words that are definitively opposed to actualism, so any time it tries to fill in the gaps it fills them in with bullshit lol.
It could be a case where fine-tuning would be useful, since we have recorded conversations I could potentially try creating prompt response pairs out of people’s questions and actually free/virtually free people’s answers… might be a fun exercise. I still doubt it will be as good as you would hope for a long time.
For the time being I’ll add a disclaimer and make the system prompt a little more strict about not filling in gaps and sticking with the source text.
I’ll test that out yea, curious to see if it starts struggling to perform with a big system prompt, but it may be better.
Another option I’m considering is a two step approach, where step one creates the responses we are currently seeing, then step two references the cited documents and pulls the direct quote snippets that most closely correspond to the initial response, and then replaces the initial response with direct quote snippets.
I put the 180 Degrees Opposite page into the system prompt, and I think it led to another improvement.
It seems like a bigger system prompt full of careful distinctions leads to it looking for more sources and writing more complete answers, definitely filing that away for future applications.
I haven’t tinkered with any models for a month or two (had to move to physical data centers instead ) so my skills have essentially rotted on the vine, but I’m pondering if even more context would yield better results. I think Gemini still had the largest context of all models a month ago - would it be possible to use that one? I didn’t like it (speaking of safeguards) but maybe with enough context it can be brainwashed.
I’d definitely be interested in seeing what Gemini makes of actualism if I have some time, its 1 million token context window could handle probably at least the entire Richard section of the site
Updated the chatbot to use the entire website (I scraped it slowly over night and added it to knowledge base)
Also I have done some work on the prompt engineering side to try and keep it in line lol, maybe a little better now still not perfect.
I do eventually want to add a ‘direct quote mode’ option, that will respond with only relevant direct quotes and links, but it might be a little while before I have the time for that.
Presently testing the chatbot using the new chatgpt-4o, huge improvements for sure, though the cost is 10x from using 3.5. feel free to play with it for a while i have loaded it up with 10$ to burn through
how can i start practicing without pure intent, i haven’t had a pce so i dont know how to access pure intent which seems crucial?
To start practicing actualism without having had a Pure Consciousness Experience (PCE) and thus not having direct access to pure intent, you can focus on the following steps:
Set Short-Term Goals: Begin by setting small, manageable goals to break lifelong habits and beliefs. For example, aim to not feel miserable about the weather, not get annoyed by other people’s driving, or not get frustrated if a waiter is slow. Everyday life is full of opportunities to cease being antagonistic and to stop feeling miserable about being here (www.actualfreedom.com.au/actualism/peter/list-af/corr32b.htm) (www.actualfreedom.com.au/actualism/peter/list-af/corr32b.htm) .
By following these steps, you can begin practicing actualism and work towards experiencing a PCE, which will then provide a direct experience of pure intent.
The utility for me is being able to pull up whatever material is relevant for me at the time immediately. And since it’s a quote bot, I know the words are from AFT and the provided links allow me to explore further. Though it does add some of its own understanding
Yea thanks for clarifying that, since I updated it a bit it often responds with direct quotes, but still also includes its own interpretation/summarization of its sources. However it did get quite good at generally using direct quotes through simple prompt engineering.
what is a pce
A Pure Consciousness Experience (PCE) is described as an experience marked by the absence of the ‘self’ or identity, allowing one to perceive the actual world directly and clearly. Here are some key descriptions and definitions of a PCE:
Touchstone for Actual Freedom: “A PCE is the one and only experience that makes me aware of and allows me to experience the actual world that lies hidden beneath the elaborate, confusing and ever-changing chimera created by the identity inside this body and a PCE is the one and only experience that can clearly guide me towards an actual freedom from the human condition” (www.actualfreedom.com.au/actualism/vineeto/list-af/corr16e.htm) .
These descriptions highlight the essence of a PCE as a profound and clear experience of the world, free from the distortions of the self, albeit temporary in nature.
) so my skills have essentially rotted on the vine, but I’m pondering if even more context would yield better results. I think Gemini still had the largest context of all models a month ago - would it be possible to use that one? I didn’t like it (speaking of safeguards) but maybe with enough context it can be brainwashed.
To start practicing actualism without having had a Pure Consciousness Experience (PCE) and thus not having direct access to pure intent, you can focus on the following steps: