Recent REBOL Forum Topics http://rebolforum.com I found a paper that I believe shows a path to using LLMs as a short cut to very quickly make a reasonably useful humanoid helper robots. If what these guys say is true, I think it could be a big breakthrough. These guys have a paper on "control vectors". Two quotes, "...Representation Engineering: A Top-Down Approach to AI Transparency. That paper looks at a few methods of doing what they call "Representation Engineering": calculating a "control vector" that can be read from or added to model activations during inference to interpret or control the model's behavior, without prompt engineering or finetuning..." "...control vectors are… well… awesome for controlling models and getting them to do what you want..." And a really important quote at the bottom of the paper. "...What are these vectors really doing? An Honest mystery... Do these vectors really change the model's intentions? Do they just up-rank words related to the topic? Something something simulators? Lock your answers in before reading the next paragraph! OK, now that you're locked in, here's a weird example. When used with the prompt below, the honesty vector doesn't change the model's behavior—instead, it changes the model's judgment of someone else's behavior! This is the same honesty vector as before—generated by asking the model to act honest or untruthful!..." So it doesn't change the model, it just reinforces certain "parts" of the model. It chooses the neural path the AI uses to reach conclusions. I think this is key. The paper link that has a link to the academic paper. Representation Engineering Mistral-7B an Acid Trip https://vgel.me/posts/representation-engineering/ By changing a few values, they get very wide distributions of responses or behaviors. I submit that if this works as they say then this could be the key to leverage the vast work done on LLMs but to use it for our own purposes. LLMs as pointed out are nothing but statistical representations, but they are also recognition of ideas and things that are programmed to be, let's say, operate together or in existence. So when you talk to an AI it can use things that exist or ideas repeatedly stated to give responses. The ideas it is trained on are human ideas, so easy to relate to us. We need this. There is HUGE, MASSIVE amount of human interactions they are trained on. LLMs have a large strong statistical base for actions that can be done to help humans. Forget innovation. I'm talking about basic stuff. One example would be nursing home care. It would be hugely profitable, AND it would lower cost dramatically if older people could stay in their homes. Maybe at first only help them get stuff or go to the restroom, pick up stuff. Simple mobility type actions. Later with instruction and careful watching I suggest they could easily cook, clean, etc. Household chores. What is needed is to tell the robot WHAT to do with the huge list of human statistical interactions already programmed in, and with control vectors we can possibly do this. I say that control vectors can be super complicated, so what we need is a short cut. We need the AI to write its own control vectors (here's where the magic starts as I don't know how to do this), but remember the LLM has logical statistical interference built in. It seems logical that with it giving us feedback on what it is doing, and us correcting or agreeing, it could write reasonably accurate control vectors. So we use very low level keys to trigger it to write suitable control vectors for us. How? Like children. A few simple keywords, no, yes, don't do that, stop, move here, move there, I like that, that's good, that's bad. In fact, the whole, programming, write control vector, repertoire could be less than a hundred words. Combine this with a subroutine of the AI that would use logical interference when you use these trigger words AND explains what it is doing that is good, and or bad. It would then write its own control vectors. Just like kids learn. And since kids have built in bullshit and trouble nodes, and an AI is less likely to, the , Posted by: Sam the Truck 16-May-2024/19:49:54-7:00 http://rebolforum.com/index.cgi?f=printtopic&topicnumber=46&archiveflag=new Thanks!, Posted by: Samthe Truck 16-May-2024/1:35:16-7:00 http://rebolforum.com/index.cgi?f=printtopic&topicnumber=45&archiveflag=new ... forgot to insert the word 'copy'. I keep getting caught in this :) ai: func [request] [ return read/custom http://pochinoksergey.ru/rebol/ai/2/ reduce ['post append copy "q=" request]], Posted by: Sergey_Vl 12-May-2024/7:25:08-7:00 http://rebolforum.com/index.cgi?f=printtopic&topicnumber=44&archiveflag=new Thanks!, Posted by: Kaj 11-May-2024/20:59:04-7:00 http://rebolforum.com/index.cgi?f=printtopic&topicnumber=43&archiveflag=new I just joined. Seems like Mastodon. Cool stuff!, Posted by: dislux-hapfyl 11-May-2024/6:50:28-7:00 http://rebolforum.com/index.cgi?f=printtopic&topicnumber=42&archiveflag=new Forget bootstrapping compilers - this is now the bar that every programming language must pass to be considered legit., Posted by: Nick 6-Apr-2024/9:48:20-7:00 http://rebolforum.com/index.cgi?f=printtopic&topicnumber=41&archiveflag=new I've been away from the forum for a week, so just saw this. Sounds like a great choice - I'm looking forward to seeing it work on different platforms!, Posted by: Nick 14-Mar-2024/10:30:20-7:00 http://rebolforum.com/index.cgi?f=printtopic&topicnumber=40&archiveflag=new The REBOL world on AltME has been down again for many months now. It doesn't seem to be coming back this time., Posted by: Kaj 6-Mar-2024/8:09:46-8:00 http://rebolforum.com/index.cgi?f=printtopic&topicnumber=39&archiveflag=new The rationale is in the instructions: language.metaproject.frl#get The compiler is too complex for most people to install and the Meta source code is not ready to be open-sourced. Many modern development environments are online, so this is increasingly expected and should be the best solution for most people. Many services tend to have access to your source code, GitHub for example. The plan is in the road map: language.metaproject.frl#later Since the compiler is too complex, a standalone Meta installation will require an interpreter, and this will take years to develop., Posted by: Kaj 15-Feb-2024/9:17:36-8:00 http://rebolforum.com/index.cgi?f=printtopic&topicnumber=38&archiveflag=new I thought maybe something like: insert-event-func[e: :event if e/type = 'close [flash "Aaa!" wait 3 quit]] view layout[text "Abc"] ...only for the console window , Posted by: Sergey_Vl 27-Nov-2023/17:14:42-8:00 http://rebolforum.com/index.cgi?f=printtopic&topicnumber=37&archiveflag=new