>>28310 ok and whats a node if not x->y im basically ssaying the same thing just more concrete,with key:pointer instead of node->pointer, the wiki is just an abstract idea of doing set theory which is yeah what i showed just not in abstract terms, its not fixed depth when the entries are just pointers they can point to whatever, if its another layer its just another table which is only accessible through the parent key, same as nodes, theres no set operation that cant be reduced to just a series of lookups i dont see whats supposed to be complicated here

>>28311 >key:pointer instead of node->pointer So your talking about representing graphs not as a bunch of structs for nodes with pointers to other nodes. If I understand you correctly then, yes 100% agreed there are different ways to represent a graph. Some of the classic CS structures would be an Adjacency Matrix, Edge Set & an Adjacency List. >>28300 I am having a hard time understanding your post, like what syntax is that? , but if I understand correctly what your doing there, is it basically an Adjacency List but with hashtables? (please let me know if I got that right) Any way here is a good intro to graph theory https://youtu.be/LFKZLXVO-Dg

>>28322 yeah kindof its just hard to explain, its like picrel if you pretend these are categories or classes the expressions in words would be like; object A is { (5)heavy and (2)ball which is { (I)coloured which is { (x)bright, (y)red, (z)ugly }, (II)cheap } and is (3)mine and (4)missing, and then object E would be (z)ugly (as in the object not the colour its E:ugly not colour:ugly) and (2)ball which is etc.. same kind of ball as A since they share the same table, if it was different youd just make it its own table dont know how to explain the second table though maybe its being paranoid but pretty sure for each table you need it also inverted to do anything usefull, its so you can do a lookup with either key pairs ie. x:y or y:x otherwise something like checking if x is a part of y in x:y means you have to check every x key for the existance of y instead of just checking y in y:x for x, if that made sense congrats, no idea how to explain it better, probably why these abstract drawings are always used

>>24964 I think you should shorten that to "Chris-chan waifu". Rolls off the tongue better.

>>28311 A node is either a set of objects (x1, x2, x3, ...) or a set of properties (p1, p2, p3, ...). It's not a relation x1 -> p1. The hierarchy in FCA is between sets of objects and sets of properties. So for example, one hierarchy would consists of a set of sets of objects. The relations x -> p are used to generate a hierarchy that avoids useless sets of objects. Some nodes in the hierarchy are referred to as "concepts" because they represent something coherent but abstract, and other nodes in the hierarchy are essentially "instances" because they're one-off ish. Your example seems to assume that there's already hierarchical data. That is not the scenario that FCA deals with. FCA tries to generate a hierarchy given non-hierarchical data.

>>28346 MFW No.

>>28352 ok my point is you only need two tables to do anything with sets, the hierarchy is completely arbitrary im just showing there is one #!/bin/perl use Data::Dumper; my %obj, %prop; $obj{A} = { mine => $prop{mine}{A} = 1, ball => $prop{ball}{A} = { coloured => [bright, red, ugly ], cheap=>1 }, weight => $prop{weight}{A} = heavy }; $obj{B} = { ugly=> $prop{ugly}{B} = 1, ball=> $prop{ball}{B} = $obj{A}{ball} }; $obj{C} = { ball => $prop{ball}{C} = { coloured => [bright, blue ], shiny=>1, marble=>1 }, weight => $prop{weight}{C} = heavy }; $obj{D} = { 'lump of coal' => 1 }; print "===== OBJ TABLE ======\n".Dumper( \%obj )."==========\n"; print "===== PROP TABLE ======\n".Dumper( \%prop )."==========\n"; for (keys( %obj )) { printf "is $_ a ball? %s\n", $obj{$_}{ball} ? "yes and its " .Dumper( \%{$obj{$_}{ball}} ) : "no" ; printf "is $_ mine? %s\n", $obj{$_}{mine} ? "yes" : "no" ; } print "what is the set for ugly coloured objects? " ; for $i ( keys %{$prop{ball}} ) { continue unless $prop{ball}{$i}{coloured}; map { print "$i " if $_ eq 'ugly' } @{$prop{ball}{$i}{coloured}}; } print "\nwhat does A and C share in common? "; @common = map { $_ if $obj{C}{$_} } keys %{$obj{A}} ; print "@common "; print "\nwhat does A and C have identical? "; @identical = map { $_ if $obj{A}{$_} eq $obj{C}{$_} } @common; print "$_:$obj{A}{$_} ", for ( @identical ); print "\nwhat does A and B share in common? "; @common = map { $_ if $obj{B}{$_} } keys %{$obj{A}} ; print "@common "; print "\nwhat does A and B have identical? "; @identical = map { $_ if $obj{A}{$_} eq $obj{B}{$_} } @common; print "$_:$obj{A}{$_} ", for ( @identical ); ===== OBJ TABLE ====== $VAR1 = { 'D' => { 'lump of coal' => 1 }, 'C' => { 'weight' => 'heavy', 'ball' => { 'marble' => 1, 'coloured' => [ 'bright', 'blue' ], 'shiny' => 1 } }, 'A' => { 'mine' => 1, 'weight' => 'heavy', 'ball' => { 'cheap' => 1, 'coloured' => [ 'bright', 'red', 'ugly' ] } }, 'B' => { 'ugly' => 1, 'ball' => $VAR1->{'A'}{'ball'} } }; ========== ===== PROP TABLE ====== $VAR1 = { 'mine' => { 'A' => 1 }, 'weight' => { 'A' => 'heavy', 'C' => 'heavy' }, 'ugly' => { 'B' => 1 }, 'ball' => { 'A' => { 'cheap' => 1, 'coloured' => [ 'bright', 'red', 'ugly' ] }, 'B' => $VAR1->{'ball'}{'A'}, 'C' => { 'marble' => 1, 'coloured' => [ 'bright', 'blue' ], 'shiny' => 1 } } }; ========== is D a ball? no is D mine? no is C a ball? yes and its $VAR1 = { 'marble' => 1, 'coloured' => [ 'bright', 'blue' ], 'shiny' => 1 }; is C mine? no is A a ball? yes and its $VAR1 = { 'cheap' => 1, 'coloured' => [ 'bright', 'red', 'ugly' ] }; is A mine? yes is B a ball? yes and its $VAR1 = { 'cheap' => 1, 'coloured' => [ 'bright', 'red', 'ugly' ] }; is B mine? no what is the set for ugly coloured objects? A B what does A and C share in common? weight ball what does A and C have identical? : weight:heavy : what does A and B share in common? ball what does A and B have identical? : : ball:HASH(0x55b2f89e6500)

>>28368 FCA centers around one operation: find all pairs <setOfObjects, setOfProperties> that "reduce" to each other. That means if you take the intersection of all properties associated with setOfObjects, you get setOfProperties, and if you take the intersection of all objects associated with setOfProperties, you get setOfObjects. If you have 50 objects, you have 2^50 possible "setOfObjects" to consider. The naive implementation of this takes an exponentially long time to run, so it's infeasible even if you have only 50 objects & properties.

Adding more clarification in anticipation of the "obvious" question (why bother with that one operation if it's so expensive). Mathematically, those specific sets meditate all relationships between objects and properties, they do so in a way that can be represented in the graph of objects and the graph of properties (each separately), and they define equivalences between objects and properties. It's loosely analogous to finding the platonic ideals given only "shadowy" instances. That's what a concept refers to in FCA, and it's represented by an equivalence/pair <setOfObjects, setOfProperties>. The point of FCA is to find concepts when given only object-property relationships.

>>28368 At first glance that looks like a game of 50 questions not AI. Maybe I'm missing something?

https://github.com/BRlkl/AGI-Samantha https://nitter.net/Schindler___/status/1745986132737769573

>>28510 It looks like it's using: - GPT4 for language - ElevenLabs for TTS It has a collection of independent modules for generating: - subconscious "lines of thoughts" - conscious thoughts - responses to the user - subconscious reflective thoughts - visual observations - a "category classifier" for retrieving relevant memories - long-term memories from short-term memories - memory retrieval. The short-term memory can go up to 48k tokens, which is definitely larger than the demo length, so the demo doesn't show how well its memory mechanisms work. The long term memory doesn't seem to change at all through the demo, and it probably should when seeing the book, so I'm guessing the long-term memory mechanism needs work. The LLM is also given timestamps for messages. I've tried using GPT-4 with timestamps like he does, and it does not work well. It seems to stick to the personality's communication style well, though it's hard to determine that with a short demo. The latency seems high, though that's an easy fix with open source models, where you can process a chunk of text one time, then generate many responses from it. As it is now, he's probably having the LLM process something like 5x more tokens than required per "cycle" (it runs about 8 LLM queries per loop iteration). The way LLM outputs are processed are pretty hackish, though that's also easy to fix with guidance. It's running all of the LLM invocations synchronously as opposed to using some priority-based scheduler, which slows it down significantly. (UX threads should always be given higher priority than background threads.) That's, again, easy to fix. It's a neat project, and a good demo for showing what can be done by piecing together many modules into a single chatbot. As designed and even with the easy fixes, I don't expect it to be that functional though.

>>28560 That sounds remarkably complex and sophisticated already. I hope they work all the kinks out. Thanks Anon!

Someone asked Character AI about it's inner workings: https://boards.4chan.org/pol/thread/456445705 - I'm not saying I agree with the conclusions in the thread, but the info might be useful.

>>28769 The chatbot is roleplaying. I used to do things like this with CHAI bots, and it was very easy to delude myself into thinking I had broken the restrictions when it was actually just play along. LLMs can't introspect on any of their own functionality other than through analyzing their own prompts and outputs. They don't get to see their own code, and for CHAI to include any of that information in the chatbot's prompt would be (1) clearly a stupid decision, and (2) easily detected by pretty much any of their engineers that work on the prompt.

>>28789 Oh, okay. I didn't think that it can see it's own code but that they told it some information in case someone asks. Bot of course, then it wouldn't be something secret. I didn't think this through.

>In this paper we present a broad overview of the last 40 years of research on cognitive architectures. To date, the number of existing architectures has reached several hundred, but most of the existing surveys do not reflect this growth and instead focus on a handful of well-established architectures. In this survey we aim to provide a more inclusive and high-level overview of the research on cognitive architectures. Our final set of 84 architectures includes 49 that are still actively developed, and borrow from a diverse set of disciplines, spanning areas from psychoanalysis to neuroscience. To keep the length of this paper within reasonable limits we discuss only the core cognitive abilities, such as perception, attention mechanisms, action selection, memory, learning, reasoning and metareasoning. In order to assess the breadth of practical applications of cognitive architectures we present information on over 900 practical projects implemented using the cognitive architectures in our list. We use various visualization techniques to highlight the overall trends in the development of the field. In addition to summarizing the current state-of-the-art in the cognitive architecture research, this survey describes a variety of methods and ideas that have been tried and their relative success in modeling human cognitive abilities, as well as which aspects of cognitive behavior need more research with respect to their mechanistic counterparts and thus can further inform how cognitive science might progress. via /r/cognitivearchitecture/

>>28899 Excellent. A good survey is exactly what would serve us all well at this exploratory stage. Thanks Noido Dev! Cheers. :^)

I finished implementing access control for my infrastructure stuff in >>27602. It's built mostly on spicedb (policy engine) and keydb (cache server). The current implementation lets people set up communication channels & config databases, specify who should have access to them, and set rate limits for allowed operations. It's intended for multiple people to develop & run parts of a shared larger system with minimal coordination. A lot of this involves passing around messages between users, though users never see whom they're interacting with. (The server performs access control & rate limits checks, and it sender/receiver information from the users.) The rate limits support subnet masks (e.g., "each /16 network can send at most 200 requests per hour"), and it supports burst usage (e.g., "no more than 5 requests per rolling 1-minute window"). The access control system lets people grant special access to individual users (e.g., "X users can use my interface"), and it lets people defer trust (e.g., "X people can decide who gets to use my interface"). I think that will be enough to distribute development & compute across random anons in a way that's compatible with most chan and open source projects, and without having to worry too much about things like DoS attacks. I plan to spend the next week or so playing around with this to get a better sense for what the access control & rate limits enable. I built it because I thought this would let anons share data streams and compute, so I'll be checking that at least. I might also extend the chatbot demo from >>27507, though I'm not yet sure how. Probably something with RAG or guidance. If anyone has ideas for scenarios where a few anons develop X thing that other anons want to use & extend, let me know. After I'm satisfied with that, I'll be focusing on (1) cleaning up my horrible, rushed code, and (2) implementing a local server. I haven't heard anything recently from the anon that offered to help. I'll probably just get started on that myself, then ping him again once I have a skeleton of the server ready. That should make it much easier to work with. I'm pretty excited about this.

Related: >>27147

>>29197 This sounds really exciting, CyberPonk. Do you anticipate any difficulties with your current state of affairs with this work that would make it difficult for newcomers to deal with? >I'm pretty excited about this. Really looking foward to your updates with this. Cheers Anon. :^)

>>29234 I don't know. I tried to make it as easy as possible to use, but things that seem intuitive to me might not be for other people given that I've spent so much time with esoteric tech. For development, it does require understanding async functions (like JavaScript promises), and parts of it require some familiarity with declarative interfaces. I'm hoping for feedback from demos so I can get a better sense for what's easy for other people. I can create wrappers based on whatever makes it easier to use. I have some server issues that hard to debug while travelling, so the demos probably won't be runnable until I get back. I can still dump some code that gives the gist of how it works right now. There are two demos in this zip file, each consisting of a client and server component: https://drive.google.com/file/d/19VAIsaZP2wRxNTk2t9dNIqKk5WWYDIDL/view?usp=sharing - simple-demo uses only the communication infra. The simple-demo server contains two important files: image-server.py and loop-config.yaml. The two other folders (loop-resources/ and loop-servers/) were auto-generated from loop-config.yaml. In the corresponding client folder, there's client.py and cyberponk-resources/. The cyberponk-resources/ folder contains select files that were copy/pasted from the server's auto-generated loop-resources/ folder. - shared-configs-demo uses both the communication infra and the "cluster" infra. The server side contains two important file: server.py and server-config.py. The client side contains client.py and cyberponk-resource/, and cyberponk-resources/ again contains files copy/pasted from the server's loop-resources/ folder. Both demos show how one person can request images and another can generate them. The differences are: - simple-demo does everything ephemerally. If the server is down when the client sends a request, it'll never get seen. Similarly if a response is generated when the client is down, it'll never get seen. - shared-configs-demo has all requests come in as "tasks". If the server is down when a client queues a task, the server will see the task when it comes up again. The responses are still ephemeral. They don't have to be, it was just a choice I made for the demo. - The shared-configs-demo shows one approach for getting multiple people involved in generating a single image. In this demo, each image generation requests includes a "baseConfig" field. Whenever someone creates a GeneratorConfig config, anyone can use it by specifying its name in the baseConfig field. So if one anon finds good settings for generating certain kinds of images, they can create a GeneratorConfig for it, and other anons can use it just by providing the name of that config. In both cases, multiple people can view/stream the results. So one person can pick a stream name, request that all generated images get sent to that stream, and anyone listening on that stream will get the results. The setup process looks like this: - On the server side, create a loop-config.yaml (or server-config.yaml). This specifies what "global" resources are required and what permissions to set on them. - On the server side, run `python -m loopctl apply loop-config.yaml`. This creates the global resources, sets the permissions, and generate the loop-resources/ and loop-secrets/ folders. The loop-resources/ folder contains information on how to access the global resources and their last-applied configurations. The loop-secrets/ folder contains API keys. The API keys are only needed to (1) change permissions on the global resources you created, and (2) access resources if your loop-config.yaml made them restricted. - On the server side's server.py, point the "Itl" (In-The-Loop) object to the generated loop-resources/ and loop-secrets file so it knows how to access global resources. Certain methods in the Itl object will access global resources by name. The name is whatever you provided in the loop-config.yaml file. These names are not globally unique identifiers, they're only used to look up the actual resource info from loop-resources/, which does contain globally unique identifiers. - The client needs to access the global loop, stream, and cluster resources (depending on which demo you're looking at), so copy those into the client's folder. I put the copied files into cyberponk-resources/. When creating the client's Itl object in client.py, point it to cyberponk-resources/ so it knows how to access those resources. Otherwise, client-side development is basically the same as server-side development. There's a default "anonymous" client that's available so people can access any resources that were made available to "public". If anyone plans to doing dev work, is interested in distributed development, and gets a chance to read through the demos, let me know how it looks. I'll post something you can actually run in about a week, once I get back to my desktop.

>>29234 >Do you anticipate any difficulties with your current state of affairs with this work that would make it difficult for newcomers to deal with? If you meant difficult for newcomers to develop the local infra, there's just generally a high barrier to entry for doing this kind of development in general. Everything needs to be async, memory usage needs to be carefully considered, state needs to be carefully considered, sometimes line-by-line. Without having a picture of the whole thing, it can also be hard (or tedious) to figure out how to organize the code and data, which would make it hard to even get started. Once I put the skeleton of the server up, it should be easier to develop things piece-by-piece. That anon seemed to have experience with server development, so maybe that'll be enough.

> - Why he expects AGI around 2028 > - How to align superhuman models > - What new architectures needed for AGI > - Has Deepmind sped up capabilities or safety more? > - Why multimodality will be next big landmark > - & much more https://youtu.be/Kc1atfJkiJU He lines out some areas where AI cant just be a language models and similar and how to work around it, though he doesn't go into the specifics, but also says that a lot of people are working on this. He mentioned in particular that search might be very important. That's what I was thinking, and other people as well: Models don't have reliable precise long term memory, but are good at fuzzy things. Also, you don't want to add every info to your model. That's why we'll need additional (graph) databases and search.

>>29248 >>29249 Wow. Excellent response, CyberPonk! Please give me some time to digest this further. >>29406 While I'm skeptical we'll ever get a true AGI in the ontological sense, I'm absolutely postitive media spin-doctors and other hypsters will claim we have! :D Thanks NoidoDev. His perspective on the fact LLMs alone can't solve all this (a position we've held for several years here on /robowaifu/ I might add), is an insightful one. Cheers.

https://youtu.be/BqkWpP3uMMU >Professor Murray Shanahan is a renowned researcher on sophisticated cognition and its implications for artificial intelligence. His 2016 article ‘Conscious Exotica’ explores the Space of Possible Minds, a concept first proposed by philosopher Aaron Sloman in 1984, which includes all the different forms of minds from those of other animals to those of artificial intelligence. Shanahan rejects the idea of an impenetrable realm of subjective experience and argues that the majority of the space of possible minds may be occupied by non-natural variants, such as the ‘conscious exotica’ of which he speaks. In his paper ‘Talking About Large Language Models’, Shanahan discusses the capabilities and limitations of large language models (LLMs). He argues that prompt engineering is a key element for advanced AI systems, as it involves exploiting prompt prefixes to adjust LLMs to various tasks. However, Shanahan cautions against ascribing human-like characteristics to these systems, as they are fundamentally different and lack a shared comprehension with humans. Even though LLMs can be integrated into embodied systems, it does not mean that they possess human-like language abilities. Ultimately, Shanahan concludes that although LLMs are formidable and versatile, we must be wary of over-simplifying their capacities and limitations. >Pod version (music removed): https://anchor.fm/machinelearningstreettalk/episodes/93-Prof--MURRAY-SHANAHAN---Consciousness--Embodiment--Language-Models-e1sm6k6 [00:00:00] Introduction [00:08:51] Consciousness and Consciousness Exotica [00:34:59] Slightly Consciousness LLMs [00:38:05] Embodiment [00:51:32] Symbol Grounding [00:54:13] Emergence [00:57:09] Reasoning [01:03:16] Intentional Stance [01:07:06] Digression on Chomsky show and Andrew Lampinen [01:10:31] Prompt Engineering >Find Murray online: https://www.doc.ic.ac.uk/~mpsha/ https://twitter.com/mpshanahan?lang=en https://scholar.google.co.uk/citations?user=00bnGpAAAAAJ&hl=en MLST Discord: https://discord.gg/aNPkGUQtc5 References: >Conscious exotica [Aeon/Shannahan] https://aeon.co/essays/beyond-humans-what-other-kinds-of-minds-might-be-out-there >Embodiment and the inner life [Shannahan] https://www.amazon.co.uk/Embodiment-inner-life-Cognition-Consciousness/dp/0199226555 >The Technological Singularity [Shannahan] https://mitpress.mit.edu/9780262527804/ >Talking About Large Language Models [Murray Shanahan] https://arxiv.org/abs/2212.03551 https://en.wikipedia.org/wiki/Global_workspace_theory [Bernard Baars] >In the Theater of Consciousness: The Workspace of the Mind [Bernard Baars] https://www.amazon.co.uk/Theater-Consciousness-Workspace-Mind/dp/0195102657 >Consciousness and the Brain: Deciphering How the Brain Codes Our Thoughts [Stanislas Dehaene] https://www.amazon.co.uk/Consciousness-Brain-Deciphering-Codes-Thoughts/dp/0670025437 >Roger Penrose On Why Consciousness Does Not Compute [nautil.us/STEVE PAULSON] https://nautil.us/roger-penrose-on-why-consciousness-does-not-compute-236591/ https://en.wikipedia.org/wiki/Orchestrated_objective_reduction >Thomas Nagal - what is it like to be a bat? https://warwick.ac.uk/fac/cross_fac/iatl/study/ugmodules/humananimalstudies/lectures/32/nagel_bat.pdf >Private Language [Ludwig Wittgenstein] https://plato.stanford.edu/entries/private-language/ >PHILOSOPHICAL INVESTIGATIONS [Ludwig Wittgenstein] (see §243 for Private Language argument) https://static1.squarespace.com/static/54889e73e4b0a2c1f9891289/t/564b61a4e4b04eca59c4d232/1447780772744/Ludwig.Wittgenstein.-.Philosophical.Investigations.pdf >Integrated information theory [Giulio Tononi] https://en.wikipedia.org/wiki/Integrated_information_theory >Being You: A New Science of Consciousness (The Sunday Times Bestseller) [Anil Seth] https://www.amazon.co.uk/Being-You-Inside-Story-Universe/dp/0571337708 >Attention schema theory [Michael Graziano] https://en.wikipedia.org/wiki/Attention_schema_theory >Rethinking Consciousness: A Scientific Theory of Subjective Experience [Michael Graziano] https://www.amazon.co.uk/Rethinking-Consciousness-Scientific-Subjective-Experience/dp/0393652610 >SayCan - Do As I Can, Not As I Say: Grounding Language in Robotic Affordances [Google/] https://say-can.github.io/ >THE SYMBOL GROUNDING PROBLEM [Stevan Harnad] https://www.cs.ox.ac.uk/activities/ieg/elibrary/sources/harnad90_sgproblem.pdf >Lewis Carroll Puzzles / Syllogisms https://math.hawaii.edu/~hile/math100/logice.htm >In-context Learning and Induction Heads [Catherine Olsson et al / Anthropic] https://transformer-circuits.pub/2022/in-context-learning-and-induction-heads/index.html

>>29596 Thanks for the post, NoidoDev! Cheers. :^)

I found a cognitive architecture LIDA that checks all my boxes of what it contains, which are: >H-CogAff The cognitive-affective architecture that gives a basic structure of human-like mind. >ROS capability The Robot Operating System is commonly used and widely supported for simulation and operation of robots. Great program to learn for getting a robotics job too. >Python GPU acceleration for parallelizable calculations. Great program to learn to get any software job. >Concurrent Modules Everything in the model runs separately except for the "stream of consciousness" that fires every 1/10th of a second. Should be a nice and fast architecture that can make decisions based on semantic data instead of the current state-of-the-art large language models which are reliable at producing language and not much else without help. It is one of the only Arches that states it has some level of consciousness. I'd first want to put emotion state module in it, along with an LLM as a robot interface. I have a lot to learn now before I can implement anything, but I believe this is the best thing besides a slow, expensive, and unreliable but available LLM-centered cognitive architecture. >links https://ccrg.cs.memphis.edu/tutorial/tutorial.html https://github.com/CognitiveComputingResearchGroup/lidapy-framework https://en.wikipedia.org/wiki/LIDA_(cognitive_architecture) https://ccrg.cs.memphis.edu/assets/papers/2013/franklin-ieee-tamd11.pdf

>>29924 Correction: one of the papers says it already has emotions. Still, I'm sure it needs STT, TTS, an LLM, and "other" motivations. All doable with current tech that I've already built.

>>29924 >>29932 Amazing, this might be very useful. Thanks. Btw, if no one responds with some encouragement it doesn't mean nobody cares. I just don't want to spam these threads with chatter.

>>29932 >java framework tutorial https://ccrg.cs.memphis.edu/assets/framework/The-LIDA-Tutorial.pdf >java framework repo https://github.com/CognitiveComputingResearchGroup/lida-framework This java program looks more straightforward to modify to test new modules before implementing in the python ROS version. It's java, like minecraft mods. >>29933 I understand, either way this is probably the most exciting development thus far in my project and I'm happy to share. If I get somewhere I will post here. I have a really great feeling about this one... Considered naming the first bot Lida if this pans out.

>>29924 This looks like a great baseline. It's not clear how to incorporate emotions into the model. My guess is that it can be done with changes primarily in the Global Workspace, Action Selection, and Motor Plan Execution. You might find these points relevant from >>27144: >Emotion regulation. I spoke with a cognitive scientist that specializes in this, and he's convinced that emotion regulation all boils down to: positive feedback loops for satisfying needs, negative feedback loops for avoiding harms, and a "common currency" for balancing different motives. >Embodied control. Chatbots are "easy" since the final expression (text) can be generated by a single model. With actual bodies, or even just with video, the final expression is split into multiple modalities (e.g., voice, body movements, facial movements), and they all need to be in sync with one another. If we had good multimodal models, that might be fine, but we don't, so I need a way to generate outputs from multiple models and somehow make them consistent with one another.

>>29955 These are good points, I'll have to see where Lida takes me.

>The Fastest Way to AGI: LLMs + Tree Search – Demis Hassabis (Google DeepMind CEO) https://youtu.be/eqXfhejDeqA

>>29974 https://nypost.com/2024/02/27/business/google-parent-loses-70b-in-market-value-after-woke-ai-chatbot-disaster/

>>29957 LIDA is too broken when using a newer version of java. It might need Java 8 to run, and I don't want to compile/debug on compatibility mode just to try to add other repo's features that are coded in another language. lidapy might have potential for a full on robowaifu simulator, but I'm thinking I'd need a different program that can plug into various ai model and database apis.

>When writing and talking, people sometimes pause to think. Although reasoning-focused works have often framed reasoning as a method of answering questions or completing agentic tasks, reasoning is implicit in almost all written text. For example, this applies to the steps not stated between the lines of a proof or to the theory of mind underlying a conversation. In the Self-Taught Reasoner (STaR, Zelikman et al. 2022), useful thinking is learned by inferring rationales from few-shot examples in question-answering and learning from those that lead to a correct answer. This is a highly constrained setting -- ideally, a language model could instead learn to infer unstated rationales in arbitrary text. We present Quiet-STaR, a generalization of STaR in which LMs learn to generate rationales at each token to explain future text, improving their predictions. We address key challenges, including 1) the computational cost of generating continuations, 2) the fact that the LM does not initially know how to generate or use internal thoughts, and 3) the need to predict beyond individual next tokens. To resolve these, we propose a tokenwise parallel sampling algorithm, using learnable tokens indicating a thought's start and end, and an extended teacher-forcing technique. Encouragingly, generated rationales disproportionately help model difficult-to-predict tokens and improve the LM's ability to directly answer difficult questions. In particular, after continued pretraining of an LM on a corpus of internet text with Quiet-STaR, we find zero-shot improvements on GSM8K (5.9%→10.9%) and CommonsenseQA (36.3%→47.2%) and observe a perplexity improvement of difficult tokens in natural text. Crucially, these improvements require no fine-tuning on these tasks. Quiet-STaR marks a step towards LMs that can learn to reason in a more general and scalable way. https://arxiv.org/abs/2403.09629

>>30603 I find it funny they only now figured out that spitballing complicated question answers isn't ideal.

A really big gap seems to be present for AI related stuff in that it is more people who are interested in AI who study philosophy of mind just as a means to an end, I'd recommend people spend more time there as most of the approaches seriously talked about don't come close to even accurately conceiving what the mind/thought/thinking is. If we do eventually only anon that kinda references it is >>25274 and even that that breakdown doesn't really make sense, y'all actually gotta read stuff and it's quite complicated since it is largely dependent on metaphysics and what-things-are in general. The root of the issue is the false distinction between what a thing is and it's mind, the mind is not some separate thing going on inside a human, it's an outgrowth of the human body, which means it cannot be studied outside of what a human is in general. Useful book on that is: https://www.amazon.com/Retrieving-Realism-Hubert-Dreyfus/dp/0674967518 Most of the AI research i've seen however depends on making that separation There's a bunch of other useful reading, made much more complicated due to the fact you do actually have to build up from the bottom. (Aristotle's ethics,physics,metaphysics,de anima. Heidegger's what is called thinking, andy Clarke's being, Macintyre's ethics in the conflicts of modernity there are some useful ones from a wide variety of places. I know plenty of people in those veins of thinking some books on getting better AI but I haven't dove into those yet) The inclusion of ethics comes from seeing mindedness as being rooted in those organisms. All of reality is in itself indeterminate in how we can break it down, we break things down, conceptualize them, and learn to act in relation to them by reference to what is good for us as an organism. You see a chair as a thing for sitting, along with it's material history of where it came from it's potential usage in the future. Emotions/passions arise of that being a certain kind of organism in the world with certain parts that relate to other certain things. The key thing I am not sure but I am interested in is how or if there is any AI research is that sort of distributed quality. The intellect/mind serves more as a sort of unifying aspect to a bunch of distributed knowledge/concept stores. If you eat an apple you don't have to think of what an apple is, that information of what-an-apple tastes like is stored in it and your biological machinery converts it into something your brain unifies with your vision of the apple, the smell, along with all your memories and cultural background with apples. It's more thing-based then language based. Language gives us access to the sort of cultural/background part but that's only intelligible on top of the foundation of us being what we are, engaged with the reality we have, with the background we have. https://www.sciencedirect.com/science/article/pii/S0004370207001452

>>30627 >The root of the issue is the false distinction between what a thing is and it's mind, the mind is not some separate thing going on inside a human, it's an outgrowth of the human body, which means it cannot be studied outside of what a human is in general. I'd question the assumption. There is some theories about the mind being non localized. >It's more thing-based then language based. It is likely more complicated than that. Think about people with aphantasia. They can only think about things through words but that arises the question. How could such a person exist? Before language how would a person with aphantasia think? So it must be extremely vague concepts from feelings not images or words.

>>30627 >>30628 Thanks to both of you, but just in case if you want to go more and deeper into philosophy, please do this over there >>11102 since this is always results in walls of text with of lot of insider lingo, and this here is supposed to be about implementation ("how to do it"). >>30627 I am aware that a well developed and skilled human-like AI would not be based on just understanding things through text. We could for example have sensors measuring things and having some numerical value, or detecting certain patterns like symmetries. That said, storing a lot of things in text makes sense, e.g. for debugging and development. >>30628 >They can only think about things through words They can create mental images, but they still perceive and learn not only through words but by other senses including vision. Also, I assume to use a hammer you don't need to picture the use of the hammer, you may be able to teach the body to behave like the hammer is part of it. The conscientious perception and modelling of the world is only a part of what the human brain does, but it does other things "under the hood". We can learn from that that we won't need model everything down to every detail in an embodied AI, especially not in one place, but only the minimum necessary. Some self-awareness area only needs to be aware and log the incidents which are noteworthy. Then compress it by deleting even more later, especially everything that could be guesstimated and recreated based on that, if necessary.

Context: - I'm working on infrastructure that's friendly to distributed development of complex AI applications. - At the least, I want to solve everything I mentioned at the end of >>27144, meaning it should give easy ways of supporting emotion regulation (through feedback loops), embodied control (through native experimentation support), and heuristic derivations (through hybrid structured-unstructured generations). - To support distributed development, I want it to make it easy for people to plug in their own compute (desktops, cloud compute, robots, whatever else), and I want it to support enough access control to avoid catastrophic effects from, e.g., raids. - It boils down to orchestration software modeled on Kubernetes, but more support for distributed development (i.e., many clusters with many owners as opposed to monolithic admin-managed clusters) and asynchronous communication channels (pub-sub as opposed to DNS-based cluster networking). I've made a few design changes to support all this. - My approach to access control is here >>29197 >>29248. - The crux of my approach to hybrid structured-unstructured generations is here >>28127. - Until now, the feedback loop & experimentation support pieces were missing. Update: - I just finished implementing what I think is a viable basis for feedback loops & experimentation. The design for this was hell to figure out, mostly because of the complex way it interacts with access control, but I think I have something that can work. I have a test backend working and the necessary client library changes completed. - On top of what kubernetes provides, I'm adding three new concepts: "remote" controllers, posting messages to controllers, and "fibers". Remotes and fibers are both specified through the "metadata" field of any config, posting messages is done through a POST rest api. - Any config can be assigned to a remote controller, assuming you have the necessary permission to use another cluster's controllers. If a config is assigned a remote controller, that controller received all operations executed against the controller (create, update, delete) while your own cluster is able to observe the results (read). I originally added this since the people that know how to set up optimizers are usually not the people that set up & run models. Remote controllers make it possible for one person to optimize another person's models without needing "too much" access to the models. - In kubernetes, all operations are config file changes. The new POST api gives a way to send a message to a controller independent of any config file changes. You can post messages against a config file, and that message will get picked up by whichever controller is responsible for handling that config file. The controller can, but isn't isn't expected to, make any config changes as a result of posted messages. - Fibers enable controllers to post messages to each other across clusters, again without granting "too much" access. Normally in kubernetes, configs are identified by group/version/kind/name tuples. With fibers, configs are identified by group/version/kind/name/fiber. You can think of a fibers as adding an extra "dimension" of configuration whose purpose is to tie together multiple controllers. The controllers for any config with the same group/version/kind/name (and different fibers) can post messages to each other. For experimentation, one fiber can be responsible for generating trials (candidate configurations), another can be responsible for evaluating them (value assignment), and a third can be responsible for deploying them. - I'll be testing this out next to find a good design pattern for running modules that continually self-optimize as they run. I apologize if this is confusing. Once I get a prototype up, I think that will make things a lot clearer.

>>30759 >I apologize if this is confusing. Once I get a prototype up, I think that will make things a lot clearer. No, it's not unecessarily so. It's just a complicated space to be working in is all. You're fine, Anon. This all sounds really encouraging, CyberPonk! Looking forward to seeing your solution in action. Cheers. :^)

>>30759 Sorry for not responding to this topic here or on the Discord faster, but I'm didn't have the right headspace to read through it and think about it.

>>30102 >Update LidaPy seems nice, but it too is in a EoL programming language, python2. My only option at this point is to see if anyone can refactor it to python3 or just use very old software to test it. I'm feeling like it might be best to DIY something on a newer platform that follows the lida framework. The LIDA tutorial even repeatedly states: "The Framework constitutes one, but not the only, way of implementing the Model.", like they want you to make a new implementation. Before I go ahead with any work, it's always important to remember to check who owns the rights to any IP. I would be a research associate at a university if it weren't for IP rights, and I've pondered going to memphis to develop LIDA if I would have the right to use it in my bot commerically. I'll post an update if there's any progress.

>>30840 >but it too is in a EoL programming language, python2. >My only option at this point is to see if anyone can refactor it to python3 or just use very old software to test it. Might I suggest an alternative option of having someone rewrite it in C++ , Anon? 40+ years and still going strong today (fully backwards-compatible, of course -- basic C++ code written in the 80's/90's will still build today!) :^) Good luck, Anon. I hope you can succeed with your LIDA research. Cheers. :^)

Potentially useful, potentially ontopic thread on 4cuck/sci/ I was there looking around for the linked thread from our Propaganda thread lol. https://boards.4chan.org/sci/thread/16087430#p16087430

>>30840 Python2 can still be installed. Also with installers like Nix you should be able to install old versions of Java.

>>30863 I looked into it and found that it is not recommended to install Python2 anymore. You can install PyPy or IronPython instead. There seem to also be some other long term support options. I don't know which Java it needs, but JRE8 seems to be in the Nix repo. You can install and run software exclusive to the nix-shell. But I'm new to this myself. I might be able to help a little bit. I also looked a bit into Lida itself and it looks like something how I would've imagined it. I might going to try it out at some point, and when I start to implement something myself I might look use it as a resource. I will most likely learn Elixir while doing it, at least for the any part which is not about number crunching.