Hey, long-time fan of Steve’s work. I recently regained a years-dormant fixation on the subject of AI, despite the best efforts of contemporary examples (LLMs etc.) to make it unlikable and boring. I blame the funny turtle streamer.
So, I spent the last couple of months processing a lot of information. Reading up on the theory behind current LLM stuff, a lot of Wikipedia, whatever science papers I can find related to my own thoughts/observations, like non-linguistic forms of intelligence, the brain being a physical machine that changes structure, concepts like brainwaves applied to AI (LinOSS seems interesting).
And, of course, I’ve been reading blog posts and comments from Phantasia (and Steve’s old WordPress, etc.).
Something I’m not fully clear on (and hoping it isn’t rude to ask!) is whether some of the outside ideas/research address Steve’s past criticisms about the shallowness of mainstream AI research, whether certain ideas overlap, etc.
Like “world models”, for example. The premise of modeling and predicting the states of the physical world seems similar to the premise Phantasian brains, but I don’t think they’re the same.
I also figured it’s time that I join some communities with people who actually know/talk about the subject of AI in some way, instead of burdening a couple Discord contacts with my entire stream of consciousness, random scientific papers, and excerpts from my cluttered notes.
Obligatory mention: I sometimes do YouTube. I’ve struggled to actually be productive for many reasons (I’ve rewritten this post several times) but I’d be happy to contribute on the “content creator” front if people have ideas.
I sometimes do YouTube.
If you do any youtube video meantioning phantasia in any kind, drop me a link!
I will promote it across all platforms i am! Best would be something technical.
But i take what i get!
i am excelent at beeing annoying and impossible to ignore, a perfect promotor!
P.s. start the video by talking about “The british guy that raised an AI daughter” and then remove neuro sama image with the words “No, the first one” and switch to Steve and Lucy, that way i can ignate the swarm beacon!
Haven’t said much in a while on the reasoning that I don’t have much to say, but I realize it doesn’t hurt to share or think aloud a bit more than I’ve been.
I feel like I’m starting to build out my intuitive understanding of Steve’s ideas.
For some context, my interest still leans further into AI than AL.
In some ways, a computer desktop/browser is my “world” (I’ve been taking more walks lately, but that’s not where I “live”), which also makes me wonder whether the same could be done for AI.
Of course, the mind I “project onto” computer screens was still mostly trained in the real world, and I don’t lose awareness of (or at least “access to”) my surroundings and body. But to say words, I don’t have to consciously pilot my fingers on the keyboard. To click a link, I don’t need to look at the mouse and think about how I’ll move it.
That reminds me of the faint childhood memory of not having yet learned to coordinate with the mouse, which I partly attribute to left-handedness (every mouse is on the right, and even if I try to switch sides with an ambidextrous mouse, then the keyboard is weird!) and would’ve once just called “developing hand-eye coordination”. Now it feels like there’s something more interesting there.
But, AL/AI.
Steve seems to have a much better understanding of brains and intelligence in general than most (saw “learning how to learn” spoken of as a doomsday prophecy from a Microsoft executive recently…they think their competition is “every PhD”, not “any mammal”), so I’ve tried to consider how that understanding might be applied from the opposite direction.
A more efficient, adaptable, and “real” intelligence; but still in the more abstract domain of language, visuals, games, etc. rather than an “animal”.
That’s not to say I think this is necessarily even possible or relevant, but it’s been interesting to consider. One of many observations is that it really seems like most contemporary/generative AI just has the whole model as one set of weights (my understanding of that is also incomplete), where “brains” seem to have very notable sections and structures. Everything with a brain is bilateral, and vertebrates have brains with at least 3 sections (according to Wikipedia). Everything being one monolithic array also seems to be part of why LLMs require so much memory.
The monolithic architecture (still not sure this is necessarily accurate) does seem suited for what LLMs are useful for, but not for something you wanted to be “intelligent”. I also wonder if an architecture based on multiple smaller models/maps could parallelize differently using multiple relatively low-performance/memory GPUs. Or maybe it’s a system that doesn’t parallelize well at all, and would best run on a many-core CPU architecture. The scale of LLMs is definitely *hugely* excessive.
(…in my very incompletely-supported unprofessional opinion, but that’s what I have to work with.)
(Also, I did make a video, but it’s not really related to Phantasia. More importantly, it’s bad and I don’t want to share it! I plan to do better, but my CPU cooler seems to be broken so I won’t be able to do anything for the next 3 days. The terrible CPU throttling is my excuse to stop writing and go sleep. Except I went back up and wrote several more sentences after writing that.)
@genesis Got something. It’s basically my most recent understanding of how LLMs “act like people”, and dispelling what I feel is an easy place to assume “magic” exists. I don’t know if I’d call it “technical”, directly, but it’s about the technology. I figured out “where the intelligence lives”, in a way.
Steve’s perspective was relevant to my understanding and directly quoted a few times, so I included the address of his old WordPress and, of course, phantasia.life. A-life perspective was useful, as was my heavy exposure to the concept of “maps”.
Naturally, I made the “foreshadow Vedal -> It’s Steve” joke (and later, briefly, actually tied my point to Neuro/Evil).
(And, terribly sorry if sharing/promoting this is not appropriate. 😓 And for the production value – I’m aiming for “published, not perfect”, and graphic design is my burden.)
@johnologue nice, new toy to play!
Last one was keeping a video in the algorythm for an extended period of time (anti norn video, right now 2 months and still growing)
Now i can play to figure out how to feed a video into the algorythm! Wish me luck…
…and don’t worry, any video that can become a potential swarm beacon is welcome. We need the swarm. I am sure Steve doesn’t mind AI fans who love cute stuff and brag about trowing money at random things! This is exactly what we need!
@johnologue from the view numbers this morning compared to those now, the algorythm manipulation was easyer than expected.
However the low like rate is concerning, without some natural likes, maintaining momentum is nearly impossible. This means you just need to find your style and frontload some drama into the forst seconds.
Trying not to lean too hard into “optimization” upfront. The important thing is making more videos, practicing (training my personal neural networks, one might say)…content comes first. I do try to include some joke related to “liking” in most of my scripts (I think I missed that for this), I plan to utilize cold opens where suitable, etc.
I’ve given things a lot of thought; that’s why I never get anything done! 😆
For now, getting over 100 views is a win; that’s a lot of people when you’re not thinking in “internet numbers”. Or if you consider the activity on this forum! I imagine the best way to liven it up is slowly. Bring in a few more people who like to talk, and their conversations help set expectations for people who want to join in.
It’s hard to start a topic if you don’t know what you’re supposed to talk about, or to materialize into an ongoing conversation if it doesn’t exist!
@johnologue How to optimize, the first minutes are the most importand ones. There you had little images, that gave nothing for the eyes to see. You need a hook at the start, maybe runthe script truh an LLM to optimize for viral youtube contend with interesting first few seconds. Trow it in different ones and see if any gives a good advice. To learn from them – not to copie them exactly! Because if you copy from those LLM, you make yourself dependend on them – however if you understand why they sugested what, you learn from them and improve yourself instead.
Next the algorythm optimisation, when you answer comments, trow in some maybe related words – youtube uses those to suggest it to different people – why else do you think such odd first comments apeared? 😉
Afterwards something you can’t do – watch the video multiple times to increase watch time (not you,youtube knows your devices, must be someone else) – best someone who is aleady in the algorythm category you want to be (that is where i come into action)
And from there on the video needs to do it on it’s own – i can prolong how long it stays in the algorythm, by increasing the watchtimeon my end. However this only works when others interact positive.
If i overdo this on a video with little interaction, youtube would me place in a other alorythm group of les watched content. However when the video itself got good interaction, prolonging the algorythm time is no problem. So far i managed to kep his one https://www.youtube.com/watch?v=IDxFxWakhm0 in the algorythm, despite it beeing already 2 moths. When you look atthe other thins in the channel, you see the difference in views. (commenting there on a lot of comments helped there as well)
Also you need to grow into the rle of a youtuber as well, this is somehing that can’t be rushed since it is a learning experience. Saw this ina v tuber clipper hat started and slowly became more professional. An initial oost helps, but you need to do the mental growth.
But we will get you there, a youtuber loosely related to this project is something we need! Because that is a great way to stich together many communitys and grow the people aware of steves work.
This means for you, just continue growing and know you have a save haven here! We need you to grow as well!
Oh and for future contend, watch the amazin digital circus, they droped steve rands name as G*an* in episode 8 in a screenshot around 30 minutes. Soon it will have a theater release and if thee is any conectiontion to any interesting AI topic (it is the finale, there will be) you ride the wave!
