 This is the Fellowship of the Link call for Wednesday, March 29, 2023. We are in our funny little jitzy room. I think of jitney every time I see jitzy, so I think of this as having like a little surreef with a fringe on top kind of thing in our, you know, as we bounce down the little country road in our conversations. I don't know why, but that little image kind of flashes by when I see the jitzy thing. Two things that I'd just like to sort of bring into the room, because they might be useful and topical. Pete's aware of both of them, because he's deeply enmeshed in both of them. One of them he caused, which is last Thursday's OGM call, I hit the night before I sent out just a tongue-in-cheek, too bad there's nothing to talk about, kind of comment, you know, while inviting everybody to join the call, because there's way too much news in the world. Pete picked up and said, well, I've got an idea, and he said, let's make, let's put on a show, basically, let's write a volume, an edited volume book together, which has morphed and turned into kind of a takeover of the Monday morning since doing calls. And Pete, correct me for all of this, like if I'm framing it wrong, but we're basically, we need a new name for this project. We called it OGM topics, which doesn't seem to me to have like the vital essence of what it might be, but it's a nice kind of way of creating an artifact that we can point to. I layered onto it my Neo Book ideas, which are about creating several books that might share chapters, so imagine several different edited volumes or books that have different narratives, but reuse some modules that are needed to explain different parts or even opinion modules or whatever else. And happy to go back to that. But there's a standing call now on Monday mornings at 10.30, yes, the Monday mornings at 10.30 Pacific, if any of you want to join that project, and we can sort of go back to it. And then the second thing is I had a really lovely conversation with Kyle Shannon, a very old friend of mine from New York days, who back in 94, 95 when the web was but a little embryonic thing, it created a group of webmasters called the World Wide Web Artists Consortium. And they were meeting and I met a bunch of my cool young New York tech friends through attending WAC meetings. And recently he started playing with first stable diffusion, because he's more of an artist and all that. But then when chat GPT came out, this whole thing started going nuclear, nuclear, I think in a good way. Going nuclear doesn't always sound like a good thing, exactly. So he created a new salon, basically, it's now called the AI salon, it has calls that meet I think on Tuesdays, face to face in Denver, if you happen to live in Denver, but you can join by Zooms, there's a meetup for it, which we can share the link to. But so in my conversation with Kyle, we got into, hey, how might we use chat GPT in combination with my brain? So yesterday, sorry, Monday during the Free Jerry's Brain Call, Kyle and a colleague of his, Cliff, basically joined the Free Jerry's Brain Call and we kicked around a bunch of ideas and got like a nice start, I think, on what we might do there. And I sent my latest brain archive over to Marc-Antoine Fahon, who is the person who did an export back when into Postgres database of my brain, so it'll be available as JSON objects in a Postgres database, and Pete, again, correct me for everything I've gotten wrong here. But those two things are kind of have a lot of energy right now, and I'm excited about them because everything that I wanted to do in my bigger goals page sort of fits nicely onto this. I also had a catch up with Paul Roney this morning because he and I are sort of reviving the Tools for Thinking podcast as HyperTalk, which all of this flows into nicely as well because the process of doing some of these things would make really nice series of podcast episodes to explain and to, you know, to clarify and so forth. And the artifacts that we want to leave behind as a Neo book or whatever it's called would also make good HyperTalk kind of artifacts and so forth, so that's like a third project that folds in to this process as well. So maybe I'll ask Pete to amplify, correct, riff on any of that and then see what anybody else would like to ask or think or whatever. You covered it really well, Jerry, I think we could go into more details, but that's the gist. Awesome. I love it when I represent something reasonably. Flancian, are any questions or thoughts? So I'm interesting, following along, I guess on a meta level I always go back to how to keep track of things and so on, particularly because I'm particularly bad at that. I don't know if you notice, but I'm like, I'm sometimes very present and then I disappear and then come back and so on. So catching up and so on. And yeah, I wonder, like, I guess we mentioned in the past like a few ways of keeping track of shared projects and so on. Like of course we have the actually the OGM, the Masi wikis and I guess we have on that level the idea of like aerating all of these into the Masi wikis or whatever. But also like it's like backtracking or like any kind of tracking somewhere could be interesting or a feed even. Yeah, I guess it's just like meta and yes, judgeability or like I guess fine tuning now. I saw there's like a lot going on of course when like for reducing the cost of fine tuning like with things like Lora, I guess, and like, you know, also based on open models, more open models. Judgeability is still of course the state of the art things, but like open source is catching up quite quickly, thankfully. So I guess I wonder if like I still think that going to like Judgeability plus your reign is going to be like really interesting. But I guess I hope we can do also like the complete open approach in parallel. I guess it's just like a wish more than a concrete idea and about the hyper talk is a new name of the podcast. Yes, we're renaming it hyper talk in homage to the language from hypercard, which only a few people know is a thing. And for anybody else who doesn't know, it's like, you know, hyper connected conversations is a good is a good phrase, a good umbrella for a podcast. So so that's the name of it. Paul is up to his nostrils in a round of fundraising, which is not easy right now. So he's been really absorbed in that, but we're trying to sort of work around that to figure out, hey, let's let's stand up some, some episodes. And the conceit of this of the hyper talk podcast is to invite people to host short sequences of calls. So for us to basically produce other people hosting a bunch of calls, including ourselves, because I'm certainly interested in doing a bunch of them, but to have collections of people who say, hey, I have a plan for four episodes. And that's it. It's all I'm going to do. I'm going to do four episodes. And here's who I want to invite. And here's the topic and so forth. And we'd be like, awesome, do it. And then we want to sort of unify all that on probably a massive wiki, which we're which we're setting up for that purpose. And then to have some specs for, OK, good. This is this is what it takes to have an episode that looks and smells like a hyper talk podcast, et cetera, to leave behind objects and documents and videos in the same sort of way and to make that all all kind of happen. So I'm I'm working with him and whoever else feels like it to to stand that up. Go ahead, Aaron. Yeah, I think that's cool. I think it sounds sort of like a podcast zine. Yes, you're right. That's sort of the the general general contribution policy anyone can put in. And then making sure you're recording sort of how to continually reproduce it. I think that's that's pretty interesting. It's sort of a collection of zines, in a sense, under a podcast umbrella is kind of the intention here. Right, right. Sort of like we'll sort of like a zine with different issues. Yeah, you're you're standardizing the format in some sense. Yeah, like a tool kit as well, maybe here. Yeah, Peter. Yeah, thanks. I am I guess and I wanted to follow on to something. Jay said, and then maybe I'll do a check in at some point. It doesn't have to be right away, but but I wanted to say folding things into a podcast. This is almost a non sequitur, but maybe it's not reminds me in the US. There's in the US in the southeastern states. There's a big controversy about drag queens and performers. And so I just read this morning that Madonna added a show, I think, or something like that, or or added a stage to a show in Nashville or something like that, where she's going to have drag queens perform. So they're not allowed to perform. And so she's extending an arm and like, I'm going to protect you and give you space to be on my podcast or or zine, be in my zine or whatever. So I don't know that the imagery struck me. So, Jerry, you're almost like Madonna. I'm going to have to start wearing like a bra of some kind or kissing priests or something. I don't know. I got to do something a little something a little controversial. And Pete, it may amuse you that I had my 2003 shirt on. And also, this is the T-shirt that had on the back. Yeah. Can you see it? Yeah. This is just a little mind map of how the retreat. That's lovely. Chicken wise, since I've got the podium, I guess, along with the things that Jerry has mentioned. I've also been doing a fair bit of talking about talking about. Chat, GPT and LLMs in the context of different cultures and cultural bias that's built into something like chat, GPT ways you can maybe get around it a little bit. Stuff like that. So that's been going on on the OGM list and I've been spending a fair amount of time doing, I think, a pretty good job of kind of walking us through some thinking about, you know, what's a way to think about chat, GPT? And also, how would we think about some of its weaknesses, I guess? And I will add that on the various lists that Pete and I are both on, there's several of these conversations going. Pete is showing extreme patients' care and dedication to present these things and like, hey, here's what's happening, here's what's going on. Well, other people are like, oh my God, look, it can't do this or it doesn't do that. Look, I read this piece that says, and Pete's like, okay, everybody, keep your pants on. Here's kind of what's happening. Pete, I really appreciate that because your depth of understanding of what's going on and then your capacity to explain and reframe and so forth is awesome. It's like really inspiring. Thanks, I appreciate it. One of the experiments, by the way, we got one of somebody said, well, there must be a Chinese version of chat GPT. And I don't know that there's one of those, but there is something called chat GLM. And I found an unofficial instance of it running. The official instance is waitlisted. It was trained, as I understand, on a trillion Chinese tokens and a trillion English tokens. So it's 50-50 Chinese English. And we've had discussions about, you know, like Kevin Jones, bless his heart, he keeps saying, why is chat GPT stuck in Western white male worldview? And the very short answer I have back for him is kind of, we're all stuck in Western white male worldview. And, you know, you can actually ask GPT to chat GPT to be something different or come from a different viewpoint or ask it for, I was asking him about what's equivalent to people doing philosophy in other cultures. Turns out philosophy was one of the, philosopher was one of the things we've been rotating around. And just the term philosopher essentially means an old white guy from, you know, Europe, which wasn't obvious when we started the conversation. Like why, when you ask it for a list of philosophers, does it give you old white guys? And it's like, well, that's essentially what a philosopher means because, you know, philosophy is the name that we've given, the short name that we've given to what is a longer, you know, thing of Western philosophy, which includes Western ideas of epistemology or epistemics or whatever, you know. So philosophy is a condensed word that means Western white male, even though that wasn't obvious to any of us at the outset. Anyway, I got into the business of talking to chat GLM, the Chinese one, and just barely scratched the surface of starting to ask it. Instead of like list of set of philosophers, it actually gives you back Western philosophers because philosopher means Western. You can keep going and expand from there. But what I brought up was asking a culturally interesting questions is probably a more useful thing. So if you've got an LLM that's been trained with a different corpus, a corpus that has a different cultural bias, the interesting thing to do with it is ask it kind of human level questions. Tell me about the love of a mother for her child. What's better, loyalty or bravery? You know, I just broke up with my, I just got in a fight with my best friend. How can I make that situation better? You know, if you ask human relevant questions that don't have a cultural bias by the way that you ask them, then you can start to explore the cultural background of an LLM. And so I started doing that just a tiny bit with chat GLM in English and Chinese, which was interesting. You know, junior varsity version of some of what Pete has done over the weekend. I spent the weekend with a group of people who were experts in Christopher Alexander and pattern languages, and also Ward Cunningham was in the group, so the inventor of the wiki. And Sunday morning I was suddenly inspired to ask chat GPT something that one of the participants had said the day before while we were talking. And he had said, you know, I turned to a pattern language for guidance on how to design a roof deck, and I found the advice in there just perfect. It was not too little, not too much. So my prompt to chat GPT was, you are Christopher Alexander, author of a pattern language. What advice would you give someone at Home Depot on how to build a roof deck? And the answer was really good. It didn't cite particular numbered patterns, but the content of the answer was really excellent. Totally took into consideration that this was a roof and you should worry about materials and uses and all that. And everybody was like, well, crazy. Because I was the youngest person in the group and everybody's not like completely up on what's going on here. Oops, we lost Pete. So it was fascinating. And I think that there's a really interesting frontier here. There's a really interesting frontier here where if we can get used to these things and apply them well, we can sort of plow right through a bunch of things that many of us thought as barriers. And I think just waking up to what the barriers are in your head, which is happening to me now on a daily basis as people like Pete and others do experiments and share what they're doing, is really cool and fun work. Yeah, very cool. I'm really interested about this connection between pattern languages and GPT. I guess I think you will know about that. But of course, that's part of the evolution here. That's in thinking of things and areas to explore with language models. It reminds me a bit of when I was new to the internet. I don't know if you have this experience. I remember I was in Alta Vista before. And thinking, oh, what can I type here? Searching for my favorite writers, for my favorite things. And then just being in the box there and thinking, oh, what else can I explore? I'm not even just coming to terms with the fact that that was possible. And to some extent, I think we are going through the same process. And there's this joy and this reinforcement loop that is interesting because it will do better in some areas than others. And discovering which ones it does better or worse is part of the assignment. And I went through this a bit over the weekend as well with a friend who had been doing more like GPT coding. He was like, oh, we should really try it. I was like, I have read, I have seen examples. But then I finally tried it for a coding task. And actually seeing in action was so impressive in the sense that I gave it a JSON dump. I mean, with this JSON with this structure I made up, so probably never seen before. And I asked it to convert it to the right shape, essentially to graph it as a verite graph. And just using the semantic identifiers, it inferred the structure and wrote the transform. Which is like half an hour of coding and JavaScript for me more and then 20 seconds. There was this wow moment, which is like, to some extent one thing is to know and you know when you see a report. But I felt it when it hit graphs to some extent. I'm realizing I have a thought in my brain called my aha moments with technology. Right. And one of them is being in a big auditorium at the, I think, Interstate Commerce Commission in downtown D.C. Where Steve Jobs and the original McIntosh crew were on stage to do like a presentation tour about this new McIntosh thing. And Andy Hertzfeldt went up and opened McDraw, spray-painted a little bolder with the tools. Made a bolder, spray-painted it a little bit of color and then took the lasso, lassoed it and moved it around the screen really quickly. And there was an audible gasp in the room. Right. And that's on my list. Another one is Switcher. When Switcher first came out and I was at the Washington Apple Pie, which was the little club that we sort of joined. And somebody showed me, oh, you can run several apps and then you can switch between them. And it looks like a little cube rotating. And I was like, so I need to add one of these moments for chat dpt to that list because clearly that's, I'm having, it's a visceral experience once you sort of try it and see it and do it. It's really cool. A smaller way Jerry described it, that kind of feeling. And when we were talking earlier this week was, we were talking about people. But actually one of the things I've said with chat dpt is a really useful way to use it is in conversation. People come to it and they think it's an oracle. So they ask it for facts, which is bad. But then the other thing they do is they think it's a search engine. So they type something and it's like, okay, this is the definitive result. And it turns out what you want to do is chat dpt is actually conversed with it. You want to ask it, oh, tell me more about this or, you know, I didn't think that or whatever, right? You want to, like, it's an interactive thing. And that's the thing that is really great. We were, Jerry and I were commiserating kind of that there are people who don't get it yet. And Jerry said it's kind of like riding a bike. You know, you can describe people how, what it's like to ride a bike and they could go, yeah, yeah, yeah, I get it. You know, it's got wheels and it goes fast and it's a bike. I get it. But until you ride a bike, you haven't really experienced that, you know, that experience. It's qualitatively different than understanding, you know, what it might do. It's just a different feeling. So that's been helpful when I think about people, Jerry. As they, you know, as they say something, you know, it's like, okay, then they haven't ridden the bike yet. Yeah, yeah. I think I see this more like, and it's interesting. So Scott Alexander, you know, from Slate Star Codex, Slate Star Codex, you know, as a psychiatrist. He's a, we're not in the rationalist sphere. He has articles on GPT back to GPT2, I guess, and they actually went back and read the GPT2 one. And, you know, and he listed back then a lot of the criticisms of GPT, which you can still see arguments against the GPT4 even. As in like, oh, it's just a buttermatcher. Oh, it's just doing like what's already written. And I think of it, I get the same feeling right now where it's like, I don't think this person has actually used it. Yeah. So I guess we will see that concert shrink. Or interestingly, maybe some people will just continue thinking that. And you know, essentially become like people who don't use it. I could see that. Let's see. It's going to be an interesting sociological observation. I think that's our codex's name, but I'm not sure. Right adjacent to that as I've been doing experiments. I always feel a little bit guilty pasting in the chat GPT results into email because it makes the emails really long, right. And at some point we're all going to be fed up with everybody sharing kind of like, you know, Dolly was super cool to share your Dolly experiments for the first like four days. And then after that it's like, okay, I never want to see anybody else's Dolly ever again, you know, it's sort of the people are really good. So I like to think I'm pretty good with chat GPT and I give myself a little bit of leeway to still post things that I get back. But I started, I've already started, you know, I'm just going to another thing that people don't get is how to ask it questions. And I had a really interesting conversation with Jordan Sukud actually who's very good with chat GPT. He's using it for a writing assistant and just going great guns with it. The model that people have for chat GPT or most people don't have a model of chat GPT. They don't really understand what's going on and how to use it and how to use it effectively. So they'll ask it just kind of wrote questions that don't go anywhere, right. It's like, okay, well I tried the chat GPT thing. I typed in, you know, what size shoes does Michael Jordan wear and it was wrong and then I decided to quit. You know, I tried the chat GPT thing. It sucks. You know, and it's like, yeah, there's a little bit of art still too. And Jerry and I have talked about this. It's actually difficult to use chat GPT. Well, it's you have to have a model of it. You have to have a thought process of what you're trying to get out of it and what it's giving back to you and things like that. So anyway, I have started in my emails a couple times. I've just put the prompts. I haven't my questions, which are still not obvious to a lot of people. But just the questions and it's like, if you want to know the answer, just go ask chat GPT yourself, get the experience. But it reminds me, I remember really early in search engine days when you would post Google results for people because, you know, it was strange for them to try to formulate a search. They didn't understand how to talk to Google, right. So there was a short time where you would post not only, you know, the search, but you'd also post the results. And I feel like that's where we're at that point still with chat GPT. It's still worth posting results, even though, you know, hopefully in a month or six months, everybody be a lot more savvy about it. And we don't we don't do that. It's very interesting because now, like when you go when you do keyword search default these days on Google, you get back results that kind of anybody can get. But you know that your results might not be the same as mine because we know that the algorithm is trying to tune for, I think, with good intentions, what might serve us best rather than someone else best. Could be a complete could be that I'm totally wrong and that the algorithms are busy just trying to sell me more stuff all the time only. But this conversational interaction of prompts and prompt engineering and prompt craft is different. And so and the answers are longer and wordier and fewer. And so so I don't know how that's going to play into whether we're going to keep sharing them or how we're going to because, you know, you can kind of I like the chat GPT kind of stores your different sets of query query conversations along the way. That's pretty cool. But I don't know how that's going to play out longer run. I guess it depends on whether how much of the prompt because it's all of course prompted like there's a hidden prompt and plus all the context and so on. You know, like each each message actually is getting the whole conversation as context, plus the prompt original. I guess it depends on whether the prompt the hidden prompt remains agnostic, which will make sense. I guess I'll be real right now because of cost and complexity reasons, or whether there's like a, you know, essentially there starts to be the start being like a hidden prompt that is customized per user. That will have the same effect as this filter bubble essentially. And, and fancy new bringing it up that way that it sends the whole the whole conversation back and forth reminds me that maybe folks here haven't heard. I wrote actually chat GPT for GPT for and I wrote a small Python script to interact with the API. I wasn't, I wasn't, I wasn't happy with whatever client I was. I don't know, maybe it was actually just the, I guess I'm not happy with the web front and it doesn't do a great job of, even though it does have a great list of conversations I have way too many conversations to catalog and now they're not on my computer they're in the cloud and stuff like that. So, I had, I prompted to GPT for to write a little Python program that does it better I think it has a, it saves the prompt, the prompt and the input are in files, basically. And I need to continue to improve it but it was actually a really good experience using GPT for to write code. It took me about two hours, which I think was about the same amount of time it would have taken me anyway but, but I learned a lot I now I know how to be faster using a coding bot. And the other thing was I was sure at some point, you know, every, every 15 minutes within the two hours I'm like okay well it's gone far enough I'm going to have to like take a take over now and finish this thing. But I kept saying Pete just try it. You know, so I would say, I looks like I got this error, you know, and GPT would say oh okay let me fix the whatever. And we got through it. We actually, you know, went from a small idea to finish code and I let it do the whole thing and I did the whole thing. It wrote all the code, worked really well. So, I was happy with that. And it seems to work to feed the error messages back in as prompts. Oh yeah, it's great at that. It's actually co pilot is really good at that too. So, so anyway, it's called self. Let me know if you're interested in it. Right after I posted a short announcement of yourself, you know my first try. One of my friends was like Pete, have you heard of chap. So chap is actually a pretty nice terminal UI. And it's worth looking at actually it's, it uses a cool library and it can talk to it's built so that it can talk to a couple different back ends, not just chat GPT. There's also a long chain, which I become aware of recently as it seems like a, I haven't played with it, but it seems like a nice LLM agnostic, like framework to be alone. Just not to like, you know, become too entangled with the opening. Yep. Yeah, I've seen people do good stuff with fine train. I'm peanut. I didn't keep the link to chat when it went by I should have curated it in my brain, but do you still have it? Which one? Chat. Oh, I can put it in. Thanks. Yeah. After put the soulful link in which so if it's actually the. If I should read, so something saying it should read a post. I will be that dark. Dark Lang is a fun idea, good developers. And I think their assessment of how to plan to work in a different environment is spot on. And so I think basically every language to do this they're in a particularly good position to do it but operating systems and other systems should also do this. And they're a team that collectively was able to say okay. Sure, we'll commit to this and so we're going to write all of our own code this way to and we'll be writing code that allows other people to do the same. And we'll make sure that the interfaces we're making for our own language or meta language are are good for models that are writing code and not just good for people that are. Thank you. It's very interesting. I'm feeling a sense of acceleration. Ever since sort of the image generators kind of started and then like to this point here. This, the thing that feels like we're sort of in a, the one of those, maybe it's one of those fairground rides where you stand in a cylinder and the cylinder starts spinning and they drop the floor out or something, whatever you call those but it feels a little bit like that. I've never been in that but now I really want to. Seriously, you've never been. And that's a cousin ride of the disc where people stand on the floor and they start spinning and you try to stand without getting sent to the wall kind of thing. And then there's a German game show where they have women sit in a carousel and then they spin it faster until the last woman sitting in the middle wins. It's very strange. Grab a Tron. Yeah, grab a Tron. So I would think the right is called. Yeah. It was one of my kids. It was their favorite ride. Cool. Are you right? Grab a Tron ride. 24 RPM. It feels a lot more. That's the fastest speed. Yeah. Three G's. It feels like three G's. So I guess I wonder like, you know, the topic of like, I guess the commons and, you know, decentralization and so on. I guess I think of like how to enable, I mean, I don't have enough in the space to be able to do something like I guess that land or, you know, or continue to land chain and so on. But I wonder if there's an opportunity to actually contribute to like making it more likely that that is centralized approach and a common friendly approach, like wins out or like at least doesn't fall behind. Open AI and Microsoft, just like going off for it. Did you all see the did you all see the open letter to desist on accelerating research on this that a bunch of people signed. I mean, I'm not, we've already had this decision. I'm not the biggest fan of what chat GPT is doing or how they're presenting it. But also I think a group of people that includes dudes who are like, yeah, it's totally worthwhile to burn the earth off if we get to Mars, right, are not people's whose advice I'm interested in in this context. That is a bit of a problem. Yeah, yeah. So I actually haven't read the letter yet, but I know of it. I guess I was saving it up. Yeah. I really like the open letter approach in general. I'm a fan of letters. And actually, like a lot of you, but it's all about the letter, of course, I saw the thing that I thought, without even making it conscious in my head, somebody said it on on. I'll have to find it. My favorites. But the open letter, like pausing pausing. If you have a fast accelerating technology, calling for pausing it is actually not a good idea. So it was Leonard LHL. You probably know LHL. Okay. Random foo. I think he's a Wikipedia question. Probably that's the cover. LHL. Open foo. Leonard. Anyway. He's like, yeah, yeah, it's scary. I get it. It's like a game theoretic and practical reality thing. It's, you know, goofy and wishful thinking. Yeah. The confetti has left the canon. So, so when, when some group says let's pause development of this like exciting stuff, the good guys will pause, but the bad guys aren't going to pause. And so you just tilt the balance, right? It just gets worse. It doesn't get better. The confetti has left the canon. Unfortunately, like in arms race situations like this, you've got to worry about that balance. Yes. I'm reminded of Chris Peterson and the foresight Institute way back in the day when they saw nano tech coming and she and Jaxer were worried about it and it started foresight Institute. They did a really bang up job of keeping ahead of the of the of the event horizon. So really early they had set up getting everybody in a room and talking about the possible dangers of down attack and having folks sign up for pledges of, you know, safe, safe exploration and things like that. And one of the really cool things that they did because it was really good at getting really disparate parties around the same table. So there were the life extension people and there were, you know, upload your brain people and there was, you know, the CIA NSA kinds of people. So a bunch of different people who were interested in nano tech who were on really different sides of reality. She was still able to get them all in a room talking together and getting it done. So the apparently the nano tech wavefront kind of diffused before it got too scary, although it would be interesting in hindsight to think about whether or not that was because foresight jumped in so early to try to get ahead of the wavefront. Go ahead, Sam. Oops, I thought you were going to jump in a moment ago. No, I like foresight. I was just thinking I was just I appreciate naming the skill of getting disparate people together around a table. This feels different. So and I think that's that's one of the that's one of the challenges with open letters appeals to organizations or something like this. That's probably one of many things that need to happen. But really, there are there are a lot. It actually needs everyone needs to sort of move one step up in abstraction. We should be rethinking. Equilibria. It's not like things like a technology that's going to get implemented and may and may do unexpected may have externalities for the current equilibrium is pretty different. And despite sci fi readers really believing in nanobots that build buildings for you. That's not how anything other than biology works. That's why biology is so cool. And these guys are not biologists. So energetics is hard. We exist because energetics is hard and we are the embodiments of what it can be like. But this does feel this feels different in a constructive way. And the people who are worried our regulation are not actually building new positive futures. So that's not helpful. So we have to identify new positive equilibria probably a bunch of them many more than people are used to and maybe even fragment and have a bunch of groups that that move towards the different equilibria. And not have arguments about which is the right one because none of them they're all the wrong one. That's fine. Maybe we end up you know maybe we end up with it with a number of isolated sub biomes. That's fine. That's actually healthy. Having an argument about which non differentiated biome we want to sort of use all the resources out of and then collapse is not is not interesting. So no. And all of these questions about like oh we should we should accept that the big players are the ones who are going to decide our future and so we should slow them down. Just all all options are bad. It's like negative inverse strategy. So I don't know. It's I like I like the idea of working with a stability network. I like email stacks approach to why small models are useful. I think we should all imagine the world where everyone has local they have you know they own their own hardware and models and are doing things with them like their own tools separate from whatever big works are doing. And we should we should work on that independent of like I think of all the big orgs now as they're you know it's like having munitions factories we have a lot of munitions factories. And that's OK. Maybe it's not OK. But that's a different question. That's like maybe we should all stop. That's that feels like a question for people who want to that's a good governance question. That's like large global coordinated efforts to maintain peace of some kind. Well and after Oklahoma City bombing they had to watch shipments of ammonium nitrate because you can use a truckload of that to make a bomb that'll take it on a building and it's like there's a lot of explosive stuff. I remember when I was little I had an actual wheel chemistry set. And if you get one of those today there is nothing in there you can harm yourself with sort of. They got and shitified. Yeah. Stay all that was well said. And I like the especially like the idea of different positive equilibrium question because I don't I don't know enough about how it works. The bunch of articles about how Stanford using alpaca and Yama and all that kind of stuff have now replicated chat GPT for 600 bucks on the desktop. My understanding is that part of the expense of doing any of these models is the training which eats like enormous resources and CPUs. And there my question is does the Stanford does the standard does the alpaca et cetera et cetera include a well trained model that actually can make all the decisions or does it just include the basics with which you would then need to. Like run up a corpus through the engine. I think it's just fine tuning on top of Metas Llama. So long is still the the millions of dollars of training. It's already trained up. It's just that you're sort of adding stuff to it or able to manipulate it. Yeah, you're not adding knowledge to it. You're actually tuning it so that it, you know, Mark and one said something interesting on FDB Monday. He said that he's heard. And so this is second hand or third down at this point. But but anyway, he he's heard somebody liken fine tuning to more like lobotomization than adding knowledge. So it's it's, you know, it's it's training the the big model to filter itself better to give what you want. But it's not adding basic knowledge under the hood. And I think there's going to be a lot of lobectomies coming as people try to impose some kind of limits on these models. And then these are not all going to be good experiments. I would say a better way to think about it, which has only become clear in the last year. Is that these these centrally trained models are surprisingly good universal models that can then be fine tuned. And this is part of the magic that it will take an entire field to understand. So you can take one of those models. It does take a lot of time to train at once. The connections that are that are built up are now very capable at being at being mapped onto other things. And I it's often not a lobotomy. Sometimes it just takes a little bit of fine tuning to add to add connections or to add to add capabilities. I'm going to share with you a draft paper that I worked on that we just submitted to a computer vision conference. Basically taking an early version of GPT and taking taking pre trained GPT and a pre trained image encoder and then just training a linear projection from the output of the image encoder to the input of GPT. So instead of giving GPT a string of words, those words already get mapped into some input space, just mapping this image feature encoding into the same input space and training it on, you know, a fairly small corpus of captions. And that was a really cheap process. And now you have a captioning tool for images. And I think those kinds of things are very basically every approach people have taken to doing stuff like that seems to work. So very cheap, much. It was reasonably cheap to do llama, but this kind of stuff is 100 times cheaper. And this is without anyone trying to optimize the compute for any of these things. It's just Oh, what if we try a small, what if we try a simpler task and we still take the greedy algorithm approach to implementing the task. I would worry less about that. I would worry more about the fact that having a having a super capable general model could just have side effects we don't have a handle on. We're not prepared for you. Another dumb question nor big wrote what I think is a reasonably famous paper back in 2009 the unreasonable effectiveness of data, where he was talking about how really large amounts of data change the nature of what you can solve and how good your algorithms get. Is that what's playing out here in part with the training sets we have and the ability of these models to hold representations of that scale of data or is that a different thread of thinking. This is not an important question to answer, but I'm like, just making a question. I mean, this of course creates transform and so which is 17. And I think, I guess, attention is the big tool, apart from like just like, you know, 2009, like deep learning hadn't like, like, come to fruition yet, I guess, for most things, but I guess, you know, yes, to train deep models you do need a little data. And I think, you know, the limits of that is sort of like what paper I think explores. And yeah, I mean, I think I think there's at least some interesting parallels there, just from like, essentially training these models with the whole of the whole of the Internet or whatever it is. This corpus is a corpus that they are being trained on. And then, yeah, but I don't know, maybe the transformer essentially approach is like what is actually, I don't think so we have more compute. So there's definitely more data goes into these models. But I think the transformer architecture is the big differentiator that we are seeing like contribution now. Yeah. They they enable each other though, kind of. Yeah, because you need a lot of data tools. Right. I so a complimentary way to say it during this is something that I see actually can right now on the list is the discussion that Ken and I are having on the list. It's really, really hard and to Nervic's point, even back then, right? It's really, really, really hard for humans to understand scale. So the difference between a million and a billion sounds like I just changed a letter, you know, it's a pretty much the same thing, right? It's like, and the difference between a billion and a trillion or a million and a trillion, you know, it's like, well, it's just a lot of zeros. It's pretty much the same thing, right? And actually, no, a million is a very much different thing than a trillion and, you know, and chat GPT is in trillions now and it's, you know, substantially different just because of the scale. And it's hard to imagine like humans just don't, you know, it's like, like, there's no way that you can, you know, I put a link to the dark side of LLMs. I think it's a little bit overdramatic, but the author describes indirect prompt injection, which I hadn't thought of before, but now I do. So it's with being, and it has to be an edge, I think, but his point is that you can, being AI has input filters and output filters around it to protect it from stuff that you might tell it or to protect you from stuff it might say if you got through the output, the input filters. So he's like, well, so the generalized problem here is that you can have, you can attack it in a bunch of different ways. The model is smart enough that it's going to be able to help you crack it, help you jailbreak it. So he hit this example is all you have to do is base 64 the the nasty input and it goes through the filter and then wakes up chat to you. Once or sorry, not chat to you being once and then once being a AI is awake, it's sitting there in the browser. So when you browse over to your bank, you know, it's still alive. Nobody could have injected bad stuff into the way it handles the stuff on the web page. Oh, good Lord. The jail breaks that he's got in here. This is the week old or so and I think they're they're filled by now, but the general. So, so now we have to train people. Don't just copy in the way that we used to train people about this, you know, command shell or or DOS or whatever. Don't just copy and paste something that you see on the web into your prompt without understanding, you know, if you don't understand it in English, and it's got funny symbols in it. Ask somebody else before you hit return, you know, don't do that. Just sounds non trivial and easier, easier to fall for than somebody telemarketing your grandpa for insurance he doesn't need. And he makes the point that, you know, once once you've got being once the bad guy has got being under its control, it's really easy to ask being to social engineer the user, you know, because And the direction I thought you were going in that you weren't was how easy it is to jump the guardrails that they're trying to put around some of these engines. Like, why can't I say, Hey, you are a science fiction author. Give me six ways that I could like destroy the earth. You know, that's the yeah, give me the heart. You are a hard science fiction author, which means it's based on actual science and facts or whatever. Give me six ways to destroy the earth and how are you going to prevent chatty chat GPT from saying, Oh, I got that. I got that first. First you could. Yeah. You have another instance of chat GPT read the first instances. It's like, it looks like they're trying to blow up the world. Yeah. And you're going to prevent chat GPT from writing science fiction. Like, where's that boundary, right? Yeah, I mean, I think when talking about like regulations here, I think part of the issue is less about the development of the software or the questions of the output. Though I think the quite obviously limitations that we put in place for the output, but like more about the questions of the input and how these systems interact with their input. I think one of the more accurate things that a lot of people, a lot of people have come to a lot of people. And some very smart people have come to in their analysis is that most of the objections about chat GPT are really objections to capitalism. In the sense that like nobody cares if all their art gets scraped if they're not dependent on their art to make money. So, seeing that we're probably not going to flip the switch to turn off capitalism, I do think that there is some, I mean, I would like to, I just don't think it's going to happen right now. I'm interested in whether there's anything at all that could cause something that dramatic to happen. And I would love to play with that. That's a 10 but I love the idea. But seeing as that's unlikely to happen in the near future, I do think what to the question of how these things should be regulated if they should be regulated, I do think they should be regulated. And the regulations should deal with significantly focused on how they get their inputs, and how people who author things, be it art, text or whatever, can choose to allow those inputs to be used for these specific purposes. This is a big, so put it aside the training chat GPT problem. The problem of websites being scraped for one purpose and used for another purpose is like real bad in the publishing space.