Episode 275 - Connecticut Chronicles, Columbia Conferences & Questioning Bots
Aaron talks to Max about his move to Connecticut. They review Columbia University's data science day. A touch of Twitter news and a tweet we thought was interesting on rationality.
Probability Distribution of the Week: Multivariate Gaussian Distribution (or Normal Distribution)
Links
Viper: ViperGPT: Visual Inference via Python Execution for Reasoning
J.P Morgan: Artificial Intelligence Research
Twitter: Perry E. Metzger “The way to become better at rationality isn’t primarily about knowing Bayesian reasoning or decision theory or anything like those. Those are usually more interesting to academic philosophers than to people actually trying to be less wrong.”
YouTube: Why π is in the normal distribution (beyond integral tricks)
Related Episode
Episode 253 - Make Better Decisions with Helen and Dave Edwards
Episode 269 - Image Recognition Technology for Health with Susan Conover
Episode 273 - Stop Making AI Boring
Transcript
Max: I am recording on my Zoom.
Aaron: Okay, and it looks like I've got Audacity going here.
Max: All right. So we are ready to go. All right, so let me, I'm going to play the sound. So I'll take us out of the sound so you'll still you'll know when to when to start talking, right? Okay. Okay. You're listening to the Local Maximum episode 275.
Narration: Time to expand your perspective. Welcome to the Local Maximum. Now here's your host, Max Sklar.
Max: Welcome, everyone. Welcome. You have reached another Local Maximum. And today we are joined by Aaron. Aaron, how are you doing?
Aaron: I am doing well. Glad to be back on the show.
Max: Alright. So, it's been a while I'm pretty exhausted here. I don't know why. You could see I'm in a new space. You're in the same space, although behind you is the podcast desk. So all right, this is my Stamford, Connecticut Local Maximum studio. This is also where I'm working every day. How’d I do?
Aaron: It looks good, sleek and modern and clean. Unlike my workspace, which still has boxes that haven't been unpacked from when I moved here, mumble mumble years ago.
Max: Oh, well, it's only that you can't see all the boxes at my feet. Because I'm at my standing desk right now.
Aaron: Shows that camera placement is important.
Max: Yeah, yeah. That is very important. Yeah. And right. I got the standing desk. This is interesting. We're going to try to do the podcast standing today. And yeah, I've been here nearly a month. I know you.
You were supposed to question me on Stamford before we get started. We've got some more topics. I'm gonna talk about my time at Columbia Data Science Day. Some Twitter fun and probability distribution of the week and we're gonna wrap up.
Aaron: Awesome. Yeah, so let's let's dive right in. So you are now officially in Stamford, you have been for almost a month now. Right? Like you said, this is the first time recording from at least from your new studio room. Right? Are you fully settled in? Are there still some big milestones for things you need to do to get the place feeling like home?
Max: Yes, there. No, I'm not fully settled in. I think part of the problem is you move in on day one. And then you even unpack largely unpack on day one. And it's like, okay, great, or day two, and you're like, great, okay, I'm here, I can start working. I can do whatever I want.
But then over the first few weeks, things start accumulating where you're like, oh, I don't know where this is. I don't, I have a toothache. I don't know where the pharmacy is. I don't know where to go. I'm missing something. And I don't know. And so it's sort of like an unsettling feeling that comes around or like, I'm dealing with the building where my old building let me do X, and my new building won't let me do X. And it's like, well, I didn't know that, so things like that, that come up. And so that's sort of frustrating.
I have to, I have to register the car eventually. But I want to keep that Life Free or Die license plate as long as possible. And yeah, still exploring. So I started a new job, very small startup, like seven people. Actually, it's weird. So when I first decided to move to this area, it was like, well, all the opportunities that I'm seeing are in New York City. Turns out my job is remote. So I really don't have to be near New York City, after all.
But I did feel like I was traveling down here way too much. And having family down here. And also having lots of industry events and stuff in New York. I thought it'd be good to come down here. Just try to get my career restarted. If that, if that makes sense at all. I mean, I know there was lots of, yeah.
Aaron: We'll talk more about Columbia Data Science Day. Yeah, but but but so you've been into New York City multiple times since you've moved down?
Max: Oh, yeah. I would say like once a week at least.
Aaron: Awesome. It's certainly a heck of a lot easier than when you were up in New Hampshire. Yeah, and I did it. What would you say is the thing you miss most about New Hampshire at this point?
Max: So I miss being able to come over to your house. And I miss having you come over and do the podcast that's one for sure. I definitely. I definitely miss going to some of the like New Hampshire events that all the friends I met in New Hampshire from New York, the Free State Project events. And it's interesting.
So now I'm in kind of a more urban area. And we could talk about this and sort of, so I've, I haven't been in a house since I graduated college. But I, I sort of miss being able to walk outside and have all the nature right there. And but I also had lots of stores and stuff. I still lived, everywhere I've lived, right near a mall, and this is no exception.
Aaron: New Hampshire certainly wasn't rural. But you did have the what is the rail rail trail right behind you. And so it was easy to get away from the mall, even though it was conveniently located and not feel like you're living in a city.
Max: Now look, there are some nice parks here that I walk outside to so that's good. And maybe I just need to do a little more exploring. I didn't find the rail trail in Salem on day one, but downstairs here is so on the good side. It's like hey, I can go out 10 at night 11 at night there are bars open, people having fun, whatever, stuff is happening, but it's also like New York where there are some like weirdos about, you gotta like kind of watch out.
But it's, I have been able to drive on the good side here. I have been able to drive to some parks, like state parks and even Devil's Den, even all the way to Weston and to beaches. Actually, I'm finding those things are a lot more accessible here by car.
So whereas New Hampshire, yeah, all that stuff exists in New Hampshire, but I have to drive like an hour. Whereas here I have State Park. I tried. I went to the Mianus. I think it's called Mianus State Park today. Have you ever been there?.
Aaron: I’m not familiar with that one, no.
Max: It's in Stamford, Connecticut. And, and I did some hiking on the trails along the river. And I kind of overdid it a little bit. I went too far in and I was alone. And I was listening to a podcast. And then. So I was like, oh, no, it's gonna get dark.
So I was getting confused because it was, the trails were confusing. I was almost at the end, and there was a narrow part in the path. And then a deer walks out and just stops right in the middle of the path. So I have to kind of like tiptoe around the deer being like, I'm not here to hunt you. It's okay. Make sure that, but so that was a little harrowing earlier today, but I think it turned out all right.
But I do miss, I feel like there's a positive and minus, there's a plus or minus when it comes to the wilderness like there's some things are in New Hampshire, I had more stuff that I could walk like, I'm in a much more urban environment, right in my immediate vicinity. medium distance, like 10, 20 minute drive. I have really nice places here.
But New Hampshire had much more to do in like a longer drive, if that makes sense. Like you go skiing. I don't know, but I feel like it's more just, I miss not paying taxes. That's kind of nice. I've already started paying state taxes. Yes, I forgot. Yes. This past week was tax day. Yeah. Well, I didn't pay state taxes this week, but I'm used to having something get ringed up, and not having it add a few dollars. So that's kind of nice about New Hampshire.
But yeah, I just think it's just a very different culture. And it's, it's almost like, I need to have a step in both. It's like I like both areas. And there's nothing that'll satisfy me, it'll always be grass is greener on the other side type thing. So maybe, maybe in the long run, I've got to get a vacation home up there or something, or just crash your place.
Aaron: So I got a couple other quick ones for you. So with the move itself, which, thankfully, that is over now, like you said, there's still a lot of little things to get squared away. But what was the toughest part of the move?
Max: The toughest part of the move that is…
Aaron: Or most stressful part? Maybe?
Max: Yeah, I'm kind of turning up blank because I'm, I feel like the fact that I had to move so quickly because it was such a, I decided on my job at the last second. So I didn't want to move right away. And like I think the fact I think searching for apartments was very stressful.
Aaron: You did move pretty quickly on that.
Max: Yeah. Then I think there was the stress of having the movers come in and I did have them pack this time, but it was kind of like, Ooh, they're going to be going through all my stuff. That was a little bit stressful. And then I, I felt like, I feel like it's a little better now but for the first part, I was having trouble figuring out how my furniture will be organized.
This place is a little smaller. New Hampshire has more space associated with, the apartment itself is not much smaller, but it's still I feel like the bedroom is larger, so like the living room, the couch doesn't quite fit in there. And I think and I was also trying to make my podcast desk fit in the living room. But it didn't work.
So all the furniture, all the all, both the podcast and the workspace are in this room, I kind of wanted to separate it, it didn't work out. And then I put my Ottoman in the bedroom so it kind of worked out. So it was hard to make it feel comfortable.
Aaron: Nothing is lost or damaged in the move? No, no headaches with that?
Max: No, no everything that I searched for I found but there were a few things that were like, oh, like I left it with so and so or like, yeah, so but I'm pretty sure nothing got lost, everything that I looked for. There are a few things that took a while to find but everything I looked for got found.
I built this new bookshelf over here because the New Hampshire apartment came stock with a bookshelf. So I had to put together a lot of furniture. I got very, it was, do you enjoy putting together furniture, Aaron? Because on one hand, I enjoy it, but…
Aaron: I think my wife enjoys that process more than I do.
Max: Really? Okay.
Aaron: She finds the assembling IKEA style furniture to be meditative.
Max: Wow, yeah, me too. Me too. But there was one issue with this, this desk that I'm here now, I ordered a desk online. And I put it together. And it was kind of rickety. It was large enough. But at the end of the day, I decided you know what, I'm going to return it.
So I deconstructed the whole desk box, I had to get it back in the box, I had to take it to the UPS store. And now I got this other desk, which is that you could see right now, which is so much better. But it was just like and then this desk takes another couple weeks to come and I'm still working at the kitchen table and it's not very comfortable. And I had to figure out my chair situation all that so it definitely took a long time to get comfortable.
Aaron: Okay, so you mentioned a couple of things that you like about the new location you're in. Yeah, you also mentioned when we were talking before we started recording that you're right near a train station which is super convenient for not only the city and going elsewhere, but have you found a spot I'm thinking like a restaurant or coffee shop or a kind of a hidden gem in your new neighborhood that you want to give a shout out to.
Max: So maybe not a hidden gem yet. So one is the the train station here that I'm near in Stamford downtown, Stamford. It is both Metro North and Amtrak so I can get into New Yorker or whatever anywhere in Connecticut really kind of use it as a subway, if you want it to although usually would drive.
But it's also an Amtrak station, so full Amtrak station so you could use it to get to Boston you can use it to get to Philly, DC, all that. So that's really cool. I'm looking forward to using that occasionally. Around here if I go downstairs there's just a ton of stuff that's open late. There are a few bars, there's a Greek pizza place that is open late that has been really useful. You can walk, I went to a sushi place tonight, a Japanese place, a little takeout. There's Curly’s Diner, full diner here, which is okay, of course there’s Dunkin Donuts. I always live near Dunkin Donuts. It is kind of a necessary necessity.
There's just as much here as there is in New Hampshire and Massachusetts. And yeah, I don't know if I have a favorite yet. I just have some places that are convenient that I'm trying but there are a lot of places that are convenient. Oh, I did find like if you go up close to the Merritt Parkway, so I've got to drive.
Then there's one part of town with like, I think, a better kind of lunch selection because it's got a Rye Ridge Deli. It's got a Layla’s Falafel. Those are all very good. So, I'm gonna explore that a little bit. But so yeah, and New Greenwich. Oh my god. I went to Greenwich. Have you ever been to Greenwich, Aaron?
Aaron: A couple of times? Yeah.
Max: I felt like it was like the Metropolis version of the Hamptons. It was like, it was insane.
Aaron: It definitely makes you feel like a schlub. No, no matter what your socioeconomic class is.
Max: Oh, yeah. I go and try to buy things for the apartment but then they were trying to sell me like a $300 shower curtain. And I'm like, I don't know if I want my shower curtain to be $300 some things maybe but that just doesn't doesn't make sense to me.
Aaron: At least they didn't give the you the whole like, Oh, if you have to ask then then you don't want to know the price.
Max: Of the shower curtain. Right? And then there's a whole foods there that's kind of like a normal Whole Foods and there's a Whole Foods. Yeah, and what I like about this building is that I keep forgetting that I have a car.
Because it's like you go to the second floor, you can get your car from the driveway, from the parking garage, and just drive right out. And within five minutes. It's not like being in New York where you have an in-building garage where it takes forever to get out of New York. But here you drive five minutes you're in the wilderness.
I guess it's obvious but one thing to keep in mind. Like there's a North Stamford which goes as high as you know Weston and Wilton. So literally North Stamford is all just like, it basically looks like rural Connecticut that we both know.
Aaron: Well, changing gears a little bit. So you've been at this new job for a little bit longer than you've been in Stamford. What is it like working for someone else? Again, you're no longer solely employed by Local Maximum Labs.
Max: Yeah, well, so I am definitely. So I'm definitely enjoying picking up new problems, I think I'm gonna get to build some great causality models that is going to help customers drive revenue, which I'm really excited about. Startup is winware.ai, I'll just be, we don't have to talk about it a lot right now. But obviously it’s on my LinkedIn and all that. And so I'm kind of excited to tackle some real world problems, because I feel like even the side projects I have are kind of fed with these real world problems like newmap and stuff. So I'm excited about that
I'm sort of trying to think real hard about how to get myself and this might maybe we could do an episode on this, because I think this would be an important episode is how do you keep as productive as possible during the day? How do you keep focus during the day, especially if you work alone?
Aaron: And when you solve that problem? I definitely want to hear the answer.
Max: Yeah. So I have one thing that seemed to help to start out which, for me, it works. I'm not sure exactly why it works, I have a suspicion. So I put in, not headphones earplugs. There's something about the ambient noise around that, like, drives me crazy. So maybe that'll help. I'll try that. Try to rely more on that this week. And we'll see how it works.
Other than that, I think try to figure out how to put together the schedule, better to be most likely to be most productive. I'm having this problem where I'll take a break in the afternoon. And I'll be like, Oh, I'll do it the evening. And then, and then I have all the stuff to do in the evening. And then I ended up going to bed late.
Aaron: I was gonna ask because I know you do have some colleagues who are in different time zones, and in fact, different continents. Has that presented any challenges or everybody's schedule is all over the place? So it doesn't really play into it directly?
Max: Yeah, I think the challenge is just slower. Like, if you just have someone working right next to you, it's great. Because you can above I need somebody next to me to bother all day to ask how to do my job. But I think that just more frequent communication is the key. So but yeah, there are definitely times when it's like slow.
Aaron: Well, enough about that. Let's dive into kind of the main topic of the day. You went to this Columbia Data Science Day. Was this last week?
Max: Yeah, it was last week. It was at Columbia so I only had to take the express train one stop from Stamford to Harlem. 1/25. So that was great.
Aaron: And this was your first big conference in how long? Since the pandemic?
Max: No, no, no, remember ML Conf that I went to a few weeks earlier. And that was episode one, or sorry 273, where I basically ranted about how I didn't like the conference.
Aaron: That was your first conference back. And I remember because I listen to that episode. And I commented you afterwards that I like angry, angry Max. Yeah, that's a good episode.
Max: So I'm less angry today. This conference is way better. First of all, it had way better sandwiches and way better lunch. And I get it. Not everyone has the best lunch. But what I felt like for ML Conf, when you're paying $250, and it's sponsored by Amazon and Google, you better have a damn good sandwich for lunch. And they didn't have it.
Is that wrong? I don't know. But since there was better food and coffee available for this one, which I just think Columbia does a very good job with. And the speakers were all very interesting. I think they're really talking about the kind of cutting edge of AI, which you don't have to do to be interesting. I thought NormConf was very interesting, where they talked about, the idea of NormConf was talking about very normal problems in AI.
But like, I just felt like they were, they made a much better effort to keep it interesting.
Aaron: Yeah, I was gonna what, well, mission statement is probably too buzzword-y, but like, what was the focus of the Columbia Data Science Day? And was it meaningfully different than what ML Conf was trying to do? And just one was more effective than the other? Or were they aiming at different targets completely?
Max: Yeah, because Columbia focused more on researchers and research. Now, the keynote was Manuela Veloso So she was a longtime researcher at Carnegie Mellon, and actually played some of her videos of robots playing soccer, I believe. And the one takeaway on that was, she showed the one from like from 2007, with robots playing soccer really well, then one from 1997, where they're really bad. And she's like, the moral of the story is, you got to have a lot of patience. Because that was a lot of years.
But she now works at JP Morgan Chase, and is doing is doing automated trading. And I thought that her approach was very interesting. And I was very surprised by what it was. Because it's like, okay, you have numbers, you have time series data on what all the prices are, right?
And so I think, okay, why don't you build a model on that time series data directly? But what she did was, she said, look, you look at all these traders all day, and you see that they're looking at charts. So here's saying, why not take the image recognition algorithms, and recognize images on the charts as displayed, and so that was, that was very curious to me, and very provoking to me, because I'm like, well, that's you're just adding information.
The numbers are just raw information, the chart is just how it’s presented to humans. But apparently, image recognition on the charts. Well, she thinks it's a good idea. I assume there's some data to back that up. And so that is, that's wild.
Aaron: Yeah, so it's, it's definitely more it's the OCR version of interpreting text, rather than doing an analysis on the actual underlying text itself. That's yeah, that's faster. Exactly. Oh, and and I'd be really curious how, because one of the things that that I don't know if we've talked about it, specifically on an episode or not, but you can do bad statistics, but even more common is people representing data in an ineffective manner, or, or taking good data and representing it in a way that that is somehow intentionally or not misleading.
And I wonder how that impacts these kinds of algorithms' ability to scrape that data if, if it's been presented in a less than optimal or perhaps biased fashion even though the numbers are still valid.
Max: There's so many ways to present these charts. And I assume that no matter how you present it to the machine, so long as it's uniform, it can kind of figure it out, but there's got to be better ways and worse ways of doing it. I would assume. And yeah, it's it sort of reminds me so there's a lot of image recognition in some of the posters too. The posters were a lot of like master's students and people like that.
But there was a lot of like glaucoma detection, there was a lot of like, medical kind of stuff. And so not only was it a image recognition for medical diagnosis, kind of like we did, we did an episode in that. Yep. somewhat recently.
Aaron: With the founder, right?
Max: Yep. Yes, Susan Conover. Yeah. Episode 269. But yeah. So I looked at one poster, where they were actually that, what they were actually doing was saying, hey, the experts are better at this than the machines. Why is that? They would hook up a, like, eye tracking glasses, to the experts who are looking at these images, and also doing it to like, regular people looking at these images, and trying to figure out what the experts are focusing on even if the experts can't say, and I'm like, that just sounds crazy, is crazy enough to work. Like if it works, like isn't that that's, that's, that's very clever like, it's like this unknown knowledge that they can't even articulate.
Aaron: Oftentimes, humans are able to do something, but not able to articulate exactly how we're why they did it. It's the hidden knowledge or dark knowledge is stuff that's learned through repetition and apprenticeship and can't just be put in a textbook. And maybe there is a way to capture that. That is, in fact, fascinating.
Thinking back to what you were saying about the algorithm trying to read charts and parse data that way, what one of the things that popped in my head that might be a positive use for that is we've got that long list of worst practices, let's say, for displaying data. And if you could have it go through scientific papers, whether this is in published journals or submissions, and correct for all those things. So if it's data that's being shown on different scales and different charts and automatically realign that to force best practices, that could be a neat use for something like that.
Especially if you don't necessarily because I know with a lot of scientific papers, they won't provide the raw data that they used for their initial dataset, or if it is available, it's not necessarily included with the paper itself. So that might be a reasonable use case there. Sorry. Just kind of yeah, my, my gears were turning on that while we were chatting.
Max: It is fascinating. It's like it, I want to think about that more like why isn't the underlying data set sufficient? There's some way that I'm thinking that I need to update and I'm not exactly sure what it is. I kind of have an inkling of it. But I'm not exactly sure
Aaron: Or the frequent use case where you see a chart on the internet somewhere, and they don't cite exactly where they got the data from. And so you're kind of trying to reverse engineer the data from what they've displayed.
Max: Yeah. Yeah, that is that that would, that probably would be very possible with optical recognition, probably pretty easy. To be honest there.
Aaron: And I would expect it can do a lot better than me eyeballing it and saying, Well, that looks like a 12.5.
Max: Yeah, I'm sure just from the but it so long as the chart is like, kind of uniform, like kind of created uniformly, like, so long as it's created through some software, it should be able to find out figure out from the pixels, pretty much where everything is, although you would have some very sophisticated optical, or not optical but image recognition.
If it's like, well, yes, but the chart could be produced in any way. Part. We don't know exactly how the chart is produced. We don't know the colors. We didn't know the line scheme. We don't know the points. And so every charge is a little bit different. But I think that AI should be at the point where that's a solvable problem.
Aaron: Yeah. Yeah. So there were some other speakers you wanted to talk about. Right?
Max: Right. Right. So and these were all the panels, the panel is a good way of doing it, too. That was, you see, I liked the panels better than the speakers. They all spoke for, like 5, 10 minutes, and then they got to go on a panel, and they get questions. So one of the highlights for me was Carl Vondrick, who was working on something called Viper GPT.
And so this kind of takes your ChatGPT which is very bad at math and stuff like that. And also questions about images and things like that, and then it sort of takes your question and turns it into code. So one example is like, how many cupcakes are in this image? Well, it would take that question and then write code being like look at all these different windows, identify cupcakes and then like, add them together kind of a thing.
So it would take your question and then translate your question into code, then tell you what the code said, and then execute the code, and then tell you the answer. So I feel like that makes the AI a lot more transparent, and probably a lot more correct in what it's doing.
Aaron: So that's definitely one of the big things that a lot of people are looking for. I don't know if transparency is the right word for but but like, you're kind of legibility for what are essentially black box models that if there is some sort of thought process going on here? How can we better understand how it's getting from input to output?
Getting some visibility into that is, is key to figuring out what we can and can't do. And you're where we should be skeptical on where we should be trusting these tools.
Max: Right, right. And so, in my opinion, the most dystopian one was from a Show you, and I don't think she meant it as dystopian, I actually, she's a, from what I can see a very fantastic natural language processing researcher, and I believe I've reached out to her to be on the show, I'd certainly reach out again, but her demo, showed a bot trying to convince you to donate to charity.
And I felt a little manipulated by it, I have to say it, I'm like, well, what if you start to train, if you start training these bots to convince the, the human to do something, and project that out into a future that is, it doesn't even have to be general super intelligence. I feel like you can stumble on heuristics that can get people to do really dumb things really easily.
Aaron: This is why you need a bot to interact with that bot that will convince that bot that you're not a good target.
Max: Right, right. And so it was it would say something like you should give $1 to charity, and then the response was like, no thanks. And then it was like, well imagine if your children were starving, wouldn't you want someone to help them? And then the response was, I think, Well, where are their parents? And then it says, like, well, these children don't have any parents. So maybe you should give just $1 and help them out. And then they're like, Okay, I'll help them.
But then you do that you realize, oh, my God, I don't know where the money went. How do I evaluate this charity? Who's asking me this thing? And so you are, but okay, it's one thing if like, charities try to compete on persuasiveness of their bots, but then companies are going to start doing that, and governments are going to start doing that. And intelligence agencies are going to start doing that, if they aren't already.
Aaron: I was gonna say, what is going to make us think that this is not already being used as widely as it's at all feasible? It's absolutely the type of thing that is going to be leveraged from day zero.
Max: Well, we've all interacted with bots, right? I mean, and so but the idea is, the technology is at an inflection point now where it's going to become really convincing. Really fast.
Aaron: That's, that's one of the things that I think was was thrown out as a possibility for when when the ChatGTP really showed up on the scene, they're saying that if you can fine tune something to be the ideal arguer for insert, philosophical or political stance, and then you send it off to Reddit or wherever.
And no longer do you have to spend 12 hours a day arguing with people about your topic du jour, you have a tireless bot that can do it, and can fine tune the best arguments to best convince people of this and flood the zone with that.
Max: Yeah, I think that, as I've said before, the only saving grace there is that they'll also be bots to try to convince you otherwise. So we might get a steel manning of arguments in some way. But it might also not be presenting their best arguments. It might be presenting the most manipulative arguments on both sides, which already feels like what's happening
Aaron: Well, depends what your objective is? If you're trying to get someone to do a thing, that may be the best argument for your purpose.
Max: Right. Yeah, yeah. I'm just saying though, there'll be manipulative arguments on all sides of every issue. So hopefully, hopefully, it'll be an arms race type situation where some evil robot doesn't take over the world.. I recently saw an interview with Elon Musk, I think that was the one on Tucker Carlson where he said, well, the host said, Well, what's the problem with these AIs? And then Elon Musk said, well, the pen is mightier than the sword, as they say. So these can be the biggest weapons possible. Now, I don't know how that looks in practice. I really, I, I really have trouble imagining what he's talking about. But we'll see.
Aaron: Were there any other speakers you wanted to mention?
Max: There are a lot of speakers there. There are a lot of great ones. I think we can move on from here. Yeah, I think we can. I think we can move on from here. I guess the main news on, we could just skip that. The main news on Twitter this last week was that I got a checkmark just as everyone else was losing theirs.
Aaron: You're feeling blue?
Max: Yes, we're gonna try that out. And then is it called bluesky? Or it's not called dark sky? That's the one.
Aaron: I think it’s bluesky. I haven't been following it closely. But my understanding was, this is something that former Twitter CEO and founder tried to launch internally. And when that didn't go anywhere, he kind of took the idea with him to have a more open source. I don't think that fully captures it, but.
Max: Sort of decentralized?
Aaron: Right, so that it would be an infrastructure that not only that Twitter, the, the corporation is not the only one that can interface with the infrastructure for microblogging, for tweeting that you can have multiple providers, but on the same infrastructure and service, so it wouldn't be that you have to go to what is it, Substack Notes versus Twitter versus whatever Facebook's attempt at competing with that that same market space is that they'd all be able to interface with each other in some way, in a lot of senses, similar to a Mastodon type approach, but probably a little bit more user friendly. Have I captured it correctly?
Max: Yes. In the initial reviews are that this is better than Mastodon. I haven't tried it yet. I would like to try it. If you'd like to see more discussion about it, please let us know at maximum.locals.com, localmaxradio@gmail.com.
I actually do want to mention the thing I said before, which is somebody asked me at lunch and I had very interesting discussions at lunch, someone asked me, how do you get someone to admit the downsides of their creation? I think I was going on my rant about how all the AI for good talks are evil.
And one possibility I came up with is ask them, What is the worst thing that someone can do with what you're building? Not that you're doing it? Oh, no, no, not that you're doing it? What's the worst thing? What will your worst enemy do with it? I think that will get people thinking.
Aaron: Which is a thought experiment that I think many of our listeners will be familiar with, from a political context that when passing a law you have to think not, not what will I do with this power that could do good.
But if my opponent or the worst possible person to step into the position of power can wield this, what could they do that we wouldn't like? So, from a technology perspective, that's probably, I don't know if that thought mechanism that that thought experiment has a particular name or not.
Max: Yeah, I was gonna ask us to have a name. So yeah, interesting. Alright, before we get to probability distribution of the week, I want to mention Perry Metzger, who you've probably interacted with online a little bit. He had an interesting tweet storm on rationality, which I'll post on the show notes page, localmaxradio.com/275.
He writes, “The way to become better at rationality isn't primarily about knowing Bayesian reasoning or decision theory, or game theory or anything like those. Those are usually more interesting to academic philosophers than to people actually trying to be less wrong.”
And I think he's talking about day to day decision making or even personal decision making even of a large nature. And then another quote in the tweetstorm you'll have to go online to read it fully yourself. “The practical methods to improve things look much more like reading a reminder every morning about the importance of intellectual humility, and much less about futilely trying to calculate prior probabilities for matters where no information exists to allow such calculation.”
Now, so I, I partially agree, I think, obviously, I've always been a big proponent of Bayesian reasoning, and it's helped me solve problems, but it's always been, it's always been in a situation where it's like, Okay, I'm gonna decide to solve this really hard problem in a Bayesian manner, like Bayesian-ism is kind of like the big gun that I'm, that I am, that I'm bringing out.
But most of the times in life, you don't have the time or the energy to bring out the big weapons. So I feel like this is right. And this reminded me of the episode, the interview I did a few months ago, called Make Better Decisions with Helen and Dave Edwards, which is about I think, just that.
So I thought that was just an interesting idea to play around with it. We're bringing up a lot of ideas to play around with today. And maybe we'll, we'll follow up on them later. All right. So any comments on that? Or are we ready for our segment?
Aaron: I haven't read that particular tweet storm. But I know that the last several months have had.,..Rationality has been maybe under a more intense magnifying glass than it has been in recent memory between the FTX fiasco. And now, Yudkowsky is much more in the spotlight with his stances on recent AI developments. So hope, hopefully, this will be a crucible that strengthens the best aspects of that intellectual movement and burns away some of the cruft.
Max: I mean, the rationalist movement. All right. Yeah. Well, we'll see. Hopefully, there was something I wanted to add to that. I don't know what but I just, I just feel I just wanted to say I was thinking about doing a show on specifically, those arguments with, with Yudkowsky and all that. I don't know if I did an episode on it. I might have mentioned it a few times. But I feel like I tried to do a solo show on it. And it was just so I found it difficult.
So maybe we should do a show about that together. Or I have tried to have him on the show a while ago. But he doesn't do podcasts unless it's someone he knows personally, I think. So. Maybe I'll try again. But I moved just so he could speak for himself. But maybe we can dive into these ideas a little bit more. All right, time for our segment.
Narration: And now the probability distribution of the week.
Max: All right, the probability distribution of the week, and today we are talking about an old-timey favorite, the multivariate Gaussian distribution, or the normal distribution. Okay, so we spoke about the normal distribution, way back. Oh, shoot, I don't even know what episode we talked about the normal distribution. Do you? Do you remember?
Aaron: Oh, definitely not off the top of my head. But we have talked about it.
Max: We have talked about it. It was definitely one of the episodes that I'm pretty sure it was one of the episodes that I did with you. I believe so. Yeah. Okay. Okay. So the main takeaways from that normal distribution. It's like a bump, a famous bell curve and all that I can even write this here, I should get my mini. I can't draw it very well. Yeah, this is my attempt here, very bad attempts.
But okay, there is an analog to that in two dimensional space, three dimensional space, etc. So in one dimensional space, you have like your mean and your standard deviation, you're like, Okay, what's the average of the numbers that I'm getting? And how far do I expect to venture off from that central value? And because of the, the, what is it called the central limit theorem. Eventually, all distributions when you run them enough times start to look like normal distributions. That's why it's so common in nature.
But you can have distributions on spaces other than on a straight line other than a number so you can have a normal distribution on a plane. And a very good example of that is like a bullseye. you might think, okay, well, if I have everybody aiming at the bullseye, and they're all trying to aim at the dartboard right in the middle there, let's say it's a dartboard could also be anything else with a bullseye, like, you know.
But you'd think that the highest probability would be where they're aiming, it would be the center, and then as you go further from the center becomes less and less likely. And so, the standard deviation there, or the variance, would kind of be like, Okay, how, how good of a dart throwers are these people, right?
Aaron: So we get a distribution that is circular, with the highest density in the center, and decreasing as you move towards the edges of the circle.
Max: Right, right. And it's just like, it's just like a normal distribution. And then
Aaron: you can extend that into three dimensions. And now we're talking basically a sphere, but with that same kind of concentrated central core. And as we move out in any direction, it becomes more diffuse.
Max: Right, right. And you can talk about four dimensions, n-dimensions…
Aaron: My brain is not ready for that at this hour of the evening.
Max: You could talk about infinite dimensional spaces. And in fact, you could talk about it; any vector space can have a multivariate normal distribution. So here's an interesting fact. And you'll get this if you watch the, what’s his? I always forget what these videos are called, I'll link it on the show. But it's called, right, 3Blue1Brown, I always call it 3Brown1Blue. I don't know what I used to get there. Maybe I haven’t clicked it enough.
So one very interesting thing about the normal distribution in two dimensions is that it's like, okay, let's suppose I have that dartboard. And I want to create a continuous probability distribution on it. If I assume that x and y are independent, in other words, okay, like, Hold x constant, and y is going to be a normal distribution, hold y constant, and x is going to be a normal distribution. And it doesn't change. Like if I pick one y, and I pick another y, then the two distributions over x are going to be the same. So they're independent. That's the first assumption.
And the second assumption is that the distance from the center is all that matters. So if I'm one inch from the center of the board, then I'm going to be kind of a certain probability, no matter, like that circle, is all going to be the same problem is all going to be uniform around that circle, given that it's on that circle, and then two inches, like, yes, the uniform probability is going to be different. That might be less, but it's going to be uniform around that circle.
So given that it's on that circle, it could be anywhere on the circle in the same same proportion, if that makes sense. uniform. All right. So just assuming independence, and depending on the distance, it's just assuming those two things, it has to be a normal distribution, mind blowing. So that's the only one where that works. So that's pretty cool.
And another thing about the multivariate variant distribution, is that in the one dimensional case we only have the variance, and also the standard deviation, also the precision, but basically, all of this is based on the single number, those are just different views of the same number, like variances, the standard deviation squared, and precision is one over the variance.
But basically small variance, it's like really close to the mean, large variance could be things are very diffused from the mean, in two dimensions, things start to look a little different, like you might not have, it might not look like a circle, when you do the kind of topographical map, it might look like an oval. it might be stretched, and it might be like rotated a little bit, it might be like an oval that kind of looks like a galaxy that has been like an oblique angle or something.
Aaron: We would presumably get that by the distributions not being kind of centered in the same place that if the two dimensions are kind of off, slightly out of alignment with each other, then then we can get that stretching effect.
Max: Right, right. So you can get that stretching effect if they're slightly out of alignment. And then you can also get that rotation effect if the two dimensions of the data from the two dimensions are related to each other.
Aaron: So that's bringing some some dependence in, rather than being fully independent,
Max: Right, right. And so in three dimensional case it's not just a ball, but it could be like a stretched out ball, it could be a ball, if you can imagine could be stretched out in a lot of weird different ways. And it could have like, all sorts of different angles, it gets worse and worse, the more dimensions you go up. And so now, instead of having a single number to represent the variance, you have a whole matrix that represents this thing. It's called the covariance matrix.
And there are rules as to what that matrix can be. It can't be any old matrix. But it's generally like some kind of rotation matrix, if you guys know linear algebra out there. So I can't tell you the rules of that matrix off the top of my head, I haven't dealt with it for a while. I know it has to be positive. And I know it has to be symmetric. But who's counting?
So yeah, so that's interesting. So now the variance set of being a single number actually becomes a matrix, which is pretty cool. And then the other thought experiment is: imagine taking, Imagine taking a two dimensional one, and like taking in an oval, and having it on an oblique angle, and having it stretched out thinner and thinner.
So you have like a thinner and thinner ellipse that represents your standard deviation, essentially. And so like, a certain percent are going to be in there. And then the next oval around that, like a very large percentage is going to be in that two standard deviations away. At some point, the data starts to look like it's in a straight line. Right?
And so what's interesting about that, is that that matrix is the covariance matrix. And that's kind of the same raw materials that are used to calculate r-squared, which is like, how related are these variables? Or how well can you put these variables into a straight line? So there's kind of like, you sort of think like, okay, fitting a line to the data. And then talking about these Gaussian blobs have to be two very different things, but they're actually very related, which is, which is quite interesting, I think.
Aaron: Interesting. Yeah. I'm visualizing something that looks a little bit like a topographical map since you mentioned ovals, inside ovals inside ovals. And right, I don't know what exactly that visual connection between those can tell us, but it's what popped into my head.
Max: Right? Well, I guess, if you're looking at a topographical map here. I'm thinking and sorry, for these bad illustrations, I'm thinking something like this. Yeah. And so most of the points are going to be in the center. And then a few points, most of the points are going to be in the center here, a few points in the middle here, and then very sparse on the outside. So yeah, something like that.
And obviously, if this gets stretched out more and more, it's going to end up looking like the scatterplot has a linear relationship. So yeah. Anyway, that's just, there's so much that can be said about these and like, how to manipulate them and stuff.
But I feel like that's all I want to say for today. But that's really cool. I'm also having and I know this is additional distribution, but I'm having a little problem with the literature, where so we talked about the categorical distribution before and like, the literature also mentions it as like the, the general finite distribution and I am having, I feel like statisticians are very confused on what to call it. Some just say multinomial. So maybe I'll talk about that again.
It's like, man, sometimes statisticians and mathematicians can't come up with a single name for something. So maybe we could debate what it should be called at a later date. But alright, I think that's enough for today. Remember to check out the website localmaxradio.com. Remember to check out the Locals, maximum.locals.com and email us localmaxradio@gmail.com. Aaron, any last words for today?
Aaron: No, I'm curious to hear more in the future about some more distributions. But I think that's it for tonight.
Max: All right. Sounds good. Have a great week, everyone.
That's the show. To support a Local Maximum. Sign up for exclusive content and their online community at maximum.locals.com a Local Maximum is available wherever podcasts are found. If you want to keep up. Remember to subscribe on your podcast app. Also, check out the website with show notes and additional materials at localmaxradio.com. If you want to contact me, the host, send an email to localmaxradio@gmail.com. Have a great week.
