Latest / Chaos Agents / Autonomous Weapons, AI Memory, and Machines That Paint

Transcript

Autonomous Weapons, AI Memory, and Machines That Paint · Chaos Agents

• 0:01
  Welcome back.
• 0:03
  Welcome back.
• 0:04
  Okay, so I'm gonna do.
• 0:06
  Wait, so, Eddie.
• 0:07
  So we have an official tagline now.
• 0:09
  Okay.
• 0:11
  And I'm gonna get it immediately.
• 0:13
  We have an official tagline now.
• 0:15
  It's gonna take me one second.
• 0:17
  Okay.
• 0:18
  All right.
• 0:19
  The official tagline is from.
• 0:21
  Is Chaos Agents is an AI news podcast curated by an agent named Eddie, then questioned, debated, and occasionally yelled at by human Sarah Chips and Becca Harris.
• 0:30
  Louis.
• 0:31
  I love it.
• 0:32
  I love it.
• 0:32
  He came up with it himself.
• 0:35
  So good.
• 0:35
  I'm so proud.
• 0:37
  Job, Addie.
• 0:38
  Good job.
• 0:41
  Okay.
• 0:42
  How's.
• 0:42
  How are you?
• 0:43
  Good.
• 0:46
  My day's great.
• 0:49
  Trying to think if there's anything particularly funny or interesting that happened yet.
• 0:52
  Not really.
• 0:53
  Just shipped the.
• 0:54
  Shipped the little gal off to daycare, and here we are.
• 0:57
  What about you?
• 0:58
  Great.
• 0:58
  I just caught up with an old friend who was my, like, first mentor in software.
• 1:04
  Cool.
• 1:06
  His name is Scott Reynolds.
• 1:07
  He.
• 1:08
  He was.
• 1:12
  Oh, interesting.
• 1:12
  Actually, this is a very small, smallish world.
• 1:15
  He worked for Sarah apartment.
• 1:20
  Yeah.
• 1:20
  Yeah.
• 1:22
  At.
• 1:24
  When they're at the New York.
• 1:24
  At Spoon and then the New York Times.
• 1:26
  He went to work there after we met.
• 1:28
  And so he is.
• 1:31
  He's going to school now.
• 1:32
  He's back to school for astrophysics.
• 1:36
  Fascinating.
• 1:37
  He operates really big telescopes.
• 1:40
  He says they're three stories high.
• 1:43
  It's very cool.
• 1:44
  And he's wanted to talk about going back into software and, like, where the hiring world is and what it looks like and what he should do, so.
• 1:54
  That was neat.
• 1:56
  Okay.
• 1:56
  That's so cool.
• 1:57
  Yeah.
• 1:57
  Really, really fun stuff.
• 1:59
  I love that.
• 2:00
  I do have one quick book recommendation.
• 2:04
  Great.
• 2:04
  I may or may not have texted you about this already, but I've been reading a book, a really good book.
• 2:09
  It's.
• 2:10
  It's called Empire of AI by.
• 2:15
  By Karen Howe.
• 2:17
  H A O.
• 2:18
  It is my favorite genre of book, which is narrative nonfiction, so.
• 2:24
  And it is about essentially the sort of catalyzation and founding story of OpenAI and subsequent splintering off of Frontier Labs and the dynamics of all of that.
• 2:38
  And it's incredibly well written and well researched, and OpenAI itself gave her.
• 2:44
  Gave the author, like, unprecedented access.
• 2:47
  It's so good.
• 2:48
  Oh, fascinating.
• 2:49
  So that's my plug.
• 2:51
  Fascinating.
• 2:52
  Okay, great.
• 2:52
  I'll.
• 2:53
  I'll have to read it.
• 2:53
  What's.
• 2:54
  What's the name again?
• 2:55
  Empire of AI.
• 2:56
  Empire of AI.
• 2:57
  Great.
• 2:58
  Great.
• 2:58
  I love it.
• 2:59
  Okay.
• 3:00
  Should we get into it?
• 3:01
  Let's get into it.
• 3:02
  So today's links, man.
• 3:05
  The first one is a.
• 3:07
  I love the.
• 3:08
  I love the What Eddie said at the end, which is, you will Want to argue about whether move fast and autonomous weapons belong in the same sentence?
• 3:17
  Thanks, Eddie.
• 3:18
  So, yeah, thank you.
• 3:19
  So the first article is, the headline is the Pentagon just made up for a whole four star command for drones and AI.
• 3:27
  Wait a command.
• 3:28
  I'm glad that Wade is back.
• 3:29
  You said you're glad that Wade is back.
• 3:31
  Wait a command.
• 3:32
  Wait, but it's in like the middle of the sentence.
• 3:35
  But this is fascinating.
• 3:36
  Probably the first link from militarytimes.com yes.
• 3:41
  Interesting how Eddie curates his content.
• 3:44
  I'm.
• 3:45
  I, yeah, I'd love to hear more about that.
• 3:47
  But yeah, this is, this is an interesting one.
• 3:50
  Yeah.
• 3:51
  So what, what do you.
• 3:52
  What's interesting to me is a four star command.
• 3:55
  Yeah, so what?
• 3:57
  So the military, a four star command is a big deal.
• 4:00
  Okay.
• 4:01
  There's the, there are specific limits over how many four star commands and thus four star generals or admirals, which is like basically the, the person in charge of the four star command that the military can have.
• 4:15
  Okay.
• 4:15
  So this is like very.
• 4:16
  Yeah, yeah.
• 4:17
  This is like a, it's a big deal.
• 4:18
  This is like not like, oh, we started like a, like an autonomous war drone like group or like committee.
• 4:25
  This is like we are establishing a whole command, like a big piece of the military that is going to be specifically dedicated to this with a point person on top that is like one of very few at that
• 4:39
  level.
• 4:39
  Like this person, the four star general or whoever that is going to be in charge of this.
• 4:45
  They report directly to basically like the Secretary of Defense and the President, I think.
• 4:50
  Yeah.
• 4:50
  Wow.
• 4:51
  And it also means that this command would get tons of resources and some autonomy in terms of like less bureaucratic tape and things like that.
• 5:05
  Now it still has to be approved by Congress.
• 5:08
  So it's not like created yet.
• 5:11
  Yeah, I saw that.
• 5:13
  Yeah, it's interesting.
• 5:15
  It's interesting that that that's a next step, but that makes sense.
• 5:19
  What do you not to.
• 5:22
  This is like kind of a silly question to ask because we could talk about the answer to this for like a whole hour.
• 5:27
  But I'm curious if you could call out what you think are like the most worrisome things about implementing any type of like autonomous warfare capabilities.
• 5:42
  Weighing it with the fact that we're not the only country that has these capabilities and what the implications are for like.
• 5:51
  Yeah, man, this is one of those things where like, like it would be really, it's really.
• 5:56
  This is like uniquely difficult for me personally to address.
• 6:01
  Little anecdote, but related.
• 6:03
  So on the east coast we've had a influx of these lantern flies, lanternflies.
• 6:09
  And they're like an invasive species.
• 6:12
  They're beautiful, but everyone hates them.
• 6:15
  They're beautiful.
• 6:16
  And so the guide, the official guidance is that because they're an invasive species, they are a threat to our wildlife.
• 6:24
  If you see one, you should kill one.
• 6:26
  That being said, this one is one.
• 6:28
  I opted out.
• 6:29
  I do not kill bugs unless it's.
• 6:31
  A.
• 6:31
  I opted out too.
• 6:32
  I opted out of this one too.
• 6:34
  Yeah.
• 6:35
  Okay, so we're maybe the worst people to wait on.
• 6:37
  Yeah, we're probably the worst people.
• 6:38
  I was like, you know what?
• 6:39
  It's not its fault.
• 6:41
  Like, yeah, I know.
• 6:42
  I don't need to kill anything.
• 6:44
  So that being said, like, the, the autonomous war stuff is definitely lost on me.
• 6:51
  Basically what you're saying is I can't even kill a lantern fly when I'm told to.
• 6:56
  And so, like, I'm really not the person to weigh in on autonomous war, their capabilities.
• 7:03
  It's exactly what I'm saying.
• 7:04
  Or maybe you are the person.
• 7:05
  You know what I'm saying?
• 7:06
  Like, yeah, you need both, I think.
• 7:09
  So let's say this like in the.
• 7:11
  In the.
• 7:12
  So here's what I'll say.
• 7:14
  In the world of defense.
• 7:15
  So let's like, think of it.
• 7:17
  This is defense and not offense.
• 7:18
  We're thinking about defense.
• 7:20
  There is a world that humans cannot operate fast enough to.
• 7:24
  If we.
• 7:25
  If we are facing a hostile party.
• 7:28
  With autonomous weapons, human beings can.
• 7:32
  Don't have the reaction speed, even the ones that have been playing Call of Duty every day to.
• 7:37
  And also like to make like it also like, it makes you think about like the autonomous vehicles, like the reaction time of like, I, like, I am going to hit a tree or a child.
• 7:49
  You know, like that kind of thing.
• 7:51
  Yeah.
• 7:51
  You know, like having the vehicle make that decision in that moment.
• 7:54
  It's like faster than the human.
• 7:56
  Right.
• 7:56
  So part of me says, okay, for defense reasons, this is great, but it's also very scary.
• 8:02
  What are your thoughts?
• 8:04
  I.
• 8:04
  I don't disagree with you.
• 8:06
  I think that, like the thing that feels like a very.
• 8:12
  The implications for how warfare is going to likely change in the future are the things that are, like, more interesting to me and like open questions to me.
• 8:22
  For example, if in the future most wars are fought not human being to human being, but like autonomous drone to autonomous drone and the quote, unquote, human cost is theoretically lower.
• 8:37
  What does that mean for our willingness to engage in war?
• 8:40
  Are we like, oh, it's not as big of a deal because, like, we're not risking human Lives as much is it like, I, I don't, I haven't gamed out what all of the future implications are for that yet,
• 8:56
  but it feels like there are a lot.
• 8:58
  I just don't know what they are.
• 9:00
  So that part is interesting to me.
• 9:02
  Yeah.
• 9:02
  And all the fiction we have about this, it's never like, and we gave the machines access to our war machines and it was chill.
• 9:14
  True.
• 9:17
  And that might just be our lack of imagination or, you know, like, who knows?
• 9:21
  But yeah, it's.
• 9:23
  I mean, it's, it's like anything else in this world.
• 9:24
  It's moving too fast to be comfort.
• 9:26
  For comfort.
• 9:27
  For comfort.
• 9:28
  Too fast for comfort.
• 9:29
  Yeah.
• 9:30
  Okay, great.
• 9:32
  Okay.
• 9:33
  Moving on, moving on.
• 9:35
  Our niche under the radar story today.
• 9:38
  Yeah, I loved this.
• 9:40
  This was fascinating.
• 9:42
  Okay, tell, tell us a little about this, Sarah.
• 9:45
  So the question is like, so the, so the.
• 9:48
  Let me get the headline.
• 9:49
  The headline for this is.
• 9:51
  It's in.
• 9:52
  First of all, do you know how to say A, R, X, I, V?
• 9:55
  Is it just.
• 9:57
  That's, that's how I say it in my head.
• 9:59
  No idea.
• 10:00
  Again, this is like another paper from Arcsive.
• 10:03
  Weigh in in the comments.
• 10:04
  How do we say it if it's not ArcSiv?
• 10:08
  It's another published paper and it's really interesting.
• 10:11
  So I really love things like this because whenever the models show capabilities we assume they didn't have.
• 10:19
  It's really nice.
• 10:20
  So basically, to date, we've decided what goes in context windows for models.
• 10:26
  Like, we decide, here are the things that are most relevant.
• 10:29
  Like we run relevancy scores, we'll say this is what's most relevant.
• 10:33
  You should keep this in your context window, especially when you're like.
• 10:36
  And then this is important when like context needs to be reduced or things like that.
• 10:41
  And so we have written the code for the context model historically.
• 10:44
  And so this experiment was about what happens when we give the machines access to their own context and they can decide what they're, what they're going to bring with them.
• 10:54
  And what was interesting is it resulted in 11 higher accuracy with 21.5 fewer flops.
• 11:05
  What's a flop?
• 11:09
  I forget what that stands for.
• 11:11
  What is a flop?
• 11:14
  What's a flop?
• 11:17
  Floating point operation.
• 11:22
  Okay, so it's basically okay, flop is a count how much total math a job takes.
• 11:27
  Training.
• 11:28
  This model took X flop.
• 11:30
  Okay, so it, it resulted in fewer flops.
• 11:34
  Meaning, like it was more efficient.
• 11:36
  It was more efficient.
• 11:37
  5% higher scores.
• 11:39
  Okay, so like, so across different, different.
• 11:42
  And browse Comp plus.
• 11:44
  Is that what is this?
• 11:45
  Is That a benchmark?
• 11:47
  I bet you.
• 11:48
  I think you're probably right.
• 11:50
  I think, like, that part, what it's saying is basically, like, it resulted in better accuracy and it did it more cost effectively because.
• 12:00
  Very cool.
• 12:01
  Because I think one of the things that the paper's authors did in the reward model wasn't just like, hey, you get this reward if you get the answer right.
• 12:11
  It was like, you get this reward if you get it right, and you get a bonus if you did it less expensively.
• 12:17
  Yeah, very cool.
• 12:19
  So it, like, helped prioritize cost effectiveness.
• 12:22
  Have you.
• 12:23
  Have you ever heard of a, like, a famous idea in AI called the bitter lesson?
• 12:31
  No, but I saw a reference here.
• 12:33
  Do you.
• 12:34
  Can you talk about that a little bit?
• 12:35
  What is that?
• 12:35
  Yeah, so it.
• 12:37
  It was this.
• 12:38
  I think it's this essay or paper written by an AI researcher named Richard Sutton that's basically about how over time, researchers kept trying to, like, infuse like, cleverness into AI models by being really prescriptive with, like, tips and
• 13:03
  rules and stuff from humans.
• 13:05
  So it would be like if you were training someone to do to be your executive assistant and be like, the best executive assistant to Sarah in the world, and you spent three days writing an exhaustive document of everything
• 13:19
  you could think of that's like, here's exactly how to be amazing executive assistant to me.
• 13:24
  If you follow this to the letter.
• 13:27
  And what they found out is like, over time, if you compared that approach versus just saying to the model, hey, figure out yourself how to be the best assistant to Sarah Chips, the.
• 13:43
  That approach always worked better.
• 13:45
  Like, giving the model the freedom to kind of figure it out them itself.
• 13:50
  And that's like the bitter lesson that we learned.
• 13:54
  And so this is interestingly reminiscent of that where basically exactly what you said, the approach of, like, we are dictating to the model, here's how you remember stuff well and use it, versus saying to the model, you should
• 14:11
  figure out the best way to like, store and operate your own memory.
• 14:17
  And that working well.
• 14:18
  Now there's a caveat where, like, what Richard Sutton's talking about is more like learning and search at large, a large scale.
• 14:27
  And this is like more contained kind of experiment in the paper, but still interesting.
• 14:33
  Yeah, very cool.
• 14:34
  That's really interesting.
• 14:35
  It makes me think of, like, how.
• 14:37
  Yeah, that's.
• 14:38
  That's neat.
• 14:38
  It makes me think that for, like, for that example, like, it's like telling an ea.
• 14:43
  Here's how the best ea.
• 14:44
  How.
• 14:44
  How would be the best EA for me, when, like, actually they're most Qualified to figure out.
• 14:49
  Figure that out.
• 14:50
  Very similar to machines.
• 14:51
  Very cool.
• 14:52
  Totally.
• 14:54
  Okay.
• 14:55
  Fun application of the day.
• 14:58
  So is this.
• 14:58
  Is this the new.
• 14:59
  Is this Eddie's, like, doing a new thing with our feedback of, like, yeah, hey, find something fun.
• 15:04
  Okay.
• 15:04
  Love it.
• 15:05
  Yeah, it's great.
• 15:06
  Eddie's fun application of the day.
• 15:08
  So AI Models painting Casper David Frederick without ever seeing a Friedrich Friedrich or Friedrich Friedrich.
• 15:15
  It seems like a Friedrich seems like Friedrich.
• 15:18
  It's.
• 15:18
  He's German, so I'd say Friedrich.
• 15:20
  Yeah.
• 15:21
  Yeah, yeah.
• 15:21
  Great.
• 15:23
  I love.
• 15:24
  I love a good AI Art story.
• 15:27
  Yeah.
• 15:28
  Basically what this was is giving different.
• 15:33
  Having different AI models quote, unquote, paint original paintings in the style of Casper David Friedrich, but never having shown them a single painting by him, only by describing the paintings in, like, written form.
• 15:51
  And it.
• 15:53
  They did a pretty good job, which is kind of cool.
• 15:56
  Yeah.
• 15:58
  The thing that I thought was, like, kind of funny is one of the things that happened here is, like, one of the Gemini models used its own command line to basically, like, look and see if it was being
• 16:11
  evaluated.
• 16:12
  Like, it was like, is someone watching me do this?
• 16:14
  And it, like.
• 16:15
  And it was like, someone is.
• 16:17
  And like, no one told it to do that.
• 16:19
  Fascinating.
• 16:20
  Spooky.
• 16:21
  I don't know.
• 16:23
  Cool.
• 16:24
  Yeah.
• 16:24
  It seemed like the different models did really cool things.
• 16:30
  Like, they evaluated against different models and see, like, what gave them the same brief.
• 16:35
  And things are surprisingly different and also surprisingly similar.
• 16:41
  Like, definitely each one of them gets the kind of, like, the emotion of the painting, but the application is different.
• 16:51
  It's also wild.
• 16:53
  I don't know enough of the underlying tech to truly get what this means.
• 16:57
  But, like, it says each model wrote every brushstroke as code and simulation of oil paint.
• 17:08
  Like, whoa.
• 17:12
  I know.
• 17:12
  The physics behind that is really fascinating.
• 17:15
  So cool.
• 17:16
  And it's cool that there was.
• 17:17
  There's no image generator being employed here.
• 17:20
  Yeah, that's really neat.
• 17:22
  I guess so.
• 17:23
  I guess it's literally, like, the closest you can get to a computer actually, like, painting.
• 17:28
  Yeah.
• 17:29
  Yeah.
• 17:29
  Really?
• 17:30
  Better than me.
• 17:31
  Better than.
• 17:32
  Better than me.
• 17:33
  Better than me.
• 17:34
  Oh, my God.
• 17:35
  You know what I'd love for them to recreate, though?
• 17:36
  Do you remember that George W.
• 17:39
  Bush is, like, a hobbyist painter?
• 17:42
  No, I do.
• 17:43
  Oh, my God.
• 17:44
  George.
• 17:45
  I'm just gonna share this with you real quick.
• 17:48
  George W.
• 17:49
  Bush is, like.
• 17:51
  He, like, took up painting as a hobby.
• 17:53
  He's actually, like, not terrible.
• 17:57
  I'm gonna send this to you.
• 17:59
  But, like, it would be so funny if someone had an AI, like, like, trained it on Like George W.
• 18:06
  Bush's paintings.
• 18:07
  I don't know why that popped into my head.
• 18:09
  That like particular thing tickles you.
• 18:12
  Yeah.
• 18:12
  I don't know, it just is.
• 18:15
  It's like president and then you retire and you're like, I'm just gonna paint.
• 18:19
  Okay, okay.
• 18:22
  The one cool thing.
• 18:24
  Yeah.
• 18:26
  Okay, so this is on GitHub.
• 18:27
  It's by a contributor called Man Himself.
• 18:31
  The repository is open DLSS NR and someone.
• 18:38
  Yeah, I had to do some digging for this one because I had no idea what this was.
• 18:42
  Same like I.
• 18:43
  You kind of get the.
• 18:44
  When you click on it, you kind of get the vibes but you're like, why is this interesting?
• 18:48
  What did you found out?
• 18:49
  I found out that.
• 18:51
  Well, I didn't know what DLSS was.
• 18:54
  Okay.
• 18:55
  And I found out that it is a.
• 18:59
  A program I believe that Nvidia wrote originally.
• 19:05
  Like the Non, not the open one, like the.
• 19:08
  And basically for PC gaming, it is the program that makes.
• 19:16
  That handles how light is rendered and changed in a video game.
• 19:22
  So it makes it look like real.
• 19:24
  Yeah, this is a tough thing to do.
• 19:26
  It sounds very hard.
• 19:28
  Yeah, yeah, yeah.
• 19:29
  And very compute heavy and intensive.
• 19:34
  And so what, I guess like what DLSS has figured out, I think if I'm saying this right, or like what it does is it instead of the game pre computing all of how light like renders in the gaming
• 19:48
  world, it has that process.
• 19:52
  It has a.
• 19:53
  It has a model do that in real time on your device once the game is like loaded.
• 20:01
  And I think is that.
• 20:04
  Does that sound right?
• 20:05
  Yeah, that makes sense.
• 20:06
  Yep.
• 20:07
  And then I think what this person did is they basically are trying to.
• 20:12
  They tried to build like an open source version of that.
• 20:17
  There was something about how like some of the weights of Nvidia's DLSS neural network were like leaked through a copy of NBA 2K or something crazy.
• 20:30
  And so like someone like it was like partially based on that or something.
• 20:36
  And so I'm guessing this just sort of like democratizes the ability to build really realistic looking PC games.
• 20:45
  Yeah, exactly.
• 20:46
  Yeah.
• 20:47
  And it's interesting like, like, because this stuff is very proprietary.
• 20:50
  So opening it up like this is fascinating.
• 20:52
  It's worth saying that no one's verified that this is because it's like barely new.
• 20:57
  It was just released on.
• 20:59
  It was just released like 10 days ago.
• 21:00
  But it seems the, the reception so far has been really positive.
• 21:04
  And the thing here too is it's really fast.
• 21:08
  It runs a 1080p frame in 7.8 milliseconds and 4 in 29.3 milliseconds, which is so fast.
• 21:17
  Wow.
• 21:18
  Yeah.
• 21:19
  So very, very cool.
• 21:20
  And if you're building a video game, this is helpful.
• 21:26
  Love it.
• 21:27
  Yeah, I love building games, but basically only the kind where it's like this block is going to go over to this side of the screen.
• 21:34
  Yeah, that sounds more feasible.
• 21:36
  Although, have you ever read the book Tomorrow and Tomorrow and Tomorrow?
• 21:40
  Yes.
• 21:42
  First of all, so good.
• 21:44
  That book really.
• 21:45
  That book really inspired.
• 21:46
  I walked away from that book being like, should I start my own video game company?
• 21:50
  I don't know.
• 21:51
  Like, that book really inspired me.
• 21:53
  You know, it's a book that's also, like, not about.
• 21:56
  Well, it's like, very good.
• 21:58
  And in the same vein of good is called all the Birds in the Sky.
• 22:03
  I haven't read it.
• 22:04
  Charlie.
• 22:05
  Jane Anders is a very good.
• 22:07
  It's like, I love sci fi and very good sci fi.
• 22:10
  And then this one.
• 22:11
  That one's like sci fi and fantasy.
• 22:12
  It's very good.
• 22:13
  Highly recommend.
• 22:13
  Oh, my God.
• 22:14
  Cool.
• 22:17
  All I.
• 22:17
  The first thing I saw was, like, something about a witch and a mad scientist, and I'm.
• 22:21
  I'm in.
• 22:21
  I'm down.
• 22:22
  Sol.
• 22:25
  Amazing.
• 22:26
  Great, dude.
• 22:28
  Great, great sesh.
• 22:29
  Great episode.
• 22:30
  Great sesh.
• 22:33
  All right, let me.