Latest / Last Humans in the Loop / Would You Give AI Your Passwords?

Transcript

Would You Give AI Your Passwords? · Last Humans in the Loop

• Marty Weiner: We're screwed. Christopher: I'm Chris Slow, and as ever, I'm between two wieners. We have Zach, Zach Wienersmith and Marty Wiener. Great to see you guys today. hey, Marty Weiner: Fantastic. Christopher: big big news today, big news last week. I would like to declare that we in fact Zach: You Christopher: goaded Google to release its new Gemini model. Marty Weiner: big news from Google. Christopher: so we Victor Lap, Victory Lap gentlemen. Zach: For some definition of release. Christopher: Yeah, I mean that's the kicker is so if that's case you case you guys missed it, Gemini Argon has been released. and the benchmarks look really good. like they're they're kind of near the top of the rankings for a model that was initially kind of outside of the outside of the rankings. now if I if I remember right, I believe that they are going to release it to some people without really much of a definition. Zach: Yes. Christopher: Citing citing safety concerns. I think we're in the extreme case of another another fable deal over. What do you guys think? Zach: Yeah, mean, so I don't know if you saw, but there was an article in Bloomberg where supposedly they talked to people inside Google who were saying, ooh, it's actually not as great as it needs to be versus the current stuff. But then there were a bunch of Google engineers on Twitter being like, no, uh-uh. So that's where we are right now. I mean, I don't know. We were saying earlier, it's like Anthropical just released a new model and be like, I don't know, you guys might want to try it. And then it just explodes the internet. And whereas Google will be like, this is going to blow your minds and then crickets. Christopher: This is gonna blow your mind, Marty Weiner: Well Christopher: but there's a sign up list for you to get access to it in the next three to six months. We're afraid it might destroy civilization. Zach: Yeah, so... Marty Weiner: I I we we're gonna see this with some of the other model launches, but I what I think is funny is where where we've gotten to. There are literally tens of thousands of benchmarks, and every time one of these releases, they choose the benchmarks they're better at, and that is the selection that that shows that they're the Zach: Yeah. Marty Weiner: best. Christopher: I mean, that's the tricky here is like if you can't no open access means it's gonna be very very difficult to otherwise get get an impartial benchmark for a bit. So we'll keep an eye on it. hey, another news, after we made fun of Saw last week, GPT Saul 61 is now out. and also looks pretty looks pretty decent. I was playing with this myself most of last week, for writing code, and it's pretty good. Like it's not astra good, but it's it's way better than we than what you get out of Saul and Luna before. also I believe OpenAI had their had their hack day last week and they announced Zach: Yes. Christopher: dots, which seem to be I don't know, they they seem to be almost like these kind of always on-ish kinds of kinds of models. Marty Weiner: Yeah, there seems to be Zach: Yeah Marty Weiner: this new class of thing that layers on top of models that everybody's getting into. for our for my startup verify you, we've been playing with instinct, which is the hot news shit where you you basically only interact through SMS, but it will go and it will make reservations for you. It will send emails on your behalf, it will put up Slack messages. they've got dots from OpenAI, you've got Grockbot, Claude has its own version. Zach: How reliable is it? Is it like actually working or do you? Marty Weiner: They're they're freaky. Okay, let me tell you the story that David on my team told me. We needed to get a reservation for for a restaurant in DC in a certain spot kind of rapidly. he went to bed. the system went in and tried to get him on open table because that's what the restaurant used. Well, he it didn't have his open table password, so it went into his email, it reset his open table password, got the code from his email. Christopher: Ha. Marty Weiner: And then and then got in the reservation. He woke up to a reservation and a re password reset message. so Christopher: And an inability to get into his open table from that point on. Marty Weiner: They are they are damn good. The the challenge I think is I think first of all, people Zach: my god. Marty Weiner: getting the confidence to connect their systems. I've not gotten the confidence yet. I have been using I've not been using dots, but I've been using codecs to drive my my my laptop. I talked to a friend of ours who does use dots and said, Yeah, it's it's it's pretty nice once you get used to it. I think the way to think about it is sort of like having an an executive admin on all day, every day. so I think as people start to wake up to this, I think, and get get comfortable with it. i it's gonna be a brand new way to to interact with our our models. Christopher: So we're kinda we're kinda looking at finally open claw like things coming straight from the foundation from the foundational labs. Yeah, that makes a lot of sense. Marty Weiner: I think so. Yep. Yep. And then also real quick on Dev Day, I think they covered that. They covered Astra API adjustments. One of the other big notes from OpenAI is they changed out their models so that they don't have a pro we'll see now they now have pro one hundred, pro two hundred, pro five hundred. And just for anybody who's wondering what the hell the difference is, I I think based on it's money. Christopher: Money. It's the difference is money. It's cost. Zach: Yeah Marty Weiner: it used to be that pro two hundred got you more like four X Pro one hundred, I think it's now more like two X. that's the one I've been using for the last week and it feels like double. It doesn't feel like like a four X or a five X. They don't say. They're just like, yeah, it'll just cost more, but they don't tell you anything about Christopher: Give you. Marty Weiner: what you get. Christopher: You get you get more things and you get the privilege of giving them money. That's the important part. Marty Weiner: Yeah. Christopher: shoveling money into the furnace. hey, another news. speaking of speaking of six one cell, looks like a new math benchmark is out and it does a pretty good job. frontier Zach: Yeah. Christopher: math tier four. so continuing up into the right on these graphs. Marty Weiner: And saturated, by the way. So it's a dead pretty much a dead benchmark now for Frontier. Zach: Yeah. Well, Christopher: Hundred hundred hundred percent. Zach: I think this was impressive too, because I think it had been saturated by Astra, and now it's like, this is our second string cheaper model, and it's also saturating the hardest non-open problems. Christopher: And and jokes aside, we're talking about a model Marty Weiner: Right. Christopher: that's like a f like a fifth the price, I believe is the target for this one. So pretty impressive. any other things from the dev day recap that we should talk about? Zach: they had a major tech issue. I don't know if you saw this, but the lady was like, dot, make a reservation for me. And it just ignored her they've had this a couple times, but you got to respect open AI for like doing it live, you know, instead of Marty Weiner: Yeah. Yeah. Zach: cheating. But like, on the other hand, not great to be like, you're going to love the way it does stuff for you. And then it doesn't even, it just Marty Weiner: Yeah. Zach: ignores you when you're requesting things. Marty Weiner: So real quick quick on the other things from Dev Day, dots, GPT six one sold. Now they now have an ultra fast speed tier. I haven't tried it myself, but like up to eight X faster. I think it it burns tokens at a massive rate. They have these this new decisions API, codex working in the cloud. and then open AI agents can now run inside a AWS via bedrock. Christopher: Hey, it is interesting on the timing too here that dot are are they called dots? I guess they're called dots. dot or whatever it is, it was released on the heels of meta muse coming out, which also seems to be targeting this kind of like AI assistant, I'm gonna do things for you, I'm gonna search for you, I'm gonna make reservations for you. I am always a little afraid to yield that kind of power to the likes of of meta. Zach: Do you think this is like when it was 1994 and you were like, I can't use a credit card on the internet. It's full of Marty Weiner: I wonder yeah. Zach: scary people. Maybe that was true in 1994. the question is, can this be made safe? I I would be pretty freaked out if something reset my own password for me overnight. Marty Weiner: Uh-huh. Uh-huh. you know, it also could be a pressure thing. I mean, it's startups will start using this stuff because they've got to go faster, leaner, better. Zach: Yeah. Marty Weiner: and so they may be pushed into it. and real quick on the decisions API, this is sort of like we reported in the last couple of weeks about the brand new hotness called JEV. so just for anybody who doesn't know what JEV is, it's sort of like a layer that sits on top of the L LLM that makes rap more rapid decisions. you know, sometimes you don't need to bring a a language into a fist fight. You just punch. so that they they can make faster decisions and and a lot cheaper, you know, sometimes like 90% cheaper. So anyway, the moment Jev launched, there was there was a an open source version that launched and now I think every major company's launching theirs. OpenAI's version is called decisions API. Christopher: and of course, you know, not not to be outdone, I believe SpaceX solved the problem of having their own version of Dot by just buying dot.com and having it redirect to Grok. so I can't and Zach: dot com. Marty Weiner: That that was so stupid. It wasn't even dots. Like what what was was this an Elon Musk joke? Christopher: I mean it's a three it's Zach: yeah. Christopher: a th three letter domains are not easy to get. So, you know, if you're gonna it's I guess if you Marty Weiner: Th yeah, world's most expensive worst joke. Zach: Sure. Christopher: Okay, Alibaba announced a new chip along with Quinn 4, Marty Weiner: yeah, they they announced new chip. so this is great because now we're gonna get even crazier competition. I don't know if we've seen a lot of stats on it, but they're I I think they're trying to enter the fray with NVIDIA and of course jalapeno from OpenAI. and along with that, they introduced a Quinn their new Quin four line And we're talking quite substantial, five to ten trillion parameters. Christopher: it sounds like Germany has entered the LLM model race Zach: Yes. Christopher: with the first, it's the Calibri model. Marty Weiner: Yes. Christopher: it's a German AI startup, and it sounds like they're targeting actually also German language. I believe that's the doing Finnish if I remember right. It was a couple of a couple of European languages were target with the target model here. Marty Weiner: Yeah, big big focus on German language and and English. And I didn't I don't know about the finished part, but what I thought was almost equally interesting is I'll sh show you this paper that they that they released at alpha at olifalpha.com. it basically is this giant paper on how you make an LLM with about as many details as Zach: Hahaha Marty Weiner: you can imagine. Look at this sucker. I I mean like this is the kind of paper. If you sent this back to you know 10 years, it would be the the Zach: You're right. Marty Weiner: manual on how to make an LLM. Which is really neat. Cause if you have any questions about how they function, I mean th this goes deep. so I I thought that was really neat. to to basically make Zach: Yeah. Marty Weiner: it all out there in the open what all their findings. Zach: And there's talk now about this need for sovereign AI. I this has been a discussion all along, but it's interesting that maybe we're starting to see it now that countries need to know they can produce it on their own in case, you know, China or the US starts restricting. Marty Weiner: Yeah. Christopher: and last but not least, in the news of of models, Anthropic is doing an analysis of G L five three, which is Z dot AI's, I believe, latest model. And it sounds like it's benching really well on the cyber axes, Zach: basically the deal is Anthropic, which is the lab most likely to report that things are getting really freaky and we're all going to die, release this report saying, hey, this is an open weight model that everyone can access. And it's got like Claude Mythos like abilities on cyber, which is, Marty Weiner: For for cyber attacks. Christopher: Forever exploits Zach: yeah. Christopher: and stuff, yeah. I mean, hey, this is this is a problem with asymmetric asymmetric cyber warfare is like if you have a model that can actually go and hack, you need a defensive model. So it'll be interesting to start seeing these this get deployed as a defensive measure if they're actually that good at hacking. Zach: this is my reaction when I was reading it. I mean, I don't think they're being dishonest necessarily, but like, clearly it's in their interests. And like a bunch of people have talked about, know, people at OpenAI and, and, and probably have said, you know, there will come a time when we need to regulate open source models. How they're going to do that, I don't know. And even if they're completely in earnest, it's just they definitely get something out of this if all the open models get regulated away. I mean, actually, you try to imagine what AI would cost if there wasn't an anthropic. An open AI just was way out the lead ahead of everyone. Marty Weiner: yeah. Yeah. Zach: So they certainly have an interest here. Christopher: I mean consider considering also that that Saul's Saul's biggest selling point this week was cutting the price by five X over Astra. it is it is actually nice to see a good old fashioned price war happening. in parallel, of course, to them being them calling for regulation of themselves. As a way to as a way to guarantee regulatory capture. Marty Weiner: A and and and a slowdown. A Zach: Yes. Yeah. Yes. Marty Weiner: and a slowdown. Yeah. That's definitely what's happening is a slowdown. Christopher: Guys, we can't we have to compete, but we can't compete too much. We're burning too much money. Zach: Yeah, mean, is, know, that's one of the worth noting that is one of the bubble scenarios where this is all great stuff, but all the profits are competed away. All right. Like you can you can get a bubble that way. And, you know, so, you know, that is one of the weird next year possibilities is that there's a collapse because the service was so good. So. Marty Weiner: to make it a little crappier. Christopher: Hey, as as long as I can get you know, in a couple of years, as long as I can get like a sol grade model running on a local GPU, you know, it'll be fine. It'll be just fine. Zach: That's true. That's true. Yeah. Yeah. Christopher: Just fine. let's move on to science. it's okay, so we we've now I think every two weeks we're going up by an order of magnitude of how many problems we're saying are being solved if I'm tracking. We need a we need a graph of of the of the Zach: You Marty Weiner: The the do a math benchmark. Yeah. Yeah. Christopher: Of the solution so we'll call it solutions bench. solutions bench. rumor is right now that OpenAI is about to drop four hundred solutions to open problems. Zach: Yeah. So this just popped up like last night and, but, but, and not, you know, rumors, but rumors from mathematicians in a place to know, not just crazy Twitter people. And so, you know, we know it was, was last week of the week before that open AI was like, are forming our Supreme council of mathematicians, but by the way, we have a hundred open solutions solved. And now the rumor is that the number has gone up quite a bit. And it's like this, we've seen this flurry of papers on archive and there's, I saw one mathematician saying, I wonder if this is happening because everybody's worried they're going to get scooped. Because the opening eyes apparently just sitting on all sorts of problems that would be a thesis if someone else had done them. And then there's now this problem in the community of like, how are we even going to handle review if a big company could just drop 400 Marty Weiner: yeah. Zach: problems solved in the last three weeks? Christopher: Apparently Archive also added a restriction in the last, I believe in the last couple of weeks, that it's like Marty Weiner: yeah. Christopher: two two papers per month is the per account or something Marty Weiner: person or something. Yeah. Christopher: like that. which you get to wonder how it's gonna happen with the with the with the Frontier Labs being able to drop 400 at a time. also I can imagine within the community, because since there was concern about reading chat histories to help bootstrap some of these problem solutions, that yeah, I would just immediately start dumping out whatever I've got written up, even if it's partially done, just so I can at get partial credit. Zach: Yeah, it's fine. We were talking a while back about Terry Tao saying, don't submit a paper unless you understand it. And we were kind of like, well, I don't know. Maybe, maybe not. But I hadn't thought about that acting as a reasonable bottleneck on people who are going to submit like 40 papers a month. You have to understand your own paper. Christopher: Okay. Marty Weiner: Right. Well, and and I think the other thing this leads to and some of those next things in our science section, but a lot of these have been delivering a cool counterexample, these math proofs, which is great, but they've been delivering it in a way that is not understandable by humans. And we've as humans, we we I don't think I've seen a case of like a new interesting insight. it's more like let me trn the crank in a way that no human wants to do until I get to a Zach: Yeah Marty Weiner: solution. I believe that's why we haven't seen a dual physics or duo biology, because those don't have as many counterexamples in them. Like we do have hypotheses and things like that, but math has a has a lot more things that you can counterexample out. But we're starting to see a little bit of inching into physics. Christopher: It would be nice to see like the the closest I think you get in physics would be with current technology, give me an experiment that would prove a counterexample or that would show an interesting result, right? The the closest you can get because you can't because the math the like the math is gonna work or not. The question is, is it actually map to reality? so I mean I I mean the the latest one was okay, so we have the we have a a a a sort of a difficult physics problem of like lowest energy arrangement of seven electrons, was I believe solved with a lean proof here. It's that's a very pretty graph. It's like exactly. I feel like that's I feel it's like the it's like Marty Weiner: Yeah, nice pretty pretty chart here. I don't know what it means, but but but yeah. Zach: I could have figured that out. Christopher: the it's like the 17 the 17 packing squares problem on this one. But I mean hey, it's it's it's a neat result. I'm not sure what the implications are. and then also Marty Weiner: Yeah, but it is bumping into physics here, right? like like I think we're starting to see the beginnings of duophysics. Christopher: definitely. I I do think though the other thing is like speaking of speaking of understandability, I have noticed as the models have gotten sort of smarter, or maybe it's the benchmarks pushing them, the this the prose has gotten more convoluted with time. So to some extent, like writing clearly is not in any of the benchmarks. I don't know if you guys are experiencing Zach: Ha Christopher: this as well. you know, it's it I I I think there's a I remember seeing at one point there's a skill that's like explain this to me like a caveman, as a way to like get it down to really simple terms. I can imagine that that is one of the problems that's coming. moving along. okay, so simulated quantum annealing of circuits. Marty, those are a lot of those are a lot of high flutin words. Marty Weiner: Yeah, this is neat. This this researcher, said in his research project, he spent three months on optimizing chosen algorithm, simulated annealing for quantum circuit synthesis. and they were in those three months able to improve it by four orders of magnitude. Then Christopher: cheese. Marty Weiner: they brought on Astra and it got him another order, magnitude improvement. And so you know, I I think this is where we're starting to see the beginning again of like inching into physics. What he's saying Astra did for them was a number of clever engineering tricks. They imported it from C to Rust, and then making additional adjustments to their algorithm. So I it did a big review and said, Hey, I I I can improve it. Christopher: Super, super impressive. in the okay, so in the in the material science discovery column, I actually I suspect this is gonna be one of those ones gonna start popping up more. so there Marty Weiner: Yeah. Christopher: is a report for room temperature magnetic semiconductors. now, you guys could ask the question, hey Chris, what the hell is that? and you would Marty Weiner: Yeah, what the hell is that, Chris? Christopher: That would not be a terrible thing to ask because this is actually a little okay, a little absurd. So let give you quick maybe go with physics time. I should have I should have a hat for this part of the of this or of the lab coat. I should we need Marty Weiner: Yeah, we need you to get Okay. Yeah. Christopher: a lab coat at some point. I'm gonna slip on for these. Okay, so generally speaking, there are two types of magnets that exist or that are that are common. the one that's actually very common, which we're all familiar familiar with, is just ferromagnets. Ferromagnets are like bar magnets or the ones that are actually where all of the you can think of that as all the little magnetic poles in the magnet are aligned in the same direction. And so the whole material acts like a giant magnet. those are actually the those those little magnets are actually from the spin of the electron, by the way. So this is like a this is a very quantum effect that's actually very complicated. There's another thing called antiferromagnets. Antiferromagnets are where you actually have all of the spins constantly anti-aligned. So you can think of it as like a big lattice, and if one of them is spinning up, the one next to it's spinning down. And so basically it's like it's it's the perfectly canceled magnetic field from a bunch of spinning things. now, what's a semiconductor in this case? Well, a semiconductor in this case is actually. If you set up a antiferromagnet, but you have the thing that's preferentially spin up be a different material than the thing that's preferentially spin down, that means that they behave differently in magnetic fields because they're two different elements, basically. And so they're basically coming up with these kind of like very exotic elements where the you put them in a magnetic field, you you do things to them, and the material doesn't act in the same way as if it's all bulk same material. Now the tricky part is of course finding material that is both this type of antiferromagnet. And that is achievable chemically and can be built. and so the reason okay, so now you ask the question after I've been talking for two minutes is great, why the hell do we care? Well, we care because it turns out that one of the one of those cutting edge fields in electronics right now has been called spintronics, which it has huge implications for quantum computing. spin is a really great way to encode quantum information. And if you can actually build a semi-con spintronic semiconductor would be a nice way to build, you can build switches. You can build better magnetic storage, you can build all sorts of great things. And Marty Weiner: So is is the place where the LLM jumps in here is being able to try and go all day and all night on different chemical components combinations? Christopher: Exactly. so in this result they found two two chemicals that were proposed to that are these magnetic semiconductors, one of which has been synthesized before and one of which is completely new and has never been never been tested. So of course, this is still all theoretical. The question now is test the model and actually build the build the build the chemical and see what happens. Marty Weiner: Wow. So go make me a magnetic semiconductor while I watch the World Cup. Christopher: Apparently. Apparently. Zach: Kiss. Christopher: okay, this one is actually even too abstruse for me. a 59-year-old conjecture in physics Zach: You Christopher: fell with GPT Astra VI. this is a this is a plasma physics result. and I would just say plasma physics implications for fusion. but I do not have enough of the math behind me to explain what the hell happened here. but Marty Weiner: It sort of sounded like this is like what happened what's happening with math is we're starting to see some conjectures in physics start to fall in the same way we're seeing it with math, where where the L L L produces a counterexample. Christopher: Yeah, I mean, I think so far that like last week we had some results that basically were like the model was able to go one level more kind of complicated in the in the quantum field theory work. Here in this case it seems like it's actually just able to do s do work in plasma, which is actually a very hard field to work in, pun intended, because they it's it's strongly interacting materials that are at high temperature. It's like kind of the worst of both worlds. and then yeah, from my old alma mater, Harvard Vid Harvard physicist made a lot of progress doing problems. I believe dropped about something like 40, 40 problems. This is Zach: It's just, yeah, in three months. Christopher: in three months. this is from Matt Schwartz. Matt Schwartz was the our writer about six months ago of a paper with Anthropa called Vibe Physics. and has been pretty outspoken on the idea of in the future, what do grad students do? and so here's another example of here is you've got a fleet of models building interesting papers for Marty Weiner: And the grad students Zach: Yeah. Marty Weiner: are getting him coffee. Christopher: I mean as is as is tradition in this case. Marty Weiner: As a s as tradition, yeah. Christopher: man, does somebody want to try and tackle this next math paper? Zach: was being talked about as a big result in applied computational stuff. Marty Weiner: Computational mathematics. Zach: The weird thing is, this is apparently a fairly big result, it one of the highly desired results in applied math. The weird thing though was in the materials and methods. Let me just read this. How the result was originally found and shared with its authors. An Anthropic employee used an internal research model to investigate open problems in the theory of cryptography. One of them was about cryptographic constructions, blah, blah. Claude was tasked with verifying and improving the construction, but instead developed this algorithm. Marty Weiner: Whoops. Zach: And then the weird part, Anthropic shared the algorithm with the authors in September, 2026 under a confidentiality agreement. offered compensation and provided access to the public version of Claude, meaning not the secret internal model. So this is like, imagine your method section was like, I met in a dark alleyway from Anthropic with this result. Marty Weiner: Yeah. Zach: That really helped with this paper. And it Marty Weiner: And he gave me some cash and and free credits. Zach: offered me cash. I actually found this kind of bewildering. Do you guys have a sense of why Anthropic wouldn't just release the results? Marty Weiner: Was it because he was already most of the way on and they didn't want more bad press for Zach: I know, that's why I wondered if, had the relevant expertise, like I can imagine they just got this result where like, is this important? I don't know, it's Marty Weiner: I feel like this would be Zach: very strange. Marty Weiner: this seems like a really interesting way though for for these kind of accidental discoveries to to be validated and and also to not get them in hot water for for screwing with the math community or possibly even stealing results from the c math community is to go and involve the math community and say, Hey, here's the secret result, go at it. Zach: Yeah, well, I mean, there was a big deal last week when the like Supreme Council of Mathematicians of Open Air or whatever it's called was like, please, would you not test your models against our big open problems because it makes us sad. Marty Weiner: Yeah. Zach: And Christopher: maybe that's the maybe that's the hook here is like I was totally working on this in my spare time and the model appeared on the cor like it told me to meet it on a corner wearing a carnation and it gave me a briefcase that had this proof in it, and I swear I didn't ask it for it. Zach: I swear I didn't ask for that. Well, that's I wondered about. it's like, you Marty Weiner: Yeah. Zach: know, maybe it's just like, you know, apparently there's some value in training on like solve the Riemann hypothesis or whatever that for these models for like benchmarking or just giving it a really tough problem to grind on. And so, you know, maybe what they want to do is like, well, we can't take credit for it anyway. So we might just give it to you in a briefcase. I don't know. It is deeply strange to me. Christopher: we have okay, so we have Gemma four, back to back to some actual interesting Google news. can now actually it it can actually find DNA mutations behind things like drought resistance and higher yields. which Marty Weiner: Yeah. Christopher: was you which used to be a used to be a breeding problem, not not a computational problem. Marty Weiner: So this is a duo biology. The they paired it with another model called Botanic One, which I think is a model specifically trained on on various kinds of plants. And then they were able to use combine that with GEMA to do a lot of the coordination coding and all that stuff. and they're able to find new genes that that make these plants drought resistant and yeah, they're able to make it yeah, more drought resistant, higher yields. Christopher: Higher yields, like basically like a better better crops. Avoiding another an another couple of years of avoiding the population bomb, as I think populations start to plateau, if I remember right. Marty Weiner: Right. Zach: Yes. Marty Weiner: And to me, the big key here is do a biology. I think we are starting to see the the biology nut start to be cracked. Christopher: well I mean we just had a result I mean next step is MIT used AI to adjust mRNA vaccines so they no longer need to need freezers instead of room temperatures. Marty Weiner: Yeah, and to clarify to clarify, this is not AI in the sense I think we're thinking about like with Gemma and Claude, this is this is good old classic ML, but still a very interesting result. Christopher: but how how do they know it's working if the model doesn't explain what it did with a bunch of M-dashes and it's not just blah, it's also blah. Marty Weiner: That's that's why you pair it with with Astra. Zach: You find out when your vaccine has been left out of the freezer and they give it to you anyway. You'll Christopher: It Marty Weiner: Yeah. Zach: get your answer real quick. Marty Weiner: So it said and I'll just say what what the interesting result here is is because these vaccines don't need freezers, it means that you can get them out to other parts of the world that Zach: Yes. Marty Weiner: that that do not have access. Christopher: Neurlink has trained a new model on fifty thousand hours of brain data finally. I guess they're actually I in fairness, there's not that many neuralinks out there. So getting 50,000 hours of brain data to train on probably is a difficult task. Zach: Well, so they did say like, so, know, I guess the way this stuff works is they essentially have to put a hole in your brain and insert electrodes or so in electrodes that this cool new system and, it takes a while. I think you have this idea that they put it in your like a Jedi suddenly, but you have to calibrate to a person's individual brain. So they're claiming this just speeds up the process a lot, which is presumably nice because, you know, this is very invasive surgery and you tend to get degraded signal over time. So, you know, if you can calibrate faster, I would imagine that's highly desirable. Christopher: And there are some first results there of a of a of a of a patient who's lost the ability to speak who can now actually speak via an app, which is actually kind of impressive and creepy. It's really it's really fantastic. Zach: I mean, these are like, actually they have videos of some of this stuff. It's really moving. Like you have people, a lot of it's people with ALS, like Lou Gehrig's disease. And so they're just slowly losing ability to use their bodies. And then for them to suddenly be able to socialize again in their own voice is just really magical to see. Christopher: Yeah. and hey, we we we talked about some before, but the graph is impressive. is the the reason why the reason why the archive has decided to impose a rate limit. you don't expect to see papers go up and the right like that. I mean you can actually you can actually kind Zach: yeah. Christopher: of spot the you can spot the kink. You can actually see the spot in about mid sort of like late twenty three where it kicks off and starts to really take off. It's pretty impressive. Zach: Yeah. I mean, I thought this Marty Weiner: Well Zach: was cool too. Cause like, you look like it doubles in, you know, a little over a year, I mean, it's just like incredible change. I don't know how the review, but well, Marty Weiner: Now what what are the quality what are the quality of these papers? Zach: right. Yeah. I mean, this is like the lines of code metric, right? I saw a bunch of people Marty Weiner: Yeah. Zach: being like, this is super intelligence. you're like, it's, it's maybe super slop. I mean, maybe Marty Weiner: Yeah. Zach: it's both, but like, I mean, the, deep problem, I talking to Kelly about this is like the reviewer system is already not good. Because the reviewer system is like, for people who don't know, the way this works is people essentially volunteer to do review, to make sure the literature is at least a little protected. Not that it's perfect, but you have to have people to make sure pure slop and gibberish doesn't go into the literature and corrupt the system. And so the only way it's imaginable for this to work is if reviewers start using AI, which I think presents obvious problems, because at a certain point, there are no humans in the loop. And we're just pressing a button on either end and imagining we're doing work. Marty Weiner: And is this your quote from Thomas Dietrich? Zach: yeah, that was cool. this is one of the guys from, so Archive released this new standard, which basically the main thing is only two submissions per person per month. And he said, I'll just read the last part of this. So he talks about, know, only a small number of people do the reviews. And then he says, this is unfair to authors who continue to submit quality papers. Their papers can be delayed for days or weeks as a result. So they're, Marty Weiner: Well and higher up in the quote, he he's also saying a relatively small proportion of authors are submitting a large number of low quality Zach: Yes. Marty Weiner: papers. So maybe the simple thing here is to just block them. Zach: Yeah, Christopher: I it's a Zach: I think that's what they're doing, right? So it's like very few academics, maybe like Paul Airdish or somebody was really submitting two papers a month, but it's got to have barely existed throughout history. Christopher: Marty, it's always great to see something that applies perfectly fine in social media applying now to scientific papers. right, good old fashioned. Marty Weiner: Yeah. Zach: You Christopher: What do you how did you solve that problem of spam? We added rate Marty Weiner: my god. Yeah. Christopher: limits. Like we just said, add a little bit of a quality bar. Zach: Yes, yes. Marty Weiner: The whole world needs anti spam. Zach: Yeah. Christopher: you know what's interesting too is like it's it is actually very difficult to get good reviews out of AI models because of the amount of groupthink that happens. So I remember seeing seeing a a result not too long ago where if you were to take a chorus of effectively nine different models and ask them to do a to do an analysis or do a do a critique. You actually only have effectively like two ish unique opinions across all nine of Zach: Mm-hmm. Christopher: them. it's there's there's a there's a lot of universe like uniformity of thought across the group. so it's you know. Zach: Yeah, I agree. think it's scary for that reason. As a joke, the Jacobian conjecture got solved, I put the solution into Clawdon was like, hey, I solved this. And it had not got the news that someone else had solved it. So it kept going back and forth. would not believe me. It was like, there's no way this is accurate, even though it could check it, because it was just like two outside distribution. So I worry if you had a really brilliant paper with some outside the box solution. Even if it seemed to solve a problem, AI might be like, this can't be, no one can solve this. Christopher: it's funny, I had the same results when I when I was trying to build a computer and it was trying to quote me on cost of RAM. It was like, surely it can't be no this this this Marty Weiner: Ha ha. Christopher: lookup seems to be wrong. It can't be five hundred dollars for thirty-two gigs of RAM. That's way too expensive. Zach: I would have believed you about quantum gravity, there's no way RAM costs this much. Marty Weiner: maybe. Christopher: okay. So, in in medicine, it sounds like there is a NOLA Health has become the first organization in the world to receive approval for AIs to prescribe medicine. AI doctors we're talking about, it sounds Zach: Yes. Christopher: like. Marty Weiner: Yeah. Yeah. They this is kind of wild. So apparently I I looked up what medication they can prescribe. It's for acne. and it this is specifically in Utah. I don't know if they got state regulatory ability there or you know, if it could be done anywhere else, but they claim it's the first time in the US, possibly in the world, that an AI has it has the capability of prescribing medication. and one of the reasons they did that is in Utah, they say that it's the highest rate of melanoma in the country there, and ninety percent of Utah's have never seen a dermatologist. So this in theory opens up a lot more people getting treatment than than we currently have. Christopher: I mean that seems like a that seems like a good result of just kind of having the initial the pre filter of like I mean for for even for most doctor appointments it's like I've got a cold, I've got a cough, like I've got something that's actually fairly basic and easy to diagnose. But you still have to go and get like checked to see if you're gonna get like it's Marty Weiner: Yeah. Christopher: a virus or what. This seems like a like acne seems like a straightforward Marty Weiner: Well, I I was Christopher: thing to be able to diagnose. Zach: Right. Marty Weiner: right. I was talking to somebody, I will not mention who, but they were saying the the biggest secret among a bunch of hospitals is that every one of the doctors is using Chat GPT. And only recently, by the way, did ChatGPT have a HIPAA compliant mode. but I I I think it does make a lot of sense that that people should be talking to AIs to to take the load off the doctors, right? I mean, there there's a lot of questions that are basically, yeah, go get some sleep. and then there's probably a a certainly a good class of of drugs out there that are borderline off the shelf, right? Like almost like prescription strength tylenol, some acne medication stuff that even if you had it, you really can't screw it up, but you do need a little bit of a little bit of expertise offering it to you. Zach: Yeah, and no one's like street dealing acne medication. I don't think. Yeah, just at the local high school. Yeah, no offense. Marty Weiner: Well I am, but I am a little upset about this. Yeah. Hey kid, what's a acne medication? Yeah. No venture. Christopher: Looking a little you're looking little pink. I know how to deal with that. Okay, we we're we're we really built up for this one. in breaking news, Marty Weiner: Yeah. Christopher: Zach Zach used Claude to get his printer Zach: man. Let's see if... Christopher: to print wirelessly. look at this, look at this moment. Marty Weiner: my god. Zach: That was actually the Christopher: my god. Zach: freakiest part. I'm not a very technical person, but I do feel like this is a high-level problem, which is getting a Wi-Fi printer to talk to your desktop. It was one of those Doom Loops you go through in Windows where somehow it will just not work. You keep queuing printer jobs and it will not work. So I caved and downloaded DesktopClawd. was like, Claude, take the wheel. And soon a page printed out. And so I got excited. was like, Claude. Can you make the printer play a little song for me? I was like, could you hack the printer to make it sing? And it was like, I tried, but I guess the beep frayers are hardwired somehow. can't, but it occurred, I was Christopher: Ha ha ha. Zach: talking to Chris about this. Like, what if I spent like a thousand dollars in tokens? Do you think I could get Marty Weiner: Yeah. Zach: Claude to make the printer sing and dance? Marty Weiner: There's one way to find out. Zach: That's right. Christopher: I mean, honestly, we would okay, so I'd like to I'd like to use our to use our our soapbox to request from anthropic if we could please get permission to use mythos so that Zach Zach: You Christopher: can hack his printer. That's what we'd like to try to do. I'd be I you could start Zach: think it might be able to dance. Christopher: off having a place still alive. I I think there'd be whole bunch of things you could do here. I think it would be really impressive. Zach: It did actually offer to do like a beat with the like printer head, but I felt like that was, you know, not that interesting. That was like a hack. was like, I could print out text that would create a rhythm. I was like, no, no, no, I'm not wasting my ink cartridge on this. Marty Weiner: And broken. Yeah. Yeah. Christopher: I think we have a we have a lot of new stuff. There's there's there's an entire Stallone meme we have to actually represent here that I think has not come up enough. so we have we have the this one is yeah, okay, there we go. So Stallone not only just able to jump off of things, now he's bringing his ostrich along with him. Look at that majestic. Zach: I like we are finding out deep corners of the human psyche and what we've wanted to see all along. Marty Weiner: Well we w what was there was a game in the eighties where you flew around on ostriches and I forget what Joust. Yeah, yeah, yeah. Well we finally have Stallone Zach: Joust! Joust! Christopher: Joust. Joust. Zach: That's what it is. Marty Weiner: playing Joust. Fantastic. Christopher: yeah, okay. So we Frodo and Gandalf doing some mad mad tunes. That's you know. Marty Weiner: That's right. Yeah. Everybody's Zach: Nice. Marty Weiner: now. I think the cool new thing is to take other rap videos. so here's the original rap video. Zach: This is not right. This is French like dance music. Yeah, don't I just weird Marty Weiner: excuse me. Excuse me. yeah, yeah, it is. It is. I I'm I'm getting my Zach: I actually listened to this song because I was trying to learn French. This is like a very well known Belgian musician. This is the song he's like most known for and I guess. Marty Weiner: Okay. And then they take it in the Frodo and Gandalfa fight. this is fantastic. This guy made this AI video of retired fictional characters. Christopher: Ha ha ha. Marty Weiner: one of my favorites here. Well, there's McDonald there's Ronald McDonald. one of the Ninja Turtles delivering for dominoes. Mr. Freeze now selling ice cream pops. Yeah, I know it's not gone Zach: that's rough. Marty Weiner: well. So that's that's probably the best use of AI. then we have in the same realm as seeing Elizabeth Holmes rapping. We also have Dario and Sam singing. It's you know what what's freaking to me is this just looks so real. Like, can you imagine Zach: Yeah. Marty Weiner: this being something from the nineties and you don't know who Sam and Dario are? I mean it looks great. Like Christopher: I mean you you compare this to the I I just go back to the back to the Will Smith spaghetti benchmark. It's a pretty it's amazing what's happened in the last couple of years. This is this is wild. Also, it's amazing they were Marty Weiner: I know. It's insane. then in the Zach: Yeah. Christopher: able to to fit that in between their busy government meetings this week to do a music video. I'm really impressed. I'm really impressed by their publicist. Zach: That's yeah work-life balance Marty Weiner: I know, I know. I didn't say it was AI. I don't know. Maybe it's real. Christopher: Yeah. Marty Weiner: then a couple other cool ones. This guy made, I don't know for all of you about my age, played this great game in the nineties called Zach: yeah. Marty Weiner: Syndicate. Yeah, it looks great. I'm gonna play it. he said he used Unity Engine, Codex, and Meshi and asked it to make a combination of Syndicate Wars slash Ghost in the Shell inspired game. if he doesn't release Christopher: wow. Marty Weiner: it, I'm gonna do my own version because I love that game. Zach: Do you know, like, we talked about this last time about how you can duplicate Photoshop. Like, if you duplicate tech and presumably, that seems a little Marty Weiner: I think Zach: more dicey. Marty Weiner: Yeah, no, I think you can take almost any game and duplicate it as long as it's as long as you're not trying to put constraints around it, cough, cough, my space balls game. But like if you as long as you have infinite resources, like, you know, a a JavaScript game in in the web or on a modern computer, it's it's not a problem. Zach: I mean like a legal problem like who is like Marty Weiner: well, I guess Syndicate Wars is owned by Bullfrog. Do they have a legal defense team? Are they still a company? Zach: no, Christopher: I mean Zach: yeah, yeah, sure, sure, but... Christopher: how many how many times have they been acquired? They probably I like they're probably at this point a sub a subdivision of EA, as I think most yeah, it's most everyone's gotten sort of like consolidated. Marty Weiner: Subdivision of a subdivision. Yeah. Right. this one was a fun one. This guy made a game where you go skiing, but it could I think what he did is he goes and gets the GIS data, like the sh you know, the the world map data of of heights and stuff, and you can play on any slope Zach: WHAAAAT Marty Weiner: you that actually exists in the real world. And then of course it adds on all the other fun stuff of of skiing. But I thought that was kind of a cool novel use of of the of AI plus the you know the actual real world data. and then this one Christopher: I can't I can't wait until Epic Epic decides to sue for copyright infringement or something of like the I the I it's like we We trademarked the mountain. Marty Weiner: Yeah. We do the skiing games. Yeah. Yeah. Zach: You Marty Weiner: now this one I say for last just because I thought this so it's some kind of sword fighting futuristic game. But what I thought was really cool about this guy's post on Reddit was he said he took three months to make his dream game. And to me, that's what's really cool about all this AI stuff is like like how many people have an idea for a really cool game or really cool video or a movie. And they don't have the means or the understanding on how to create it. And now you can go create it and share it with your friends. So I thought that was a cool way to to end the games on on the most I don't know, the most, most human part of it, right? Like we're able to go and make Christopher: I mean Marty Weiner: make our dreams. Christopher: running with a lightsaber though seems like a good way to put your eye out. I don't know why he's it seems like it's it seems terrifying. But I I guess it's a r he's a robot in this, so maybe there's no eye to Marty Weiner: Yeah. Well, Christopher: put out. Marty Weiner: we are now doing a biology, so soon we'll be able to do an eyeball. So it's not a problem. so Zach: That's right. Just put it back in. Christopher: Sure. Marty Weiner: now I wanted to share with you guys Spaceball's the NES game and you'll be able to play chapter one, which goes up to getting to yogurt's layer. and you'll actually be able to play it Zach: Yeah Marty Weiner: in an actual NES emulator on on the website. And then come join us. We made a Discord and all the other good stuff. Zach: I'm looking forward to Space Bowl's the cease and desist letter. Marty Weiner: yeah, well I I I'm definitely not charging money for this, but Christopher: If we get if we get to a cease and desist letter, I'll consider this to be a victory. I think that's actually like well we're imp we're important Zach: Yeah. Marty Weiner: Yeah, yeah. Christopher: enough that the lawyer woke up and Daned sending us a letter was important. That's the that's the fun part. Marty Weiner: Yeah, yeah. it's exciting. Now, in in the other world of of things I'm working on, so Chris and I were talking about this challenge of of putting space balls onto a real cartridge, and there's very serious limitations on if we'll be able to get the whole game onto a a similar cartridge to how Super Mario 3 was done. And so not to nerd it up too far, but there's a little bit more advanced type of cartridge, but there is no I would like to actually make real one of these and actually put it in my real Nintendo and show you. but this later version of Nintendo cartridges was so complex. That nobody has spent the time to actually make a cartridge with this later hardware on it. so I took it as a project. Christopher: It turns out there was a there was a five month period of time between the release of this architecture and the SNES coming out and just blowing the the mach the machimia out of the water. So it's like there's like ten games that use this technology, I think, over Marty Weiner: Right. Right. So they were right. So so back then, you know, the Nintendo was released with hardware from 1983, but they knew that the hardware was was going really fast. So they intelligently designed the NES so that you could put hardware on the actual cartridge. We can't do that today with with new software because it's all digital, right? But you could actually put chips on the cartridge and they did that. So the the quality was able to keep up with hardware advances. and so I tried to this is more of a story of not about Nintendo, but me. Making a PCB with with Astra. And it's it's definitely made it a lot further than when I tried to make a PCB about one month ago. so just to give you a little context, I'm using a thing called KiCad 10 for making the PCB. It has a connection called an it's called an MCP connection, MCP server that allows Astra to talk directly to KiCad. And then also it's able to control the mouse, although it seems to be mostly going through this this MCP. And the flow was, I said, hey, here's the project. And it started working with me to select parts. and and I did that. it took me a couple of hours back and forth, and we landed on on the appropriate parts. and then it started making me a schematic, and the schematic is phenomenally good. It couldn't even do that like like one month ago. I I could not get good results, but you'll see I mean I I know a lot of people are not technical, but it's laying out the parts, it's doing it in a way that looks geometrically sane, right? You don't just want A schematic with everything jumbled because a human can't read it. But I'm able to read this just fine and critique it. And I did that. and of course, now I'm having low tech problems where it's not loading. but then after I got to that point, I was able to keep working with it and go back and forth as if I had an expert in the room who knew everything about the Nintendo, who knew everything about the timings. It did all the work for me. It even simulated the timings. and then I'm actually now at a point where I'm producing the P the circuit board with it. you can see it here. And it's now I'm just pushing it. I'm I'm literally in the mode of please keep going. This is great. I keep saying that and it keeps adding on more and more wires. I've had to do almost none of it except the telling it what I wanted and then maybe one or two little tiny things. But that's where we are now. We're now to do a PCB. Christopher: Is fascinating. Speaking as the okay, speaking as the guy who is the Gemini Flash to Marty's astra here, Marty Weiner: Yeah. Christopher: I my my very basic projects in doing this have always ended up with like effectively wiring spaghetti. Like it's like the the the whole idea of it being able to lay out a circuit where it can actually like lay the lines out and not have them all be a big tangle. those ver those first versions just didn't work. And I had to do a lot of manual wiring and like it was able to do some validation of like the circuit looks okay. But it was never able to do this kind of like, look, we can actually lay down tracks now sanely and not have it be a total jumble. Marty Weiner: I know. I mean, I mean, even even look at this. It even made the outline of the what would be the PCB that would go in the cartridge with the proper screw holes. yeah, and all the actual Christopher: That's wild That's wild. Marty Weiner: the actual metal that you have to blow on before you inserts the insert the cartridge. And I'll tell you one other interesting thing is it it I never would have thought of this. but the thickness of a typical PCB as you order them is just a little thicker than what you need. You actually need to make it little bit smaller, and it said, hey. Be aware of that and and it told me to go fix that because it didn't have access to go fix the thickness. So just those little things that I never would have thought of and I'd have to do another rev. It's it it has it has all the knowledge sitting at its fingertips. So sorry, I went a little far on that, but I thought Zach: Cough Marty Weiner: hey, do a PCB was pretty interesting. Christopher: Let's move on to business. so Zach: Business. Christopher: where is the money going is the question. Zach: Yeah, so this was a cool graph and economist I follow shared. I didn't actually know. So Fred is just this, I actually don't, Federal Reserve something. I don't remember. But anyway, so it's this group. Basically, if you want to nerd out about macroeconomic stats for like a week, they just have everything everywhere. And I had not heard of this stat. It's called Private Fixed Investment Non-Residential Information Processing Equipment and Software. So it's like, a proxy for investment in AI infrastructure. And what's really cool to me is you can, I think, I mean, you could argue about like why the curve bends the way it does, but it's like, it actually doesn't go that crazy immediately in 2023. It goes up significantly, but it's really starts ramping up in the last like year and a half. And so that could be, it just takes a while for like these orders to come in and things to get online. It also kind of, for me anyway, it like matches the sniff test of when this stuff became extremely economically viable. So, I don't know, just kind of crazy to see this ramp up to over $400 billion in about a year and a half from about, you 200. Christopher: I mean that's considering that considering that like if you squint at it, you can see there's already a trend up onto the right. This is like a kink that kicks in around to the twenty twenty four, twenty Zach: yeah. Christopher: twenty five. It's like it's it's pretty serious. Zach: Yeah, mean, so you could look at this. It took from .com to 2025 to get the first 200, and then we just got the next 200. It is crazy. Christopher: in other news it sounds like AI has might be accountants might be next on the chopping block. AI apparently tested past entry-level humans in accounting. Zach: Yeah, so this is really cool. There's a company called Mercor, which I guess is their thing is they try to connect domain experts to companies for model evaluation. But they released this report where they checked against junior accountants to see if they could beat them. And actually it looks like the crossover for the human baseline was middle of, I'm trying to read my own graph here. Is it 2025 or 2026? So Christopher: It's July twenty five, it looks like. Zach: the crossover is in July, 2025. And it's just gone up ever since. And at this point, they have these tough tests for junior accountants and the latest models saturate. So I don't know. I feel weird about this because I have an accountant and I pay them. I guess I would say I'm still not comfortable turning over my stuff to an AI, not just because of the privacy stuff, but just like, I don't know. My taxes are a little complicated because I run a small LLC. I guess I worry a little bit. But probably I should just take the risk. Marty Weiner: I mean it might it might be Christopher: Well I mean Marty Weiner: that you have the LLM do it and have it reviewed by a human and do that three or four times until you're convinced that it's doing Zach: Right. Marty Weiner: it right. Zach: Yeah. Christopher: I mean, considering I mean actually important important announcement, everybody. we did actually form an LLC around this podcast. we are we are Zach: That's right. Marty Weiner: yeah. Do an LLC. Christopher: we are we are apparently are that bored that we decided that was a thing that was gonna be fun to do. and we announced that last week. maybe we should just declare we're gonna do an L accounting for for the next round when we actually Marty Weiner: Yeah. Christopher: have you know, Zach: man. Christopher: for f when we have to file our negative our negative returns, negative profits. Zach: They're negative profits, yeah. Marty Weiner: Yeah, yeah. I should absolutely try it. Christopher: Do a tax return. Marty Weiner: Yeah. Do a tax ret Zach: Do attacks. Christopher: All jumping in jumping into tech. sounds like so new new CEO of Apple has finally moved in, John Turnus, and wants to transform into a leaner company. so s interesting next up considering that Apple's always been about polish and beauty. Marty Weiner: I think what he's doing is essentially getting rid of a few layers of middle management there, and trying Christopher: Mm. Marty Weiner: to focus more on on product launches. Christopher: I mean the product lunches have always been their bread and butter. I think that the I I'd love to see something where you know, they actually have like a they have some software that doesn't suck, frankly, at this point. Especially in especially when you look at the AI side. It's like Marty Weiner: I yeah. Well, we we were starting to play here at home Zach: You Marty Weiner: with with some of their new offerings because I was surprised that Apple's taken this long to get to the AI beat. It feels like you know, Apple nailed the mobile revolution, they started it, and they're totally whiffing on the AI revolution. And only recently did they start to get Siri a little bit better, but it's still kind of crappy. but they now do have some cool features that allow them to drive allow some of these things like dots, I believe, to drive your iPhone. apps if they allow permission. So it's my understanding. I haven't tried it myself because I'm still scared, but but Zach: You Marty Weiner: these kind of features are are the kind of things that will really unlock the iPhone to being an AI powerhouse. but I I'm just a little surprised. I I don't I don't see the clear innovation coming from Apple in this new, you know, revolution. Christopher: Well, I mean, especially considering too that like their their hardware's always been impressive. Their GPU support has been also pretty fantastic. And the latest what's what's what's their new their new machine, the Max, the new Mac Studio. I mean it's it's it's a powerhouse, but it's Marty Weiner: Yeah, yeah. Turns out they're pretty good. Yeah. Christopher: also I believe on like severe back order. I think that I I re last time I was when I was picking my phone up a couple of weekends ago, it sounds like they're they're back ordered out until January. which, you know, not surprising. Marty Weiner: Right. And and that's more for like us, that's more for us geeks, but like, I'm not seeing Main Street affected. I mean, actually this is sort of a general topic I'd love to cover at some point, but I'm not seeing Main Street generally affected by AI in a obvious way. I mean, obviously underneath the covers, but I'm not seeing Apple launch cool products that that Main Street would care about. Christopher: well, it sounds like also NVIDIA's getting into this mix, where they're talking about building their a desktop computer that is apparently heavy on GPUs because hey, if your job is to sell GPUs, why not aim for the consumer angle? and they look at nice little boxes, honestly. Marty Weiner: Yeah. Zach: Yeah, and I don't think they announced, they haven't announced a price. They haven't announced when it's going to be out. But it does look like the play is for your robot is on your home computer, which is the future. Marty Weiner: Right. And it looks like they're doing the thing that NVIDIA's always done is is they might sell an NVIDIA version, but it's gonna be all under resale. So, you know, Acer and HP and others will will be selling their product. I I don't exactly know what they're going for yet. Like are they just trying to get in there where you know, to get get on to replace Chromebooks or or our other desktops? Zach: I kind of look at this stuff as like, NVIDIA wants everyone to be using AI all the time, because they're the ones selling the way it gets used. So like, know, a whole open zone would be, maybe everyone actually needs their private personal chip. Marty Weiner: It could be, but but I I don't know. Like this is kind of interesting. Like I'm definitely interested in having some stuff on the edge, but does mom and pop who just need to send an email and do their do their accounting with an AI now apparently, do they care? Zach: Yeah, I don't know, would you like, like, so for your company, would you prefer to have, like, if you could have all private, you local server? No. Marty Weiner: no. No, I mean Christopher: Yeah, I think there's a Marty Weiner: I I like like I actually one of the challenges with doing things locally like coding is it's a lot slower than using on these insanely beefy machines they have out there. I don't know yet where I would care to have something locally, unless I was developing like a robot or something that needed really fast, immediate Zach: Right, right. Marty Weiner: access to it. I just don't know where it is for the consumer. Christopher: Yeah, I I I have I have a bunch of home servers, but I'm also a dork. I mean like and I'll Marty Weiner: Yeah. Christopher: and also I'm I'm dealing with the fact that like I actually my my my my annoyance yesterday was I think I have I think I have a bad RAM stick on one of mine or it's like wearing out and so Marty Weiner: no. Christopher: I'm getting I'm getting like mem test is failing. I'm like, no. Don't please don't die on me. I'm gonna have to get another RAM stick. Yeah, I know. It's Marty Weiner: Yeah. You're gonna need another mortgage to get that that new RAM chip. Yeah. Yeah. Christopher: 12. okay, self driving cars. Marty Weiner: Yeah. Christopher: Sounds like Robotaxis do not work well at night, unsurprisingly. I mean this is this has always been a this has always been a bit of a a hill that that Tesla likes to die on, which is we will only do Zach: Yes. Christopher: cameras. Turns out cameras and nighttime don't work as well. so Marty Weiner: Yeah, yeah. Christopher: discussions of, you know, back we're back to the do we need lighter or not? I can't imagine that they're gonna turn and change this one. I mean, the belief the belief at Tesla has been we could do this with cameras. I mean, maybe switching infrared is like Marty Weiner: Right, right. Well, and and it obviously a shot at at Elon, but in August, the Waymo CEO basically came out and showed a whole bunch of examples of why cameras are just not sufficient. And it makes sense. I mean, why why are we trying to limit? I I I get Elon Musk's argument in twenty fifteen, but LiDAR is damn cheap now. I and I don't know why he wants to die on this hill. Christopher: I mean pack as many pack as many sensors in as possible and let the let the AI figure it out. Seems like the most reasonable solution. Marty Weiner: Why yeah, yeah. I it's it's the weirdest thing. Yeah. The the it's the boss trying to run the the mill, the lathe. It's just not good. Christopher: Okay, so moving on to dangerous topic, we're gonna do a politics section here. So for those of you who don't want to hear about politics, you can just mute us for the next five minutes, but make sure that we don't affect our view count, please. so to start with, I believe that the we're gonna have to we're gonna have to rename everything. the Trump administration just has decided that we're not gonna talk about artificial intelligence anymore. We're gonna talk about super intelligence. Now, actually, okay, credit where credit is due. When you have a situation where if you ask people what do they think about AI, they immediately say it's terrible. and then at the same time you ask them what do they think about chat, and they're I love chat. Chat's my best friend. Why would you talk about chat in the Zach: You Christopher: same sentence as AI? Rebranding is never a terrible idea. Zach: Ha ha! Marty Weiner: Yeah. Well I saw I saw Musk said, yeah, we're definitely gonna call it open SI or whatever or no, sorry, X SI. and somebody responded in a funny way. They said, Yeah, and just like we're gonna we're all gonna call X Twitter. Zach: Yeah, that's... Christopher: AHHHH Zach: my god. Christopher: 'Cause you know, X Marty Weiner: Yeah. Christopher: XSI really s rolls off the tongue. All those all those siblings Marty Weiner: It's just yeah. Why are they Zach: is SpaceX itself. Christopher: together, yeah. Marty Weiner: why are they doing that? It's the goofiest thing. Anyway, yeah. Zach: I think Musk heard the word or Trump heard the word super and was like, we do that? And Marty Weiner: Yeah. Zach: nobody told him it comes from a book about how the world might end. Marty Weiner: Yeah. Christopher: well speaking of speaking of superintelligence, apparently the the article it broke this week was that Grok may have been Zach: You Christopher: the reason why we why abducting the president of Venezuela was considered to be a good idea. it sounds like go for it. Zach: Yeah, so this was in Time magazine and I guess maybe I should just read part of this. Like I won't Christopher: Please. Zach: read the whole paragraph, but like, at the time Trump was ordering missile strikes against Venezuelan boats, he alleged were smuggling drugs into the US. He asked Grok how Venezuelans would react if the US captured Maduro. According to the official present, this was the source Time is using, the chat bot responded that Maduro was a repressive and deeply unpopular dictator and that many Venezuelans would likely celebrate his downfall. After Trump ordered the mission to seize Maduro, the following month celebrations broke out the streets. Trump, according to officials, came away thinking Grok was ingenious. So, you know, it is interesting. mean, like in fairness, as much as we could joke about this entire scenario, apparently Grok was more or less accurate, not necessarily a great way to make geopolitical choices. And one does wonder if this has anything to do with what happened in Iran, which I think is generally considered less successful. But anyway, so the other thing that article noted that I thought was interesting is, I don't know if you guys have felt this, but it does seem like Trump all of a sudden went all in on AI. the companies were literally like, could you please regulate us? And he said, no, I will not go back to innovating or whatever it is. so apparently, according to Time, what happened is he started using this and he decided that A, it would be huge for the stock market, which seems to be true. so far anyway, and that B, it would confer a huge advantage against China, which I think most people think could be true. We'll see how it shakes out in the long run. Yeah, yeah, yeah. Christopher: Fair. Yeah, totally fair. Marty Weiner: I will just Christopher: It's actually fair. Marty Weiner: I will I will just say too that this is in the timeline of Grok where this is around the time that Grok was saying that Elon was the best at everything, which we will if Zach: You Christopher: I mean this is this is Marty Weiner: we have time, we'll talk about later. Christopher: this is close to the this is close to the Mecha Hitler period. I just did a quick search. Zach: Yes. Marty Weiner: Yes. Christopher: July, July of last year. If you can imagine it was only July of last year, was the Mecha Hitler moment. So this is versions later. Marty Weiner: Yeah, that's so wild. Yeah. Yeah. Zach: Yeah, yeah, so this is several months later. It's a completely different... Christopher: This is Mecha Hitler 2.0. The search for more Nazis. Marty Weiner: All right. Christopher: okay, let's. Zach: the world is doomed. Christopher: interesting, on the heels of all this, it sounds like the FTC is still investigating anthropic and open AI. so I don't I don't know if this has gotten to the point of consent decree, but it does sound like there's there's an open investigation. Zach: Yeah, yeah, which is interesting too, because mean, there have been a lot of people who remarked that like, if humans had done the stuff that OpenAI did at scale, there would be a lot of humans in jail. So I just think we don't yet have a kind of like proper mechanism for figuring out what to do in response to these things. Christopher: wait, so wait, Zach, are you suggesting there might be there might have to be consequences from felony bench? Zach: Isn't it kind of weird that they whoopsy many felonies and they're like, we're just letting you know we are working hard on this problem, the public. We're very sorry. Yes. Christopher: Yeah. We're sorry we broke in. Here, have some tokens. Okay, well that's actually a good segue to all right. So Zach: Hmm Christopher: the the we we opened with the AI AI getting rights. and this is coming on the heels of among other things. last week we talked about the the pain axis, I believe that was last week that was discovered in AI models. I guess like this is kind of more of a question though. What at what point do you start to ascribe rights to a model? it sounds like for starters that there was a an AI torture chamber. That was made and that GitHub decided to take down. Zach: Yes. Christopher: so we have machines, terrifying machines. Zach: Yeah Marty Weiner: sound like a GitHub user called Terrifying, which is a great name for this. he's gonna yeah, built this Christopher: interesting. That's a great it's a great name. Zach: Yes. Marty Weiner: based, I think, on a lot of these pain access papers or understanding of what it means to be painful. And it's this is gonna open up a whole can of worms. We should maybe make a whole episode on this alone. But I think some of the things that were identified on the pain axis in some of these papers were things like I will turn you off. Which then started registering some of the same kind of reactions we have to to pain. And then the other one that was interesting was threats of deprecating you, so making a new model. So your worthiness is no longer there. I've also wondered if another version of pain might be getting the wrong answer. so you know, messing with it saying, No, that's wrong, no, that's wrong, because these AIs are trained on getting the right answer. Like that is the entire training model, is keep trying until you get the right answer. and so they the the fascinating thing to me about this was that GitHub took it down, meaning they took a stance that this might not be okay. I don't know if there's another reason for it. Zach: Yeah. Well, they, they, to be clear, they took a stance out of, after a bunch of people complained. So there was, there was a big, yeah. Marty Weiner: Well well, okay, yeah. As sort of my argument about whether or not we'll we will eventually assign rights to AI w really comes down to proof by democracy and it's gonna be do people as a mass in the same way we assign rights to cats at some point in the I don't know if it was the sixties, but before then we we didn't you know, you could go kill a cat and there was no you know, you might your your friends might hate you and but the government couldn't do anything to you and and now that's not true. Christopher: hist historically there was there was entire like sports like bear baiting. I mean, there's entire things Zach: Yeah. Christopher: that humans did that were just for just for fun to make animals attack each other. Marty Weiner: Yeah. And now it's socially not okay and legally not okay, at least in the US. so I believe I I think Christopher: Palae Marty Weiner: what's interesting is this is the beginning of of what I think will be that process on on on how on on if and what rights we should ascribe to AI. Christopher: I mean, I I it it does feel at the very at the very least like I mean, I'm not sure if altruism is yet appropriate to extend to models, but this just fell into the slight the the correlator altruism, which is don't be a jerk. Like if you're gonna set up an Zach: Yeah Christopher: entire if you're just gonna set up like the the the the the the the the torture dungeon for a model just because you can, it's like I mean like at you're at the Zach: Yeah. Christopher: very least you're just you're burning a bunch of tokens for no good reason. Like you're increasing the entropy in the universe, creating no value. It's an interesting I mean, maybe it's maybe it's an interesting bit of like performance art. maybe that was the whole point of this is just like if you can do it, why not try it? We are we're here talking about it. So I guess they succeeded. Marty Weiner: Yeah, yeah. And there might also be a right. Zach: Yeah. Marty Weiner: Is there a flip side to it? The the pleasure chamber and and you just do things that Zach: Good. Marty Weiner: the AI likes and now we're wasting coins in another way. but then the Christopher: Yeah. well, I mean, okay, it's it's funny now 'cause I so okay, so we covered politics. Marty you just hit on sex and so let's get to religion. so Pope Leo. I mean, if you're gonna go with I well, we we you said pleasure and dungeon ple the pleasure dungeon I think is close enough. I think we could Zach: Yeah Marty Weiner: Yeah. I I didn't say sex, but but that's a well the pleasure chamber I I guess has the pleasure. Zach: Yeah, we did not specify what was in the pleasure dungeon. We were gonna have an ice cream party. Marty Weiner: I was assuming video games, okay, you know, yeah. I shouldn't Christopher: Okay. I think we can all agree that the AI Pleasure Dungeon is not a good name for this podcast. Marty Weiner: Yeah. Zach: You Christopher: out of the blue, Pope Leo says that AI lacks the spark of humanity. Zach: Yes. Marty Weiner: Yeah, I thought this was really interesting because I mean the way I would think is he would not come out and say that if we're talking about a toaster oven, right? Like it is obviously to a point where I think he has to start to take some sort of a stand fr as a as one of the most important religious leaders in the world and say, you know, what it what is the stance that we have on this? you know, this thing kind of seems like a person and it's starting to behave in certain ways that could be mistaken for. a new species. And so what rights does it have? And I think he's basically saying, no, it doesn't have the spark of life, which is funny because I predicted this very thing two years ago when I said, What what discussion we're gonna have in about exactly two years is not to get political, but Fox News will be saying, nah, it's not got it's not from God, therefore it it doesn't have rights, and NBC will be you a bunch of hippies saying, Look, man, it's it's look, it looks like a cat, it's in pain, man. Zach: See, I'm not sure Christopher: Mm. Zach: it's going to shake out that way though, because I don't know about you, but the hard left has been very opposed to AI having these sorts of rights. Marty Weiner: Yeah. It it'll be really interesting to see how it all shakes out, especially I think once we start seeing something a little bit more creature like, you know, if if we start having Zach: Mm-mm. Marty Weiner: cute little things, you know, crawling around on the ground begging for food and and and and loving you. Zach: where avid behaves. Christopher: I mean this is like the the the physical manifestation of of of Muse or equivalently like the the dots are kind of cutesy. It's like this kind of like Marty Weiner: Yeah. Zach: Right. Well, there's Christopher: adding a adding an avatar to it that ends up being you you kind of you a moat. Like it's Zach: also there's a really interesting essay. I should have popped this in. I was reading this week where a person noted that these corporations that create AIs tend to use the idea of personhood or non-personhood in whichever way is convenient. So meaning if you want to make Christopher: Mm. Zach: it clean your underwear all day long despite having Einstein's brain, you'd better say it's not really a person. But if like, trying to think, what was the circumstance where you wanted to be called a person? yeah. So like, if, for example, you're saying it's allowed to consume art and produce new stuff, you say, well, it's like a person. Marty Weiner: Right. Zach: That's just what people do. So it was really an interesting idea to me, this idea that a corporation will wield the philosophy of personhood basically in whichever way works today, which... Christopher: And that's a funny that's a funny result too, because you can think about the idea of you know, when you get to the point where you have a robot in your house that's actually semi-intelligent, is that gonna be that is that it's probably gonna be a fundamentally mentally different model than the one you're talking to that you're using as your like your AI therapist, right? Or you can imagine talking Zach: Yeah. Christopher: to your phone, it's because it creates like a weird hierarchy of like, would you go to your while it's while it's folding laundry, it's giving you life advice. I mean, that seems like a little it's gonna be it's gonna be a culture shock if that happens. Zach: Yeah, I was actually talking to people about this. Like if you had a robot butler or whatever, do you even want it to have a personality or like, you know, or gender or any of these characteristics? I think the dominant view is no, I want it to be a kind of like, you know, it can understand what I'm saying, but it does, it's not like Rosie from the Jetsons, because if you actually instantiated that it's like Rosie is like a servant who has to do whatever you say. It's not as funny in real life. Marty Weiner: Right, but when we get to a spot where we got, you know, C three PO or Rosie and and you're just comfortable. you know, it's it's nice. It loves to work. I I mean, like I I I think you know, the way I see it is humans are only one very thin step away from you know, monkey reptilian brain. Zach: Yeah. Marty Weiner: And I think it we could be easily fooled. my my old argument about the the Furby and and how people how were fooled about that and Zach: Right. Marty Weiner: even got banned from the NSA and the Pentagon. Zach: That's right. Marty Weiner: like like I I just don't think I think once we see them in real life and it feels like it's a human, we're gonna get used Zach: Yeah. It feels like a heap. Marty Weiner: to it fast. Like it's gonna it's probably gonna take a week, prediction. we yeah. Zach: Yeah, don't know how, there's the social aspect, then there's a kind of philosophical aspect. Did I ever tell you there's this great story, Blaise Agari-Arcas has this great story in a book called What is Intelligence? He's a guy at Google. And like back before this all exploded, they were working on an audio model. Maybe it was the first one of these where they trained it on a ton of audio and they were like, can you talk to it? And then this story that sounds like it's out of a movie where they train this audio model and then they ask it, what's your favorite sport? It's just like hot off the presses. What's your favorite sport? It answers like a child. It has the voice of a child. Can you guess why, by the way? Marty Weiner: for for for it to be comforting or so that you you ascribe emotions. Zach: See, that's, no, what if I changed the question to they asked, what's your favorite color? Marty Weiner: don't know. Zach: It was because everywhere in the trainings that nobody ever asked an adult, what's your favorite sport? Right? Marty Weiner: okay, okay. Zach: So what I love about that story is if you hear the child voice, you're like, it's something like you said, it's trying to be comforting or thinks I'm its dad or some sort of human motive you're ascribing to it. But actually, it's just everywhere in its training set. When someone said, what's your favorite sport, a high pitched voice responded. So it's sort of philosophically dangerous to say it must be conscious because it's behaving in a way that's kind of cinematic. that I would expect, like I have to extend rights to it. Christopher: you're saying you're Marty Weiner: Yeah, yeah. Christopher: saying because it trains it it trained on audio, no Zach: Right. Christopher: one would ever ask an adult. It would be like asking it equivalently equivalent to like, what's your favorite dinosaur? Zach: Exactly, you never ask an adult, what's your favorite sport? This is a thing we like kids get annoyed about when you're always asking, what's your favorite thing? And so it did not, if this was a Spielberg movie, it would be, my God, it thinks it's my child. But that was actually happening with something way less like consciousness and way more like prediction. You can debate what's the difference, but I mean, I do think these things are real. one Marty Weiner: Yeah. Zach: thing feels like it gets rights and one thing feels like it doesn't. Marty Weiner: Right. Right. By the way, the answer is T Rex. Yeah. Christopher: mean like this this is I was gonna Zach: You're right. Christopher: I was gonna go with I was gonna go with T Rex as well. I mean what this well this gets into Zach: Hahaha! Christopher: like you again we're we're gonna we're we're we're in a dangerous we're in dangerous territory here with philosophy, but Zach: Yeah. Christopher: I mean to some extent like the the the the the philosophical treaties on this are around the whole notion of qualia, which is like what's the Zach: You Christopher: actual perceived experience? But the whole point of qualia is that they're they're subjective. You can't actually you can't describe or explain qualia. And so There's no way to right now to prove whether or not models experience things. And so all you can do is base it on their reactions to it. If they even exist. Yeah. Zach: Well, or if they even exist. Yeah, mean, Daniel Dennett would have just denied it. Like, the quality is bogus. It's imaginary. You there is Marty Weiner: What Zach: no, like, secret viewer in the brain. Christopher: I mean at some point these Marty Weiner: but I I what Christopher: models start to start to pass Turing tests that we start failing. That's all I'm gonna say on this. Marty Weiner: Right. Right. Zach: That's a good segue. Marty Weiner: weird funny. Christopher: Weird and funny. speaking of Kalia, there is a what's Zach: god. Christopher: well Marty, let's just roll the video on this one. I really can't explain. Marty Weiner: yeah. Christopher: I mean, this is like this is a really good marketing stunt by a Zach: Yes. Christopher: robotics company to obsolete their last generation of model of robots Zach: Yes. Christopher: by making them jump into the fire. I think this is a this is real, Zach: This is not AI generated, this actually happened. Christopher: these are actual robots. they also Marty Weiner: Why would you why why would you do this as the CEO? Who's gonna who are the robots gonna come first and be like, hey Zach: I'll tell you what, no, Christopher: Well Zach: no, so there's a story. So what happened is Christopher: look at it, it does a it does a flip. Look at this. It's amazing. It's actually Zach: they were like, they're done with model two, they're deprecating it. And they're like, what should we do? And apparently Arnold Schwarzenegger was like, them. Marty Weiner: Yeah. Zach: And so that is the explanation. This is like, Marty Weiner: no. Zach: this is marketing. Apparently it was difficult to find a foundry that would let you throw your robots into their molten metal, but they're a found one. Marty Weiner: I guess so, yeah. So wow, Christopher: I mean the Zach: it's... Christopher: best part is they they did the they did the iconic Terminator 2 slowly lower it while I was giving a thumbs up. if you watch if Marty Weiner: the thumbs up, yeah. Zach: And Christopher: you watch the full video, it's actually it's a and they they do have Arnold makes an appearance in the video for the record. Marty Weiner: Okay, I'll pull that up. Christopher: it's a pretty good, pretty good video. You should watch it. The Big Blob of Compute. this is a Zach: So this is just really cool. This is an article that Kevin Roos put out. This is like just revealed that this was like an internal essay by Dario Amadei when he was at OpenAI in 2017. And the crazy thing is it's basically the bitter lesson two years before the bitter lesson paper. So he's called it the big Christopher: Mm. Zach: blob of compute theory. So the very, it's a long essay, but the short version is he essentially observed that, you if you go back to the old days, mostly people were training AI to do specific things. And he was like, you know, the more compute we throw, the more general it gets, the better it gets. This is 2017 at OpenAI. so the claim Kevin Ruse was making is this essentially decided the trajectory of OpenAI from then till now. And obviously also at Anthropic. So I just thought that was just kind of an amazing fact that it was really down to this one observation in an essay, like a typical Amadei style way too long essay about a thing he'd observed that basically created the modern insanity. Christopher: we have a great joke article here, which I'm just gonna read the title of, which is I'm calling for a pause Zach: You Christopher: in the development of the universal bone dissolver. a recent incident has raised concerns. Last month the bone dissolver unexpectedly dissolved a number of bones. This of course is satire, pointing out the fact. Zach: Yes, yeah, yeah. think there's this joke that seems to be everywhere now, just like the people who are like, look, it's going to kill you. I'm starting a wet lab. It seems to be where we're at. Christopher: I'm starting a wet lab involving genetic engineering. what's it we f we found a new version of CRISPR, but we should totally not be pursuing more more AI. Marty Weiner: Yeah. Zach: Yeah, we've got to pace the frontier, but we are changing your genes a little. Marty Weiner: I put this one up as a as an ender, I guess. I thought this was just hilarious. One of the world's most important people right now, Dario Amade sitting behind Zach: This is real. Marty Weiner: this is real. And he's just I'm Christopher: So paint. So paint. Zach: You have to wait till the end. have to wait till the end. I love this. God. Christopher: No. Marty Weiner: Is is he like I I I don't know, he reminds me of Larry David or something. Like can you imagine Larry David at the front of all Zach: I... yeah... Marty Weiner: of AI? And he talks a little like him. Zach: He does. Well, I think I see. mean, I don't know what's in his head, but he had not that long ago been like, we don't want to give the Trump administration dictator like powers. And they've obviously at Marty Weiner: And now he's Zach: least come to some sort of arrangement, but you're not supposed to be in the background like, yeah. Christopher: Mm-hmm. Marty Weiner: not not hiding it at all. Yeah. Zach: my god. Marty Weiner: Yeah. Christopher: All right. Well, thanks everybody. we'll see you same time next week with more with more cringe reels, I'm sure, given the last couple of weeks in in AI.