UL NO. 451: Altman Says ASI in "Thousands of Days"

big model that got released. And it turns out it was Google as opposed to anthropic, although maybe anthropic is going to do something as well. I was hoping for like an opus four or something, or at least an opus 3.5. So it turns out to have been Google and it looks like new Gemini 1.5 Pro and Flash models. So right off the bat, I mean, these seem very minor. They're minor updates to 1.5 Pro, but whatever. I'm looking

for reasons to be excited. 52% reduction in output token pricing, two times higher rate limits for Gemini 1.5 flash and three times higher rate limits for Gemini 1.5 Pro. So these are really updates to 1.5 Pro. And if I go to let's see here, if I go to Kitty and I do this yeah I'm not seeing 002. Oh there we go. Yeah 20. Okay. So if I put 20in here into fabric I'm now using Gemini 1.5 Pro version two, which is the brand new one that just

came out. So I guess we could do like a classic. Um, so this is the rivets video pipe to fabric SP extract wisdom DM all right, so streaming works. Seems pretty fast. See how these insights are? So first of all, it extracted a whole lot of insights. That's or no that was ideas. Now insights should be fewer. And these look really really good. So here's one thing that I love

about AI right now. So if I give this same challenge to Gemini and GPT four or any of the modern ones from OpenAI or anthropic, they all come up with this one as the top quote. You can't necessarily think yourself into the answers. You have to create space for the answers to come to you. And I think that is really, really cool. And all these other quotes are just fantastic. This is like one of the videos I know the best. It's like one of my favorite

videos ever. So anyway, this is the results of the capture from 1.500002, which is the brand new one. And let's look at the habits. First of all, it followed instructions really well. And by the way, the DM version of Extract Wisdom has a whole bunch of extra instructions. It's like way more chain of thought and way more prescriptive for like what I'm looking for. So that should give a pretty good advantage. And it looks like Google did really well with it. I like the way that

they're saying this. Oh, you know what I want to do my other new favourite one. So let's take that one and let's do create story explanation I love this create story explanation one. Absolutely fantastic. So it basically just walks you through like you're in a conversation about whatever the input is. So they lament the decline of creativity and deep thought. In the West. They discuss the impact

of technology, social media and constant barrage of information. One speaker shares her experience of deep diving into the history of the New Testament manuscripts. They discuss the importance of deep reading, embracing diverse interests. So yeah, the conversation concludes with a powerful image of the decline of the creative spirit. I mean, this is just a cool way of describing a piece of content. So I think I could do like another one. Neri Oxman so this is one of

my other favorite YouTube videos ever. Yeah, look at this. Oxman believes nature possesses a wisdom beyond human intelligence. She's building a company called Oxman to explore the wisdom and augment nature with technology. Her goal is to grow products, blurring the line between made and grown. See? Check this out when I go to summarize why I love this video so much, it's hard to capture it. It's hard to remember what I love about it so much, and

this does really well at that. The tech does really well at capturing what makes her so awesome and it's got advice. It's just like great. In fact, I might even get rid of this pre and post sentence. I think I might just keep that content right there. But I'm not going to do that live. But anyway. Yeah interesting. Looks like I have a little bit of a key showing. But that is not a real key. That is a HTTP Cookie header. Always like to watch my OpSec here.

Going to leave a token at some point. I'm going to leave a token to see if I'm going to leave an API key out there and and monitor to see if anyone comes in on it. Exercise for the reader. Okay. So cool. So yeah, we don't have the video of Parul and Sam Altman yet. It looks like he has not dropped it yet, but that's fine. Just go back to the show. Basically some model upgrades from Gemini and they look to work quite nicely. And I just changed my default model to it just now. So I'll see

how that does performance wise. Okay. Um, Thomas Rocca created a web GUI for fabric called fabric UI. Check this thing out. Look at that. And Jonathan, who's on the fabric team, is actually working on something like this as well. But look at this. You can select the model. You want to use a pretty good selection there. You put it in your API key. The guy's a hard core security guy, so I don't think he's going to be stealing keys anytime soon. Um, and then, look, you pick

your pattern here, and. Yeah, you got a guy to run fabric with. And again, Jonathan on the fabric team is actually building something similar. He's building like an API server and also guy. So that's going to be really cool. But this guy, uh, Thomas Rocha. Super cool guy. I've known him for years and years. Uh, he's part of the UL community and, uh, evidently just built a UI

for fabric. So that was very cool. And I just spent a whole bunch of time, like, deep cleaning on my phone, which is the new iPhone 16 Pro and deleted, probably 40 applications, phone screen cleanup, widgets, Refactors watch faces. I redid all my focus modes, which are these things up here and uh, like right here. So you got like work and personal and Do Not Disturb and all that. And all of those have individual settings which map over

to the phone as well. So it's like it's basically a giant ecosystem upgrade, like computing ecosystem upgrade within the Apple ecosystem. And I do that every year, but sometimes I only do a very little. And like the last couple years, I've only done a little bit of it, and I probably spent 5 or 6 hours this year doing a massive cleanup. And the reason I mention it is because it feels really good to do that. You've probably been feeling weight weighing on you of like, oh man,

these apps are nasty. Why do I have these? Do I have subscriptions that are still turned on? So all that to say, I encourage you to go do a cleanup. It will feel good. Okay, so I did an analysis of OpenAI's long form conversation with the zero one strawberry team. So this is the conversation.

OpenAI video with Gemini, but whatever. It's not that deep. Okay, yeah, they've been working on this for a long time. It's a reasoning model, meaning it takes time to think before answering. The team had several aha moments, like, I feel like you could just read these ten bullets and have a conversational, explanatory explanation of what's in this video. And that's why I created this pattern called Create Story Explanation. It's just a way of explaining anything in a super approachable way,

including like the most crazy, crazy. Okay, I'm going to do particle physics, so let's do string theory. Look at this. Imagine everything in the universe not made of little balls, but tiny vibrating strings. They vibrate at different frequencies, and those frequencies determine what kind of particle it appears to be. Look at this. That's clean. That is so clean. And it's different than then extract wisdom, which I don't even

know what that would do. Yeah, that. I'm curious. Take all of string theory, basically that the model knows about and send that to extract wisdom. I very curious what this is going to do. Ideas? Holy crap. It is extracting all the ideas out of string theory insights. Okay, it's doing actual insights. Yeah. Quotes. Nothing. Facts. Nothing. Habits. Nothing. Recommendations. Nice. Okay,

I'm pretty happy with this, actually. See what we get from the insights potential unified framework for understanding all fundamental forces and particles. Well, yeah, that's that's why string theory was exciting and or is still exciting. Not sure if it actually still is. Extra dimensions. A key element of string theory could reshape our understanding of the universe. See, this is really cool, right? Look at this. Look at ideas. Insights. This is really cool and powerful, but it's not as

cool as this. If you're trying to explain it to just a regular person or or heck to me on any given day, I would like to hear it this way. And then if I really want to go deep on it, like then maybe extract wisdom to pull this stuff out and definitely extract wisdom if I want to extract, you know, purely the ideas or whatever, which I actually have separate

patterns for those. But anyway, bottom line, the reason I got on to this is that I did an extraction of the techniques that were that they said they were using the zero one model from in this interview, they said we here's what we love doing. And they went around the room to all those different people and it was like, hey, what? What do you do? What what have you found that oh, one can do that GPT four can't. And so they came up with this list.

Generate code part pass error messages for debugging. Learn complex technical subjects asking it to explain concepts without hallucinations. Brainstorm solutions. Okay. Um, okay. So I think this might be my Twitter post about it. Yeah. Yeah. So I did a thread here. I'll just walk through these real quick. So some thoughts. A lot of these takeaways are similar to GPT four sonnet that can already do. But the main theme is simply asking more of the

model and expecting more out of it. And another key thing seems to be idea generation brainstorming creativity. So it's more like a creative partner. Oh, one is rather than previous models. So that was one thing they were really excited about. One of the ones was I loved the bit about asking to connect ideas. So if we go back to this list, look at this, um, ask oh one deep questions about abstract concepts to receive thoughtful, detailed answers. Um,

implement projects with little prior knowledge. Oh, here we go. Organize unstructured thoughts by asking oh, one to connect ideas and identify missing elements. That is cool. I'll just open this up, make text more creative by requesting multiple alternative ideas. So again, all of these are very thinking based and very creativity based. I generate ideas for structure and content of writing. Yeah, really really cool stuff. So that's what that was about. Um, yeah. And it's in the newsletter,

so you can go check it out. All right. Wrote a piece for AT&;T business about AI transforming real time monitoring. So this is basically a security post about many AIS. And, uh, yeah, I think you'll like it. It's, uh, a piece I've been thinking about, so I was happy to do it for AT&;T business. Who, uh, who sponsored the piece. So I'm excited to be a keynote speaker at Swiss Cyberstorm, which is a really, really cool, uh, security conference. I've

been reading more about it. They're not about, like, all the latest hacks and like, oh, we we hacked this thing. Let me show you how we can break into this thing or whatever, which again, I've done a million of those talks. I'm all for them. At this stage of my career, though, it's like I'm not finding that as interesting just because I feel like it's too easy to hack things. There are lots of kids standing at the top of a stairway, and it's like, easy to walk

by and push them. Why would you want to walk by and push them? The whole world is super fragile. It's a surprise that any of this works like I just. I'm not getting the same enjoyment that I used to get from breaking into things or knocking something over, you know what I mean? It's like it doesn't give me a good feeling. So the thing that's giving me a good feeling now is building and trying to figure out

what the most difficult problems are in the world. Um, anyway, bottom line for this thing, this conference here keep going on tangents. Um, this conference focuses on that. It focuses on the ideas around security as opposed to like the latest hacking thing. And I like a mix of both, honestly. But I prefer more the the thoughts and the trends and stuff like that of like, where is this going? Are we moving to a place where there's no trust?

How does all security change when there's no trust? What sorts of infrastructure are we going to need in that world? That kind of thing. And you could use code unsupervised learning for 15% off for your registration. And the theme for this one is AI Revolution. So yeah. Cyberstorm Swiss Cyberstorm. All right. Security Israel launched an extraordinary attack on Hezbollah using a combination of supply chain and remote triggering techniques. And I've got a whole bunch here. You know, I

don't think I'm going to go into this. It's it's political and it's a bit long and we're already like pretty far into this podcast. So I think I'm going to let you go and read that one security researcher named X, y, z three VA found a catastrophic flaw in the Arc browser That lets attackers inject arbitrary code into user sessions using just a user ID. That is, that is nasty. So presumably I haven't looked deeply at this one, but it's like presumably each user has some

sort of guide. And then there's infrastructure that controls like the guides. And then if you attack that infrastructure, you're now attacking the user who's using the browser if they're in as that guide. That is just so gross. So gross. All right. Hacker named Adkar 72 424 has leaked a massive database of 3.3 billion unique email addresses or claiming they're unique. And he said, yeah, it's not dupes. It's not old stuff. It's a new captcha. And evidently it's 21.8GB.

We're thinking about putting this into Suckless, but that's pretty huge. Suckless is already massive as it is. Chinese scientists have figured out how to use Starlink satellites to detect stealth aircraft and drones, which are designed to dodge radar. Presumably, they're basically it's background radiation coming from the satellites. And if you look up and you detect the satellite radiation, the signals coming from the satellites, and you see things missing,

you see little cutouts. Well, that's where the thing is. That's my guess. Either that or there's bounce off the thing that produces like interference, and they're detecting that interference. I'm betting it's one of those two things. Thanks to Nudge Security for sponsoring nuclei templates. Version ten is out, including new Azure Config stuff. Google is making it easier to use Passkeys by syncing them automatically via Google Password and across Chrome on windows, Mac and Linux and Android

and iOS. Support is coming soon. Greynoise has been tracking mysterious noise Storms of spoofed internet traffic since January of 2020, but still remains unknown. They're still trying to figure it out, but it includes this love Ascii string in ICMP, and they might be like covert communications or DDoS coordination. Kind

of interesting. That's a man that reminds me of like old school, like ham radio stuff or CB radio where like you, you hear this signal and it sounds like a pulse or something, and then you research and you find out it's like some Russian submarine somewhere, or some like Siberian radio station. But there's all this digging and this is like pre-internet. So it's like really hard to find information. Then you find out, oh, Jim. Oh, you got to talk to Jim. Jim knows about this. He's

in whatever Saskatchewan, Idaho. And you talk to Jim and he's like, oh yeah, well, there's this one time. I love that idea of like digging up archaic knowledge. And I find this really interesting where it's like, Greynoise has a bunch of sensors and they're seeing this love Ascii thing and they're like, what is this? We've been tracking it since, you know, January 2020. We don't know what it is. We're trying to figure it out. I think

that's cool. And thanks to Threatlocker for sponsoring the podcast. All right, AI tech. So Sam Altman just dropped an essay called The Intelligence Age. I'm going to pull this thing up. We're not going to read it. But first of all, interesting website. I mean, it's just like super got this piece of art, which is no doubt AI art, I'm guessing probably a doll image. Who knows, maybe he made it, but it's like this super basic website. It's got a date. It's IoT. Sam Altman. Com and essentially

what he does. Well, actually, here's what we're going to do. Watch this. See, I just did, uh, command a and we're going to go back here and we're going to do uh, we're going to paste that in. So I did a create story explanation of the Sam Altman piece. So here we go. Society's collective intelligence has always brought us forward and I supercharges that. AI tools will help

us solve hard problems. Imagine personal AI teams of virtual experts helping us create anything shared prosperity where everyone's life is better. Deep learning is the key. AI will become increasingly capable. AI assistants will handle complex tasks like coordinating medical care for us. It will help us create better AI and accelerate scientific discovery. It's going to need a lot of computing power and energy. There will be challenges.

The potential benefits are immense, similar to solving climate change and even with job market shifts. AI will amplify our abilities and create new opportunities. So it's basically a rah rah for AI, which is not surprising coming from somebody running an AI company. But it was it was interesting. It was a good read. And, uh, yeah, a good summary there from Google. Gemini 1.5, Pro Dash 002 and

the create Story explanation pattern in fabric. South Korea's research institute has unveiled Deja Vu AI system that analyzes CCTV footage to predict and potentially prevent crimes. Yep, basically every sci fi ever preventing crimes. Like. I'm not saying this is not possible, but I think just because it's possible doesn't mean you should run full speed, you know, carrying scissors towards that possibility. Look, a lot of good can come from anything that's even starts out a little bad.

And the opposite is absolutely true as well. Like if somebody is an obvious, obvious criminal, they're walking like a criminal. What does that mean to walk like a criminal? I don't know, maybe it means like having a hoodie on and glancing suspiciously from side to side at three in the morning while carrying a crowbar in a neighborhood full of cars that get broken into at roughly 3 a.m.

with crowbars. Right? That would be an example of like a high correlation where, like, you have little much less chance of bias, much less chance of jumping to conclusions. It's like yeah. Oh, and also they're limping on the right side in this particular group. Like puts all the loot in their right side or whatever. You know what

I mean? It's like the more you add up the evidence or whatever from like actual data and not random things of like matching the clothing or matching the skin color or matching whatever, you know, human bias based thing, wherever you can have as much data as possible and really good analysis and retroactively and currently look at it and evaluate it and make sure you're not being like, biased in a, in a in a bad way. Sure. The problem is that's not where it will stop, okay.

They will people will deploy that and it will be cool and it will be useful to so-called predict crimes. But the problem is they'll be like okay, cool. So let's open up that faucet. And now suddenly we're detecting all people who look different in the neighborhood because they're in the wrong neighborhood. Like, well, people who look like

that don't ever come into this place. And then unless they're up to no good and some I could actually learn, that doesn't mean the good ones will, but there's going to be lots of bad and halfway bad eyes in addition to the really, really good ones. So I think any time you start predicting crimes and then affecting people's lives as a result of those predictions, like that's sci fi shit and it's not a good thing. Johnny IV has confirmed he's building a hardware AI device with OpenAI.

This thing actually might crush the others. This is what I was talking about, because the synergy between hardware, software and aesthetics is really important in AI gadgets. Like something you're going to have on you, something you're wearing, whatever. Who gets that better than the, you know, British Apple guy? Not many people. Blackrock and Microsoft are teaming up to raise $30 billion for AI infrastructure, putting it into $100

billion in investments, hopefully. And they're building data centers and energy projects primarily in the US. But this is a partnership between Blackrock, Microsoft and the UAE to build data centers and energy products projects in the US. Cool Canadian study found an AI tool can reduce unexpected deaths in hospitals by 26%. LinkedIn has quietly opted users into using

their data to trade generative AI models. I think they might already be rolling this back because there was so much backlash about it, and I think in the UK they definitely said no. So I'm not sure how much that's still continuing. But they tried. They tried and they did the right thing by telling people and they freaked out. Recent study by ring rover. No ring over found that 76.5%. So 77% of recruiters preferred AI generated headshots over real ones.

Pardon the drinking there or edit it out. Okay, a lot of Amazon employees are upset about the requirement to go back to five days in the office. Basically, people are freaking out. They're like, this is super lame. Whatever. I said my piece on that a million times. Sure, I'll say it again. Apple's iOS update has RCS support, which means you're going to get to do more iPhone like things with Android users with green bubbles. So typing indications, emojis,

stuff like that. It's all going to be a lot more smooth now that you're on iOS 18. Apple's A16 mobile processors are being produced in the US at TSMC's Arizona facility. Love this self-reliance American manufacturing push. Absolutely love it. Apple's iPhone 16 now supports wireless firmware restoration. I'm guessing Android has had this since like 2002. The Apple's Watch Apple Watch's Remote app now lets you adjust volume with Digital Crown, so it's basically like a better remote control

now for your Apple TV or whatever. And somebody shares his journey of turning blog content into an e-book. Facundo Solano. Turning blog content into an e-book. Humans Rick Beato argues that music is getting worse because technology has made it too easy to produce and consume. There's a new study showing that omega three fatty fatty acids can help reduce system symptoms of anxiety and depression in mice. So if

you're a mouse, huge, huge news for you. US Department of Energy is rolling out over $3 billion to fund more than two dozen battery projects across 14 states. Love it, love it, love it. Astronomers have discovered the largest black hole jets ever observed. So basically, the black hole on its axis is shooting out stuff. These are the largest jets they've ever seen, stretching 23 million light years, which is lining up 140 Milky Way galaxies. That is ridiculous.

Voyager one, my favorite like friend. Like, I don't know, I'll talk to this guy like, hey, we're still watching you. You're doing good. You're doing good, buddy. You're doing good. 47 year old spacecraft cruising through space since the late 70s. Fired up thrusters that he hasn't used in decades. I right here. Meanwhile, the asphalt on our roads has to be replaced, like, every 45 minutes. They don't make them

like they used to. Interesting piece in the Wall Street Journal about how pediatricians might have sparked the peanut allergy epidemic. So basically they they told parents, make sure they don't eat peanuts because they're make sure you don't eat peanuts because peanuts will give you allergies. Not eating the peanuts is what gives you the peanut allergy. I think that's what it's saying. I'm not a peanut doctor, but I'm pretty sure that's what it's saying. And I've heard this

from a million places. Ridiculous. Is that my new favorite word? Probably. Ohio is directly funding private schools with taxpayer money. I'm all for more structure in schools, but I think we need to be careful about how we get there, like not with a theocracy. Hopefully, Motus is revolutionizing wildlife tracking, using lightweight radio transmitters to monitor the movements of small flying animals like birds, bats and insects. Putting a radio

transmitter on an insect that's a big insect. 50,000 animals across 400 species since 2014. All right. Discovery reCAPTCHA fish. My buddy John Hammond created a fishing tool that mimics a reCAPTCHA form and steals your stuff. RGA rip grep on growth hormone basically lets you search through not only the normal stuff you could do with rip grep, because this is a wrapper for rip grep, but lets you search through PDFs, ebooks, office documents, and even compressed files.

So it's got a whole bunch of extras built in so that rip grep is just way better. In fact, I need to install that like right now. Anyway, I'll do that later. All right. Nuclei templates. We talked about that one Dune shell a new take on the command line experiment experience. Aim to bring a cozy customizable feel that bash lacks dam vulnerable drone drone hacking simulator sci fi ideas. Someone compiled a massive CSV file containing every

sci fi idea imaginable. Eli Burzynski Bendersky Eli Bendersky talks about building LLM powered applications in go fabric. Recently switched to go. We got a sick go developer now helping Jonathan and I out and yeah, it's going really well. That's definitely the next language I'm going to be doing. Everything in is all go. Asset node talks about their approach to recon, and Paul Graham's one pager on how to start a startup. Really like the format of this thing.

Look at this thing. Pretty clean idea people. What customers want. Yeah, quite nice ideas. Don't call them LMS. Probably the biggest idea that's exploded in my mind lately is Karpathy's point about LMS being poorly named. So he basically says Transformers are general purpose computer systems and that LMS are actually sequence predictors. And crucially, it doesn't matter what the stream is. We just happen to be sending language right now. But

you can actually send anything. And it actually that takes us right into the recommendation of the week. Start reframing your thinking about AI and specifically about llms away from just the next token of text. Instead, think of it as sequence prediction and then one level after that, think of it as answer prediction. So as Karpathy talks about transformer architecture works on sequences of anything languages was, you know, language was just a natural start. It works on whatever

you feed it. It's just a sequence thing. So the recommendation of the week is to update your model of AI from specific text predictor to generalized answer predictor. And in fact, that's my new favorite acronym for AI. I'm going to try not seriously to try to convince people to use this gap gap generalized answer predictor. I think it's way better than LM. It doesn't mean it's going to win it. You know, the best acronym doesn't win. It's the one that people just decide to use. And

the aphorism for the week people are strange. They are constantly angered by trivial things. But on a major matter like totally wasting their lives, they hardly seem to notice. People are strange. They're constantly angered by trivial things. But on a major matter like totally wasting their lives, they hardly seem to notice. Charles Bukowski. Unsupervised learning is produced and edited by Daniel Meisler on a Neumann U87 AI

microphone using Hindenburg. Intro and outro music is by zombie with a Y. And to get the text and links from this episode, sign up for the newsletter version of the show at Daniel meisler.com/newsletter. We'll see you next time.