UL NO. 450: Thoughts on o1-preview and the Path to AGI

Published Sep 17, 2024, 10:43 PM

80% Chinese Cranes, Drones vs. Abrahams, a RAG kickstart, a Canary-based Security Maturity Model, and more...

Subscribe to the newsletter at: 
https://danielmiessler.com/subscribe

Join the UL community at:
https://danielmiessler.com/upgrade

Follow on X:
https://twitter.com/danielmiessler

Follow on LinkedIn:
https://www.linkedin.com/in/danielmiessler

See you in the next one!

Welcome to Unsupervised Learning, a security, AI and meaning focused podcast that looks at how best to thrive as humans in a post AI world. It combines original ideas, analysis, and mental models to bring not just the news, but why it matters and how to respond. All right, welcome to unsupervised learning. This is Daniel Miessler. Okay. What do we have here? All right. This is, uh, when CSA makes you release a patch for your end of life product because the vulnerability is so bad.

I have abandoned my child. Say it. I abandoned my child. Say it louder. Say it louder! I've abandoned my child. I've abandoned my child. I've abandoned my boy.

That's so.

Good. Okay. Uh, fabric is now supporting OpenAI's new model zero one preview. So you basically just pass it. The switch R flag, which it sends the request using user rather than system, because O1 preview is not allowing system access, so you only have to send it in. User results aren't that much different, honestly, and you also can't send the temperature parameter which we were doing before. So this raw option allows you to use O1 preview with fabric and that works great. Got an insane cookbook use case for O1 preview, so this person actually used it to. I'm going to open this up to actually create synthetic data and then validate whether the data was accurate or not. Really really cool. Really really cool stuff. And just one of the things that you could do for O1 preview that you can't actually do with other models. Okay. So let's see here. Back to kickboxing I'll be kickboxing today, so hopefully it won't be dead later tonight. I got a connect on LinkedIn button here in the newsletter. So if you want to go and connect there, if you connect with me I will mutual connect. Okay. Last week's comments on current AI advances. So last podcast, if you're listening to this one, you might have already listened to that one, but I did a real deep dive on like the current state of like the path to AGI. I talked about the oh one preview. Um, and I just did a whole bunch of analysis, but you should check that out if you didn't listen to that one. And I wrote a new essay called the Art Quality Tier List. Pull this one up. And this is basically a way of rating art. And keep in mind this is for if you're a beginner, if you're already like an advanced art person, you already know art back and forward, then like you need to teach me stuff. I'm not teaching you anything. This is for someone who is thoughtful but is not an art expert. Even though I did take an art history class in college. But anyway, I don't really remember any of it, and I've basically just not understood what art was for the longest time. And I finally came up with a definition which I think is pretty decent. It is indirect expression of something that matters to humans. And I basically got a hierarchy here of like the different ways you could look at it. So going from not art to like proto art to like hollow art, which is kind of like AI art is one example where it might actually look pretty cool, or it might actually make you feel a certain way, but the artist didn't feel anything. And there was a little bit of a caveat here. If the artist feels something, but they're using AI to try to make the art, bring it out. So there's a little bit of a caveat there. And then decent art is where the artist feels it and the audience feels it. Great art is like, yeah, and people like it. And then S-tier or Brilliant is like, it's recognized as like a world talent or world class art and that's, that's basically it. So six different levels. And I even have a methodology for enjoying art, and this might be especially useful, especially useful for beginners. So start by feeling the piece. No thinking, no analyzing. Just take it in and experience how it makes you feel. Do that for a number of seconds or minutes, however long you want to take. Now that you have a sense of how it affected you, analyze the message you believe is trying to convey. Then, once you have the message, think about how it was transmitted to you. Like was it a song or was it paint or whatever it was? And this one's really important. If you're with a friend, do this in silence. Do one through three in silence for however long you want to spend on it for each piece, and then discuss your results with them. It's a really fun way to do this, you know, get a glass of wine or whatever it is. The first time I ended up doing this, I was doing it with my partner and with my friend Sasha, and we were doing this, I think it was the the Big Art Museum in New York, and we had lots of fun, like, we all felt like novices, but we did something like this methodology and it made it a lot of fun. Okay, so that is that newsletter just went out, by the way, and security. So US is evidently super reliant on Chinese cranes. This is freaking me out, particularly from this one company, Zpmc. And this report says 80% of US ship to shore cranes are owned by this Chinese company. I thought it was going to be like 25 or 50%. But yeah, I hope someone is actually tracking this stuff in a war room somewhere. And I'm pretty sure they are doing doing war room exercises. Uh, Fortinet has confirmed a data breach after someone called. I'm not going to say that word. Something naughty claimed to have stolen 440 gigs off of their SharePoint server. And they're talking about third party risk as being the problem. Uh, GitLab release critical updates to fix multiple vulnerabilities. Got 9.9 on the Richter scale here. Enable remote PlayStation with minimal user interaction and low privileges. Lazarus Group, which is North Korea, has been targeting Python developers with a piece of malware disguised as coding tests. They've been doing this for about a year now, thanks to material for sponsoring. Mastercard is buying Recorded Future from Insight Partners for $2.65 billion, and they bought it for 780 million. So nice ROI on that one. And what I noticed here is like this motion from okay, it's a startup that does this piece of functionality. Now we're going to bring it into a platform. In this case the platform is like Mastercard. So you have good ideas and execution. And their natural home is within some sort of ecosystem. And I think this is a good way to think of startups. It's like a petri dish for features that will live inside of something bigger, a bigger platform, bigger ecosystem or a framework. Security Canary Maturity Model is a framework designed to help organizations assess and improve their security posture by using Canary tokens. I love this concept of detection maturity model. I was just talking to dropzone about this actually. So here's like a percentage of your most likely minor behaviors that you're likely to see based on who you are. And you know, what percentage of these can you detect. And thanks to Defend defy for sponsoring as well. Australia is set to criminalise doxing with penalties up to seven years in jail as part of a new legislation aimed at modernizing the Privacy Act criminalizing doxing seven years in jail. I like it. I like it. I think, yeah, it does really come down to there are times that this happens accidentally. In fact, last week I was worried I accidentally did it, which I could have just rerecorded, but I caught it and realised it wasn't doxing so I didn't have to rerecord. But the point is, hopefully they're doing it for purely malicious people, which I'm sure they are. This piece discusses how AI powered autonomous weapons are changing warfare. So a bunch of Abrams tanks had to pull back after being targeted by kamikaze drones run by the Russians. Yeah, definitely. Go read kill decision by Daniel Suarez. This is the book. It is extraordinary. Uh, what? Did I read this? I read it in 2016. Yeah, really? Really good. Russia's naval activity around undersea cables is freaking people out in the US. Uh, they're worried that they're going to sabotage underwater infrastructure through a unit called, uh, goochie. Goochie goochie something. So this unit operates submarines, surface vessels and naval drones and has been spotted near critical deep sea cables. And these carry 95% of international data. US is drafting a New York joint statement to bolster security of submarine communication cables. Focus on excluding Chinese firms from the supply chain? Yeah. I think we need comprehensive critical infrastructure dependency analysis. And like I said before, I'm pretty sure this is already happening. I just hope we're using really good red teamers to emulate the capabilities of these adversaries. US House has voted to block the purchase of new drones from DJI. So much coverage of Counter China stuff lately. Seems like leadership is getting the message here, which is great. State Department has declared that Russia's state owned RT news has become a key player in the Kremlin's military intelligence operations. I mean, I remember a friend of mine talking about how awesome he was because he was hearing stories he's never heard anywhere before, and he's like, yeah, you know, the US has really messed up. And I'm finding out how messed up the US is from this channel. And I'm like, okay, what's the channel? And it's RT and I'm like, dude, you realize this is like, this is just a Russian state media, media, right? This is like this is pretty close to propaganda. And he started watching more and he's like, oh yeah, it was absolutely. But he didn't know he was being manipulated until it was called out to him. And this is like so long ago, this, this might have been like 2015 or something. All right. Siri flash. Best Chronos Dave is a civilian radio enthusiast who's become a key figure in Ukraine's drone defense strategy against Russia. Operating from a mobile intelligence center in his Volkswagen van, he monitors Russian radio transmissions and shares his findings with over 127,000 followers, including soldiers and government officials. I'm not sure I'd be giving this guy a lot of press. It is basically putting a bounty on him. I and tech newspaper had humans and AI create novel research ideas, and then had human experts rate the ideas to see which ones they liked, and they actually liked the AI ideas better. And I think this is the way to measure the abilities of AI, not with like standalone benchmarks, but like what you do is you have I results, you have human results and then you have a rating system which is humans. And later we could use AI for that. But in the meantime we have humans doing that. And then we essentially say, which ones did you like better? And they're blinded. Of course, they don't know which one is which. So when they pick the AI ones, that's a pretty good estimate estimation that the AI ones are better. And I think that's how we should do a whole bunch of these. Otherwise it's going to be a lot of bias. OpenAI releases their new zero one preview model focused on reasoning. We've talked a bunch about this one. Um, yeah. Got my thoughts on it so far in the podcast. And the big difference I just want to point this one out. That's why I kept the story in here. Big difference is that it uses chain of thought reasoning and the fact that it actually spends time and actually tokens, which means money thinking before responding. Because before it would be tokens in and tokens out, and that was pretty much it. Klarna's CEO says that I could replace enterprise. Enterprise software giants like Salesforce and Workday claims that conversational AI, like OpenAI's strawberry, can handle natural language commands to build custom apps. 100%, this is what I've been saying Sbca that's going to happen. And yeah, this is my Fpca piece, and this is what I think future software looks like you've got, uh, see, I can find the thing. There we go. State policy questions, action. I think that is the vibe. AI powered SAR satellites are now capable of detecting aircraft from space due to new radar tech, which enables real time monitoring of air traffic, good for both civilian and military uses. Cardio tech? That's a good name. Leveraging AI to tackle cardiovascular disease. Leading cause of death. Partnering with 65 hospitals to build a massive human heart tissue. Multi-omics data set to identify new drug candidates. Super exciting because the AI needs data to form its model of the world. And this is a whole bunch of that data to model it, because then you could figure out that's kind of the punchline here. You can figure out if drugs might work or that's the plan anyway. If you have a really good model of how the heart works, you can have drugs, potentially synthetic testing of drugs to see if they will work without doing trials or without doing as many trials, or for as long or as expensive. Salesforce just launched Agent Force, a suite of AI powered agents designed to enhance human workers across various business functions. Yeah, yeah, definitely designed to enhance human workers. Not designed at all to replace them. Yep, yep. Waymo's latest data shows that human drivers are responsible for most serious collisions involving its driverless cars. 16 out of 23 severe crashes being rear ended by human driven vehicles. Tesla Cybertruck is spiking in the truck segment, while electric truck segment with 61% of sales surge in July, a surge of 61% in July did way better than the R1t's and the Ford F-150 lightning. I think we didn't. We just see a story saying the lightning was crushing it. I don't know why there's a big jump right now, but I have seen a whole bunch more very recently. Anecdotally, USPS has rolled out a new delivery vehicle. Let's see if we can get a picture of this thing, which I saw. Yeah, look at that. Looks like a duck. Looks like a duck, Bill, doesn't it? Yeah. Evidently the drivers love them. They're very, very practical, but they look very strange. Dmitri Greenberg has managed to run Linux and Ultrix on a business card, turning it into a tiny computer, eight kilobytes of Ram and 32kB of flash storage. New studies showing that debunk bot, an AI chatbot can effectively persuade users to abandon conspiracy theories. The bot made significant progress in changing people's beliefs, challenging the notion that facts and logic cannot combat conspiracies. I mean, this makes sense to me. So this is what I wrote you. You can convince. If you can convince somebody of something, you can convince them as well. This is why I'm optimistic about having AI on us all the time, 100%. It can be Orwellian, or it could be a defender or protector or tutor or coach, etc. that's always with us. Community college had to cancel its CS career fair because no companies reached out. Super sad, but super expected. If you have people coming out of college with four year degrees and master's in CS and they can't find jobs, what do you think? Do you think companies are rejecting? Right? The companies who would come to the job fair, they're rejecting the master's degree. People, but they're going to go to community college. I don't think so. This is why we need human 3.0. Future is connecting directly to individuals not going through these credential forms. Google has officially killed off cache links, which is sad because that's how I used to see old versions of the site. And you can see the site even if it was down, like, whatever. United Airlines is partnering with SpaceX to bring free Starlink Wi-Fi to all of its planes. I'm a united person. I'm so happy that I'm a united person right now, man, because I fly JSX as well and the Starlink is so good. It's better than a lot of my friends. Wi-Fi at their house. Ukraine just launched its biggest drone attack on Moscow yet, hitting the region with 144 drones. I still don't understand how Ukraine could possibly be winning this war. Completely insane to me, but in a good way. Sweden is increasing how much it's paying migrants to go home. It's now up to $34,000, roughly paying them to just leave NASA's advanced composite solar sail successfully been deployed. Its ultrathin solar sail is now in low Earth orbit means it's visible in the night sky from various locations. It could be as bright as serious. Sirius is the brightest star out there. The dog star. It is very bright, so I can't wait to go find this thing. And I wonder. I mean, it would be crazy to see something as bright as Sirius actually moving, so I can't wait to see this thing. Got a comment? They're calling it the comet of the century. I feel like these astronomer, uh, journalists are always saying it's the brightest. You will not see another solar eclipse for 900 years or whatever. And then it turns out, well, not one that comes directly here. Not one that falls on a Friday. Not one that you could see while standing upside down or whatever. And then you look and you're like, okay, when is the actual next eclipse? Oh, nine months from now or whatever. Anyway, this is going to be a cool one. It's going to be bright. So I'm going to try to remember these times. I'm sure they'll say more about it at the time so I can go take a look. US is closing a trade loophole that e-commerce people like Timo and Shine, shine, Shine have been exploiting. Loophole allows them to ship goods directly to customers without paying tariffs. So the US government is going after that. There's a leaked PDF that details Mr. Beast's unique company culture, and it got a one page summary of it here. Pretty, pretty smart business stuff. Although he is, I think, being investigated for some some stuff. See here. Uh, person says sunlight cured their migraines. It's not a study. It's really just somebody talking about it. But I figure most people who have migraines have tried everything else. Why not try this? It's basically sunlight getting lots of sunlight, I think in the morning. Love this concept from Laura Hogan. Be a thermostat, not a thermometer. That's that's cool. I mean, it's cool if you don't even read the article. Be a thermostat, not a thermometer. Content driven development is a strategy for making progress on side projects by focusing on small, shareable pieces of work so you finish a thing you share, you finish the thing you share. In 1913, Vienna was quite a place to hang out. Hitler, Leon, Trotsky, Tito, Sigmund Freud, Joseph Stalin all hanging out in the same city. And by the way, this is before Hitler. And I guess Stalin turned evil. And whoever else in that list was evil. I'm not sure how evil Trotsky was, because that's how poorly read I am about Russian history. Discovery. Merkle map CLI command line tool lets you search and enumerate subdomains using the Merkle map API. Okay, 71 terabyte zfs NAS 24 four terabyte drives. It has lasted without a single drive failure for over a decade, thanks to a strategy of keeping the server off when not in use. Cheating. I feel like it's cheating. Doctor Mordechai Guri has unveiled a new side channel attack called Rambo, which uses radio signals from a device's Ram to exfiltrate data over Airgapped networks. I was like, I know this is Israeli. I know for sure, I guess Tel Aviv, but it was actually a different Israeli university. But yes, it was Israeli. Israelis are the side channel goats. Six techniques I use to create a great user experience for shell scripts. This was a great read. User friendly shell scripts with techniques like comprehensive error handling, colorful output, and progress reporting. These things really do make a difference. They make it feel like an actual app. Two people made a tool that lets you transfer playlists between Apple Music and Spotify and YouTube music. Those are kind of the three big ones. Nothing is a timer that celebrates the art of doing absolutely nothing. And how to create a rag pipeline with pinecone and Semantic image search CLI ideas. I love it when experts completely disagree about a really important thing. Like for example, I have no idea if China is thriving right now and is about to take over the world, or if they are about to have this demographic collapse and their economy is tanking, and they've got the housing crisis and they're just in major upheaval, and she is facing pressure and time is running out. And they have brain drain going to the west. And they're super screwed because they're slowing down AI and they're going to get passed up. I don't know if that's true or if they're about to invent their own chips. They're about to not need Nvidia. They're about to take over everything because the West is stupid. And I could find the world's best experts who tell me both of those things and tell me that the other set of experts is completely wrong. I find that fascinating, and I've read a whole bunch of books. I've read probably 12 deep, like large books about China or dedicated just to China and its development, probably the most important one being like Chip Wars. Although that was more about chips than it was about China. But it definitely talked about China a lot. But I read a whole bunch that were dedicated just to China. And I mean, it didn't tell me the answer. I mean, I have all these opinions and I listen to these opinions of, like, the smartest people, but the fact that I still don't know, after reading all those books and seeing all the data and listening to all these smart experts and I still don't know, that is exhilarating to me. I'm like, wow. I mean, I love the idea of like and I'm happy with myself for not like picking one and just saying, oh, this is what I think is going to happen. I tend it's like strong opinions, loosely held, you know, and the recommendation of the week actively guard against age related lock in, which starts around 30. I'm guessing I made that up, but roughly around there, maybe even late 20s. And for some people it doesn't kick in. I've been very resistant to it, although I am writing this also to myself, because you actively have to fight it. Listen to new music. Read new books with new ideas in them. Talk to new people. Go to strange restaurants. Try new foods. Don't let your experiences reduce into a tighter and tighter death spiral. Variation keeps your mind young and the aphorism of the week. Choosing not to read great books has the same effect as not being allowed to. Choosing not to read great books has the same effect as not being allowed to. Unsupervised learning is produced and edited by Daniel Miessler 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 miessler.com/newsletter. We'll see you next time.