21:02
Uh, no, you haven't tuned into the wrong channel. I'm I'm doing a traveling visiting thing and I'm here with Alex and Alex does a lot of really cool AI content, like a lot. Like a lot more way more AI content than I do. And it's so cool cuz he's got all the toys. I may have the toys, but I have Wendell visiting me for the knowledge. No, I'm just I too am an internet rando, but we can go on this journey together. It's a lot of fun. We're going to We're going to chat. We're going to talk shop, but it's a little bit meta because it's not about the AI, it's about the game. For love of the game. I don't know.
AI has come a long way in the last year or two. You're making a lot of content. Yeah. A year ago, would you have ever thought that we would be where we are now? I mean, no, not here. Definitely not. Uh, it's it's growing so fast and the amount of work that I get done now with AI is has exponentially grown. It's not only the amount, but also I think the quality. It's raised the quality bar for me and for my audience. Uh, you know, I run a YouTube channel also, just like this. So, it's like uh, it helps below. You should check that out cuz it's cool stuff.
Yeah. Thanks, uh, it's it's just the productivity boost is incredible and I'm thankful for it. I'm glad it's here. I've been uh, happy with the results where it's like, I did this work. Help me look over this work because I think there's a mistake and then it spirals or it's like, oh, let's do this or this is a different way to Oh, I learned some stuff. This is really good. I think that's a really effective way to use some of the tools. Um, and so yeah, it's grown a little bit into my workflow as well, just looking at uh, how it helps with just doing real work and that's on and off camera. So, Yeah. it's it's sort of wild. What are some cool stuff the
projects that you got coming up? Do you have anything you want to plug or anything you recently done that's like, "Oh, you got to check out this video. This is an unexpected result." I think uh some of the coolest things I've tried lately have been large machines. I just built a cluster of uh four RTX Pro 6000s. Um and just experimenting with bigger and bigger hardware and different kinds of hardware. So, the way models behave on something like this. This is the Dell Pro Max. It's a GB10 unit. And it's like the DGX Spark, basically. It's very
different than the way it would behave on something like this. So, just getting the getting my mind wrapped around that stuff is next to impossible if I didn't have help from AI to do it because I come from a web development background and mobile app development. So, I have no idea what I'm doing when it comes to AI, but I'm learning as I go and just running it on different hardware allows me to dig in, see what works, see what doesn't. It's really cool. It's It's been my experience as well. We just like number formats. It's like, "BF16." And it's like we both got time with a DGX Station. You know, that's a GB300 standalone desktop machine. And again, it's like it's mind-blowing that
that is a machine that just sits on your desk and it's really limited by just like the North American power outlet. Right. Which is sort of nuts. That's where we are. So, North America, you have to keep up and upgrade all your outlets. We got to go to 240. We got to have 240 everywhere. Europe is laughing at us in British. They turn their kettles on. It's like, "Our kettle uses more power than your DGX Station." Not untrue. So, what are you working on? Uh just more of the same. We're We're doing We're getting ready to go into Devember. So, like every Not every year, but fairly often we do this thing called Devvember where it's it's it's mostly
for students, but also people that want to do development stuff. And it sounds like it would just be December, but it's really about 3 months. And you do a dev project. But this year it's different cuz AI and AI tooling. We've already had, you know, we've had a lot of posts on the forum that I would say are maybe like AI psychosis where it's just like somebody's talking to the AI and it's like I want to build this thing and it's like okay, that's great, but and the AI is telling them all the stuff, but it's the AI whispering sweet nothings in their ear and it's it's not a good situation because the AI is kind of like glazing and it's like you got to very gently let them down. And it's like why So, this year's Devvember probably going to have a lot
of that, but it's going to be let's build the awesome home lab unplugged thing. So, it's like while the tokens are cheap, let's see how much how many services we can get to run offline and how many durable objects that we can have that exist from now until the end of time. Like let's build some software to help us run things offline. And so I'm getting I really like Devvember cuz that'll also run on like potato class hardware. It's like we don't need good hardware. We don't need anything. We can just build this on potatoes while everybody in their ivory tower is off doing other things.
Like Raspberry Pi class machines. Yeah, yeah. Well, Raspberry Pis are too expensive now, too, but we want one of those cast-off, you know, six and seventh gen Intel business class desktops that nobody wants. It's all we [snorts] need. I mean, yeah, this you know, you can build software for old hardware, too. And yeah, the older something is, the more knowledge, I think, the LLMs have about it, right? Doing something that's on the brink of discovery is always harder for it because it doesn't know what it's doing. So, you kind of have to uh cross your fingers and hope that it works. But stuff like Raspberry Pi, stuff like Arduinos,
LLMs can crunch through that like anything and that's a good place to learn especially for people that are starting out with development, too. I really think that if the powers that be understood how much power an individual already has running even something that'll run on a Mac or a GP10, if they really understood how powerful that was, especially in the context of older hardware, they would be racing to outlaw AI. I give you a really good example. It's like here's a set top box that doesn't, you know, talk to a TV [clears throat] or HDMI or anything else anymore the way that it should. It's really just a software or firmware problem. These
models are so sophisticated now that you can just hook it up to the model and say, "I'm having trouble doing this. Here is a firmware file from this company's website where you can upgrade the firmware to fix this problem, but it doesn't fix all these other problems and they've abandoned it. Let's reverse engineer the firmware and turn it into something else." And like you can actually do that now. That's insane. that's just Yeah, I feel like you know, even just writing applications that I would normally just to go through my day to help me with little things here and there, Yeah. I just let AI do that now and it doesn't matter. It's not production level stuff, but it helps me get through my day, which is just incredible. It's
like so much help. It helps me visualize things. It Yeah, I could be glazing AI for a long time, but I'm just happy it's here. Uh let's let's flip that around into turn like software [clears throat] as a service. Is it like is it going to be apocalyptically bad for companies that are selling software as a service? I think so for exactly those reasons. Putting a little different spin on it cuz again, it's like if people understood how powerful this is, they'd take it away. A lot of the like I'm going to pick on salesforce.com a little bit. A lot of the reason for the expensive salesforce is not the
thing that it does, it's that here's the design patterns for good CRM. And good CRM really does drive business. So, there is value there. But, a lot of the complexity and machinery of that depend on all the customization, all the web stuff, and all the custom workflow, and like building a system on top of the system, which is an order of magnitude more complexity than just take these LEGO bricks and assemble it into a tool. So, like you know, getting through your day with AI assistance to put something together custom in Python or shell script or whatever, and it's like you might not understand all the minutiae.
You know, that's too much of an ask to ask you to do that, but you also don't need something as simple and easy and intuitive. I say intuitive, but like what Salesforce offers, you could just build the thing to do it in another language, and it'll get the job done, and actually be less complex as well. what happens to all the Salesforce consultants at this point? They're going to become AI consultants. Ah. Yes. So, people that are resisting AI, what are they really doing? I don't know.
Uh you know, it's There are aspects of So, Cuz there's some people really fighting it, you know? That is like Take those same consultants, right? They are really destined to lose their position. A consultant gets paid a lot of money uh to do things like, I don't know, do a SharePoint plugin or a Salesforce plugin. Yeah. Uh but now well very often. Yeah. Yeah, I've seen some of the code. And I've written some of that code, unfortunately. Sorry about that. Hits a little close to home. Sorry. My bad. But, you know, you can either embrace that the AI world, and as a consultant be able to now use AI to help you get that job done faster, get more clients, your client jobs done faster,
charge the same amount of money, huh? There's not a moat there anymore though, which is the good thing. It's actually good for both of you. There's not a moat that is there to protect the turbo sales and as a customer, there's not a moat that keeps you locked in. It's like this is your business process. If you know what your business process is and the work you're trying to get done, you can just take the work, [clears throat] move somewhere else and there you go. And I think that's a large part of the content that both of us do is like you have to think for yourself. Don't abdicate thinking. And you put the thing together and then you can do the stuff. And that's very empowering and that's
very awesome. And if those sales people do that, then they're not going to have any trouble finding employment. Exactly. And I just want to add to that, don't abdicate thinking where it matters. Cuz I abdicate a lot of my thinking these days to things that don't matter, which is great. That's the thing that you want to do is get rid of the things in your brain that don't matter because normally you'd have to do all those things, but now you don't. That's a lot of the videos that we've done in the past is like here's my how-to, but then here's all the suffering that I went through to give you this how-to. That, you know, a lot it since the time immemorial like the system administration part of that
was called yak shaving. Yeah. It's like oh, I have to go and it's like I just want to do the thing. And so the AI can make a lot of the yak shaving stuff go away. So they're like conceptually it's like this is the design pattern that I need for a secure login. This is the design pattern that I need for handling this data. How is the best way that we accomplish this? That's where the tooling starts to get really interesting. Or even taking a step back, what is the right design pattern that I should implement this thing in? Or give me four design patterns that would fit this scenario best. Yeah. And if you know the answer that question or ask that question and get an answer from the AI, you have leveled up your understanding.
Exactly. So you kind of entering into it at the level where you are and learning on top of it, but also having some help doing that. So it's like very targeted learning I'd say. Yeah. Yeah, if you treat it that way. So I you've So from my own personal experience for the last couple years, I never thought that I mean basically like Star Trek nailed what the computer experience was as science fiction. It basically has evolved to be that and it's been surreal uh sort of living and experiencing that, but you know, I don't know what's what's your experience been with the evolution of all this stuff? Cuz we have hardware that is not for mass consumption, not yet. It's too
expensive. It is too unobtainable. That will change just like electricity. It's like, "Oh, we don't need electricity for lighting. We have natural gas lighting." And it's like, electricity's probably better fit there, but yeah, I Like you said, Star Trek, good example. You walk around the decks of the Enterprise and you're saying, "Computer, do this. Computer, do that." Why is this like this? As you're walking, right? You're walking to the turbo lift and you're talking to the computer and the computer is answering you and doing stuff. This is where we are right now and it's not even the 24th century.
Yeah. Yeah, it's sort of nuts. And like all things, we're going to make lots of mistakes. Those AI psychosis people, those are also in executive positions as well. It's like, "Oh, the AI says I can just replace all the people with this." And it's like, No. Right. So There are people complaining that AI is losing them some jobs, right? Yes. And that's very unfortunate. Um There's a lot of displacement of technical jobs, but Yeah, [clears throat] it is very unfortunate. So what can these people do? Um learn AI? I think it If you have time and it's not really learn AI, it's look at the technology and look at the problems that people still have to solve and then start solving those problems. There's never been more entrepreneurial opportunity, I think, for
But you got to have that kind of a mindset. Yeah. There's the entrepreneurial mindset that some people have and some people don't. But the people that have it, they'll they're going to dig into it and learn it. The people that don't have it need to work somewhere, right? But they can all actually level up while they're waiting as well. They can level up by digging into the technology that exists right now, as well. But anyway, I was getting to how did you feel about AI when it first started? Cuz I was deathly afraid of trying cloud code when it first came out. I did not want anything controlling my computer or giving it permission to do that. Now, I'm just yoloing everything.
I'm yoloing more than I probably should be, but I'm also not sending a lot of things to the cloud. I've taken a lot of steps to erect barriers for what the cloud models can see. And so, like, if I do have a problem that I want to ask a cloud model about, and it has sensitive stuff in it, I have other models that are creating, you know, sanitized work-alike versions that are submitted to the cloud. Cuz I just don't I don't trust that's not being added to a training set. I feel bad for those mathematicians. Like that one mathematician, um was it Navier-Stokes? There is a compelling case there that his notes, where he was working on it, was used in the training set for the next model. And the next model completed
the solution, but there is some uh some strong evidence, I think, that person that had started that had a pretty good start on it, and the AI finished it. But would the AI have gone down that route without that guy's notes? So, it's kind of like, what happened there? Why doesn't that guy get credit for it? So, is this the recent math problem that wasn't solved in many years that was solved? Yep. 70 plus years or something? Yeah. But there was a mathematician that had a solution that was really close to what AI came up with.
Okay. And the AI kind of grabbed that solution and worked on it? The evidence suggests that, but we don't know. We don't know for sure. Yeah. Well, too bad for this guy. I think Terrence Tao is also another famous mathematician and he sort of had a good [clears throat] write-up on it. It was like this these are really powerful tools, but sometimes the AI can give us a solution that isn't useful for advancing the state of the art. It's like, well, it gave us an answer and that is new knowledge, but usually on the path to a new answer, we have all sorts of new mathematical tools and now we don't like we have a solution, but we
don't have any new mathematical tools. And so in that way, the mathematics thing has a kind of been robbed because they don't have what they would normally have from a normal process. So I think that will also evolve. I think eventually it's going to be Vulcan education pods. I'm going to have my own AI that is like, all right, let's learn the equations. Like how well do you know the equation? It's like how good is your understanding of this? Cuz, you know, when I whenever I've I've had to teach students or, you know, help somebody learn some kind of a concept, it's you know, you ask them a bunch of questions designed to really test how deeply they actually understand something and I think that is going to be the really amazing thing that is an unlock
for humanity for AI cuz it's like, you have to see this in your brain because there's parts of your brain that the machine doesn't do. Mhm. Yet. Unless we learn these concepts, the complex things and keep learning them and keep our brains active, it's just going to atrophy, right? Yeah. So that's the unfortunate thing. That's the balance we have to kind of figure out. We're still too early, I think, but I can see that happening even to me, you know? I'm I'm already getting kind of lazy with certain things. So, um
But this is also nothing new. I mean, think about everything all the way back to the beginning. Before there were computer scientists, Yeah. it was an extension of math and mathematics. I think about like Turing and the halting problem and those aspects of things. And it's like [clears throat] that is more a mathematical exercise than a mechanical exercise. It's an abstraction thing. We've been abstracting all throughout history, right? You're talking about mathematics, then we are talking about low-level languages, then higher-level languages, and we don't bother learning lower-level stuff nowadays. We're just using the higher-level stuff. This is kind of like another abstraction layer.
Yeah, think about how much labor would be required if we still did all the work at machine language. It would It's an obscene amount of language. One. Zero. 001. No, thanks. Yeah, it's not great. But also, you know, as a web developer, I'm sure that you know, it's like I'm doing this you know, this login form for the 752nd time. It's like, "Oh, you can use a library for that." And then it's like the library doesn't do enough for you because you need a little bit of customization. And then you have people to come up with frameworks, and then the framework is, you know, it's like, "Are you all in on the framework, or are you
bolting things on the side of the framework?" And a lot of those, you know, tech debt complexity questions are now kind of out the window because you can have the best of the design pattern, the raw metal, no framework, but you still followed the design pattern that is like computer scientists generally agrees is the best pattern perfectly. And so it really is the best of both worlds in terms of low code overhead, low maintenance, this kind of thing. Also, security. I think that there are a finite number of security bugs. And like right now, it is uh a bloodbath in terms of security bugs. But that is the last 30 years of technical debt coming due from human review and everything like that. I don't
think the surface area for these security issues is infinite. You don't think so? No, because the as the AI finds and closes all these holes, I think the security posture overall is going to be better. It's going to be better if things stay static, but I would tend to disagree and I'd say that AI is going to uncover new security holes that you're going to need to patch. And I think it's just going to keep growing and growing. I think the only way that the security issues grow is if the AI code itself grows in an unchecked and non-deterministic kind of a way. But once the AI gets to it's like this is going to happen.
Well, I don't know. I mean you can't stop it. You can't stop It's too late. You can't stop this. You get to this design pattern for a login form and [clears throat] it's like we can do this. You know how I think we actually solve it? I think we actually get to formally verified compilers. I think that AI is so good that we will get to a point where it's like the language and the way that this is implemented is formally verifiable. So if I build my system in this then we'll we'll be able to do that. Um But who's going to follow all the rules? The AI would have to do it. Well, the AI would have to be told to follow the rules, but not everybody's going to tell their AI cuz now everybody can train their own model. You don't have to follow the same rules that
everybody else. You can get into the kinds of problems with machine trust that you know, Ken Thompson talked about in reflections on trusting trust, which is an amazing paper by the way. Um and it's like do you trust the AI or the compiler not to intentionally insert a back door? I think some of the security issues that were uncovered are actually intelligence services that snuck in very subtle flaws in cryptography and things like that and AI is like oh there's a problem here. We should fix this and it's like is this an innocent bug or is this an intelligence service very placing a very careful very subtle bug that almost nobody would catch. So maybe the tooling gets like that. Maybe that kind of thing happens, but
in a scenario like reflection of the trusting trust where you need multiple compilers all agreeing, I think with custom training and custom models and being able to ablate the weights on your own and do that kind of thing. I think it's going to expose more of those kinds of issues in a way that was never previously possible before. It's like, "I got to build my own C compiler in order to do this." Mhm. And it's like, "Well, you can take this model, this model, and this model and they can all agree on a C compiler. We can check and see if the compiler's got its own back door."
Maybe. Yeah. I mean, sure, of course. But, I think more holes and just going to keep opening up and it's just a never-ending avalanche. I have a another [clears throat] question for you. Can we switch topics for a second? Benchmarking. I do a lot of benchmarking and I'm not talking about benchmarking models, but what are your thoughts on benchmarking agents and harnesses? So, this is really this is something that I've struggled with and one of the best things so far that the community has come up with, thanks three, is power tool. And power tool is point-blank looking at the models and the how the model runs because you've run into the same problem
that I've run into, which is when I'm benchmarking it, it's like, "Here's my tasks and I want you to accomplish this task." And it might be build a breakout clone, it might be find a bug in this repository, it might be this, that, or the other. And if you're benchmarking like that, then you look at it with like an NVFP4 model, four-bit, it's like, "Is this really as good as like a BF16 model that would run on something fire-breathing like station GB2 GB300?" And you look at the results from that type of benchmarking and it looks really good. It looks like it's able to Oh, this four-bit model was able to accomplish this task. But, if you look at the actual tokens predicted, the
number of tokens that it took, and the time, and that kind of thing, you see a lot more signs that those lower quants really start to struggle. There's a lot of people that are running, you know, like two or four-bit quants and the quality output of those models it generally is nowhere near as good. KV cash quantization also can have really severe impacts on the output. And I don't think anybody is doing the qualitative analysis on those sides of the benchmark. It's just quantitative. Yeah. It's like how many tokens per second? Yeah. Not is the result valid or not? Yeah. Or the result can be valid. Like these little Pro Max GB10s like produce killer breakout clones. But if you look at like if you step through it
and it's like this is you know, we went through a million tokens to get to this really amazing breakout clone. But you look at point blank every step of the way and step through everything that it did. You know, it did weeks of work in hours. But you can tell that the model is not like it struggles a little bit and it's like oh, I made a mistake. Let me go back and try that again. Or oh, that didn't work. Let me try that again. And it's like oh, we turned reasoning on. Reasoning really helped this versus one shot. And the harness makes a big difference in those kinds of things. So like the second part of your question is also really important because Nvidia has come out and said oh, even a really dumb
model with a really good harness, you can get a great result from. And it's because the harness forces things like scripted tests and uh scripts that will run through and do an evaluation. Or like code linting. Now the AI is being subjected to the code linting just like people. It's like you told it to do that. And it thought it did that. But it was actually an hallucination which may or may not be related to quantization. And so you have all this machinery to stop it from doing that. But in doing that, you also spend more tokens to have it do the thing whereas a smarter model with deeper context and everything in BF16 might have been smart enough to avoid those things to begin with. And so like that kind of qualitative analysis,
it's frontier. You and I can probably design and set the standards for that because nobody's doing that. And that's what PowerTool is meant to do. And it's a it's just toe in the water. It's literally like baby's first qualitative analysis. And in so doing, it melted my brain because it was like wow, this is not what I expected at all. I got to try that. I haven't seen power two yet. I'm going to have to try it out, but I did start working a little project myself. Um [clears throat] work in progress, but I'll show it to you later. Okay. When it's done. Or it's closer to being It's a work in progress right now. I don't have like a thing set up, a demo set up.
Um but it's a it's basically a way to um to analyze the work of agents and harnesses and also validate the work. So it's like a quality and quantity. But there's a middle step also. So right now there is definitely concrete ways to measure tokens per second, obviously. There's ways to measure quality of the output. But the benchmarking that's in the middle, not agents yet, just in the middle, just analyzing models is something that takes into account both the speed and the quality. So that's that's one step. The next step after that is the agent work and how to analyze that. And I'm working on that as well. So maybe we can uh Yeah. collab.
There's there's literally like you just unless you are working with a specific workload, you don't know. And that's what led to the work. It's like okay, I want to actually use this thing to solve this problem and it's like this is just an hallucination. The AI is glazing me and telling me that my ideas are amazing, but like this code is crap. It's like stop glazing me. We just need to get the work done. Um That's a very good idea. We should do that. Let's get on it. Where do you think we're going to be in a year? Where do you think we're going to be in five years?
Oh my gosh. If you asked me a year ago where I we'd be in a year, I would be way off. So uh don't ask me about the five, but I imagine we'd be flying cars. Probably. Um in a year I'd imagine we're just growing exponentially and the open source models looking forward to those being better and better and being as good as the frontier models right now like Claude and GPT open source. Yeah, I think open weights open source this whole legal or whatever framework to control the AI. No, that is regulatory capture. That is not in your best interest. Like whatever it's like even if you have the most benevolent person in charge of the government
sanctioned AI, I promise you it will not end well for the individual. Yeah, and I don't want to get political here, but like how come China's outlook on AI is 93% positive? Yeah, versus the US. US is like 36% and China is government run. So, how is that possible? How is this happening? China's government run, but they're also you should release the weights. And remember Quinn didn't release the weights for one of the their models and then and then the Chinese government was like, "We think you should release the weights because this is trained on the stolen corpus of all of human the works of humanity." And they said, "Okay dokey."
Meanwhile, you know, we've got Anthropic and OpenAI saying, "Well, yeah, we did steal the works the collected works of all of humanity. But, we really think that this should be government regulated." It's like I think you know, IBM I think IBM is the only company that has model that they've trained only on works that they are out of copyright that they have copyright agreements you know, a license to the thing and their lawyers will indemnify you. So, like if you sign a deal with IBM their lawyers will be in your court should somebody sue you for the contents of what their model produces. Huh. What's the name of the model?
Uh I think Granite is one of the Granite. That's right. So, it does They make smaller models though, right? Yes. Well, I think there's some larger models that are available on the enterprise side, but I thought it was very interesting that was interesting take, yeah. So, I don't know. Also um I think about the future and I worry about Einstein's quote which is I don't know what you know the weapons or something like the weapons of World War you know are probably unimaginable but I can tell you what the weapons of World War will be fought with and that is rocks and sticks. Ooh. Oof. That Einstein always come up with clever stuff.
Yeah it's just like okay in a year but five years is like I don't know. I think we'll figure this out like you know I have faith in humanity to figure it out. I do not have faith in leadership but I have faith in us collectively and I have faith in us collectively when we get off our butt and say look this is lunacy we just need to stop this and do what's reasonable. Doing what's reasonable and that's uh the words we should live by. I guess. I don't know. It's fun. Well thanks for having me out and thanks for letting me tour your studio and have fun. Welcome anytime.
Thank you. Good cool.
