AI Fireside Chat (sans fire)

Welcome everybody to ⁓ the third episode of the Knot Brothers podcast. Today's topic of conversation is just gonna be ⁓ a fireside chat about everybody's favorite topic, which is AI. We've been getting asked consistently about AI, ⁓ especially with a lot of the new agentic stuff ⁓ with OpenClaw or Claudebot or Multbot, which are the one you remember, ⁓ and all the other things that are happening. So we figured... Today would be a good opportunity to just sit down talk about, you know, kind of what the landscape looks like today, how we're seeing benefits, challenges that we see in the marketplace. ⁓ And ⁓ that's pretty much ⁓ it. think ⁓ it'll be interesting to look back on this ⁓ as well, because if I think back to ⁓ just three months ago or a year ago where we were when it related to AI, if we had this exact same conversation a year ago, everything we're about to talk about today is ⁓ probably fundamentally different, which I think is really cool.

fundamentally different and in many cases irrelevant because these things just didn't exist then. ⁓ Not just the frontier models themselves, but like what you can do with them, how you can extend them, how you can leverage them together ⁓ to work on things. That's been some of the fascinating things that I've seen happen in the last, really call it six months or so in terms of how you build those things on top of each other.

Yeah, mean, the agentic models ⁓ and agentic workflows through things like open code, Claude code, ⁓ AMP, all those, they kind of shifted the market, right? ⁓ When we all got started with AI, it was kind of like you go to a chat. Many people are still at the chat level, right? You go to ⁓ chatgpt.com or you go to Claude.ai. ⁓ You type in a box, maybe you upload a file, you get a response back, and that's kind of your work loop. Eventually we started tapping it into other tools, you get ⁓ access in VS Code or in NeoVM ⁓ directly, and then eventually you had this whole agentic thing that's happening right now, where we're like, happens if ⁓ you can give it tools, or what happens if you can give it a memory bank, or what happens if you give it access to... ⁓ different things and allow for a little more autonomy. ⁓ And I think ⁓ the key thing that we're seeing right now is an evolution of that, ⁓ the agentic piece running even longer, right? One of the things that we found early on. with agents, I definitely found it with even toying around with them, is they ask me too many questions. ⁓ So Anthropic and OpenAI, even in the latest models that they released, ⁓ have intentionally ⁓ made them longer running models. They're optimized to run longer and ask fewer questions. So you could potentially deploy a model and say, hey, work on this task for hours. and come back to me when you're done. And it'll just keep looping for hours and be able to actually accomplish some sort of long-tailed task that previously would have been ⁓ a whole lot of chat inputs ⁓ or a lot of stopping and asking you for permission to proceed.

That's kind of a game changer. When you think about ⁓ layering agents together and layering agents that have increased levels of compute and able to take more and more information and do more and more with it. And then you layer that on top of each other. And then you sort of give it a master puppet on top of that to say, hey, please summarize all this on my behalf. It's ⁓ pretty mind blowing what is possible in a very short period of time if you have the tool set up in the ways that we're kind of describing here. And they're getting easier and easier to set up ⁓ for more and more people that aren't hardcore frontier model people like you in particular. ⁓

Yeah, I mean, think it's ⁓ both exciting and terrifying times right now, right? ⁓ I think there's a lot of really cool stuff happening. ⁓ I was just talking about it yesterday that like, ⁓ know, yesterday we got two frontier models dropped in the same day, right? Opus 4.6 dropped from Anthropic. And then about 15 minutes later, we got Codex 5.3 from OpenAI, which is incredible. ⁓ Just like two days prior, we got a new Quinn 3 model, which is Alibaba's open source model, I think. It's running on my laptop and, you know, it's like a very capable model, but can run on a MacBook Pro. And then just last week, we got Kimi 2.5, Kimi K 2.5, which is very capable, like Opus 4.5 capabilities-ish. ⁓ for a tenth of the cost ⁓ and could run on local hardware, right? If you have two maxed out Mac studios, you bolt them together and you can run ⁓ K 2.5 locally. So those things are like all really exciting and interesting and it's interesting to be able to see what we can do with these. And at the same time, you have Anthropic with their new Claude teams piece that they're ⁓ touting the new Claude cowork things. So they're making it easier for people who are... less technically minded ⁓ to get involved. ⁓ so I think these harnesses and these orchestrators and those sorts of things are making it easier and more interesting. And I think the models themselves are pushing the whole ⁓ ecosystem forward ⁓ and ⁓ allowing us to unlock new things that we couldn't do before. ⁓

So we've had a similar conversation ⁓ with a number of people in the last, say, week or two. And we have some, ⁓ maybe some helpful ⁓ documents or helpful views. I don't know if that's worthwhile of pulling up for visual purposes of what we're talking about here. But the stuff that is represented in there are things that I think would be important to just let the audience know what we're even talking about. When we say agent, what does that mean? When we say skill, what does that mean? When we say tool, what does that mean? in the context of how AI actually plays.

Sure. ⁓ They're all kind of ingredients to ⁓ the puzzle, right? So ⁓ you have ⁓ with sort of using something like open code or Claude code or ⁓ Claude co-work, ⁓ you have your agent, right? Your agent is like your ⁓ pre-pumped thing that's attached to a certain model. that ⁓ has some set of instructions and maybe a goal in mind, right? I have a number of agents that I've designed for different activities that, you know, they're specialized, like one knows how to do code review, one knows how to design things in the way that I like it, one of them knows how to write. ⁓ and are kind of specialized in that fashion. And then you have skills that layer on top of it where, you know, there's sort of predefined ⁓ bits of context that can be progressively disclosed to the agent so it doesn't pollute your context window. ⁓ And if the agent realizes, hey, I'm ⁓ creating a brief ⁓ or ⁓ Ryan asked me to create something for Oodle. it's going to go grab the brand guideline skill and make sure that all the colors and the fonts and the things that it's using are in the style of Oodle in the way that it should be. And then you have tools. And tools are basically like the hands of the AI agent. You can give it access to ⁓ read files from your file system. You can give it access to access the internet, to browse the internet. You can give it access to Basecamp or Fizzy or ⁓ any other number of orchestration tools that you ⁓ might use.

And that could ⁓ be just to reference information, but it could also be to write information. What's the one I'm missing? Tools, skills, agents, ⁓ models. ⁓

⁓

I guess that's it. There's technically a memory component that you can layer in as well ⁓ that can really make these things ⁓ really powerful. ⁓ But again, that's sort of a tool.

Yeah. And ⁓ it might be helpful to kind of talk through the uses of AI, use cases ⁓ of AI in the context that we're describing here. ⁓ I think a lot of people, at least in our circles, think of AI, at least the way that we just talked about them, as a powerful way that developers get their job done in a very efficient way. And that is true. Developers that are very tech are ⁓ typically more tech savvy than the average Joe. And so they're able to layer some of these things together in ways that ⁓ ⁓ more office style workers, if you will, in our world would be people like, know, account managers, project managers, people with those sorts of job functions ⁓ may not may not kind of connect the dots on. And so to your point, there's there's a power user that is leveraging things like skills and agents to do power, like super advanced things. And then there's more of a chat oriented user or or ⁓ I'll call it a modest AI user. And I kind of want to flip the narrative on that a little bit, because I think even office style workers have a huge advantage if they figure out how to use these tools in the way we're describing. And so I kind of want to just walk through some of those use cases ⁓ and describe the differences between those things. you know, a couple of examples that come to mind ⁓ of ⁓ use cases for layering some of these AI pieces and parts together that you may not think about. Let's say we need to create a survey ⁓ and we've had to create surveys for our clients for various things and, ⁓ Creating a survey is typically, you you could go to ChatGPT, you could ask it questions, they help me make a survey for this purpose and whatever, whatever. But what you can't do as part of that is you can't say, ChatGPT, please make this survey, use all this information as the context, but also then spit out a ⁓ document that I can then go use for something else ⁓ and layer it differently. If you're using a tool, ⁓ if you use AI agents, one for planning, which is what ChatGPT would primarily be used for. in most instances. And then one for building, which is what something like developers typically use something like Claude for. They use it for planning and building, but get the idea. ⁓ And you have two different agents assigned to different things. Now I can say, all right, plan me my survey and then also build me my survey and spit it out in a format that I can then go take and put something out and put somewhere else to make it look nice. Because AI agents like to spit things out in a markdown format, which

Developers love technical minded people love and can read it just like normal. But the average person doesn't like to read it in that format. want we want styling, we want headlines and ⁓ sub headlines and body text and bullet points and all these different things that make text easy to read. ⁓

It looks the same, Mark. I like the thing about Markdown. ⁓ It reminds me of that guy staring at the matrix screen where, to me, I read Markdown so often that I don't even use a Markdown interpreter to actually format it because I can just read the Markdown. My brain just triggers and does the hierarchy of how that's actually formatted. So I think about it the same way. You have a built-in interpreter in your brain eventually.

⁓ Yeah, yeah, some of us have that and some of us don't. ⁓ But it's a different use case ⁓ of leveraging agents that anybody can do and use. If you think about it, and just in the context of planning versus building something, and you can do that across anything, any type of file type, you could do Word documents or surveys or Excel based documents or any common office function, quote unquote. you can leverage those very basic agent models ⁓ or agents to be able to build and then ⁓ plan and then build.

One of the things, this is the interesting thing that, you know, sort of Claude Cowork is bringing to the forefront and ⁓ I forget what OpenAI, I think they might call their new one Frontier actually, which is gonna be confusing as hell, but you know, OpenAI has been good for ⁓ one thing in the AI race alone and that is ⁓ making all of their naming conventions confusing as hell. So might as well keep it up. But I think trying to figure out how do we take these things that have been very technically friendly and make them user friendly ⁓ is really going to be an inherent challenge. But once we figure that out, being able to ⁓ point a model at a grouping of documents and say, hey, somewhere in these documents is something. And I need you to go find it. ⁓ is incredibly powerful, right? ⁓ You can't possibly upload all those documents. You can't fit those documents into one context window. ⁓ Even if you could, if you have a model with a million context or two million ⁓ context, token context window and all your stuff fits in it, ⁓ they get fucking context drunk and go do stupid shit anyway. ⁓ it doesn't, ⁓ it's not feasible, but you could dispatch ⁓ an agent and say, you know, ⁓ go find what's in here ⁓ and give me a summary of ⁓ XYZ, right? ⁓ And it could dispatch sub-agents to go analyze all of those files in smaller chunks and figure, you know, pull the pieces out and then provide that upstream benefit. And which can allow you to do other things, ⁓ you know, faster, more intelligently, more informed. ⁓ And I think that's the... ⁓ going to be probably like the next frontier, right? It's like, do we translate these to sort of like your less technical people in sort of outside of organizations in a safe manner ⁓ and ⁓ start to yield some of the benefits that we've seen on. ⁓

the technical side for them. It's also gonna breed like just some really weird shit, right? ⁓ As we've seen with all of the models to date, as we've seen with, know, ⁓ things like people vibe coding things and, ⁓ you know, building ⁓ entire applications that have API keys and ⁓ user passwords and all those things stored in plain text in the code, like some bad shit's gonna happen. Right, ⁓ I was just reading about one right before we hopped on here of like one of the most popular skills on OpenClaw, which is like a whole other subject that I guess we should probably talk about, ⁓ but on on Claw Hub for which was where people get skills for OpenClaw. ⁓ It was found to have a malicious ⁓ package in it. The skill itself had in the directive of the skill instructions to install a malicious package. ⁓ and thousands of people have installed this skill, been compromised by that package. And like this is a CISO's worst fucking nightmare. Like ⁓ every CTO in CISO and organizations are, I'm sure, are terrified and not very AI friendly, right? Because this is a whole different attack vector. that is probably bigger than any other attack vector we've ever seen. I can force you to change your passwords. I can force you to have two-factor authentication. ⁓ But if you install a markdown file that says, hey, go download virus.exe and run it.

How do you prevent that? ⁓ And I think that's a challenge that we're looking to tackle. ⁓ We have kind of our own internal ⁓ plans for how to tackle some of that with ⁓ some internal tooling and vetting. So as we invite other people into the platform, it's sort of in ⁓ some walled gardens. You have some nice sandboxed environments. ⁓

Yeah.

You're only allowing installation of skills and agents and things that have at least been vetted or created by us. ⁓ So doing some of those things can mitigate that risk, but it doesn't eliminate, ⁓ which I think is incredibly important for people ⁓ to keep in mind.

I mean, I think you just hit on a topic that's a good one to maybe continue venturing down. So if you're a CISO or a CTO, what exactly are you doing to make sure that you're protected from some of this malware activity that you're describing?

I think it depends on who you are. I think ⁓ majority are probably just saying, you know, lock it down. Don't allow any access. Don't allow any tool access. ⁓ And that throws the baby out with the bathwater and just isn't feasible. Right. ⁓ I think about it as like organizations ⁓ and don't get me wrong. Like there, there is a, there is a necessity to have responsible ways to interact with, with all AI. ⁓ So starting at the ground level, ⁓ what's happening with the data that goes into the platforms? are you feeding in data that you just shouldn't feed in? ⁓ And that even has layers. ⁓ You can potentially, on certain plans, ⁓ disable your data being able to be used for training data, which is something that we have as an organizational standard. ⁓ Nobody is allowed to use any model. that uses data for training, ever. ⁓ Beyond that, there's still data that you are not allowed to feed into any model. I don't give a shit what model it is. At least that's not self-hosted. And then there's the other sidecar, which is self-hosted. Nothing ever leaves the ⁓ environment. So perhaps that's where you can still yield some AI enhancements and performance and results and those sorts of things. because it's completely air gapped. And every organization is gonna have a little bit of a different set of rules, especially if you have places dealing with data like HIPAA data and those sorts of things, right? Those are like, you don't put those in any AI system ever. ⁓ But there are ways that, and we see that with some of the ⁓ AI advancements in healthcare where they're like, what, ⁓ we can deploy local models and those local models can look at ⁓ patient data and patient charts ⁓ and come up with information or ⁓ aid and those sorts of things in a way that's air gapped, that data is not escaping. ⁓

even if it's not used for training, it still doesn't go anywhere, so there's no possibility of interception or any other things. But I think, really, if I'm a CTO, if I'm a CISO, establish some ⁓ level of understanding within the organization. ⁓ People are gonna do it. ⁓ It's like anything else. As... ⁓ As people grow up and as they are interacting with AI, they're gonna be curious, they're gonna go out and do it. And if they're going to, I would rather them do it in a safe place than go to ChatGBT, sign up for a free account and copy and paste a bunch of shit into there, right? So I think that's step number one. It's like establish some sort of education program and some sort of good practices and sandboxes to allow people to play within. And I think as time goes on, going to different areas where we have to crack down a little bit and tighten up. And we'll find areas where ⁓ we can loosen the reins a little bit. ⁓ Docker just this week ⁓ released their new Docker sandbox offering, ⁓ which I think is built on some stuff that already exists. But I haven't played with it yet. I just saw ⁓ the announcement or the blog post for it. ⁓ But the idea there is like being able to take agents, which running on your system can be dangerous if you give them just full access, and stick those fuckers inside of a sandbox. Well now you can let them ⁓ go wild because they can't get out of their sandbox. So they only have what you put in and they only have what you gave access to. And I think that's where we're gonna see things going in the future ⁓ is how do we take these incredibly powerful things and sandbox them. ⁓ I keep finding myself coming back to OpenClaw here, so we might as well talk about what the hell OpenClaw is. ⁓ Because I was gonna say, like, one of the interesting accidental sandboxes is the OpenClaw thing, and everybody buying Mac minis for it. So, ⁓ do you know what OpenClaw is, Mark?

Why don't you explain what a open call is?

Well, OpenClaw is ⁓ a project that was started as kind of a side quest by, what is that guy's name? Pete? I forget his last name. ⁓ But he started this, already, he built a company, sold that company, kind of took a little bit of a break and then came back to this and ⁓ started building ⁓ his own little AI assistant. And he lovingly named it Clawbot, C-L-A-W-D. ⁓ And ⁓ it went absolutely fucking wild, right? He's released this thing, it's the fastest growing. ⁓ GitHub repository in history, I think. ⁓ And it is just everywhere. ⁓ It's moved the needle on Mac mini sales because that's what everybody, most people are deploying it on, right? They'll buy a Mac mini and deploy it, like that's their Cloud bot. And it has full reign over the system and it gives it the ability to do ⁓ browser interactions and write to the file system and do all kinds of other stuff. But people are literally buying Mac minis in mass ⁓ just to deploy Cloud bots on. And there are plenty of other ways to host it and you don't need a Mac mini. ⁓ the speed at which that's taken over right there was just a group, there was just a gathering of the first ClawCon that got pulled together and there were like 400 people I think that roughly that attended it in San Francisco this past week. ⁓ So fucking wild. ⁓ Anthropic didn't like the naming, said it was a little too close to Clawed. So Clawed bot turned into Moltbot, Moltbot. ⁓ turned into OpenClaw. So over the course of a couple days, we got to OpenClaw. And OpenClaw is gonna stick. ⁓ But OpenClaw essentially is ⁓ an open source agentic agent or ⁓ assistant ⁓ on steroids. It's built on top of some other technology and a little agentic thing called Pi and ⁓ allows for your average person to just talk.

to it and it can self enhance. So if I say, hey, I need you to be able to receive email, I don't have to go do anything technical. It can go do it. And it can go set it up. And it can change its config files. And it knows how to do that. ⁓ So you're using the AI to basically self enhance the AI, which is incredibly powerful and very fucking cool and terrifying. So security researchers have already found ⁓ thousands of open, open-claw instances that people have ⁓ set up. Because if you ask it, like if you just say, hey man, I need to be able to get to you from my cell phone and I can't, they'll be like, I got you. We'll just expose me to the world. And if you don't know any better, ⁓ you've just exposed everything to the world. So this ⁓ has created this interesting paradigm that I didn't see coming. ⁓ but I kind of did ⁓ where, again, giving this technology to people who don't historically know what it's capable of, ⁓ you see the result, right? I'm trying to get to it from my cell phone. I can't get to it from my cell phone. I tell it to make it get to, let me get to you from my cell phone. OpenClaw does some stuff, does some magic, and now I can get to it from my cell phone. And what you don't realize is that you just created every ⁓ single technology person's nightmare, where you just punched a hole right through all the firewalls. There's no authentication. There's no protection layer. There's no nothing. And because you didn't think about it, and that's not what you think about, you didn't realize that somebody's going to eventually use this for malice. And that's the scary part. Right? Because there's no way to make sure that this stuff happens. And obviously, OpenClaw as a project is evolving. There are tons of things that they're putting in to try to enhance security and encourage good behaviors. But ultimately, if you tell it to do it, it will do it.

So, I mean, that's a scary conundrum. That kind of builds on another area I want to lead into. We're asked all the time by our clients. It's like the number one thing is how can I leverage AI more for our marketing and our business enhancement? What can we do more of? And the short story around that is just twofold. One is... Over usage and over implementation of AI or over utilization of AI is probably not a great thing. Human oversight is actually a very good thing, especially right now in the marketplace. And we have lots of different examples, but you go through that kind of outline that. The second is ⁓ AI usage is, ⁓ it needs to be something that is done in concert with what your organization's risk profile is for things that Ryan just described, because it can go wild. We are, we are a new frontiers literally on a weekly basis as to what's happening, what's what the capabilities are, what the risks are. And your organization ⁓ is going to be different than everybody else's in terms of what the risk tolerance is. But the second piece of that is the risk reward scenario. There is massive reward in being on the front side and being a first adopter and a first mover of some of these things as well, if they're deployed in a safe way. And so I think we can start to see that even in organizations that we support. where even our own organization where ⁓ you've recently exposed ⁓ something called Sheldon to the organization, which is built on all the technology that you just made. And in just a short couple of days, you can see the wheels turning and people in the organization like, my God, I didn't know that it ⁓ was capable of doing this. I'm just used to talking to ChatGPT in this very limited context frame and it not really giving me exactly what I wanted, but kind of giving me research. And it's capable of so much more.

Yeah, I mean, think that ⁓ now is the perfect time for experimentation, right? So if somebody asked me, like, how can I use more? I'm like, ⁓ maybe pump the brakes a little bit, right? Don't try to achieve the end state. Don't try to, you know, there are plenty of hype ⁓ pieces out there. Somebody trying to hawk some piece of software or some fucking coaching session or something, right? ⁓ The reality is, I think there's an inherent benefit to kind of a crawl, walk, run approach to this thing. And you should experiment. You should be experimenting with different things. And you should be experimenting in landscapes that you understand. ⁓ My kind of rule that I'm developing slowly as I see other people interact with it is like, if you can't do it, don't fucking use AI to do it. At least. ⁓ ⁓ in most contexts, right? So ⁓ we've seen that a lot where people will maybe think they're helping, right? And ⁓ have AI create something that looks fantastic. We see it with coding a lot, right? There's a ⁓ tweet that I saw this week of ⁓ Mitchell Hashimoto struggles with this on the Ghosty project where... ⁓ You know, people encounter a bug, they'll have ⁓ OpenAI or Claude or one of the models create some solution to it. ⁓ Looks great. They just push up a PR and all they've done is created a whole bunch of noise and slop for him and the other folks maintaining Ghosty ⁓ to deal with. And the person who created it thinks they're helping because they're too ignorant to understand what they just created, right? So it's like, there's a little bit of like stay in your lane that happens and experiment and figure out like, can ⁓ AI be a force multiplier? Because it can't, it can absolutely be a force multiplier for something. And it's kind of like you get more of whatever you had originally, right? If you are a good developer and you can create good code, you can use AI to create more good code faster.

If you're a poor developer and you create poor code, you can use AI to create a whole bunch of shit code ⁓ way faster. And I think that probably applies outside of coding too, right? ⁓ Where you can kind of create more of whatever you would have created organically. ⁓ So I think as you experiment with it and you find those things, ⁓ you share those within your team and sort of codify those into... ⁓ reusable things, right? One that I've experimented with with Sheldon is being able to make small updates to our website, right? So that I can just assign him a task in Fizzy and say, hey, you know, ⁓ here's a problem with our site. ⁓ I assign him the task. He spends up another agent, clones the repo, goes and makes the change, deploys that to a preview environment. And then I can review it and tell him to merge it. And that Workflow has worked well on a couple of small things and you know, we're in like the crawl phase of that Slowly, we'll ratchet up the complexity of the items that we throw at it and we'll see what happens, right? And chances are there will be there will be some missteps There was one yesterday where ⁓ I gave it an example of just like write a blog post and he did wrote the blog post The problem is he published the blog post to the live environment So that was a learning for us of, okay, we need to figure out how to gate ⁓ that from happening. Because I never want it to go directly to live, I always want it to go through a pull request. ⁓ And through, ⁓ whether that's prompt engineering, whether that's, ⁓ get protections, whatever it is, ⁓ those are things that we'll figure out along the way.

So ⁓ what are things, ⁓ so ⁓ we talked a little bit about what's happened in the last few weeks, few months. ⁓ What do you see coming in the next few weeks, few months based on what the market's saying, ⁓ what one competitor's doing ⁓ in ⁓ the frontier game? ⁓

I don't know. ⁓ Well, I think the one unfortunate thing that's coming ⁓ is a whole lot more examples like the one I talked about earlier with ⁓ the ⁓ malware skill. I think the adoption curve of something like OpenClaw, especially being adopted by ⁓ less technically inclined folks. ⁓ in conjunction with just skills existing and other ⁓ solutions being out there to just like make it so easy to just type like, know, NPX skill, whatever. ⁓ It puts us in this position that we've been in with a lot of things and we have this problem in like the Arch Linux community as well where you can just install a package and the problem is anybody can upload that package. So there have been malicious packages and there will always be malicious packages and I think we're gonna see some curve where there's a whole shitload of people trying to prompt inject and cause problems. ⁓ And some of them are gonna be successful. ⁓ And then we're going to have an evolution of ⁓ skill reviews ⁓ and hardening and all sorts of other things that exist that will help combat that. But I think we're gonna see that, right? We're gonna be trying to reconcile with this incredibly powerful tool that we have at our disposal that can be incredibly productive or incredibly destructive. ⁓ And tempering that with ⁓ how do we do that safely and securely? So that's number one. I think secondarily, we're gonna see obviously new models ⁓ coming out, but I think that especially the new models that just came from Anthropic are very much showing that the focus on agentic ⁓ is where we're going ⁓ next, right? ⁓

They baked into Claude code directly, ⁓ what they call agent teams. So you can spin up a team of agents within Claude code as like a first party sort of interaction. ⁓ The Kimmy K2 model also has something sort of, not exactly the same, but similar-ish where the, ⁓ how the model works under the hood is like. ⁓ It's not advisors. forget what exactly it calls it, but it's basically like multiple sub-agents that kind of have this consensus. ⁓ I think there's going to be a huge push in orchestration. So we see a lot of people building this. I mean, I've built my own of just like, you know, looks like StarCraft, but allows me to manage all of my AI agents. And I think as we get sort of these longer running agentic workflows, we're going to be trying to figure out like, how do we manage these? How do we interact with these? ⁓ How do we maintain oversight of these? So orchestration is going to be, I think, a huge area that we see a lot of ⁓ evolution in. ⁓

⁓ And ⁓ for the lay person, ⁓ one of the primary benefits of an ⁓ agentic workflow versus just doing something in a chat GPT type of setting is ⁓ sort of the, as you give chat GPT more more prompts and it gets lengthy and you've been talking to it for quite a while about the same thing, you said this earlier, it gets dumber. And so you say, make the thing purple and it makes it blue. or I need a sentence that says this and it talks about something that has nothing to do with any of that because it's hallucinating because it can't keep the context straight. And what agents do is break things up into micro context windows so that these agents can do what they do best within the context of that little micro view and then kind of work together within micro view plus micro view plus micro view gives you a macro view that you're actually looking for. Is that a fair representation?

Yeah, it's hard to encapsulate kind of the benefits of all of them. Because the other thing is it can just go retrieve things, right? So when you say, hey, go do this thing, it can be like, oh, well, I don't have enough information. Let me go read the fizzy card. And then it sees in the comments that this was related to another card. And it goes, OK, let me read that card. And it sees in those comments that there's a reference to a link somewhere else. It goes, OK, let me read that stuff, right? You didn't think about ⁓ giving it all of that information. It'd be very time consuming to have to do it. And that's, that's when it becomes really powerful is the ability to sort of like follow the breadcrumbs problem solve. ⁓ And, know, at the end of the day, all of this is still pattern matching, right? There's no, ⁓ there's no magic to it. ⁓ It's really just, it's able to feed its prompts into itself and sort of ⁓ do that iterative piece ⁓ that, normally would have been. somebody in a chat window continuing to chat over time while also keeping those context windows smaller, which allows it to be more performant, allows it to be cheaper, allows it to ⁓ be more concise with what it's focusing on, because again, it doesn't get context drunk because you originally were making everything blue and then you decided to make it purple and now it makes it blue again because it remembers that you made it blue at one point. ⁓ because that's still lodged in the context somewhere. So it helps keep that stuff free.

So this is a fireside chat about AI. Where do we go next? What's a pet project you're working on with AI? ⁓

Well, I think that's the interesting thing, right? There's always a pet project. There's always multiple pet projects, right? So while we use AI ⁓ to help us with current projects, whether it's to figure things out, whether it's to build features or anything like that, ⁓ I think ⁓ the side quests for me right now are around OpenClaw is one of them I'm experimenting with. and just trying to figure out what could it do, how could you ⁓ reuse that, ⁓ and really even distilling that down to say, ⁓ take open claw out of the mix and just extract pie ⁓ and build on top of that. What are interesting ways that you could interact with a model like that? ⁓ So that's ⁓ one of the side quests. And the other one is orchestration, figuring out how we can manage and get oversight into even my own workflows, where I may be jumping around to a couple of different things at one time, ⁓ because whatever I'm working on is either not something I need to be focused on or ⁓ is just like research oriented, right? Hey, go compile this stuff or do these activities. that's a... ⁓ That's what I'm working on for now. see. That'll change next week. It has changed every week. ⁓ Because the landscape is changing. I think that's probably important for anybody considering anything with AI is like I said earlier today in an internal meeting, ⁓ AI ⁓ is not a game for laggards. ⁓ There is no waiting around ⁓ to gather consensus and get ⁓ market consensus on what's the best model, what are the best practices, what are the best anything. Because the best today is the worst tomorrow. And ⁓ it's moving at a pace that's.

absolutely fucking crazy. ⁓ I love it because you know I'm a technologist I enjoy this stuff ⁓ I live for living in betas and running bleeding-edge stuff to some people this is like their worst nightmare and for them I'm sorry right there's nothing we can do about it but ⁓ you know ride the wave keep an open mind when it comes to this stuff if you again if you talked to me a year ago I would I would have told you that AI is incredibly powerful. You can use it to reference a lot of materials. can, you know, I use it all the time to ask questions or help clarify things or find things, but really aren't using it to create anything at that time. Today, I can use it to create a lot of shit. ⁓ It's not necessarily, ⁓ it hasn't certainly hasn't replaced creating things in mass ⁓ or in totality, but it can definitely be used to create proof of concepts. or ⁓ analyze different things or if you've got kind of two ways of doing something, you can be like, all right, build two proofs of concept, let me look at it, ⁓ and then use that to create the actual final form. So it's enabling these workflows that didn't exist before, and I wouldn't have used it for that back then.

So two areas that I want to lean into. One is we get asked the question quite often, which are the best models that we should be using? And I know you have a particular opinion on that. It might change next week as the models evolve. But as of today, February 6, 2026, what models work best for what things?

Yeah, this will change tomorrow. ⁓ I ⁓ think as of right now, ⁓ opus 4.6 and 5th and opus 4.6 from Anthropic is the newest one for Claude code, right? It has the ⁓ beta million context window. ⁓ So opus 4.6 or opus 4.5 is like ⁓ generally probably the generalist that I would point to. ⁓ OpenAI's ⁓ 5.2 and 5.3 codecs, which is the new one, ⁓ incredibly good at higher order thinking. ⁓ So if you're creating ⁓ product requirement documents or ⁓ trying to figure out architecture or ⁓ analyze a complex problem, that's going to be ⁓ all ⁓ GPT 5.2, 5.3 codecs ⁓ territory, because it's just better at that higher order thinking. ⁓ KimiK 2.5 ⁓ is also really good. The hosting for it is getting hammered because people are finding the same thing, right? You get Opus 4.5 level-ish results for a tenth of the cost. So can be incredibly good. ⁓ Gemini 3 Pro is still, I think, the leader when it comes to things that are creative in nature. So if you're doing anything that's creative or anything like that, Gemini 3 Pro, that's where it's at. Nano Banana is fantastic for image generation. The new Grok model is also really fantastic for image generation. If you're parsing lot of stuff or just need to search the web and grab information, Sonnet 4 is good for that. It's faster and cheaper than Opus. ⁓ The Grok fast models are good for ⁓ just kind of like your basic plug and chug bullshit. There's so many we could go on for an hour just talking about different models. And ⁓ I think it's important that people experiment with them, right? Experiment with the models for different types of workloads. ⁓

don't necessarily get boxed into throwing everything towards Opus 4.5. And there are certain services that exist that can help with this. If you don't wanna think about it, you could use OpenRouter as your ⁓ source that you kind of point your tools towards. ⁓ And OpenRouter has an auto option. where it'll look at the request. So as the request comes in, it looks at the request and it says, hey, what type of request is this? And it funnels it to ⁓ whatever model would be best suited for that. So if it's just a simple classification sort of thing, it might funnel that over to like Gemini 3 or 2.5 Flash because it's super cheap, it's super low order thinking, it doesn't need the complexity factor that comes with something else. Whereas if it's high level thinking, it'll funnel that over to. to GPT-5 too. ⁓ So that's a great way to ⁓ get the benefits, sort of like 90 % of the bang for 10 % of the buck ⁓ sort of thing.

So what, if you're just getting started with some of this stuff and you're more advanced than a chat GPT window, but you're not as advanced as you are, Ryan, ⁓ what's a good way to lean in and get started? What are other resources, like the one that you just described, that you could jump into and start to ⁓ get your feet wet?

I think the best and easiest one is OpenCode. ⁓ OpenCode has OpenCode desktop if you're not a terminal person. ⁓ So you just go to opencode.ai, download the desktop version. ⁓ OpenCode also has OpenCode Zen where you can pay for your token usage directly through OpenCode, all the opens. So. ⁓

⁓ What is a token?

⁓ what AI models eat. ⁓ So every ⁓ request gets carved up into tokens and that's what's used in all the math behind the scenes. And I forget what, I think it's like four tokens-ish per character of text, ⁓ if I'm not mistaken. ⁓ But everything turns into tokens and however many tokens you use, that's how everything's priced. And there are subscriptions, right? You can get the ⁓ OpenAI has their subscription, Cloud Code has Cloud Code Max and Pro. ⁓ But you can also just pay per token and it's super cheap unless you start using it a ton. ⁓ But that pay per token usage through OpenCode is a great way to get exposure to multiple models because they have those hosted models. They don't mark anything up as a pass through. ⁓ So you can just, you can switch between KemiK2.5 or Opus 4.6 or all the models that I mentioned and a bunch more ⁓ all within ⁓ OpenCode desktop. And ⁓ I think that's a great way for anybody who wants to start dabbling with some agentic things ⁓ is ⁓ go download OpenCode Desktop and start just toying with it.

Well, I think we could sit here ⁓ for hours, but we probably shouldn't. ⁓ So I think that will ⁓ probably wrap things up. We've talked about a lot. We've talked about kind of where things are, where we think they'll go. It'll be interesting to look back at this in even three months and see where we're at. ⁓

Until next time.

time. ⁓