Valentino Stoll  00:00
Hey everybody, welcome back to another episode of the Ruby AI podcast. I am your host today, Valentino Stoll, joined by another host, Joe.

Joe Leo  00:13
Hi, I'm Joe, I'm another host. Joe Leo of Death Method and Ruby AI podcast fame. Valentino, OpenAI's rogue models roamed the internet for 4 days. Did they ever come and knock on your door?

Valentino Stoll  00:28
I didn't see it.

Joe Leo  00:30
This was a real, like, hide the women and children kind of moment, and you have to stand there with your keyboard ready to do bad—

Valentino Stoll  00:38
I've got to hear that. What happened? How did you see it?

Joe Leo  00:41
This was a great headline from Politico that I just pulled up a minute ago. "OpenAI's rogue models roamed the internet for 4 days and staged a second attack." And I love this because, like any good crazy internet AI story, it keeps giving. What started with hugging face, disclosing an attack, and actually doing a really great job of documenting the attack and exactly what happened and where there were weaknesses in their infrastructure, the security vulnerabilities, what they did about it. Really fantastic. It took another 5 days before OpenAI was basically like, "Yeah, yeah, that was us," and it was an accident and we didn't mean it. Because it was just while they were trying to train their models, the AI just jumped out of the sandbox. It was given a task, which was to essentially follow a series of steps to try to determine specific security vulnerabilities, and it found the best way to find those specific security vulnerabilities was to not follow any of those steps.

Joe Leo  01:45
So pretty smart. But the reason that it's so great is because now it's been over a week and the headlines keep coming. Now it was revealed to the rogue AI, the bandits, also molested Modal Labs and potentially some other companies, and so who knows who else. I don't know. Are they still out there?

Valentino Stoll  02:08
They've got to be out there.

Joe Leo  02:09
It seems like they have a life of their own. The first security incident that I have felt very viscerally," said Sam Altman, who's also always great for a headline and a quote.

Valentino Stoll  02:20
I wonder, too, if this has any relation to their recent open-sourcing of their security CLI.

Joe Leo  02:30
Yeah.

Valentino Stoll  02:31
Are they related?

Joe Leo  02:33
That's a great question.

Valentino Stoll  02:34
Maybe. Do you install it and run it on your code? I don't know. I was messing around with it today and it's like, hold on, like, this uploads code. Just so you know, anything you, you know, run on it, it's going to, like, send and submit to OpenAI. And I guess we're kind of doing that anyway, in a way, but I was like, eh, maybe this isn't the right tool.

Joe Leo  02:53
It's true. That's a good point. And actually, and so there's, I'll have a lot more to say about that when we get into RubyConf because I attended the security teams workshop on day two, which was very interesting. But first, I was reading Jameson O'Reilly, who's this super, super smart security researcher, and he was writing about the incident. It's interesting because it dovetails on what you were saying. So when Hugging Face tried to address this threat, they first reached for commercially available models and did exactly what you just said, like copying and pasting security data into the models, and the models were falling over. They were either refusing to do the task because they're like, this is hacking, or they were taking actions based on it and sending things off the rails. And so what did they have to do? They needed to reach for open weight models, which is really fascinating because they're all the rage these days because the White House is talking about them.

Joe Leo  03:54
Probably the first time I ever heard about Hugging Face was from you, Valentino, and I mean, this is years ago. But that's kind of where we're at right now, where these open weight models are having a moment in the sun. I think anybody who doesn't know anything about open weight models thinks they're all from China. And so there's a lot of misinformation going on around open weight models. But Hugging Face, of course, is the haven for open weight models, and they needed them to stop this security threat that OpenAI unleashed on them.

Valentino Stoll  04:26
A lot of props to Hugging Face, right? First, like, they didn't have to do anything. They could have just been like, okay, stop, and we'll move on. But I get it. They don't want it to happen again, right? Like, all of us shouldn't want it to happen to us either. And so hopefully something happens from this, but I'm not confident anything will.

Joe Leo  04:45
Well, like what? What would you like to see happen?

Valentino Stoll  04:48
Yeah, it's a disclosure, right? Like, so if there was a HIPAA violation, you have to, like, publicly announce to your customers that you've disclosed information.

Joe Leo  04:58
Right.

Valentino Stoll  04:59
Ultimately, it's like kind of the reverse of that. Something damaging has happened in a public discourse, and now, like, how many other people are affected, right? Like, I don't see any numbers. Who's affected? At such a large scale as this probably is, to not hear anything is a little concerning, right?

Joe Leo  05:21
It is. You're exactly right. And even the fact that it took 5 days from Hugging Face's disclosure. And I totally agree with you, right? Because Hugging Face, there's an old saying in sports, when somebody does something either horrible or detrimental to the team or detrimental to society, they always get the athlete up there and they're in an interview and they say something like, "Well, I'm sorry for what had happened." You know, as if they were just kind of a bystander, but they're sorry that it happened. And the equivalent in the tech community is like, well, there was disclosure that may have impacted some accounts, right? Which helps nobody and kind of sort of exonerates them. Like, well, you know, it might have, right? Which is incredibly frustrating because it sounds like you don't actually know what happened. And we really, if we're depending on you, or even if we keep some credentials with you, we really want you to know what happened. And then it also, it kind of puts some distance between them and the incident.

Joe Leo  06:19
Well, like, hey, something happened here. I don't really know what, but it's all back now. Just change your password. And Hugging Face did not do that. They went very deep into exactly what happened, took ownership of the parts that they needed to improve, and I'm sure it's already 13 days later, and I'm sure they have improved all those things.

Valentino Stoll  06:36
Even from a security perspective, anybody that's in any compliance industry, they have to disclose the event to whoever. If I've got HIPAA compliance and I see, like, a security event happen, I have to, like, work with that company. Like, are they just inundated with requests right now? They must be, right? For any compliance company. And maybe that's what the delay is, is they have to handle all these requests first.

Joe Leo  07:04
I mean, that could be. Yeah, it could be.

Valentino Stoll  07:06
But it's a little funny. And it makes you wonder, too, what might be happening internally. It's gone rogue and done publicly, right?

Joe Leo  07:18
Yeah.

Valentino Stoll  07:18
What is happening on the inside? That's probably equally concerning.

Joe Leo  07:24
It could be really messy. You're right. It could be concerning. We don't know those things yet. Well, of course we're here to talk about AI. It's an AI podcast, but we're actually here to talk about RubyConf, which I attended and really enjoyed out in Las Vegas, Nevada, in the 110-degree heat of July, and had a really great time. And, you know, some things that I'd like to focus on here. The talks, of course, and I'll talk about some of the highlights there, including the keynotes. And also, you know, I got a chance to talk to a whole bunch of people, I guess not everybody, but almost everybody that is either on the Ruby Central board or contracts for them or volunteers. And that was also, you know, really illuminating. People always talk about the hallway track being the best, and it's funny because I didn't even think that's what I was doing. But then I came away from this and I was writing up my notes and I was like, man, you know, all these conversations I had outside of talks were every bit as interesting as the talks themselves.

Joe Leo  08:27
Yeah, so happy to dive into it. Where do you want to start?

Valentino Stoll  08:31
Yeah, I love RubyConf. I don't know, people are just so nice. I don't know, I always get excited. I was really bummed I missed this one, to be honest. It happened during my transition and.

Joe Leo  08:42
Tell us about your transition.

Valentino Stoll  08:43
Yeah, yeah. So I have accepted and working at ZAR now. I left Gusto and moved on back to my Ruby roots, which I'm pretty excited about.

Joe Leo  08:54
Yeah, we're excited about that, too.

Valentino Stoll  08:57
It's wild. It's as wild as Obi's talk at the conference.

Joe Leo  09:02
Obi came up to me after, and I met his son. That was nice. And he said, "Yeah, your boy's starting with me on Monday." And I was like, man, he's really in for it. It's going to be great. So we waited a week because I wanted to give you a chance to settle in. But I'm sure that a week is not really enough because that talk was really something.

Valentino Stoll  09:18
Yeah. It's only been a week and I am onboarded fully.

Joe Leo  09:22
You're onboarded.

Valentino Stoll  09:23
I have made changes to some core system components, like the mobile app. I'm familiar with all the tooling. Just smooth onboarding to the point where now, like, I can just get handed tasks and pick them up. And to be honest, most of the time it's not me picking up tasks.

Joe Leo  09:44
What do you mean? Who's doing it?

Valentino Stoll  09:45
We call them legates, but our internal bots. And I've set up my own bot and his name is Jubal. And he delegates all my work for me and makes sure that I stay on task. And he delegates to some of the other bots that we have when the work is appropriate for it. And if it's just a quick bug fix that needs to be done, it'll just get done. It's fantastic. It's the future.

Joe Leo  10:12
Amazing. It really is. I mean, we may as well start there because Obi's talk came on the morning of the third day. So it was the early day keynote. And I made a note here. I think there were only two talks I went to that were not at all about AI. His was obviously about AI, but it was the most into the weeds. This is exactly what we do with AI. This is what works. This is what has not worked. Here are all of the tools. Here are the repositories. Here are the agents. Even for me, who we've had Obi on this show and I've read the stuff that he's written, you know, over the last year, it still was pretty mind-blowing for me.

Valentino Stoll  10:52
It's hard to visualize it, which I think is part of the problem, right? It's like, you think, oh, like, you can think about the stuff that you maybe hand off to a coding agent and, okay, you maybe work with it in different ways, but maybe there's some planning up front, hopefully. And then that planning turns into actual work that gets done. And maybe you're not even coding anymore, like many of us aren't. And so you have these workflows. And then just imagine that there is a loop on an agent that can handle that from natural language. So like, whatever planning that you do, whatever context that you need in order to reorient yourself, in order to figure out what problem you're trying to solve, all of that is encapsulated into a focused thing that does the same work. And multiply that and expand it and change about how these things can ultimately become employees of different kinds to handle specific kinds of tasks, share the organizational knowledge.

Valentino Stoll  11:57
Like, you can do wild stuff with it. And especially think about something as simple as, why did this bug happen? Just go gather that information and you have all of your systems available that can pull in that information and figure out, okay, what happened to the app? What was the user doing? What specific error happened and what was it trying to do? Like, all that can be reasoned about by a thing autonomously and surfaced to, like, that organizational fabric that has work triaged. And if there's a bot responsible for that work to be triaged, it gets distributed and handed out. And if there's another bot looking for work to do and it notices that it can handle that particular context and task, then it just picks it up and it works on it. And it submits a PR. And, you know, you go and you review it and if it looks good, you just ship it. And that workflow is, there's a lot there, right?

Joe Leo  12:56
Yeah, that's a lot.

Valentino Stoll  12:57
But it's.

Joe Leo  12:58
What you just said, yes. That covers a lot.

Valentino Stoll  12:59
But it's so smooth, right? And like, for the known cases where you have these workflows, having that automation in place to just handle those, like, daily things where they just pile on top of the work that needs to be focused on by a human, mind-blowing the amount of work you can get done.

Joe Leo  13:19
Yeah, I was really impressed by that and continue to be impressed by it. So part of this, I think, and you'll correct me if I'm wrong, but part of this really is the knowledge graph and the quality and quantity of the knowledge graph and sort of the institutional knowledge of the system. Because it sounds like, from what you're describing, you really can't have all of these agents communicating with one another unless there is some way that they can gain knowledge, especially historical data, to make predictions about what is happening in real time.

Valentino Stoll  13:49
Right. Yeah. I mean, it's definitely critical to have that knowledge base that keeps track of different kinds of knowledge even, right? That doesn't go stale. It stays fresh. The advantage of having it always on and all these agents communicating is it keeps it fresh, right? Because it's always being used and always being referenced. And it becomes obvious if anything is not true, right? Or not accurate. And so the more that you use it, the higher quality you get out of the knowledge base, right? Which is a little funny, I think, because usually it's the reverse where you add more pollution to the data and suddenly it becomes blurry, right?

Joe Leo  14:28
I think that's the fear. Let me say, that's a fear for me. You know, I talked about this with Beatrice. We both know Beatrice works at DepthMethod. She works with me. And I told her about some of this stuff from the talk, including the fact that there are like 37 agents that take on all these different employee roles. And Beatrice, of course, is back office. She's like, well, I'd like some of that. Let's get some more employees in here. So I'm not inundating her with work. And the fear becomes like, okay, well, then how good is my data actually, right? And how good can I actually automate things? Not can I automate them, but are they even automatable based on the information that I have available for an agent to peruse? I can imagine there's a lot of trial and error there. I can imagine that Obi didn't just have something right out of the gate that was so effective, but had to kind of tune it over time.

Valentino Stoll  15:13
Yeah, you know, it's kind of like a culmination. If anybody's familiar with Obi's work, he used to run a company called Olympia, which was ultimately creating an AI agent, was called at the time, into a persona. So it has its own prime directive and it has a personality. Before Soul was a thing from OpenClaw, right? He described what this entity is, what it does, what it can do, what attributes it has. And then on top of that, like, you just plug in all the capabilities you want it to have. So like you're mentioning, like an employee, it's very much that. Do they get access to Google Drive? Well, they get their own account. Then they connect and you can track everything that they're doing or gate access. And that's ultimately what ends up happening. And the more that you could treat something like it like an employee, you run through the same stuff. I think really it is just experimentation. Like, you got to try it. Reminds me of the early days of ChatGPT.

Valentino Stoll  16:15
I remember the early advertising. And granted, like, you know, ChatGPT when it first came out wasn't great. But people were just like, oh, yeah, it, like, makes my own coffee. It built this, like, animated Pixar movie, right? Like, just complete nonsense at first. But like, you know, because they were trying to just drum up advertising. And maybe people were underwhelmed at first, right? But like, use cases start to emerge, right? And you realize, wow, this thing wrote a draft article for me and now I can just go back and make edits and, like, reshape it. It saves me a lot of time, right? And if you have, like, just very common tasks that you want to do, right? And just started small with just, like, things that you need help with text, which now, like, everybody regrets doing because it's hard to read anything and it's like 20 pages long. But now we're more advanced and you can literally hook these things up to code and let them code and kind of reshape how they are in real time even.

Valentino Stoll  17:14
And it changes a lot of what you can do with it.

Joe Leo  17:18
Yeah, I am sure. And I'm sure you're very happy to get back to it. I never left. And so it's very nice doing these things in Ruby and in my kind of back office and automating things. And I do have plans for an agent kind of breeding ground that we're calling Arkham Asylum. And that was inspired by Obi's talk, you know? And so even a services company, so if you're out there and you're like, well, I don't think my company really qualifies. Even a services company, I think, has a lot to gain from this. I think there's a lot of automation. There's a lot of stuff that we're doing manually or even just semi-manually, like chatting by hand with a lot of these agents that are not talking to each other, right? You know, my agent is not talking to Beatrice's agent and they should. And so that's kind of the goal there. But that's my own thing. I'll talk about that some other time. One thing that you hit on that I think is really interesting. At the conference, there emerged these kind of two schools of thought around how to think about and frame the AI agents.

Joe Leo  18:14
And I think it's less about what to say about them and more about how to wrangle them, right? So you had Obi and you've already talked about this a lot, very much on the personification of the agents. Give them human qualities, give them names, wrangle them as if they were human. And there were a number of people on that side. On the other end of the spectrum were Jessica Kerr and friend of the show Scott Werner, who basically were saying, don't think of your AI agents as people because it's limiting you and how you orchestrate, how you try to get them to accomplish their tasks. And Obi even kind of said this because he was on the third day, so he had seen them back and forth. And he was just like, I love this. I love personifying. I love giving them personality. I think it's really fun. And so there's part of that too. It's what's fun for you? What's most effective and efficient for you versus what might be limiting you in how you think about your AI agents?

Valentino Stoll  19:17
Yeah, I think it goes back to also the conference made a great point. I feel like we're in an identity crisis with naming again. I think part of the problem is like everybody's just like, oh, just call it an agent and that's it. Well, really, an agent kind of signals a human shape. At least at first it did. And now it's kind of stuck where you say an agent and you think workflows like human tasks, it taking on these automations. And yeah, I think the other side of the argument is also accurate. Personifying an AI workflow, it's not always the right choice, right? Like, because you maybe have something that is truly like programmatic automations where you're now just applying it in an AI lens. And it's like, is that an agent or is it very strict and rigid and like follows like a built pipeline path where it is more programming centric or similar, right?

Joe Leo  20:11
Or hopefully more deterministic.

Valentino Stoll  20:13
More deterministic. So technically it is an agent because it's doing the same abstractions. But ultimately, like, is it the same thing? I feel like we need something else to identify. Okay, this is personification.

Joe Leo  20:26
I think you're right. So Jessica's talk was really interesting. It was the first night. It was the first existential talk that I got, which it really was, okay, so who are we now? And I think that might even have been the name of, yes, that was the name of the talk. And so her argument against personification, I think, ran something like, well, you're really unhappy about not handwriting code anymore or you're disappointed or you're confused about what that means for you. And you don't want to think of your job as hustling a bunch of other programmers around, which is what personifying all of your AI coding agents would mean for you. And I think that she personally tries to distance herself from that analogy because it's more helpful for her and her job to think about orchestration of particular task completers than it is to think about wrangling a bunch of programmers. And she's going to give all of them names, which I think is interesting, but it's a different use case probably than what Obi was talking about, to be fair.

Joe Leo  21:26
But that was something that she had mentioned. And in Scott's talk, which when I asked people about their favorite talk, Scott's talk came up a few times. And the reason I love Scott's talks, and we've seen a bunch of them, is that what Scott is attempting to do is to synthesize everything he is learning about AI, which is a lot. It's a lot more than I'm learning. And distill it all down into a framework for understanding, which is exactly what you were talking about, V, because he's also thinking like we kind of are in an identity crisis and an identification crisis. We don't know what to call these things and we don't exactly know what to do with them. You know, anytime I see Scott talk, it's this iterative process of like, here's what I understand about AI today and here is how we should be thinking about it. And that continues to evolve and change. But he was really looking for a way of shifting into object-oriented programming or coming at object-oriented programming from a new lens.

Joe Leo  22:23
And instead of thinking about agents as human-like task completers, thinking about them more as prompt objects. That was his thesis that he was going off of. And it was interesting to think about that way and thinking like that actually fits into a framework of object-oriented design, perhaps better than any of the or many of the object-oriented languages that came on the heels of object-oriented design's discovery or writing.

Valentino Stoll  22:52
I love what Scott has to say. We got to have him on again. You know, I played with the prompt objects library. It's awesome. Definitely go take a look at it if you haven't yet by Scott. I believe it's under the sublayer GitHub. Brings an interesting point. Prompts shape a lot. Today, like maybe as like a programmer, you think of it in a Claude or OpenAI or OpenCode, whatever you're using for your like coding agent, maybe you're thinking of like each thing that you type is a prompt, like kind of reshaping this blob that you're working within.

Joe Leo  23:28
Yeah.

Valentino Stoll  23:28
But there's also something you could do in advance too, right? Where you're front loading the scope that you want to focus on. I think of it as two things. In an object-oriented sense, you have this thing, which is your prompt that you then can message, right? It's sending messages to other objects.

Joe Leo  23:49
Yeah.

Valentino Stoll  23:50
And so what does that mean when you start nesting that or graphing it? That's kind of like where I see kind of these agent systems like coming into play of are there blobs that objects live together in? And what does that mean? Right?

Joe Leo  24:11
Right.

Valentino Stoll  24:12
Do they become organisms? Right? Is this biological? Is that even something we can apply as a concept to? We're kind of getting back to like the origins of object-oriented programming being like biological driven. I don't even know that that makes sense. They're almost taking their own shape at this point. And I start to think about maybe other concepts. What else can we apply here from our own understanding that might make sense? Like, is that something we want to continue trying to press? Maybe we need to come up with a completely novel concept of what we want to exist rather than trying to like reshape what we know. I'd like to ask Scott to investigate that one. You know.

Joe Leo  24:55
Yeah, I just think we'll pull back and change and tell him.

Valentino Stoll  24:57
And pick his brain a bit.

Joe Leo  24:58
Tell him he's not done iterating yet. He's got to come up with something new.

Valentino Stoll  25:01
But yeah, I love that. And I love Jessica Kerr. I love her talks. So I'm going to have to check that one out.

Joe Leo  25:08
Well, it would be great to have her because I might be butchering what she has to say. And I hope I'm not. You know, one thing that I also thought was really great from her talk was that she sort of extolled the virtues of human connection. You know, and saying, hey, we've kind of lost that. And she was also talking about how we might start to get that back. And so it was a very human appeal, a very human talk. And Jessica, she always speaks with a lot of emotion and there was no shortage of that in this talk. It was a really compelling talk. I would love to hear more from her.

Valentino Stoll  25:37
To that point, having worked at ZAR for just a little over a week now, it is really funny to watch the discourse in our chat.

Joe Leo  25:48
They just talk to each other.

Valentino Stoll  25:49
We talk to them, they talk to us. It becomes very human. A lot of the times you read people's messages and it'll just be like them having banter. I thought we were being playful here, you know? Like, you know, and like it'll just be like some of the conversations will just be like clearly wanting that connection, right? Coming from both sides. It's a little like Uncanny Valley in a way.

Joe Leo  26:15
Yeah. Yeah.

Valentino Stoll  26:16
But like it's even funnier to see the conversations that the bots have with each other. They do seem to communicate differently to each other when they know, oh, this is a person.

Joe Leo  26:29
This is a person.

Valentino Stoll  26:30
Because they know that under the hood.

Joe Leo  26:31
We got to dump that a little bit.

Valentino Stoll  26:37
So it'll be interesting to see how that evolves internally. And if we start to optimize that, do they start just communicating in a different language or something?

Joe Leo  26:47
Yeah, that would be interesting.

Valentino Stoll  26:48
I don't know. I'm hopeful that maybe it draws out more humanism as we start to lean in heavier to it. It's exhausting to talk to AI. And so I actually personally find myself reaching out to people and be like, hey, you want to chat? I just need a break from like the instant onslaught of texts, right?

Joe Leo  27:08
Totally agree.

Valentino Stoll  27:08
It definitely reads different. Even when you text like you can just tell instantly. And maybe that'll change, but I don't know. I feel like in this moment, at this level of like autonomy, it seems to draw out more of that human connection. But I don't know. Will that last? I can't tell you yet.

Joe Leo  27:30
Yeah, things are changing. Still changing really fast. One thing that Jessica talked about with the human connection I sort of used as a springboard into speaking to a bunch of the Ruby Central folks because what she said was basically we need to get together like in person. She herself has been working from home for years. She doesn't think that that's the best thing anymore. She would like to start seeing her fellow employees more. She thinks it's important. These are all things that I've long agreed with and I still agree with, but I don't actually have much faith that that's going to happen from people's employers. It's very comfortable and it's very easy to hire across the country, across the world, assemble teams that are global and just have them connect from their own homes. It's cheaper and it's potentially easier. And I just don't see that changing anytime soon. But one place where we could still get together is at conferences, both regional and national. And so I talked about that with a number of people at Ruby Central.

Joe Leo  28:29
The fact is that everything is really still up in the air. Is there a RubyConf next year? It doesn't look like the RubyConf this year. Nobody knows the answer to that. And, you know, I think that's okay. This isn't a slight on anybody. There were some really significant changes in Ruby Central beginning of this year. And they have to deal with the fact that they are really light on funds and they're kind of just raising again from some of the big Ruby companies in the community, which is great. They're sort of starting from scratch in a lot of ways. It seemed like priorities number one and two in some order are RubyGems and conferences. Whether or not that's RubyConf as we know it today is an open question.

Valentino Stoll  29:11
Yeah, I'm curious. How did people take that?

Joe Leo  29:15
I don't exactly know. I talked with some people at the conference about it. One idea that was floated was, hey, maybe just like the college football championship rotates between some of these different bowl games every year. Maybe it rotates through different regions. And so like next year, SF Ruby is RubyConf. And then the year after that, Mountain West Ruby is RubyConf, which I thought was an interesting idea. I floated that. Once I heard it, I floated it by a couple of people. They seemed like, yeah, okay, maybe. Certainly maybe think we need to get GoRuko back running. We're going to need one in New York, obviously. But apart from that, I mean, I think there was not a huge amount of surprise. I mean, if you're at RubyConf and you've been paying attention to the Ruby Central announcements over the past six months, maybe even a year, you're probably thinking, well, things aren't looking great, you know, for RubyConf. And so I don't think it was met with a lot of surprise. There wasn't a lot of enthusiasm about that fact, but I think it was more of a reality check.

Valentino Stoll  30:20
I know they were trying to push more smaller conferences.

Joe Leo  30:23
Right.

Valentino Stoll  30:24
Which, you know, has its ups and downs.

Joe Leo  30:27
Yeah.

Valentino Stoll  30:28
Running a huge conference is very difficult. Trying to get and orchestrate that many people, that large of a venue, that many odds and ends.

Joe Leo  30:38
And it's three days long, which is long.

Valentino Stoll  30:40
Trying to broker the hotel.

Joe Leo  30:41
In Wellsworth's only two days. A lot of his conferences are on three days.

Valentino Stoll  30:44
True. Yeah.

Joe Leo  30:44
Yeah.

Valentino Stoll  30:44
I totally understand their pain. If a smaller, wider net is attainable, go for it. What does that mean for the distribution of knowledge, right? Like I didn't watch every GoRuku. There are some great talks that I know I'm missing out on that I don't really get to even know about until I go to a wider conference and they're like, oh, you would love this. You know, I'm just talking to people from a larger area and they're like, hey, you might be interested in this talk. And I'm like, wow, I didn't even know that existed. And maybe that's more of like the problem that Ruby Central could solve is, you know, making that easier to surface.

Joe Leo  31:26
I totally agree with you. And I think that what do I want? What I want is for RubyConf to keep going so I can keep going to it. What I know from running GoRuko is that everybody else has got to want it too. I had this interesting talk with a few people at breakfast one day where we're talking about sponsors for the conference. And the best reason, in my opinion, Jim Remchick is going to disagree with me, but my thought is that the best reason to sponsor a conference is because you're hiring. RubyConf, not any conference, but RubyConf. And that's because a lot of people that go there are individual contributors, sort of junior all the way up to sort of principal engineer. DEF Method, we've been supporters of the conference for sure, but we don't do big sponsorship. We don't do a booth because we're not usually hiring that much that we would need to go and sponsor a conference. And because the VP of engineering and CTOs are typically not there or they're there in much smaller quantities.

Joe Leo  32:27
So then fast forward to this decade, really, but especially since 2022, and a lot of these companies are not hiring so much. In fact, although there were companies that were hiring, they canceled the hiring track hallway because there just weren't enough companies to participate. So that leaves it with, okay, sponsorship dollars are potentially down. I don't know Ruby Central's finances, so I'm not sure, but they are potentially down. And then, you know, you're doing the conference in July. You're not going to see the same numbers as you would in the fall, for example. And so you're looking at potentially a shortfall. And I know that's what they had at RailsConf. They were not making money. I don't know if that's what they have at RubyConf, but if you're facing that and you don't have money to cover that, well, then you kind of have to say, well, it sounds like the people have spoken. Right? You know, Ruby Central is supposed to sort of be embodying that. And if people aren't putting up the numbers in the form of attendance or sponsorship, then you got to find a new thing to do or stop doing it.

Valentino Stoll  33:29
The value of conferences used to be the connections you make, but it also used to heavily weigh on like the content. And I guess if the content aspect of it is distributed enough where, you know, you can almost get the same knowledge out of something online or a post or some other material, going to a conference used to be about the talks. And now it's being more heavily like about the connections, especially as like AI becomes more prevalent. You can use it as your own trainer. And so like where does that put the talk content? I could totally see that not working well for the advantage of the conference, you know?

Joe Leo  34:15
Well, I think you and I are going to differ on this a little bit because for me, I felt like conferences in writ large, technical conferences stopped being primarily about the talks for me a pretty long time ago. It was well, well before AI. We have different career paths. At some point, I stopped writing code every day and I started managing people and building a business. And that had different challenges. And so going to the conference and learning about the next cutting-edge technique might have been fun for me, but was not as necessary as going and meeting people and making new connections. When I go back to RubyConf now, and you know, there hasn't been a RubyConf since 2024, and just going and seeing some of the familiar faces that I've known now for 10 or 15 years, and it doesn't even have to be somebody that I know really well, but just know and can say hi to and can have a quick conversation with, that was the highlight for me. Secondarily, and this I think is also maybe specific to me, I can watch a talk online and I can learn something.

Joe Leo  35:14
But if I would have watched Obi's talk online, I would not have been inspired to go and do something new. That's just me. I go there live, I look at the person, I look at what's there, I kind of feel the energy in the room, and then I can leave there feeling like, hey, I want to do something different. I'm not going to say that never happens, but there's not the same chance of that happening when I go and watch the talk online.

Valentino Stoll  35:32
That's fair. That's fair.

Joe Leo  35:34
That's me.

Valentino Stoll  35:35
I think back to my very first RailsConf where I watched Lightning Talks for the first time and saw Corey Haines get up and showcase how he runs his tests, full test suite end to end in under a second. That was memorable because it was really only like it was under a minute of a talk because he just went up and said, okay, if you thought this was great and it would be advantageous to you, go to this link and here's you can read how I did it. If you're just perusing Lightning Talks, maybe you're not going to like realize how impactful that is.

Joe Leo  36:08
Yeah, right.

Valentino Stoll  36:08
Because I went in after it and I was like, how do I do this? You know, how do I make my test suite run under a second? And, you know, you try and like just figure it out and like it did make me like do a bunch of stuff that, you know, if I'm just watching a.

Joe Leo  36:23
Stop looking at the database for once.

Valentino Stoll  36:24
And then I started following him and like finding all this other content that he does and like really awesome. If you haven't read the Four Simple Rules of Design by Corey Haines, this great book. And yeah, I guess I would have lost that from just like an online Lightning Talks session, right? Especially if like, you know, with a talk that short, if you're like doing other stuff on your computer while you have it on in the background, right?

Joe Leo  36:50
Right.

Valentino Stoll  36:51
Which, you know, I'm sure lots of people do, not just me. Yeah, I totally get the engagement aspect. I guess I hadn't thought about that. How do you convince other people that?

Joe Leo  37:03
You know, with GoRuko, it was advantageous that our last year happened to be right before COVID hit, but we weren't losing money, but we were just barely making ends meet the last two years. And in the last year, it really almost didn't come together at all. And we were talking about are we going to have to refund everybody's ticket? Are we going to have to sort of break a promise to our sponsors because we don't have enough attendance? It was really touch and go. It was scary. And we didn't have any other mandate. You know, we weren't Ruby Central. We existed to put on that conference. So if we didn't make money, that was the end. So we really our hand was forced. And I wonder if Ruby Central's hand is forced here. I think probably not the same way GoRuko was, but that doesn't mean that those constraints aren't real.

Valentino Stoll  37:48
Yeah, true. Yeah, I don't know. I still miss RubyKaigi too. Just an incredible conference. And I feel like the more distributed the conferences get, only the better. Culturally, people just do things differently. Just from like dropping in a different culture, you get exposed to so much more of how people do things differently that you would just never even see. And I feel like having these like big global conferences is really important. So whatever they can do to like keep that up, I feel like please don't isolate people.

Joe Leo  38:25
Yeah, yeah.

Valentino Stoll  38:27
I'll support however I can to keep that happening.

Joe Leo  38:30
Well, so let's talk about some of these talks. I'll give the highlights. These are my highlights. If I didn't go to the talk, I can't really talk about it. Although I do have some ones that I wish I would have seen. But so right off the bat, I went to a talk by David Gillis. He talked about his tool. The tool is called Ready. And all Ready does is it makes all of your Ruby CLIs way, way faster. So it doesn't spend all of this time bootstrapping itself before it runs. And so it is really handy. It's a nifty little tool and you can get its open source on GitHub. The reason I want to highlight it is because it was one, it was very smart in how he went through analyzing the performance of all these CLIs and what was movable and what was not in terms of reducing the time and demonstrating a lot of it live, which really worked well. It was really smart and it was one of the only talks that was not about AI that I went to.

Joe Leo  39:31
Scott and I were talking one time and at one point during the conference and he's like, maybe we should have like a dual track. You know, it could become like an AI conference. Part of it is like Ruby AI and part of it is like traditional RubyConf. And I was like, what the hell are the non-AI talks going to be? Like there were almost none, but this one was and it was really good. Really sharp. Another really sharp talk from Cameron Dutro. Cameron is over at Cisco Meraki. He's a really sharp guy. He did a talk about implementing the YARV virtual machine. He wrote a Ruby interpreter in JavaScript and I really loved it. I almost never go to talks like this, but in his description of his talk, he's like, so why do we need a new interpreter in JavaScript? There are others. He's like, we don't, but I learned a lot doing it. So come and I'll show you what I learned. And it was really good and really fun to watch. So that was a cool one.

Valentino Stoll  40:30
That's really funny. I love Cameron. I've interviewed him a few times and yeah, he is just very comical. He's done some great stuff. I don't know if you ever used his Quby tool that he built over at GitHub, but really neat.

Joe Leo  40:45
I don't think so.

Valentino Stoll  40:46
Like Kubernetes helper for Ruby ultimately.

Joe Leo  40:49
Yeah, yeah.

Valentino Stoll  40:50
He's done some really interesting stuff. I'll have to take a look at that.

Joe Leo  40:54
Yeah, check it out. It was fun. He's not done all of it, but he's implemented a lot of the language.

Valentino Stoll  40:59
That's wild.

Joe Leo  41:00
Covered Scott already. That was a highlight. Brandon Weaver gave a talk. It's called We Who Remember Magic. This was a really insightful talk. So I'm going to try to summarize it the best I can, but he compared and he first took a look at how we as Rubyists were viewed by the community, the engineering community at large, say 15 or so years ago when I was first kind of getting into the language. Ruby's been around for longer than 15 years, but when Rails came on the scene and it started to really get popular and Twitter was getting built on Rails and LinkedIn and stuff like that, there were a lot of engineers looking at us saying, well, you're not really engineers. Ruby can't really scale. It can't really do this. And you know, what he was saying is that there were real valid critiques in what these basically Java and C# engineers were saying to us, but there was also a lot of sort of fear-based territorialism.

Joe Leo  42:00
Like, hey, you know, you can't come in here and, you know, start encroaching on our territory. We're the ones who make web apps. We're the ones who make things. We're the ones who make infrastructure. And Ruby can't do all that stuff. And of course, Ruby can do all that stuff. So sort of legitimate qualms aside because Ruby had to get better and Rails had to get better and things had to improve. But there was also a lot of just kind of fear mongering there. So he took that and he compared it to what we tend to do today with non-engineers who are playing with AI code gen tools and building apps and saying, hey, that's not real software development, right? That's not real engineering. We are the real engineers. And so it's kind of turning it around on us and saying, hey, are we being unfairly gatekeeping and territorial about engineering when really we are at a time when we need more and not less and we need people to come in from different walks of life and from different experiences, which is what Ruby did.

Joe Leo  43:03
We need those people to come in and to reinvigorate or to continue to push our engineering community forward. And I really liked that perspective. I think that that was accurate what he was saying and it made me think and I really liked that.

Valentino Stoll  43:18
Yeah, that's interesting. I remember my first experience with Ruby was a Spread Commerce. It was pretty wild seeing some of the syntax of like the splat. What is that? I remember seeing it and being like, what is this thing even doing? Like, how do you Google search double star, right?

Joe Leo  43:41
The splat operator, yeah.

Valentino Stoll  43:42
How do you say also like the ampersand, right? Notation. How do you Google search what is this doing with an ampersand?

Joe Leo  43:52
Yeah.

Valentino Stoll  43:53
That is definitely magic that I take for granted now, which I actually enjoy, but at the time it was a really big pain point as to what it was doing. Yeah, I feel like a lot of magic is definitely lost on maybe, I don't know, creativity or the AI doing its own things. I feel like I read a lot of Ruby and be like, I remember a conference talk where somebody presented this method of doing something and this looks like shit. That's like shit, but like more like where is like the enjoyment from reading this? You know, that was the purpose of Ruby was to enjoy it as a person. And so what does that mean now? I don't know.

Joe Leo  44:35
I think that also might have rubbed people the wrong way, right? And then and now, right? With Ruby, you know, a bunch of stodgy old Java developers saying, what do you mean you enjoy this? It's not supposed to be. Bro, this is programming. I've been writing in C++. This isn't supposed to be an enjoy, right? And the same thing can be said now where it's like, what do you mean you're having fun building this app? Building apps is supposed to be annoying. I think there is something to learn from that. Alicia Rojas, at Telos Labs, she gave a talk called Convention over Hallucination, which was the closest thing to Obi's talk that I saw. And she even referenced, she had a tool that Telos Labs built and she was like, yeah, Obi's tool. It's just kind of funny. But she said, so her CEO came to all their engineers at some point, I think this was like December of last year and said, you don't write code anymore. This is not your job. Your job is now that you manage these agents, which was a kind of big blow at the time to a lot of people's egos.

Joe Leo  45:35
And so she set about, and so the company set about building a lot of these tools that they've open sourced and you can go check them out. And I have, and I've used one of them already, which is their Rails template. And these are things that we use to ensure deterministic outcomes and non-deterministic piece of software, which I really liked. The biggest takeaway there was because I kept thinking wrongly that any kind of this is how we do AI better talk was going to inevitably come down to you need to put in better markdown files, or you need to put in better prompts, or you need a better knowledge graph. While I think all of those things are helpful, it bothers my engineering brain because I'm like, well, you could do all those things and still not get the result you want. That's possible. But what Alicia was actually doing was using off-the-shelf open source software, Ruby Cop, you know, things like that, and building in rules based on where software, where her agent had gone wrong, saying, okay, now that I've seen this happen, we're going to build in rules to make sure this doesn't happen again.

Joe Leo  46:33
And so there was deterministic outcomes every single time. It was just another linter to ensure that you were getting the result you wanted, which I thought was great.

Valentino Stoll  46:40
I really liked that a lot. It makes it easier to bitter a lesson, right? Like, can you remove any one of those linters on a new model and it'll work without that linter, right?

Joe Leo  46:51
Oh, yeah. That's a good call. Yeah.

Valentino Stoll  46:53
Because I'm always thinking that I build all these things to make it better, but like, do I need any of them? Often you just got to be like, all right, start a brand new session with nothing in it, isolated Docker container, and be like, all right, I'm going to run on a fresh agent and see how it does. And you've got to be doing that all the time. And I don't have a good solution to how do you regularly do that and check.

Joe Leo  47:17
Yeah, just blow it away. I know. It's hard.

Valentino Stoll  47:19
Because once you get the app, you're like, I guess who cares?

Joe Leo  47:22
Right. Yeah. You move on. I know this happens all the time. It happens to me too, where I'm like, man, that took a while. I don't want it to take a while the next time, but I'm also done. So maybe I'll just move on to the next thing. You know, and this is the same, you know, looping argument that goes on in our heads as developers every day. It's like, well, do I pause to make my tools better or do I go and get the job done? And yeah, we have to make good choices there or not and just live with the outcome. Either way, we got to live with our choices.

Valentino Stoll  47:49
This reminds me too, like there's a latent space episode where they interview somebody that worked on Symphony from OpenAI. For those who don't know what Symphony is, it's like an orchestration management for like getting codecs to build and manage its own work task list. And you can like fan it out and it'll just like chew through whatever you line up and automatically automate the whole like completion goal-driven pipeline. They use this to some aspect for ChatGPT and other OpenAI related projects. You start off with just the specification and then you use basically your agent harness to build the app. And so like there is no app really that they manage, but what they find is it's very difficult to keep up that specification because any new model, any new thing runs the specification different. And so like even after you've gotten the specification really tuned, like they went through this iterative pattern of trying to like make sure the specification was solid and perfect.

Valentino Stoll  48:54
When a new model comes out, it generates the specification different.

Joe Leo  48:59
That's fast.

Valentino Stoll  49:00
And so like I think of this too, like, okay, like conventions are definitely like the right approach of like trying to channel the things in. But at one point it's like, okay, then you're just managing your conventions. Maybe that's better. And I think that's probably true because what we've been finding at Zara is like, okay, you ship something and if it's not working, something's going to pick it up and fix it. Oftentimes you can ship the fix faster than anybody even notices it. That happens a lot and that is actually faster. We actually see less and less of it occurring, right? And so at one point, if you can hook up something that is improving your conventions too, then what are you doing?

Joe Leo  49:46
We're back to Jessica Crews, right? Like.

Valentino Stoll  49:49
What is our purpose here? Yeah, I get it. Is that going to be like the future of our next few years of conferences? Questioning our identity and role?

Joe Leo  49:59
Maybe. You know, it moved faster than I thought. So 2024, Scott and I were talking about this. 2024, Scott spoke at Ruby Conference. It was November 2024. So we're only talking about 18 months-ish, 19 months between Ruby Conf. And in that conference, there were very few AI-focused talks and a lot of the AI talks were very elementary. And 18 months later, every AI talk is really like, here's how it's really working in the world. Here's how you can get better results out of it. Here's how you could do this, do that. Here are the tools that we've built using it. Here's what we've accomplished using it. Even Obi and Dave, Dave Thomas and Matz, both were talking about new things they had done with AI and really interesting stuff. So that moved quickly. Hopefully we get through the existential dread quickly. I don't know, but it's real. People are feeling it and it's real. You know, some people are excited, some people aren't, but there's no questioning that it's different. It's a different world today.

Joe Leo  50:58
This was interesting. So I'll close out my discussion at Ruby Conf with this. So Dave Thomas gave the closing keynote. It was, I wrote here, the other non-AI talk that I went to. In truth, there might have been three if you squint, but really there are two real ones. And Dave, he was fired up, pissed and vinegar, like I haven't seen him in a while, talking about what not to do in Ruby, namely all these different reasons to not write classes in Ruby and said, hey, this is an object-oriented language and we're kind of ruining it by making everything a class all the time. He went from like pretty obvious examples where things should really be a singleton all the way to like some pretty specific and nuanced examples. He closed his talk. This is the most interesting thing to me. He closed the talk with this recommendation of how one should approach programming, building programs in Ruby, which goes all the way back to what he had recommended from like 12 years ago maybe when I took like an object-oriented design and Ruby class with him.

Joe Leo  52:00
And it was kind of like start in one file, write your test in one file, write your objects in one file, start there. Don't create an object at all. Don't create a class, you know, at all. Build from there. Let the requirements of the outcome that you want drive your design until it is too complicated to be in one file and then split it into two files. And I thought about that and I thought, how am I going to get my AI to do that? And I'm thinking that and then I walked out because there was like an ice cream social after that and every single person I talked to about that talk said the exact same thing to me. Like I was walking there thinking, well, how do I get my LLM to do that? And maybe I could like snapshot what he was saying and put it into like a context, you know, a markdown file somewhere. And that I think just was like so full circle about like where we are today. This incredible advice from a master of languages and of object-oriented design and of Ruby giving this talk and us all saying, okay, great.

Joe Leo  53:09
I don't actually know when I would apply that in real life. I need to do it.

Valentino Stoll  53:14
You know, that is great advice and it reminds me of the late Jim Warrick hearing him give talks.

Joe Leo  53:20
Yeah, Jim did too. And I've done it and it produces really interesting results. It doesn't lead you down where you think it might have if you were to set it up on your own.

Valentino Stoll  53:28
Yeah. And you know, people think, well, then how do you circle back to like dependencies and things like that? And Rake was really, or is, I don't know if people still use it, but it was really great for like the simplicity of its build structure, right? Where you just like reference one name and have it namespace and it would do the thing before, run that task before the one that you were dependent on. And it was just files, right? It was all just, sure, you wrapped it in an object, but the concept you didn't really need to. Like you could just require the file and have it do the same thing. And I think ultimately that's what it was doing under the hood.

Joe Leo  54:11
Yeah. And when you had it in non-Rails projects, it was really fast. It was all the rail loading in all Rails context and it's slow.

Valentino Stoll  54:16
Forget the application dependency. I was always trying to figure out, okay, how do I make this work without the application?

Joe Leo  54:23
I know. I know.

Valentino Stoll  54:25
And to be honest, to Dave's point maybe, I was always trying to find out, I don't know how many iterations I went through of this, of how do I load the minimal amount of Rails as a framework in order to run anything? Can I do this without Rails?

Joe Leo  54:40
Yeah. Yeah.

Valentino Stoll  54:40
Right? It was like always the first thing.

Joe Leo  54:42
Well, right? To your point, one of his biggest examples was, first of all, he said, if your class name names a Michael Feather's design pattern, it is wrong. It is wrong. It's not a class. But he said, stop inheriting from active record. Just include the stuff that you want, which I remember trying once. And it's an interesting thing. I don't know if I necessarily recommend it. I mean, it'll make things faster and it'll make you more intentional. Then every time you need a new thing, you end up like kind of having to go through, scan through the file. Maybe it's easier now with AI, but.

Valentino Stoll  55:16
Yeah. You know, like thank God for Xavier who makes site work. That was the greatest thing, like lazy loading dependencies. That introduced such a performance gain, but it's still like, all right, how many thousands of files classes are there in Rails?

Joe Leo  55:31
Oh, yeah.

Valentino Stoll  55:31
And you still like, yeah, if you reference active record at all, like that you need the everything. And how do you get just like, all right, I want to test like just this database table, not all the methods, just the ones needed for this class. Yeah, you can't do that. And so like I definitely sympathize with Dave's frustration, right?

Joe Leo  55:51
Yeah.

Valentino Stoll  55:52
I made like this thing called Rails Lightning three different times. Basically lazy loading constants from Rails in different ways. And maybe AI can do that now, right? Like maybe that is the answer. And like everybody's Rails new starts with bare and almost nothing. It can keep track of in the agent's perspective what dependencies it can include, right?

Joe Leo  56:16
What's needed.

Valentino Stoll  56:17
Then over time, it could build up just the bare minimum what's needed. I feel like maybe that's the right approach.

Joe Leo  56:23
Oh, that's cool. Maybe that's a Ruby AI podcast open source project.

Valentino Stoll  56:28
I've got to make an agent that can just start taking our ideas here and making repos for us.

Joe Leo  56:35
I know. Oh, that's a good idea. That can be an Arkham agent. We're going to work on that. Yeah. The first Arkham agent is going to be a COO that I'm going to make into a total jerk. It's just going to yell at me for doing things manually.

Valentino Stoll  56:51
I can't believe you wasted your time emailing this person.

Joe Leo  56:56
You're killing us. You're bleeding us dry. Your time is valuable. Stop doing that.

Valentino Stoll  57:01
It should just start sending you a bill.

Joe Leo  57:04
Yeah.

Valentino Stoll  57:05
All right. You wasted X number of hours at your rate. You owe me $200.

Joe Leo  57:11
Yeah. Yeah. All right.

Valentino Stoll  57:15
Yeah. Let's leave it there. That'll be our next thing is we should reach out to all these people that gave the talks that like most remarked to you to have them on.

Joe Leo  57:23
Yeah. I've got a short list now and we'll also, we'll add you on the LinkedIn when we talk about you. So this is your invite now. Don't wait for us. Just reach out. We want to talk to you.

Valentino Stoll  57:34
Yeah. Totally. Because you know, we can only talk to each other for so long.

Joe Leo  57:37
I know. Well, I could actually, it turns out we could talk for a long time, but what we really need is we need to get out of the echo chamber. You got to get on the show and tell us what we're doing and thinking wrong. Nobody wants that. Nobody wants that. Somebody come on the show and tell us where we have our heads up our asses about. That's what we really need. All right. Well, thanks everybody. It's great having you and it was great talking to you, Vi. Talk to you again soon.

Valentino Stoll  58:00
Stay human.