Interview #12: Mike Alnakhaleh

Noah Weisblat: So I'm here with Mike. Me and Mike met a couple weeks ago at the Now Foundation event where we were both panelists. Mike, do you wanna introduce yourself and tell us a little bit about yourself?

Mike Alnakhaleh: Yeah, of course. Thank you, Noah. And yeah. Hey everybody. I'm Mike A. I'm currently working at Cardinal Health at this time. I'm currently doing a contract there through a local contracting company. I also do AI consulting as it comes up, so if there's any work for that available, I'm happy to do that. And in general, I would just say I'm a fairly enthusiastic software engineer, software developer, who's really got a good focus right now on developing web applications through the use of LLMs or in a way that pulls in LLMs or image models or different types of AI models, maybe even all the way up to like classical machine learning models.

Noah: Very cool. And obviously the AI space has been ever-changing over the past two years, but just tell me about your general thoughts on the AI space, where it is right now.

Mike: I think I'm of a similar mind as I. I think it's both simultaneously overhyped and underhyped, as far as things go, like depending on who you're talking to, especially. But yeah, that's really where the industry is because of the open questions of where things are going, where they're at right now, that whole situation.

Noah: Yeah, that's a good take. Talk to me about which tools you're using right now for AI-wise.

Mike: Right now I've primarily been using quite a bit of Copilot at work, which I'm not in love with, but it is what it is for Visual Studio Code's Copilot Chat setup. So that's all right, but it I could do better with things like Cursor, and in a lot of my side projects, I tend to use that quite a bit. I have been dabbling a tiny bit with N8N just because automation is something I feel like I should at least know about, and I've had experience with QA/QE-type work in the past. So it was nice to dabble a little bit with that again, with more of an engineering or developer focus in that process. And beyond that, I've mostly just done a lot with agents in the past. I had worked on BabyAGI back in the day and mostly just been trying to keep up with the agent space without having to like, get too committed to one individual framework, to be honest.

Noah: Cool. When you talk about AGI, there's a lot of different definitions roaming around. There's the one that came from the podcast the other day. Where do you define AGI?

Mike: That's, I think that's an excellent question, because I personally, I feel like I have a pretty conservative definition of AGI in the sense that I feel like my definition hasn't radically changed. 'Cause I operated under the understanding back in the 2010s—there was this idea of AGI being like an AI that can do a variety of different tasks about as well, if not better than a human. So I would argue we've already achieved AGI personally, and what we're really trying to pursue is ASI (Artificial Superintelligence). But although most people I think define AGI with different factors, and I think we're relatively close to achieving those factors, but we can go into that if you would like, or we can go through another question.

Noah: Cool. Yeah, let's dive a bit deeper into that, 'cause I don't think we talked about it beforehand, but I think ASI is something that's pretty interesting. What do you see happening when ASI comes along in relation to the workforce? I know we're already seeing shifts with the workforce with some layoffs happening because of AI, or loosely related to AI. Where do you, and I personally, I think the actual change will come when we do start to see that ASI come into the picture. How do you view workforces changing once AGI or ASI fully becomes involved?

Mike: I see this as sort of almost a continuum process. So let's actually take a step back from ASI, let's go to what I think most people would practically define as AGI. So I think the current models we have are not quite at what that definition is for most people, but I think we are like, maybe I think two at most, two or three different breakthroughs away from what that would be for most people, especially if we were to add like a continual learning system with these LLMs, we were to make them fully embodied through robots like humanoid robots, or we were to be able to make them more consistent, reliable, and operate in over a longer period of time. So say for example, instead of on the order of hours, we're talking days, weeks, months, having them continue to operate without a significant performance drop-off. I think all three of those things are achievable. So I think we're rapidly approaching what most people would define as AGI within the next few years. So in that process, I think you're absolutely right. We're already gonna start to see a gradual pickup in things like unemployment or automation leading to lower jobs coming in over the long term. In the short term, I think a lot of that which has been happening right now is more due to economic factors than the actual issue of AI replacing people in the workforce. And I think a lot of employers are still trying to figure out internally and on a sort of organizational level how to apply AI for their business needs. And I think that's where a lot of the potential value is, at least on the enterprise side of things. But I think a lot of startups are really starting to crack into that space to make that more viable. Does that answer your question? I feel like it went a bit off from the ASI point, just 'cause I feel like it's harder to really conceptualize for a lot of people as opposed to what I think most people would be okay with defining to be AGI and what could be the effects from that outside of what we're already dealing with in the current macroeconomic situation.

Noah: Yeah, that's a good answer that encompasses it into what people are gonna actually deal with and see. With enterprise AI, I've talked to somebody last week who was very big on RAG (Retrieval-Augmented Generation) and that use case for enterprise, being able to search something at the click of your finger, any documents within the company. Obviously that's a huge use case for enterprise. Would you say that's the biggest use case for enterprise, or are there other use cases with AI in enterprise that you could see as bigger right now?

Mike: Absolutely. Absolutely. That's an excellent question, Noah. I wanna start with, yeah, just RAG. So thinking about this in the context of an internal ChatGPT that you can just use semantic search on to get an idea of what your documentation and your general like internal infrastructure around certain processes looks like—like standard operating procedures, technical documentation, HR stuff, things along those lines, that type of documentation data. I think that's gonna be a primary use case for enterprise that I think is probably easiest to approach from a developer standpoint and from a value proposition standpoint from the actual enterprise side. So from the business side, people are like, "Okay, yeah. Being able to search through our data in a way that feels almost conversational as opposed to hoping we get an acronym correct on a keyword search that explains what's going on." Because there's so much jargon in enterprise that a lot of people don't realize, so much like technical terms and language or processes that are just assumed that you know right outta the gate. So having a tool where you can just pull in that information, even if it's just from SharePoint through a Microsoft Copilot agents, that's enough on its own for a lot of people to get a significant amount of value right out the gate. And I think that's usually a good first project for a company, an enterprise organization, trying to get into AI to start with. That's a very easy one, easy value-add, easy value prop, and it's not difficult to set up from a development standpoint.

Noah: That makes sense. It's a good answer. All right. Let's pivot a little bit, talk about kind of some of the new developments with AI. Humanoid robotics. I'm fascinated by robotics, I'm interested in robotics. They recently launched a new robot the other day. My grandpa actually ordered one. But yeah. Tell me about Neo and humanoid robotics and where you see everything there right now.

Mike: Okay, so I, first off, I'm really excited for humanoid robots. I feel like that's gonna be one thing we really have to nail down if we want to say that we have AGI, for a lot of people I feel like embodiment is a core element of that. And I think Neo is in a very interesting situation where they absolutely need to get the robots out because that's how you really improve the autonomy element. That's gonna be a major element of being in the home and actually getting used to that in an environment. Even if on release, there's a significant amount of initial teleoperation. If the company can also expose that to users to where, "Hey, I have a robotic body that I can pilot and teleoperate with a VR headset." That in and of itself, like even if you don't include autonomy, there's a lot of value in that, right? There's so much utility you can get on something like that. So if we just take away autonomy, just having that as the value proposition for a lot of people. And then giving... once they've gotten that, it's okay, maybe if you opt in to certain anonymous sharing of data or maybe like weights if you wanna fine-tune the model, if you're a bit more technical. I think once you can establish that niche or that market base, you can really expand upon it as the autonomy improves and grows. And I really hope that's the approach OneX is taking, because I think initially you're gonna rely a lot on teleoperation with these robots. And I really think there's a lot of value-add, assuming you go in with that expectation in mind because ultimately the LLM in the long term or built-in for the movement of that robot will improve with the data that comes in.

Noah: Yeah. I think that's what a lot of people are missing, that it's going to improve over time. Yeah. Any other robotics use cases that you can think of or that come to mind outside of that kind of in the home humanoid robot?

Mike: Oh yeah, totally. Even if we just completely ignore the customer side, right? Even if we just completely assume or B2B the whole way through, so much value in applying humanoids and there's a lot of immediate utility to applying humanoids in the manufacturing process. That's part of the reason why we're seeing Amazon's layoffs. Now granted, not all of those layoffs were associated with people working in their different logistics centers or hubs, but a good portion of that was due to the automation that is starting to occur because it's a lot easier to set up automated robotics in a controlled environment like a logistics hub or like a warehouse of some kind or some sort of manufacturing center. It's much easier to develop and automate robotic workers in that controlled environment versus a house or a variety of different houses that can have so many different like unknown variables that you could run into. I think there's gonna be a much greater adoption initially in factories. And then once we really start to see the prices drop for a lot of these leveraging these tools, we're gonna also start to see humanoids on a consumer side start to become more viable and available and autonomous.

Noah: That's awesome. I'm excited for it. Likewise. Talk to me about Atlas. So when Atlas, the GPT browser dropped, I was super excited 'cause I like using GPT, it's up there with Claude for kind of my favorite LLMs to use, and I was messing around with it a lot. And then I did a lot more reading on prompt injection and the potential threats of that. So talk to me about Atlas at first, then what is the threat of prompt injection with the agent mode on AI browsers?

Mike: Absolutely. So to get into context, Atlas, the Atlas browser is specifically, if my understanding is correct, a ChatGPT browser that implements agents from the ground up. Correct? So that's, that was my understanding going in and from what I've seen from initial reviews, 'cause I haven't personally tried it myself. I was also worried about those potential security concerns early on. I'm a big security guy, which is why I'm taking these approaches with things. But with that, I was like, "Eh, I wanna see how they handle like prompt injection," because that was a big concern. And then the Brave browser just mentioned, "Yeah, this is an incredibly insecure browser. You guys need to be careful here." So, now that all being said, as long as you're not adding in personal credentials or any like credit card information, stuff like that, any PII, then you can use these browsers just fine. And I'm sure it's just like a Chromium-based browser. Yeah. But yeah, I didn't even know. But that's how you spin up a browser like that, right? Yeah, I wouldn't, I was... So if that's the case, then it's just relatively easy to start to think about, "Okay, I can definitely see this implemented in Chrome, or I can definitely see this implemented in a browser like Chrome." And I'm curious what Google's response is likely going to be in the long term, in addition to what is gonna be done by both Perplexity, OpenAI, and other people wanting to set up an agent-focused browser, how they're going to thread that needle for both security purposes as well as usability, because those two things tend to go at odds with one another. So I'd like to see how they thread that needle and balance it all out.

Noah: That makes sense. Okay. One or two more questions. Are you a, obviously you're a software engineer. Are you Cursor, Claude, Code? Do you use any of these tools? Which ones kind of your forte?

Mike: I generally prefer Cursor and Windsorf. I've had great pleasure using both. I've heard good things about Codex recently, so I have some friends who've used Codex a lot more now because of the cheap T5 update. And I've heard good things about Claude Code across the board. I've never really heard any complaints, but I just haven't personally tried them myself. I'm usually pretty happy with Cursor and Windsorf. Sure.

Noah: Okay. Agents, AI agents. When would you recommend using kind of building out an AI agent versus more like a simple automation? Where do you see the differences between those two and the value in going full agent mode?

Mike: That is an excellent question, Noah. I'm gonna first start with N8N as just the use case for an automation. I think when you want something with a clear, consistent outcome that you can control, and it's more of a quantitative thing versus being a qualitative thing. Like for example, if you want to include an LLM to do a qualitative analysis of data that's coming in where you maybe want to summarize it or translate it or expand upon it in some capacity, then I think LLMs are really good at that element of that automation pipeline. However, with that not needing to be the case or requiring certain deterministic outcomes, you really wanna lean more on the automation side at that point. And this is why I would include just a pure automation if, say, for example, I just needed to know the number of data points that I need to say find within a certain CSV file, right? Something more deterministic like that. I just wanna set up a simple like function or automation that pulls from a CSV file that I just plop in somewhere, right? But agents, I would say, again, more of a qualitative analysis where you're doing something to the data that doesn't really require it to be strictly set or strictly typed and something where there could be mistakes, but it's not a huge deal if that is the case. Like, I wouldn't want something that is too mission-critical right now, which is why I would lean more on automation like you've said.

Noah: Okay. Interesting. Cool. All right, Mike, I appreciate your time. Any plugs? Where can we find you?

Mike: Yeah, I'm pretty active on LinkedIn for the most part. Definitely feel free to reach out. I'll be sure to hand that link off over to Noah. And I do consulting work occasionally. If I'm open to opportunities for that, definitely reach out. I cannot guarantee that I will commit to a project, but I can guarantee that I will at least try to give advice so you get some value out of it regardless.

Noah: Excellent. All right. Thank you, Mike. Thanks for coming on.

Mike: Yeah, very well. Take care.