I am empowered by these tools, and that's how you create real value for society and the economy. I'm focusing on decentralization, individualization, and spoke models—talking about making yourself better at your day job and, even better, re-engineering the workflows of your life. That's where the real value boost is. The third takeaway, and the reason for this presentation, is that I sincerely believe the barriers to achieving all this have collapsed. The tools have already been invented; it's now about getting people to understand what's out there, assemble their own tools, and put themselves on a completely different trajectory.
Now for the fun part—my adventures. My personal agent first came to life almost exactly three months ago. I got caught up and opened the door, but the media gave me my job. I knew that wasn't practical, because security was an issue. Then someone introduced me to NanoClaw, and I think we'll hear from Gabriel after me. As a geek and a tinkerer, I like stuff I can grasp. NanoClaw has a very short code base that even an amateur like me can read and understand. It's containerized, and, as a surgeon, I know there's no such thing as a routine operation—things will go wrong, things will break. When they do, you want the risks contained. The containerization and understandability were vital for me. It's simple: go to GitHub, download it, and you're set.
Another attractive part is that there are no conflicts, because you rely on the LLM to handle all spoke tailoring and customization. Everyone running an instance of NanoClaw is running an individualized system. This is both good and comes with complications. Let me tell you what I did with it. NanoClaw provides the platform to communicate with my agent. That's not rocket science. What I was really after was how I could use it for my daily life.
Let me give you an idea of my daily life. This month, I’m visiting 12 countries. I’ll have to meet hundreds of people and understand each country's economy, geography, culture, history—war and peace. I need to know people as individuals, not just from a brief. There’s a huge cognitive overload on every deployment. The question is: how can I turbocharge this process so that, if I need a fact or background, I can get it anywhere, and even go down the rabbit hole if needed?
It’s all about coping with this overload. LLMs are useful for analysis, extraction, expression, and certainly for drafting briefs, speeches, and formulating answers to questions—including parliamentary questions. Three months ago, it included old debates and empowerment. It was extremely impressive to see both the questions and answers generated with AI. With respect to my colleagues in parliament, some AI-generated debates are far more incisive.
My agent communicates with me through WhatsApp, using software called "Bailies." This might not entirely comply with what Meta or WhatsApp want us to do, because it simulates how we use WhatsApp in browsers or on laptops. It’s a pseudo terminal, in a sense. The real frontier for people like me is memory. Fortunately, I discovered an obscure piece of software called Neement. I’m still unsure about the name, but it’s a memory system with wraps—it has entities, edges, causality, temporal relationships, and so forth. I didn’t want to be confined to keyword searches. The ability to run a local language model with embeddings means I can go beyond basic search.
With these elements, Whisper is the easy part. I didn’t want to only type; I wanted to be able to speak, and for the agent to speak back. My dream is for my agent to answer supplementary questions in real time, though I'm unsure about the legality of that. But if it happens, you heard the idea here first. Now, I can curate material, speeches, transcripts—especially my own contributions—get it into the system, have it digested and extracted, and put it into the memory database.
Around the same time, Andrej Karpathy introduced supervised weak-key generation for LLMs, so I added that too. For the user experience, I use Obsidian, partly because it works with Apple iCloud, which gives me a personal cloud. All the wikis extracted from my personally created database are available to me instantly. Remember, the key is personal understanding. I have a memory system, a communication system, an analysis system—all integrated nicely.
Last week, I found this setup incredibly useful—creating profiles, traveling, drafting speeches, even today's presentation and the slides were AI-generated. It turbocharged the pace at which things can be done. As a practitioner—not an engineer—but someone with a day job, it's genuinely useful. I can honestly say I've never switched it off. NanoClaw moved from version 1 to version 2, and when version 2 was released, I left version 1 running on another computer. My most-used agent runs on a Raspberry Pi that’s at least two or three years old, with just 8GB of RAM. This demonstrates accessibility, personalization, relevance, and practical use.
The barriers are gone. I assembled all this without writing code for Claw, Bailies, Neement, Whisper, or the credentialing system. If I keep saying it's about coding, that's inaccurate—it's about assembling tools. I didn't write any of these tools, but I read through the code. NanoClaw requires approval every time you grant BASH access to the agent, so I scan through it. It helps to understand what's going on, even without editing or writing code directly.
My approach has been to learn by doing. It's not enough to just read headlines and summaries. If you're interested in something, get your hands dirty and learn by doing. Since the barriers to entry are now so low, everyone should embark on personal experiments. Claude came up with a quote—I'm not sure who said it first—but I agree: you cannot govern a technology you’ve only been briefed on. You need to get your hands dirty to understand both its potential and its limits and problems.
A few other points: there are constraints. Depending on LLMs, and frankly, the current prices for AI engines, we know we're enjoying the effect of a subsidy. Tokens aren't cheap, compute power is limited, electricity prices have risen, and laws do not always help. We should avoid using LLMs for every problem or every step of a solution. Like the old saying, "If you have a hammer, everything looks like a nail." There are good economic and design reasons to use LLMs, but deterministic systems and expert rule-based systems still have a role. Personally, I believe a neuro-symbolic system is needed, not just LLMs, and I sympathize with Yann LeCun's perspective.
LLMs are great, but that's not how we've solved things in nature. The human brain has fewer computational layers than many current large language models. As an eye surgeon, I see cortical computation for vision and language is often based on much more efficient structures than today's AI systems. Ultimately, these are pattern recognition systems with attention and memory. What looks like simple abilities gives rise to conceptual understanding and language. Therefore, approach this with humility, do your best, and improve your productivity in the long term.
We are among the most privileged generations to witness this technological revolution. Tools matter more than models. I want NanoClaw to make all models first-class citizens. There are reasons for this, which we can discuss later. Security is also important. Even if someone hacks my system, the worst they’ll get is my phone number.
