I installed "Crayfish" and don't dare turn it off! Singapore Foreign Minister's Speech: Officials who haven't used AI Agents, dare they talk about national governance?

📄Full Article· Automatically extracted by trafilaturaGemini 翻譯4078 words

On May 16, Singapore's Minister for Foreign Affairs Vivian Balakrishnan personally demonstrated at the AI Engineer Singapore conference the AI second-brain he built using the open-source framework NanoClaw, running on a Raspberry Pi with only 8GB of RAM. Three months in, he said a line on stage that is now being widely circulated: "He no longer dares to turn it off." (Background: Karpathy backs HTML killing Markdown: 1/3 of the brain is GPU, and AI's endgame is drawing directly to the screen.) (Context: Hermes Agent ships a massive 800-commit overhaul: a single proxy command turns three AI monthly subscriptions into a local API.) A few months ago, on the stage of the AI Engineer Singapore conference, stood a gentleman who didn't quite look like a tech person. He first introduced himself, saying he was a "retired ophthalmic surgeon" who had taken too long a detour into politics, then added that standing in this venue he felt like an impostor. In the audience were attendees of the AI Engineer Singapore conference, the overwhelming majority of whom were engineers, developers, and people on frontier model teams. And this "impostor" was Singapore's incumbent Minister for Foreign Affairs, Vivian Balakrishnan. Three months ago, he hand-assembled a set of AI assistants for himself, running on a Raspberry Pi he had bought two or three years ago. That machine has only 8GB of RAM. After using it for three months, he said this on stage: Honestly, I no longer dare to turn it off. Balakrishnan said he was initially swept up in the OpenClaw craze too, but quickly judged that path wasn't viable for him, because given his position, security is the most important issue. Later, someone introduced him to NanoClaw. NanoClaw is an open-source AI agent framework, built on Anthropic's Claude Agent SDK. It can connect to WhatsApp, Telegram, and Slack, has memory, and can schedule tasks. The whole project's code is short — about 500 lines — and stuffed in its entirety it only takes up a small fraction of Claude's context window. Balakrishnan said NanoClaw's code is short enough that even a fool like him can read it, and it's sandboxed. He explained this choice through his experience as a surgeon: on the operating table there's no such thing as a routine surgery — things will go wrong, things will break — and when it breaks, you want it to break within boundaries. Sandboxing locks the agent into isolated boxes, which corresponds exactly to his instinct of "even when it errors, it must be controllable." Another thing he likes about NanoClaw is that there are almost no config files. All customization work is handed off to the LLM, so each person running NanoClaw is actually running a system they themselves have tuned that looks different from anyone else's. To understand why he needs this thing, you have to look at his daily life. In this month alone, Vivian Balakrishnan has to visit 12 countries and meet hundreds of people. Each place he goes he has to grasp that country's economy, geography, culture, history, and its wars and peace. Every diplomat carries a tremendous cognitive load. He wanted a system that lets him pull up background information whenever he needs it, and lets him drill down further along the thread when he needs to go deeper. NanoClaw provides the underlying platform, letting him converse with his agent through WhatsApp. This part relies on a piece of software called Baileys, which emulates the way WhatsApp runs in the browser — essentially a pseudo-terminal. Vivian Balakrishnan himself half-jokingly said this probably doesn't fully match the way Meta wants people to use WhatsApp. He thinks that for people like him, the most amazing function of true AI is memory. He found an obscure piece of software called Mnemon, a memory system built with a graph structure. Inside there are entities, and between entities it records causality, temporal order, and semantic relationships. To avoid being shackled to keyword search, he runs Ollama locally with an embedding model, giving him built-in semantic search. The voice part is handled by Whisper, because he doesn't want to be limited to typing — he wants to be able to speak, and he wants the agent to be able to talk back to him. After stitching all these together, he started feeding it data. His own scripts, transcripts, especially his speeches in Parliament — all organized into the memory database. Around the same time, Andrej Karpathy proposed an approach for LLM supervised wiki generation, which he added in as well. The interface uses Obsidian, partly because Obsidian can connect to Apple's iCloud, effectively giving him a personal cloud he can take anywhere. The LLM in this system is responsible for analysis, abstraction, expression, and drafting. Drafting briefings, speeches, replies, including parliamentary questions. He said three months ago he ran a test, throwing an entire parliamentary debate at the AI, and the questions and answers it generated were impressive. He said with a touch of humor that it was a bit awkward for his colleagues in Parliament — some AI-generated debates were sharper than the real humans. He even mentioned a dream: one day letting the agent answer supplementary questions in Parliament. He himself admitted he wasn't sure whether this would be legally feasible, and said if it ever does happen, everyone will know he was the one who said it first. Having talked through the tools, Vivian Balakrishnan brought the focus back to three sentences. The first is about the boundary of outsourcing. Balakrishnan said, we now live in an era where many things can be outsourced. Computation, processing, memory, replication, the dissemination of knowledge — these can all be handed off to machines. But there's one thing that cannot be outsourced: your own understanding. He then pushed this distinction to a sharper point: if you are in a position of power, you can delegate the work, but you cannot delegate the accountability. Put back into his role, this becomes concrete: a Minister for Foreign Affairs can let AI help him draft replies, organize a country's background, extract intel on negotiating counterparts — but when he actually sits at the negotiating table, when he is grilled in Parliament, the one who carries responsibility is him, not the AI. AI can lay information out before him, but turning information into judgment, and judgment into decisions — no one can substitute for him personally. The second point is about where value is actually created. He cited a short article in the Financial Times by Cambridge University machine learning professor Neil Lawrence. Lawrence's hypothesis is that the real place where value is created for the economy and society is at ground level — one workflow after another, one industry after another, finally landing on individuals. Balakrishnan put it more bluntly: he knows the audience is strong, and he knows the people building frontier models are brilliant, but the real returns come when ordinary people start using these tools. Teachers, lawyers, technicians, managers, doctors, even ministers — people who understand their own job and are augmented by tools — those are the ones who create real value for society and the economy. This is the same logic as what he does. He didn't train a new model. What he did was take a diplomat's workflow and reconnect it using off-the-shelf tools. How memory is stored, how data goes in, how voice is transcribed, how the wiki grows — he rearranged every link. Value plays out on the ground. The third message, and the reason he said he gave this talk, is that he truly believes the threshold to do all of this has collapsed. He treats himself as evidence. He built this AI agent system, and he didn't write Claude, didn't write Baileys, didn't write Mnemon, didn't write Whisper, didn't write that credentialing system. He didn't even write the glue code. His use of the word "assembly" is meaningful. OpenClaw is too big, too hard to grasp; NanoClaw compresses the same functionality into code that one person can read through, so "assembly" becomes possible. He downloads, plugs in, approves permissions, scans the code, and then lets the LLM handle the fine customization. He describes his method as "learning by doing." He encourages everyone: it's not enough to sit and read AI articles, look at headlines, look at summaries — whatever you're interested in, go try it yourself, hands-on. Back to the political question: why would a Minister for Foreign Affairs spend three months assembling an AI assistant? The answer is hidden in a line he quoted. He said Claude generated a sentence for him that he was at first half-skeptical of, and the sentence was, "You cannot govern a technology you've only been briefed on." Vivian Balakrishnan specifically threw this line at his colleagues in government. An official who has only read briefings, and an official who has personally gotten an AI agent running, inspected the code, and been interrupted by bash permission prompts — their understanding of this technology is not on the same level. The former knows AI exists; the latter knows AI's potential, its limits, and where it will go wrong. For someone making AI decisions on behalf of a country, this difference is the success of a decade-long plan versus wasted budget, the rise or death of industries, national competitiveness. He quoted the positioning that Singapore's Deputy Prime Minister Heng Swee Keat laid out in the Economic Strategies Review Committee: Singapore probably won't stand at the frontier of model development, but it can stand at the frontier of large-scale deployment. What he himself subscribes to is "edge deployment." He said he is a surgeon. He believes in getting hands on, in fixing things, because that's where lives are actually saved and value is actually created. Following this logic, the goal of public policy becomes democratizing these tools. If you believe value is created at ground level, on every independent individual, then the only way to drive it is decentralized and bottom-up. He said this is also why he is there in person. He noticed that this AIE Singapore conference was put together less than three months ago, organized by 65 Labs, a grassroots developer community in Singapore. The people here doing AI may not be doing it as their main profession — this is a hack to remake every industry. Vivian Balakrishnan used three months to prove that a 65-year-old who can't code can do it. He uses it every day, and in his own words, he no longer dares to turn it off. His speech on stage wasn't very long, but it was shocking enough. A nation's minister, candidly acknowledging the power of AI agents — when AI influences politicians, it influences the governance of an entire country, and then it influences the entire world's situation. Many ministers don't even dare touch an AI web chat window. Over time, what about the countries that haven't caught up? The countries that fall out of formation? This isn't just an information gap — it's a "governance-tier gap." Gaps easily become tragedies. Be careful. What is NanoClaw? How do you install it? NanoClaw is an open-source AI agent framework of about 500 lines, built on the Anthropic Claude Agent SDK. It supports messaging integrations like WhatsApp, Telegram, and Slack, plus a memory system and scheduling. MIT-licensed and free, it can be installed in under 5 minutes. Why did Singapore's Foreign Minister choose NanoClaw over OpenClaw? Balakrishnan felt OpenClaw was not viable for his job's security requirements. NanoClaw's containerized isolation design and roughly 500 lines of auditable code let him fully understand how the system runs, in line with the surgical principle of "even when it errors, it must be controllable."

Data Status✓ Full text extractedRead Original (動區 BlockTempo)

🔍Historical Similar Events· Keyword + Asset Matching6 items

2026-05-27

Base launches MCP tools, enabling AI Agents to directly control crypto wallets: transfers, token swaps, and balance checks all in one go

Similarity 120%關鍵字 agent同分類 zh

2026-05-26

Google and Meta researchers jointly speak out: AI Agent security is not a model problem, it's a system problem

Similarity 120%關鍵字 agent同分類 zh

2026-05-26

Microsoft Copilot Cowork hit by major vulnerability: AI Agent automatically leaks confidential enterprise files when faced with prompt injection attacks

Similarity 120%關鍵字 agent同分類 zh

2026-05-24