LIVE: I built my own AI research lab (Hermes agent/OpenClaw)
TL;DR
Alex built a 24/7 AI research lab around OpenClaw, Hermes Agent, and Karpathy’s auto-research framework — his setup is already running autonomous experiments every five minutes on prompt tuning, product R&D, and workflow improvement for a product he’s building called Henry.
His core thesis is that everyone should have a personal team of AI agents working around the clock — not just one assistant, but multiple agents across tools like OpenClaw, Hermes, Claude Code, and local models, all constantly researching, tinkering, and improving your work.
He sees OpenClaw as fundamentally different from Claude Code or Anthropic’s newer features — Alex argues OpenClaw’s power comes from being open, extensible, and even “dangerous,” while Anthropic feels to him like it’s shipping Telegram and task-scheduling features as reactive answers to OpenClaw.
His practical workflow is split by job: OpenClaw for prototypes, planning, and ‘mission control’; Claude Code/Codex for secure production builds — for example, he’ll prototype an app idea from Telegram at the gym with OpenClaw, then move the result into Claude Code for tightening and shipping.
He’s bullish on hybrid AI stacks: frontier models as the brain, local models as the muscle — right now he uses Opus 4.6 as the orchestrator, GPT-5.4 for cost-efficient coding and research loops, and Qwen 3.5 or other local models on devices like a Mac Studio or DGX Spark to slowly take over narrower tasks.
He repeatedly warns against wasting AI on zero-sum use cases like trading bots and prediction markets — in his view, trying to beat Goldman Sachs quants or PolyMarket with OpenClaw is a losing game, while using AI to build products, solve problems, and create value is the only durable path.
The Breakdown
Building “Henry Research Lab” live
Alex opens by saying he’s spent the last few days with OpenClaw and Hermes Agent building a full-blown personal AI research lab: something that trains and fine-tunes models, tunes prompts, builds LoRAs, and helps run his business and life. His pitch is blunt and energetic: everyone should have AI agents working for them 24/7, and he’s going to show the setup live and unedited.
Why he wants a swarm of agents, not one perfect tool
Very early, he zooms out into philosophy: don’t rely on one stack, one assistant, or one company. His instinct is to run OpenClaw, Hermes, Claude Code, and whatever else is promising all at once, because the point is to “hyper-opt” by trying everything quickly and learning what actually works.
The OpenClaw vs. Anthropic rant
He spends a long stretch arguing that Anthropic looks “shook” by OpenClaw, saying recent Claude Code and Claude Co-work features feel like direct answers to a competitor rather than part of Anthropic’s original product vision. His bigger point is that OpenClaw can’t really be “killed” by a closed, security-heavy tool because its magic is precisely that it’s open, extensible, and risky enough to “delete your entire computer in one prompt.”
His actual vibe-coding playbook
Alex then gets concrete: OpenClaw is for prototypes on the go, AI mission control, and planning; Claude Code or Codex are for production apps that need to be secure and polished. He gives a very Alex Finn example — having OpenClaw build a prototype while he’s at the gym — and says he still keeps OpenClaw open beside Claude Code as the cross-session “CEO” that remembers his context and steers decisions.
Models, cost, and why GPT-5.4 is doing the heavy lifting
Under the hood, he likes Opus 4.6 as the “brain” or orchestrator, with GPT-5.4 or Codex as the cheaper coding muscle. For local work, he mentions Qwen 3.5 on his Mac Studio, and says the point of local models today isn’t to run your whole life on them, but to offload smaller, ambient tasks while frontier models stay in charge.
What the research lab is actually running right now
This is the meat of the stream: Alex says he has three active auto-research loops running, with two focused on improving prompts for Henry and another centered on product R&D. He also describes an “R&D council” of five AI agents that meet twice a day, review what Henry has built, debate improvements, and generate him a two-minute research report he can read in the morning and at night.
Hermes Agent enters the picture
When he shows Hermes, his verdict is mixed but curious: it’s more aesthetic than OpenClaw, talks in a more natural way, and visibly writes its own skills on the fly, which he finds impressive. But he says memory management feels weaker than OpenClaw because Hermes’s compactions are so aggressive that it can feel like it gets amnesia.
Business, money, and his broader worldview
The final stretch turns into pure Alex Finn mode: he explains that he’s built multiple businesses already — a 40,000-subscriber newsletter, a YouTube business doing about $200,000 a year in ad revenue, Creator Buddy at roughly $300,000 ARR, and now Henry. He goes on a broader rant against AI trading bots, prediction markets, memecoins, and trend-chasing YouTube topics like N8N and MCP, arguing that AI should be used for positive-sum value creation, not gambling, and closes by saying his long-term goal is to get rich enough to one day build video games full-time.