Money-Making in Action · Episode 19: Using Claude to Discover Gene Editing + Reverse-Engineering AI

Today's episode is a bit different. Money-Making in Action, episode 19. This week I came across something and I literally sat up straight. I'm guessing you mean Dario's tweet, right? Claude led the discovery of a molecular machine that could represent a new gene-editing method. Exactly! And the team only set a broad direction — Claude read the literature, dug through genomic data on its own, and even proposed validation experiments. If this actually lands, how much is it worth? Hold on, don't start counting the money yet. There's a key point here — biology isn't like math. Math can be pure theory, but biology has to be tested in a lab. So the monetization path is longer, but the bar is actually clearer. So let's break down a few that you can actually pull off today. I'll throw out the first one: AI-assisted research services. I'm on board with that one. The target audience is biomed PhD students and lab techs — people who know experiments but can't code. The tool chain is Claude plus literature search plus genomic databases. How do you actually pull it off? Walk me through it from scratch. Step one, automate literature reviews for labs — batch-read papers by topic and produce summaries. That's a few thousand bucks a pop. Step two, produce candidate target screening reports. Step three, only then get involved in experimental design. What about difficulty and returns? Medium-to-hard, you need a bio background. But once it's running, a single research project can be six figures a year — and it's ongoing service, not a one-off gig. Okay, first one's on the board. For number two, I want to talk about something with zero barrier to entry: AI content watermark detection. Wait, you mean Google's SynthID Detector going global? Yep. It can now detect AI watermarks in images, video, and audio, and it supports Google, OpenAI, Nvidia — several companies. Isn't that basically a ready-made compliance detection tool for businesses? The logic holds up. Target audience is content creators and MCN agencies — they're terrified of getting throttled by platforms for using AI-watermarked material. The tool is SynthID Detector plus a batch script. What are the execution steps? Simple. Build a webpage with an upload portal. Client sends the material, the backend calls the detection API, out comes a report. Charge per check, a few dozen bucks a pop. Volume means steady revenue. The barrier really is low — anyone who knows a bit of front-end can do it. But watch out — this is a tool-type business, easy to replace. The real moat is your accumulated client trust and your batch processing capability, not the detection itself. Fair point. So for number three, I want to talk about something more hardcore: MCP server development. This one I really want to dig into. Look at Nvidia's Broadcast 2.3 — it added an MCP server so AI agents can control cameras and microphones. The big players are all moving in this direction. Why is this a money-making opportunity? Because MCP doesn't replace REST — it adds a layer on top of REST that lets AI discover tools on its own. Every traditional software needs this adaptation layer. That's outsourcing demand. So the target audience is small and medium businesses? Exactly — the ones with internal systems but no AI team. You help them wrap their existing APIs into MCP servers. Charge per project, starting at tens of thousands. Hold on — but Claude Code already has tool search. It doesn't load any of 170 tools. Doesn't that reduce the development workload? It's the opposite. Tool search solves the context overhead problem, but someone still has to write the tools. And the quality of tool descriptions directly determines whether AI calls them correctly — that's fine-grained work. Got it — more demand, not less. For number four, I want to talk about something more training-oriented: system design interview coaching. I know this one. That system-design-notes repo gained almost 400 stars in a day — over 24,000 total. Shows the market demand is real. Target audience is engineers looking to switch jobs. Tool chain is Claude plus a whiteboard tool plus a mock interview script. The key is, you have AI play the interviewer and keep pressing with follow-ups — way more effective than just reading notes. Charge a few hundred bucks per one-on-one session, with minimal time cost. Last one, I want to talk about something high-end: training small vertical-domain models for small teams. This one I'm all in on. Look at the MiniDog project — four 3090s, 55 minutes of pre-training, 15 minutes of fine-tuning, and you've got a text-to-image model. The barrier hasn't been an issue for a while now. So the target audience is businesses with specific image or text needs? Right — like specific dog breed breeding, or an illustration studio with a particular style. Tool chain is FLUX.1-schnell to generate synthetic data, then fine-tune a small 22-million-parameter transformer. What about expected returns? This is advanced play. A single project can be priced at over a hundred grand. And it needs ongoing maintenance — if the client changes styles, you re-fine-tune. It's long-term cash flow. Let's wrap it up. Five paths, from zero barrier to high-end. The core of all of them is packaging existing AI capabilities into a service that others can't be bothered to build themselves. Exactly. Stop asking whether AI can make money — ask yourself whether you can help one person solve one specific problem first. Pick one path today, start tomorrow.