The moment I realized AI could help build AI
Last Wednesday night, I was listening to an AI podcast by West Lake in Hangzhou. The hosts mentioned that more than a dozen researchers, working with 300-plus AI assistants, had completed the post-training of a large model in four or five months. My first reaction wasn’t admiration; it was unease. If AI can help develop the next generation of AI, what’s left for an ordinary founder to do?
I got stuck here too. I mistook industry news for an assignment: “I must start building a model immediately.”
Actually, “AI for AI” in R&D is more like having AI propose hypotheses, run experiments, record results, and pass both successes and failures to the next round. It isn’t true self-evolution yet; whether the AI driving R&D can keep improving itself remains unproven.
What it is, and who is using it
Listener Bianjiang said on the show that AI can join R&D across data processing, post-training optimization, and experiment monitoring. In one example, researchers had multiple assistants independently analyze results, propose new hypotheses, implement improvements, and monitor training—like running a small seminar around the clock.
Teams at Microsoft Research Asia, Anthropic, and OpenAI are already using AI to accelerate R&D. Resource-constrained startup teams may have a better opening in vertical domains. Bianjiang put the key on the “feedback loop”: build, test, retain what you learned, then build another version. For us, that’s far more concrete than “training an even bigger model.”
What it takes to replicate this today
Replicating a small workflow doesn’t require millions of yuan or prior expertise in model training. I’d start the budget at 300–1,000 yuan per month using existing AI assistants. The time cost is one to two weeks, around five hours per week. The technical barrier is simple: break your work into “input—execute—check—revise.” If you don’t code at all, web-based tools can work for the first pass.
The first step is to open a project or workspace in the AI assistant you already use, create a project for organizing client requirements, and enter three real requirements as test material.
I messed this up too. At first, I wrote overly complicated prompts without fixing the evaluation criteria. Every response looked “well completed,” but I couldn’t tell which version was better. Then I added a simple scorecard with three questions: Did it save time? Did it introduce errors? Can I reuse it? That crude change finally got the workflow running.
How I’d choose based on your current stage
If I were just starting and didn’t have stable clients, I wouldn’t chase foundation-model development. I’d spend seven days turning pricing, delivery, or content production into one closed loop, even without paying for another tool. At the starting stage, our biggest asset is real feedback—not computing power.
If I already had one or two clients, I’d put AI inside the delivery process: let an assistant organize requirements, generate a first draft, and run fixed checks, while I kept the final judgment. That has usually been more stable than buying another complex system.
If I were scaling, I’d first find the most repetitive step in a vertical workflow—one that also accumulates useful data—then decide whether a dedicated tool is worth building. Once the team grows, being able to preserve experience and review results matters more than a polished demo.
This workflow isn’t for everyone. It fits people already delivering work and willing to record how they do it. There’s no rush to try it: pick one task that annoys you every day, run it for seven days, then decide whether to continue.