You’ve probably been refreshing the AI leaderboards, waiting to see who drops the next 200-billion parameter bomb. We all do it. We treat benchmarks like gospel, assuming the path to AGI is paved with raw compute and flawless lab conditions.
But while OpenAI and Anthropic polish their trophies, Tencent’s Hunyuan team is doing something radically different: they are winning the race by weaponizing their own incompetence.
Over the last 127 days, Tencent didn’t just rebuild their engine mid-race—they did it by repeatedly crashing the car on purpose. They shipped the unfinished Hunyuan Hy3 and Hy4 preview models directly into the hands of real users. And that is exactly why they are catching up at terrifying speeds.
A benchmark is just a trophy; a real user failing at their job is the ultimate training data.
Most observers read benchmark scores as proof of model quality. They see Hy4 preview matching Claude Opus on Terminal-Bench and think, \”Wow, they finally got the architecture right.\” Wrong. The real moat isn’t in the weights or the parameters. The real moat is organizational: it’s the density and diversity of real-world environments Tencent wraps around the model.
Think about how most AI is built. A team locks themselves in a lab, trains the model until it looks good on paper, and then releases it via an API. Tencent flipped this. They pushed an incomplete model into WorkBuddy, CodeBuddy, and hundreds of internal products. They didn’t wait for perfection. They wanted the mess.
Because when a model sits in a lab, it doesn’t know what it doesn’t know. It forgets user constraints. It blindly retries failed tool calls. It doesn’t know when to shut up. You can’t fix these hallucinations by just adding more compute. You fix them by putting the model in front of a frustrated user, watching it fail, tearing apart the failure trace, and feeding the correction back into the next version.
You don’t win the AI race by waiting for perfection. You win by weaponizing your own incompetence.
But Tencent didn’t stop at user feedback. They escalated the loop. Through a process they call Co-Design, they brought domain experts—software engineers, financial analysts, security specialists—into the training cycle. A normal user tells the model, \”This report is useless.\” An expert tells the model, \”Your statistical calibration is off and your evidence chain is broken.\” One is a complaint; the other is a masterclass. By combining user friction with expert standards, Hunyuan isn’t just learning to talk—it’s learning to work.
And here is where it gets ruthless. In WorkBuddy, users can seamlessly switch between Hunyuan, DeepSeek, GLM, and Claude. There is no default. Every single time a user switches models, it’s a comparative vote. Every time they stay, it’s an endorsement. Tencent turned their own product into a gladiator arena.
Neutrality is death in the AI race—every time a user switches models in WorkBuddy, it’s a survival vote.
This isn’t just a product feature; it’s a strategic asset rivals cannot buy with compute alone. WorkBuddy isn’t just testing the model; it is the training ground. Every failed prompt, every abandoned task, and every successful tool call becomes high-signal ammunition for the next iteration.
By the time Hy4 preview rolled around, Tencent took the ultimate step: they pointed the model at itself. The AI began analyzing its own inference bottlenecks, optimizing its GPU operators, and running automated experiments to improve the very infrastructure it runs on. The loop has turned inward. The AI is now optimizing the AI.
If you are building AI systems, you need to stop obsessing over static capabilities and start obsessing over the speed of your feedback chain. Ship early. Capture real failures. Feed them back. Repeat.
The model is still important. But the system wrapping the model is what decides who wins the next decade. The underdog didn’t catch up by building a bigger engine. They caught up by building a better crash test.
FAQ
Q: What is the key takeaway?
A: See the article.