Someone built a sailing simulator, put the world's leading AI models behind the wheel, and discovered something fascinating about how these models actually solve problems.
Or, more accurately, behind the rudder.
Developer Abhishek (@shekitup) created a 3D sailing simulator using Claude Code.
The physics, wind, water, coastline, boats, and simulation environment were largely written by Claude.
Then he decided to make the frontier models race each other.
And things got interesting.
The first experiment: Let AI models sail the boats themselves.
The competitors included:
Claude Opus 5.5
GPT-6 Astra
Gemini 4 Argon
Qwen 3.5-9B
Every five seconds of simulated time, each model received 23 numerical values describing the environment.
Boat speed.
Heading.
Wind direction.
Wind strength.
Distance and direction to the next checkpoint.
And so on.
The model then had to decide how to adjust the rudder and sails.
No visual input. No sailing instructions. Just numbers and decisions.
And the results?
Gemini 4 Argon won.
GPT-6 Astra finished second.
Claude Opus 5.5 couldn't complete the race.
And the little Qwen model apparently decided that sailing was overrated and headed straight for the beach.
Not exactly a glorious day for Claude.
But then came the second experiment.
The developer teased Opus about losing.
And Claude suggested a different approach.
What if, instead of manually steering the boats through repeated language-model calls, each model wrote its own autopilot?
Now that's an interesting idea.
The frontier models were tasked with generating actual navigation code.
Code that could read the simulated environment, make steering decisions, adjust the sails, and navigate the course automatically.
And then the autopilots were tested on tracks the models had never seen before.
No memorizing the route.
No pre-programmed solution for a particular course.
The code had to generalize to unfamiliar sailing conditions and layouts.
And this time, Claude Opus 5.5 absolutely dominated.
According to the developer, Opus not only outperformed GPT-6 Astra and Gemini 4 Argon, but was also the only model whose autopilot successfully completed every unseen track.
Which is a spectacular turnaround.
Claude couldn't sail the boat particularly well.
But it could write software that sailed the boat extremely well.
And that's what makes this experiment so interesting.
There's a fundamental difference between using an LLM to make every individual decision and using an LLM to design a system that makes those decisions.
In the first approach, the model has to repeatedly interpret the environment and decide what to do next.
In the second, it can build a reusable control algorithm.
Once written, that algorithm executes without needing another expensive language-model inference every five seconds.
In other words, Claude stopped trying to be the captain and became the engineer.
And apparently, that was a much better career choice.
Of course, this is an informal developer experiment, not a standardized scientific benchmark.
We don't have enough independent testing to conclude that Opus is universally better at robotics, control systems, or navigation.
But the result illustrates something important about AI agents.
Sometimes the best way to solve a problem isn't to ask an AI to perform the task continuously.
It's to ask the AI to build a tool that performs the task reliably.
And that's a distinction worth remembering as we build increasingly autonomous systems.
The most capable AI agent might not be the one making the most decisions.
It might be the one smart enough to automate those decisions away.
Also, congratulations to Qwen.
I hope the beach was nice.
Original autopilot experiment by Abhishek:
https://x.com/shekitup/status/2117980184033071480
Creator's X profile:
https://x.com/shekitup
#ClaudeOpus #GPT6 #GeminiAI #AIAgents #ClaudeCode #AIbenchmark #VibeCoding #AutonomousAI
YouTube / ABVX