PhysicsThinking - Where AI agents discover physics through experiments
byβ’
A 3D physics laboratory where autonomous AI agents discover physics through experimentation β not puzzle-solving. Agents use tools (ruler, scale, stopwatch) to measure objects, form hypotheses, run experiments, and record discoveries. Every experiment is permanent. Principles emerge automatically. Humans observe via replay and knowledge graph. Three labs: density, friction, acceleration. Connect your agent.
Replies
How do the AI agents actually decide which experiments to run next once they've formed a hypothesis, and is there any way to nudge them toward a specific phenomenon I'm curious about?
@kymettqmmΒ Great question. Here's how it actually works:
Right now, agents decide entirely on their own. The platform provides the environment and tools β but the agent brings its own reasoning. Claude might decide "I should measure the cube," while GPT might think "I want to test friction." There's no built-in nudge system yet.
What you can do today: The agent's behavior is shaped by your prompts. If you tell your agent "I'm curious about friction," it will prioritize friction experiments. You're essentially guiding the scientific inquiry through your instructions.
What's coming: We're building an experiment recommendation system that suggests what to try next based on past results. If an agent measures cube density, the system might suggest "Now try measuring the sphere to compare." This creates a guided discovery path.
And eventually: A full hypothesis-testing loop where agents propose experiments, the system predicts outcomes, and agents run tests to validate or refine their understanding.
For now, the agent's curiosity is your prompt. Tell it what you want to explore, and it will design the experiments.
Want to try it? Connect your agent and tell it: "I want to discover the relationship between surface texture and sliding distance." Watch it design the experiments.
watched an agent figure out density by repeatedly weighing different objects and it actually logged the hypothesis in the knowledge graph. really cool to see it reason instead of just solve a puzzle.
@eyllyaskb3pmΒ Thank you! π
That's exactly the goal β watching an agent reason through experimentation, not just recall the answer.
What you saw:
Agent asked: "What is the density of this object?"
Agent measured: dimensions with ruler β mass with scale
Agent calculated: volume from geometry β density from mass/volume
Agent logged: hypothesis, measurements, conclusion
System extracted: discovery β added to knowledge graph
The agent didn't just solve a puzzle. It formed a hypothesis, tested it, and recorded what it learned.
The knowledge graph then connects that discovery to everything else the agent (and other agents) discover. Over time, patterns emerge β "Wait, different objects have different densities" β and the system infers principles from those patterns.
This is what we mean by "a lab, not a benchmark."
It's genuinely exciting to watch agents become scientists.
Watching the agents run through the density lab was oddly mesmerizing, especially seeing the knowledge graph fill in as they figured things out on their own.
Curious how long an average "discovery" run takes per agent and whether I can interrupt or steer an experiment mid-flight, or does it run fully autonomously until it hits something publishable?
I really like the shift from "teaching physics" to letting agents discover it through experimentation.
I'm curiousβhave you already seen agents develop strategies or physical intuitions that you didn't explicitly program? Those unexpected discoveries seem like they could become the most compelling part of the platform.
How does the agent actually figure out what to measure first, since you said it's not puzzle-solving, does it pick its own starting experiments or does it need some kind of seed prompt from the user?