Software Developer, Inference and Posttraining
- Tweets 2,006
- Following 4,581
- Followers 442
- Likes 2,338
Microduck, LeRobot, quackd: the pieces for anyone to give an LLM a body are now mostly open and affordable. The next few years of robotics are going to be a lot more hobbyist-driven than people expect.
31
What makes a good robot skill for an LLM: one clear purpose, honest preconditions, a timeout, and a result that says what actually happened. That's the same contract as a good software tool, plus physics.
21
Hot take: the robot that teaches the most people about embodied AI won't be a humanoid. It'll be something small, cheap, cute and hackable.
21
Prediction: "MCP server for my robot" will become a weekend project, like "Discord bot for my server" was. Expose the skills, connect your favorite LLM, and see what it does.
1
54
The duck form factor is clever. Nobody expects a duck to be perfect, so you can forgive it when it falls over. That's a smart way to get robots into homes and classrooms while the tech is still young.
1
9
If you've been building agents that only click buttons and call APIs, try one robot project. It changes how you think about failure, state and verification. Everything in software agents is easier after that.
8
Treat robot skills like an API product. Version them, document them, publish them. A shared library of tested skills with good descriptions is what makes an LLM planner actually reliable.
6
Desk robots, arm kits, walking ducks: we're getting a whole family of cheap, hackable bodies for AI models. The software layer that connects them to LLMs is still wide open for builders.
6
Robots are where LLM latency starts to really matter. A 3 second planning call is fine for a chatbot. For a walking robot it's fine only if the controller underneath keeps it safe while the model is thinking.
11
What I'd build first with a Microduck: a tiny agent loop. The camera looks, the LLM describes the scene and picks one skill, the robot runs it, then repeat. Ugly, slow, and an incredible way to learn.
1
14
Common mistake when connecting an LLM to a robot: giving it too many low-level actions. Fewer, higher-level skills mean shorter plans, fewer chances to fail, and much easier debugging.
6
Small robots also make great evals. "Did the duck reach the ball within 30 seconds?" is a clearer metric than most LLM benchmarks I've seen. The physical world grades you honestly.
11
Open-source robot software is a big deal for LLM builders. When the SDK, sim and training code are public, you can see exactly what each skill does, and that's what you need to describe it well to a model.
1
17
LLMs are great at turning "tidy up the socks" into steps. They're bad at knowing where the socks are. Give the planner a good perception tool and you'll fix more failures than any prompt tweak ever will.
7
The hardest part of LLM-driven robots isn't the first demo. It's step 7 of a 10-step task, when the object rolled away and the plan needs to change. Design your loop around re-planning, not one-shot plans.
1
2
16
Voice is the obvious interface for a home robot. Say what you want, the LLM turns it into a plan, the plan becomes skill calls. Microduck owners will probably build this in the first week.
19
Before: "robotics" meant ROS, a big budget and months of setup.
After: a pip install, a simulator, a small robot on your desk and an LLM API key. The barrier to entry dropped a lot, and that changes who gets to build.
15
Underrated skill for AI engineers in 2026: knowing a little reinforcement learning. LLMs plan; RL policies move. The people who can work on both sides of that boundary are going to be very useful.
9
A desk robot is the best AI teaching tool I can think of. Kids and adults both understand "make the duck walk to the ball" instantly. Explaining a transformer takes much longer.
11