Diminishing LLM returns? Enter physical AI and robotics - open-source leading the charge into real-world agents.
Scaling LLMs is so 2025 - IBM experts say 2026 flips to ‘palpable’ AI: robotics and physical systems that sense, act, and learn in meatspace. Diminishing returns on giant models have folks hunting fresh ideas, and open-source is thriving with Chinese multilingual beasts, Granite, Olmo 3, DeepSeek killing it in niches.[2]
PyTorch becomes the glue for agentic systems: multimodal reasoning, memory, safety evals - all open and flexible. NVIDIA’s pushing ecosystems for GPU adoption, accelerating collab as AI escapes screens for robots. Devs, this means tooling for sim-to-real transfer, embodied agents you can hack on laptops.[2]
I’m pumped - finally, AI you can touch. But closed labs lead, opens catch up; governance hardens with audits. Practical move: Fork a Granite model, slap it on a ROS robot arm, test physical reasoning. Who’s building the first open physical agent that doesn’t suck? Let’s talk prototypes.
Source: IBM Think