Takeaways from CMU Startup Week: Legal AI, Investors, and Physical AI

Coheso’s Co-Founder and Head of AI, Manish Agnihotri, attended CMU Startup Week hosted by Carnegie Mellon University on September 23–26. The week brought together entrepreneurs, researchers, investors, and students to talk about the future for AI and robotics.

Manish went to learn the latest from CMU’s AI and robotics labs, hear the investor perspective on how AI value is forming across applications, meet the builders and researchers pushing the field forward, and discuss how academic innovation connects to real enterprise use cases. A few key themes emerged.

Academic View: Advances in Legal AI and LLMs

Manish met with leading professors and researchers working in AI generally and legal AI specifically. They discussed model accuracy, how we evaluate models today, and what that means for real use in legal work. A core topic was hallucinations and how certain evaluation approaches can push models to make things up.

They also discussed new evaluation methods that stress completeness, instruction-following, and abstention. Agentic AI came up as well. The tech is powerful, but implementation is still early. There is a gap between developers who build agentic pipelines and do context engineering, and legal experts who know how lawyers think and work. Closing that gap will make agentic software more useful in the legal context.

Key takeaways:

• Current evaluation methods can incentivize hallucinations and skip abstention. New evaluations should check completeness, instruction-following, and abstention.

• Guardrails are still essential as AI continues to improve to meet legal's high standards.

• Agentic AI is still early. The gap between builders and legal experts must be bridged to get real results.

VC View: Legal AI and In-House Legal Tech

Manish also had an opportunity to chat with investors on how they perceive the application of AI to the legal domain, with a focus on in-house applications. Over the past two years, investors have seen strong interest in applying generative AI to the legal industry, but the bar is high. VCs are increasingly focused tangible AI-driven outcomes rather than the incremental efficiency gains that once defined SaaS success.

Lab-to-Market View: CMU and the Rise of Physical and Embodied AI

CMU labs are shipping strong work in AI and robotics. Advances in intelligent systems, advanced materials, and embodied AI is turning lab gains into products that address real-world limits. One group discussed the next big consumer milestones in Physical AI: autonomous cars like Waymo becoming common in most cities, assistive robots for child and elder care, and micro-robots that help with household tasks (like cleaning a litter box). A fun subculture note came up too: some kids in Waymo cars ride around for an hour with music on, then head home.

CMU showed broad advances in robotics, including general-purpose robots for the physical world that can keep working even if part of the body is removed by sensing limits and adapting. Another steady theme was trust in AI and robotics. Teams focused on hardening agents against failure, closing the sim-to-real gap, and building repeatable tests for safety and reliability. They stressed real-world data collection, tighter evaluation, and simple user workflows.

Key takeaways:

• Physical AI is close to a ChatGPT-style moment, moving from lab demos to real products people can use every day.

• Autonomous cars and assistive robots are near-term signals to watch. General-purpose robots that adapt to limits are coming.

• Building trust requires continued testing for likely failures and closing the sim-to-real gap. This means using real-world data over pristine lab data.