← back
Navigating AI's Frontier in 2025 - Grace Isford, Lux Capital
Original: Navigating AI’s Frontier in 2025 - Grace Isford, Lux Capital
Takeaway
Agents will only ship when teams attack compounding error directly — through data curation, eval, and step-level reliability, not just smarter base models.
Summary
- Lux Capital frames 2025 as a 'perfect storm' for agents: reasoning models (o1/o3, DeepSeek R1, Grok), test-time compute, $500B Stargate, France/Europe AI Summit, closing open/closed gap.
- Live demo: OpenAI Operator failing to book a complex SF↔NYC flight illustrates that agents fail via cumulative tiny errors, not just hallucinations.
- Five error types: decision, implementation (CAPTCHA blocks), heuristic (wrong criteria), taste (personal preferences), and the 'Perfection Paradox'.
- Math: a 95% vs 99%-accurate agent diverges to 60% vs ~99% over 50 steps — compounding error is the central design constraint.
- Five mitigation strategies including data curation across multimodal/proprietary/agent-generated data.
agentsreliabilityvc-outlook
Original description
Lux Capital Partner Grace Isford discusses the AI Frontier in NYC, sharing 10 'hot takes' on where the industry is going from a technical perspective. She covers key considerations for technical leaders building agentic systems, as well as the second-degree implications of LLMs for human behavior. Recorded live at the Leadership Track Session Day from the AI Engineer Summit 2025 in New York. Learn more at https://ai.engineer and purchase tickets to our next event, the AI Engineer World's Fair, in SF June 3 - 5 here: https://ti.to/software-3/ai-engineer-worlds-fair-2025 About Grace As a Partner based in our New York City office, Grace invests in companies innovating at the nexus of the computational sciences – data, AI and ML infrastructure, open source software, network infrastructure, developer tools, vertical software applications and more. Before joining Lux, Grace was a principal at Canvas Ventures where she sourced 10 investments. She got her start as a campus scout while attending Stanford University. Prior to Canvas, Grace worked on the LP side at the Stanford Management Company, in product at edtech startup Handshake, and in growth equity at Stripes Group. She earned a Bachelors of Science and a Masters of Science in Management Science and Engineering from Stanford, where she was a Mayfield Fellow and served as Co-President of Stanford Women in Business, the campus’s largest pre-professional organization for women. In addition, she is on the board of the Stanford Technology Ventures Program, the university’s entrepreneurship center, and is an active member of All Raise, focused on accelerating the success of female and non-binary founders and funders. Grace is originally from Connecticut, although has lived in Tokyo, Japan, and aspires to re-learn Japanese. She’s an avid runner and cyclist.