Runway CEO Siqi Chen: AI's True Penetration Rate is Still Underestimated
Silicon Valley entrepreneur and co-founder and CEO of financial software company Runway, Siqi Chen, publicly stated that artificial intelligence is currently in an "extremely early, underestimated, and underhyped" stage. He argues that there is a significant gap between tech professionals who heavily use AI agents and ordinary users. The former have integrated multiple MCP (Model Context Protocol) tools into their agents and continuously update their personal knowledge bases ("gbrain"), while the latter mostly remain at the basic level of manually copying and pasting answers from ChatGPT into Word Copilot.
Chen's "gbrain" refers to a personal AI agent memory system open-sourced by Y Combinator president Garry Tan on April 5 this year under the MIT license. It organizes users' knowledge into self-referential knowledge graphs in plain text Markdown files for AI agents to read, write, and reason. According to public information, Tan's own "brain" library contains tens of thousands of Markdown notes, thousands of "person" pages, and hundreds of "company" pages. The project received about 5,000 GitHub stars within 24 hours of launch and has now grown to about 14,000.
Another keyword mentioned by Chen, MCP (Model Context Protocol), was launched by AI company Anthropic in November 2024 to address the repetitive labor issue of needing to develop separate integration code for each tool used with AI models. Subsequently, OpenAI officially adopted this protocol in March 2025 and integrated it into the ChatGPT desktop client, followed by Google DeepMind in April 2025. In December 2025, the protocol was further donated to the Linux Foundation's Agentic AI Foundation for unified maintenance.
In contrast to "heavy users," OpenAI itself disclosed real usage data for ChatGPT: after analyzing the dialogue content of its approximately 700 million weekly active users, the company found that about 49% of conversations were for "Q&A consulting," 40% for "task completion" (such as drafting copy, programming, planning), and only 11% for "personal expression and exploration." About three-quarters of the dialogue content focused on practical advice, information queries, and writing, while programming and complex tasks resembling autonomous agents remained niche in overall usage.
Chen's entrepreneurial and investment background is also an important context for his judgment that AI is in an early stage: he graduated with a degree in mathematics and computer science from the University of California, San Diego, previously served as general manager at Zynga, and later founded several startups including hotel booking services. In May 2020, he co-founded financial planning software company Runway, serving as co-founder, CEO, and CFO, which completed a $27.5 million Series A financing round in July 2023. Chen is also an active angel investor, with over 40 disclosed investment projects, including recent investments in several AI-related startups like Sentra and Hyperspell.
From the perspective of investment and market perception, Chen's remarks essentially provide a counterargument to the debate on whether AI is being overhyped: if, as he describes, the deep usage of connecting multiple MCP tools and allowing agents to automatically update personal knowledge bases is still limited to a small group of heavy tech professionals and early developers, while the vast majority of ordinary users and corporate employees are still using AI in the most basic and least efficient way of manual copying and pasting, this indicates a significant gap between AI's technological capabilities and its actual penetration rate and productivity conversion rate that has not been fully priced by the market. For investors holding AI-related assets (whether AI concept stocks in the secondary market or AI application startups in the primary market), this "cognitive lead over popularization" gap is often seen by early movers as a signal of long-term value being underestimated rather than a short-term bubble burst.
Additional information: It should be noted that Chen's mention of "long-term online, deep heavy use of agents" and "manual copying and pasting ChatGPT answers" are based on his personal observations and statements, without specific user ratios or research data sources attached; the OpenAI official usage data cited above is an independent public statistic, provided only as a reference background for the current real usage patterns of AI.
ABAB AI Insight
Chen's entrepreneurial background itself is a history of repeatedly capturing the "gap between technological capability and mainstream adoption": he experienced the phase of social games rapidly penetrating the mass market through viral growth at Zynga, then transitioned through several consumer or SMB software startup projects. Runway, founded in 2020, targets the gap where financial teams still rely on Excel and processes lag behind technological capabilities, and secured $27.5 million in Series A funding in 2023. His judgment that AI is "underestimated" continues the same narrative he has bet on repeatedly over the past decade: technology is already ahead, but behavioral habits have not caught up.
The capital path supporting this judgment is not just the statements themselves: as an angel investor, Chen has intensively invested in several AI infrastructure startups such as Sentra, Hyperspell, ego AI, and Emergent over the past year, indicating that his statement about "AI being underestimated" is not merely a public relations move but aligns with his actual funding allocation direction. The two specific pieces of evidence he cited—Garry Tan's open-sourced "gbrain" memory system gaining about 5,000 stars within 24 hours of launch and currently totaling about 14,000 stars, and the MCP protocol being adopted by OpenAI and Google DeepMind and incorporated into Linux Foundation governance—are direct signals of continuous inflow of developer and infrastructure-level funding and attention, but this funding and attention are currently still concentrated within the tech circle and have not yet translated into the daily habits of ordinary end users.
This "infrastructure running ahead of user habits" model has repeatedly appeared in past technology cycles: in the early internet, tech geeks relied on RSS subscriptions and self-built scripts to combine various services, while ordinary users were still using search boxes and emails; in the early mobile internet, developers had already built a rich application ecosystem, while most users were still just making calls and sending texts. In the current AI cycle, OpenAI's own disclosed data—49% of 700 million weekly active users using it for "Q&A," 40% for "task completion," and only 11% for "personal expression," with three-quarters of conversations focused on writing, information queries, and other basic scenarios—just confirms that even with a large user base, the vast majority still treat AI as a "smarter search box," while only a small group of tech professionals have moved into deeper uses like agent orchestration and persistent memory systems.
Structurally, this is essentially an early-stage, unevenly diffused "technological substitution": disruptive general-purpose technologies often see tools and protocols (like the MCP standardized by the Linux Foundation and the rapidly developer-adopted open-source memory system GBrain) mature first, while ordinary user behavior waits for a "forced switch point"—for example, software like Word Copilot embedding AI capabilities directly, eliminating the need for users to manually copy and paste—or until a new generation of user groups naturally takes over before real change occurs. Chen's argument essentially states that the current market is pricing AI based on already visible mainstream behaviors (manual copying and pasting of chatbots), and once protocols like MCP are standardized and tools like GBrain spread from the tech circle into default consumer software, the productivity enhancement potential released may far exceed current market pricing—this is also a lag that has occurred repeatedly before explosive repricing in historical general technology cycles (electrification, the internet, mobile internet).
ABAB News · Cognitive Laws
- Infrastructure always runs ahead of habits; this is a fixed script of technology cycles.
- The more people copy and paste, the more it indicates that the dividend has not yet been truly realized.
- Protocols connected by a few will eventually become the default settings for the majority.