Written by: Techub News Compilation

Introduction

In the latest episode of the No Priors podcast, AI pioneer Andrew Ng had a deep conversation with the host lasting 42 minutes. As the co-founder of Google Brain, co-founder of Coursera, and managing partner of the AI venture studio AI Fund, Ng has made significant achievements in both academia and industry. He gained attention in August 2025 for proposing the concept of "Agentic AI" and joining the Amazon board. In this dialogue, he systematically elaborated on his views regarding the growth trajectory of AI capabilities, current implementation bottlenecks, and the impact of technological transformation on the entrepreneurial ecosystem, providing industry practitioners with insightful foresights.

Summary

• The growth of AI model capabilities no longer relies solely on compute power; it has shifted towards diverse paths such as "Agentic Workflows," multimodal model construction, and application development.
• The biggest bottleneck in the current application of "Agentic AI" is not the underlying technology, but the lack of engineering talent capable of systematic error analysis and evaluation processes.
• AI-assisted programming has significantly increased engineering efficiency, shifting the bottleneck of core iteration cycles for startups from "engineering implementation" to "product decision-making."
• In this era of rapid AI technological iteration, a founder's technical intuition and learning ability are more important than past experiences, making technical product leaders more likely to succeed.
• Small, efficient teams with a deep understanding of AI tools may achieve productivity far beyond that of larger traditional teams with high coordination costs.

Growth of AI Capabilities: From Singular "Scale" to Diverse "Vectors"

When asked about the future sources of growth for AI model capabilities, Andrew Ng stated that society's perception of AI has been shaped by a few companies with strong PR abilities, which propelled the narrative of "scale is everything." However, he believes that the "juice" that can be squeezed from the "lemon" of compute scale has become increasingly limited, and advancements will come from multiple vectors.

Ng highlighted that beyond compute power, "Agentic Workflows," multimodal model construction, and specific application development will all become significant sources of progress. At the same time, new technologies like diffusion models, initially used for image generation, hold exciting unknown possibilities for text generation in the future. He summarized that AI progress will not be a singular path, as many smart individuals are pushing the frontiers from various directions.

"Agentic AI": From Concept Proposal to Implementation Bottlenecks

As the proposer of the term "Agentic AI," Ng explained his intention: A few years ago, the industry spent considerable time debating "whether something is an agent," and he observed the existence of a "spectrum of agency" from high autonomy to low autonomy. To stop the debates and concentrate on building, he advocated for the popularization of this term. However, he also admitted that he did not expect marketers to spread this label everywhere, resulting in hype outpacing actual business progress.

So, what truly obstructs the implementation of Agentic AI applications? Ng believes that while there is room for improvement in technical components (like computer usage, guardrail evaluations, etc.), the current biggest bottleneck is "talent". He observed that the key difference in team performance in the market lies in whether they have mastered a systematic error analysis process based on evaluations. Experienced teams can continually analyze what works and what doesn't and make targeted improvements; in contrast, inexperienced teams tend to try things at random, leading to inefficiency.

Ng used examples of multi-step workflows in automated document processing, compliance checks, and database queries to point out that while building such processes, a lot of external contextual knowledge (e.g., whether getting the invoice date wrong is crucial or whether excessive interruptions to the CEO are a concern) typically still depends on human product managers or engineers for thinking and decision-making. This knowledge is proprietary and does not exist in the pre-trained datasets available on the internet, so human involvement remains essential for now.

AI-Assisted Programming: Transforming the Core Loops of Startups

When discussing the most economically valuable applications of Agentic AI, Ng highlighted two clear areas: one is question-answering chatbots (like ChatGPT), and the other is coding agents. He personally favors Cline, believing it demonstrates high autonomy and practicality in planning multi-step tasks and executing plans. In contrast, some computer usage scenarios (like automated online shopping) remain at the demonstration stage.

Why are coding agents so effective? Ng believes it's not because the foundational research challenges are smaller, but because the economic value of coding is clear and immense, attracting substantial resource investment, and developers themselves are users with excellent product intuition. He even mentioned that some leading foundational model companies have openly stated they use AI to write large amounts of code. More excitingly, AI models are generating training data for the next generation of models through Agentic workflows (as described in the Llama research paper).

Ng particularly opposes the term "Vibe Coding," preferring "AI-assisted programming." He emphasizes that this is not an easygoing, vibe-based task but still an intense, high-density intellectual activity, akin to "rapid engineering"—AI enables us to build serious systems at a pace far beyond before, yet it is still fundamentally engineering.

This increase in efficiency is changing the form of startups. Work that once required a six-person team three months can now potentially be prototyped by a two-person team over one weekend. This has fundamentally altered the core iterative loop of startups—building software, gathering user feedback, deciding how to improve the product. Coding speed and cost are no longer bottlenecks; product decision-making (deciding "what to do") has become the new bottleneck. Consequently, teams increasingly rely on intuition and deep customer empathy for rapid decision-making.

Founder Profile: Technical Intuition and Learning Ability are Paramount

In the face of rapid technological iteration, Ng believes that many practices from 2022 may become entirely outdated by 2025. He often asks himself whether any practices are the same as in 2022 and examines their ongoing relevance. In this context, the profile of founders has changed significantly.

He observed that founders who are more likely to succeed today are those who are quickly adopting cutting-edge technologies and being tech-driven product leaders, rather than merely possessing business acumen without technological insight. Without a good sense of the limits of technology, it is challenging to think through company strategy. He reflected on the early pioneers of Silicon Valley (like Bill Gates and Steve Jobs), all of whom were technical experts, and posited that this deep understanding of technology has once again become a key differentiator during the disruptive moment of AI.

Ng contrasted this with the mobile internet era: Today, everyone has a general understanding of what mobile applications can do, so one can start a business without being very technical. But the boundaries of AI capabilities are evolving rapidly, and having this "frontier knowledge" is a significant advantage in itself. He even shared a hiring example: he chose a recent graduate proficient in AI over an engineer with ten years of experience but limited use of AI tools, and the results proved to be an excellent decision. Certainly, the best engineers are those who possess both a decade of experience and a deep mastery of AI tools.

Team Size and Efficiency: A New Balance Empowered by AI

As AI tools enhance individual productivity, the relationship between team size and efficiency warrants re-examination. Ng sees that some of the most productive teams are small in size, composed of excellent members, fully empowered by AI, and have extremely low coordination costs. He predicts that many teams may be smaller in the future than in the past.

However, he also cautioned against the trap of "over-pursuing small size." In the 2010s, some companies prided themselves on maintaining small teams and profitability, but this could lead to stagnant growth and missed market opportunities. The key is to analyze market dynamics: If it is a winner-takes-all market, quick action and market capture are necessary, and one cannot be content with being "small yet beautiful." Lessons from history, such as Slack versus Teams and Sketch versus Figma, serve as important reminders.

Ng shared an interesting insight: AI not only enhances efficiency but may also alter capital allocation. He received two requests within a week: one was to increase the human budget, which he rejected; the other was a request for budget to purchase AI services to accomplish similar tasks, which he approved. The logic behind this is: in the AI era, one should consider "hiring AI" rather than "hiring more people." However, such decision-making relies on a keen intuition about technological potential.

The Future of AI Investment: The Boundaries of Automation and Human Judgment

As a managing partner of AI Fund, Ng was also asked which jobs in investment firms might be automated. He believes that tasks like deep company research, competitive analysis, preliminary market research, and LP reports are highly suitable for the introduction of automated tools. He personally often utilizes various AI research tools.

However, aspects involving human judgment and relationships are difficult to automate. For instance, assessing an entrepreneur's leadership, communication style, or personal qualities often relies on subtle nuances and background checks captured in face-to-face interactions, which are hard to structure and input into AI models. Moreover, persuasion work based on trust relationships (such as recruiting key talent) heavily relies on connections between people.

Regarding how to assist technically oriented first-time founders, Ng believes that the value of venture capital firms or venture studios lies in their experience of having gone through more "iterations" and being able to provide experience-based intuitive insights on fundraising, grasping technological trends, and accelerating product-market fit. Simultaneously, building networks for peer exchange and complementary teams for founders is also crucial.

The Next Five Years: A New Human Empowered by AI

Looking ahead to the broad impact of AI in the next five years, Ng firmly believes that the capabilities of individuals embracing AI will far exceed the current imagination of most people. Two years ago, who could have predicted that the productivity of software engineers could be enhanced this significantly with AI assistance? In the future, individuals across all industries and those managing personal affairs will become incredibly powerful and efficient through AI. This is not just an enhancement in productivity but a massive expansion of the boundaries of human capability.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。