Andrew Ng on "ambient programming": Don't be misled by the name. AI programming isn't easy.

Key points:

• The progress of AI will not solely rely on "scale." In the future, its driving forces will come from model expansion, autonomous workflows, multimodality, and the application of new technologies, rather than a single path.
• Andrew Ng believes that the biggest obstacle to the application of autonomous AI is not purely technical but the shortage of relevant talents and engineering capabilities in error analysis and evaluation - driven development.
• Currently, the most clearly - defined and mature autonomous AI applications in terms of economic value are AI programming assistants (such as Claude Code) and general Q&A assistants (such as ChatGPT).
• In the current era of rapid AI transformation, technical founders with in - depth intuition and understanding of technical capabilities have an advantage over those with only a business background.
• The key for excellent product talents lies in their outstanding ability to empathize with customers. They can integrate information from multiple sources to build a mental model of users and make quick decisions accordingly.
• In the next few years, AI will profoundly change the working methods of all functions. Individuals who make good use of AI tools will have their productivity and potential released unprecedentedly.

Andrew Ng, a well - known scholar and professor at Stanford University, recently visited the investment podcast "No Priors" and shared his in - depth insights into the future development direction of AI capabilities.

Andrew Ng is a godfather - like figure in the AI field. He co - founded Google Brain, the online education platform Coursera, and the venture capital firm AI Fund. Recently, he proposed the concept of "Agentic AI" and joined the board of directors of Amazon.

In the latest exclusive interview, Andrew Ng pointed out that the driving force for AI progress will come from multiple paths such as model expansion, autonomous workflows, multimodal models, and the application of new technologies, rather than relying solely on scale expansion. He believes that the biggest obstacle to the implementation of current agents is not the technology itself but the shortage of talents who understand error analysis and evaluation.

He also emphasized that AI is reshaping the entrepreneurial paradigm: the significant improvement in engineering efficiency has made product management the new bottleneck, and "technical founders" with in - depth intuition about technology are regaining their advantages.

Looking to the future, Andrew Ng believes that individuals and teams that make good use of AI tools will release far more potential than currently imagined, profoundly changing the working methods of all industries.

The following is the essence of Andrew Ng's latest exclusive interview:

01 The Next Stop for AI Evolution: "Walking on Multiple Legs"

Question: You focus on such a wide range of fields. Perhaps we should start with the most core question: Looking to the future, where will the improvement of AI capabilities come from? Will it rely on larger model scales or more efficient data processing?

Andrew Ng: Future progress will not come from a single direction but will be driven by multiple aspects. There may still be potential to be tapped on the path of scale expansion, but it is indeed more challenging than before. Currently, the public's perception of AI is largely influenced by a few companies with strong public - relations capabilities. When people mention AI progress, they often first think of "scale."

However, in reality, real breakthroughs may also come from other dimensions. For example, there is a lot of room for exploration in autonomous workflows, the design of multimodal models, and the implementation of various applications. In addition, there are new technologies like diffusion models. Initially, it was mainly used for image generation, but can it be extended to text generation in the future? These directions are also very exciting. So, the development of artificial intelligence will never rely on a single path.

Question: You were the first to propose the term "Agentic AI." What were you thinking at that time?

Andrew Ng: When I decided to use the term "Agentic AI," not many people were using it at that time. My team even advised me not to create new words. But I insisted, and unexpectedly, it became popular later. I proposed this concept because a few years ago, people often got into arguments: "Is this an agent? Is that not? What exactly counts as an agent?"

In my opinion, being an agent is a matter of degree: some have strong autonomy, can make plans, conduct multi - step reasoning, and complete complex tasks; while others rely on human prompts and have relatively weak agent characteristics. Instead of getting entangled in definitions, we should admit that different systems have certain agent characteristics to varying degrees. Only in this way can we truly focus our time and energy on research and development.

So I started promoting the term "Agentic AI." However, I didn't expect that just a few months later, a large number of marketers quickly seized this label and spread it everywhere, and "Agentic AI" became popular. The market hype was faster than expected, and the actual business progress was also continuing, but obviously not as fast as the hype.

Question: What do you think is the biggest obstacle to the implementation of real - world agent applications currently?

Andrew Ng: From the perspective of technical components, there are still some areas that need improvement. For example, the "computer - using" ability can be successful in some scenarios, but the failure rate is still not low; also, guardrail mechanisms and evaluation systems, such as how to efficiently and systematically evaluate models and promote optimization, remain a major challenge.

However, I believe that the biggest obstacle to the implementation of autonomous AI is actually talent. The difference between many teams lies not in the technology stack but in whether they can truly conduct evaluation - driven error analysis. Mature teams will continuously break down problems, analyze what works and what doesn't, and then make targeted improvements; while less - experienced teams often make random attempts and progress more slowly.

In reality, many enterprise workflows could be automated by agents, but the lack of talents with the corresponding skills and supporting tools makes it difficult to achieve engineering and scale - up. Building agent workflows often requires integrating external knowledge, which mostly exists in people's minds. Unless there is an AI in the future that can "interview" employees or even "see" the screen, human engineers will still need to be deeply involved in the next one or two years to promote the implementation of workflows.

Question: So currently, the more realistic path is still for people to collect data, establish feedback loops, etc. Are there any other challenges?

Andrew Ng: That's right. For example: many companies have a similar process - a customer sends a document, which needs to be converted into text first; then, for compliance reasons, a web search may be needed to verify the reliability of the supplier; then check the database to confirm the price; and finally, archive it. This is a typical multi - state agent workflow and can be regarded as the next - generation RPA (Robotic Process Automation).

However, the problem is that once the process goes wrong, the consequences can be serious. For example, what if the invoice date is extracted incorrectly? What if the verification request is sent to the wrong person? It's almost impossible to be perfect at the initial stage of launch. So which errors have the greatest impact on the business? For example: "Is it disturbing the CEO too frequently? Which verifications is he willing to handle personally?" These background judgments often require experienced product managers or engineers to make decisions. In the future, agents may be able to handle these complex situations independently, but at least for now, it is still quite difficult.

Question: But this knowledge neither exists in the pre - trained Internet data nor can it be easily extracted from manuals automatically.

Andrew Ng: Yes. Building agent workflows highly depends on proprietary data rather than general Internet knowledge. This makes the process complex and often frustrating, but it also represents a real opportunity. Therefore, I still believe that this is a very worthwhile and promising task.

02 Towards "Bootstrapping": Two Paths for AI Self - Evolution

Question: If we only look within the scope of autonomous AI, what is the best example you've seen so far?

Andrew Ng: In this field, some AI programming assistants have impressed me deeply. From an economic value perspective, there are currently two particularly prominent paths: one is Q&A - type applications, such as OpenAI's ChatGPT, which has clearly become the market leader and shows real high - speed growth; the other is programming - assistance agents.

Personally, I like Claude Code the most. It shows strong autonomy in planning ability. It can understand what the target software is, generate a task list, and execute it step by step. This ability to plan multi - step tasks and implement them according to the plan makes it one of the agents that truly have autonomy and can be put into practical use at present.

Of course, there are also some directions that I think are not yet mature, such as certain "computer - using" scenarios - like online shopping and web browsing. Although these applications perform well in demonstrations, they still have a long way to go before they can be put into large - scale production.

Question: What do you think is the reason for this difference? Is it because the task standards are not clear and the operation variables are too large? Or is it that the programming scenario has a better training set or clearer output specifications?

Andrew Ng: I think there are mainly two reasons: on the one hand, engineers are naturally good at making complex systems work; on the other hand, the economic value of programming assistants is extremely direct and huge. This has attracted many smart people to engage in this field. They are both users and have a natural intuition for the product, which has promoted the rapid development of this direction.

Question: When do you think models will be able to effectively achieve "bootstrapping"? For example, let a coding agent write the code for a model by itself?

Andrew Ng: I think we are gradually approaching this goal. In fact, some leading foundation - model companies have publicly stated that they are using AI extensively to assist in writing code.

Another direction that excites me equally is using agent workflows to help the next - generation models generate training data. For example, the research paper on Llama mentioned that the old - version Llama can generate complex questions through "long - term thinking," and then use these questions to train the new - version Llama to solve problems more quickly. This idea is very interesting. It shows again that the progress of AI never relies on a single path but is promoted simultaneously from different angles by countless smart people.

Question: I remember you didn't quite approve of the term "vibe coding" before and preferred to use "AI - assisted programming." What's the difference between the two?

Andrew Ng: Yes. "Vibe coding" sounds like I just need to relax and accept all the modification suggestions from AI. Sometimes it works, but it's far from the whole picture. In fact, when I spend half a day or even a whole day programming, it's a highly mentally - consuming and in - depth task. To be honest, after a whole day of AI - assisted programming, I often feel very exhausted.

Therefore, I prefer to call it "rapid engineering." AI can indeed help us build serious systems and develop mature products at a much faster speed than before, but in essence, it is still solid engineering work - just completed more quickly.

03 The AI Efficiency Revolution: 2 People + a Weekend = 6 Engineers + Three Months

Question: Do you think this is changing the nature of startups? For example, how many people a team needs, how products should be built, and how processes should be designed? Or is the basic entrepreneurial method still the same, but now everyone has more powerful tools and the efficiency has become higher?

Andrew Ng: Based on my experience in entrepreneurship and observing the industry, rapid engineering and AI - assisted programming are indeed changing the way we build companies, which is very exciting. Tasks that might have taken six engineers three months to complete in the past can now be achieved by me and a friend in a weekend.

Interestingly, what has really changed is not just the development efficiency. The traditional entrepreneurial rhythm was like this: first, develop the software, then let the product manager conduct user tests, and decide how to improve based on intuition or data. Now, with the significant reduction in coding speed and cost, the bottleneck has become more and more concentrated in the product management stage.

In other words, we can implement ideas more quickly, but the difficulty lies in what exactly to build. In the past, it took three weeks to build a prototype and one week to collect user feedback, and this rhythm was acceptable; but now, a prototype can be built in one day, and waiting for a week for feedback seems too slow. So I've found that teams are increasingly relying on intuitive decisions. Of course, we still collect a lot of data to improve our understanding of the market, but the more crucial thing is to have deep empathy for customers to make the right choices in the high - speed iteration process.

Question: Have you seen any tools that can partially automate the product management process? I know some people are trying to use AI robots to simulate real - time user responses and even build simulated market environments or user groups.

Andrew Ng: There are indeed some directions being explored. For example, Figma demonstrated the great value of design tools in product development during its IPO process; some people are also trying to use AI to assist in user interviews, or as you said, use agent groups to simulate user behavior. There are even some academic papers researching how to calibrate such simulation systems.

These are all very promising, but they are still in a relatively early stage. Overall, the help these tools provide to product managers is far less significant than the acceleration effect that programming tools bring to engineers. So currently, product management has become the most prominent bottleneck in startups.

04 The Ideal Founder in the AI Era Must Understand Both Technology and Products

Question: Over time, do you think the image of the "ideal founder" has changed?

Andrew Ng: In my opinion, many methods commonly used in 2022 may no longer work in 2025. So I often remind myself: How much of what we're doing now still follows the 2022 approach? If the answer is "a lot," then we must carefully examine whether it is still suitable for the current situation. In fact, many work processes from 2020 are completely out of place today because the pace of technological evolution is so fast.

Judging from the current trend, I believe that founders with generative AI technical capabilities, those who are technology - driven and also have product leadership, are more likely to succeed than entrepreneurs with only a business background but lacking sensitivity to the direction of AI technology. Unless you have a clear intuition about "what AI can and cannot do," it is difficult to formulate effective strategies and even more difficult to grasp the future development direction of the company.

Question: This is actually a bit like a return to the old - school Silicon Valley style. Look at Bill Gates, Steve Jobs, and Steve Wozniak, or the real pioneers in the semiconductor and early Internet eras - they were all technology geeks. I have a feeling that we once lost this characteristic, but now it's obvious that technology companies still need technical leaders.

Andrew Ng: In the past, we may have placed more emphasis on the resumes of entrepreneurs. For example, "This founder has had one or even two successful exits, so let's support him again." But in a period of drastic technological change - and AI is the fastest - changing variable today - the real advantage comes from a deep understanding of the technology itself.

Take mobile technology as an analogy: Nowadays, everyone clearly knows what a mobile phone can and cannot do, what a mobile application is, and what GPS can achieve... Everyone has a clear idea. So you don't need to be a technical expert to judge "Can I develop an app based on this?" But the development pace of AI is much faster: What can voice applications achieve now? What stage has the AI engineering process evolved to? What's the difference between foundation models and inference models? - Mastering this knowledge is a huge differentiating advantage today. This advantage is far more crucial than "understanding mobile applications" back then.

Question: In your opinion, what other important common characteristics do successful founders have?

Andrew Ng: