Andrew Ng: How can small teams using AI beat big companies?

“More and more people are writing code, but there are still very few who can truly use AI to write code.”

On November 12, 2025, at the just - concluded Snowflake Developer Conference,

Andrew Ng said:

Let AI help you write code. Stop manual coding.

The founder of Google Brain now uses AI to write code every day. He even joked that if there is no Wi - Fi on the plane, he can't program.

His core judgment: The beneficiaries of this paradigm shift will not be large companies with the most resources, but teams that dare to start with small scenarios.

Why?

With AI - assisted coding, a prototype can be built without a ten - person engineering team; with an open - source model and private data, properly organized, the effect is better than just increasing parameters; it's not about the scale of parameters, but about hitting real needs.

Today's competition is not about who can build a stronger model, but about who has started using AI to do actual work.

This is where the opportunity for small teams lies.

Section 1 | The Weapon of Small Teams: Win in a Small Scenario First

In this conversation, Andrew Ng first pointed out a misunderstanding:

“We've incubated many startup projects in the AI Fund. The hardest part is not controlling costs, but creating a product that users really like first.”

He said that many entrepreneurs are worried about the high cost of using models and high inference costs right from the start, but they've got the order wrong.

The problem to be solved first is not how much money to spend, but whether you've created something that people are willing to use and stay with. If this step isn't working, optimizing costs is putting the cart before the horse. Often, when we really find the right product direction, even if the costs start to rise, we can usually find ways to bring them down.

So, for small teams wanting to build AI products, the first step is not to seek financing or burn computing power, but to find a clear and specific small scenario and achieve results in it.

Based on this idea, he emphasized two things in particular:

First, retain options: When we design the architecture at the beginning, do a little extra work to leave interfaces for switching models in the future.

Don't lock in a particular model or platform right away. Let the system have room for replacement. Use GPT today, and you can switch to Claude, Gemini, or the open - source Qwen tomorrow. As long as the business logic is clear, switching models isn't difficult.

Second, control data: Many SaaS products today actually create a data silo within your organization. The longer you use them, the more locked in you become.

He reminded that the real core asset of an AI product is actually your own data. If this data is stored on someone else's platform, you have to pay to access it, and you need to apply for an API key, then what you're building isn't a moat, but a wall. Of course, you need a place to host your data, but if you can control your own data and let the supplier operate on your side instead of sending your data over, you'll have more initiative.

This also explains why an open - source model plus your own data is especially important for small teams.

Because you don't have the budget to sign expensive APIs, and you can't wait for closed - source models to iterate slowly. Open - source models are flexible and controllable, making them more suitable for quickly validating a small scenario.

Don't aim to be the biggest first, don't pursue perfection. First, find that first real task that can be put into use and run smoothly.

In the era when AI is entering every industry, the opportunity for small teams never lies in doing everything, but in doing small, doing fast, and doing real.

Section 2 | The Coding Threshold Is Disappearing, and Everyone Can Be a Developer

Another key reason why small teams can quickly validate scenarios is that the development threshold is disappearing.

Andrew Ng shared a small detail at the conference.

The night before, he was flying back to San Francisco from the airport. Before boarding, he received a notice that there would be no Wi - Fi on the plane. His first reaction wasn't that he couldn't watch movies, but: Oh no, I can't write code on the plane.

The whole audience laughed when he said this. But then he added:

“This made me realize that I've become very dependent on AI coding tools, such as OpenAI Codex and Claude code.”

And this is becoming the norm.

In the past, writing code was a skill for engineers. Today, in Andrew Ng's view, using AI to write code has become an ability that ordinary people can pick up.

His core view is clear:

Don't write code manually. Don't use old methods. Let AI write it for you.

This is Vibe Coding: You just need to tell AI what you want to achieve, and AI will write the first version, which you can then modify or fine - tune.

He said: Now is the best time to develop the products you like because you can complete them in less time and at a lower cost. People who can use AI for coding aren't just programmers, but also CEOs, product managers, and marketers, who can all complete tasks faster.

This is why he believes that development ability is becoming more widespread, as natural as using Excel or drawing software.

The real threshold has shifted from being able to write code to daring to start.

He even said that the unemployment rate among computer science majors in universities is rising, not because the market doesn't need people who can write code, but because universities haven't adjusted their courses in time to teach students how to use AI to write code.

He himself is also facing this problem: Even I can't recruit people who really understand using AI to write code.

So, in his view, whoever can master using AI to write code earlier can implement their ideas faster.

Now, many developers pay hundreds or even thousands of dollars a month for AI programming tools. Why? Because the output speed has increased several times.

Moreover, Andrew Ng found that when developers use AI to solve their real - world problems, the results are the best.

This also explains why small teams are more likely to benefit from AI programming: They aren't writing code for others' needs, but for solving the real problems they encounter. It's not because they're smarter, but because their goals are clearer, and they're more willing to collaborate with AI.

For everyone who wants to build a product, the question is no longer whether I can program, but:

“Have I started using AI for programming?”

Section 3 | Agents Are Tools for Work, Not Showpieces

In the previous section, Andrew Ng talked about how AI - assisted coding enables everyone to start development. In this section, he takes the topic to a more practical direction: what agents should do and how they should work.

In the past few months, we've seen too many concepts, demonstrations, and marketing slogans about agents.

But Andrew Ng's focus is simple: AI isn't for demonstration; it's for doing work for you.

He said that what's truly valuable isn't creating a chat box that can talk, but whether you can use agents to solve the most unpopular tasks in a company.

The example he gave is very specific and realistic: I'm particularly concerned about PDF files. They're the most common and hardest - to - use data type in every company. Because every company has a large number of such PDF documents: financial statements, medical records, contracts, logistics lists, all scattered in the corners of the system.

“Our company is working on agent - based document extraction, which can automatically identify fields and extract structured data from these PDFs.”

He said that this isn't the future; it's a process that's already in use. Some companies have connected AI agents, which can automatically recognize multi - page financial report tables, fill them into the database for subsequent analysis.

It's not about replacing people or enhancing cognition; it's just about doing work. The work that used to require people to click page by page and copy - paste line by line.

The host also told a joke related to this topic: The most powerful PDF search engine in my life is Command + F.

Andrew Ng continued:

“Really. Sometimes we still have to download PDFs and search for tables ourselves. Now, agents can directly extract the tables and hand them over to analysts, or even directly invoke subsequent workflows for processing.”

This is the current situation. This is a redefinition of AI tools:

It's not an impressive demonstration;

It's not a system with multiple layers of architecture;

It's a virtual assistant that can handle the annoying tasks in your real - world business processes.

He believes:

The next stage of AI is to truly utilize unstructured data such as PDFs, audio, emails, and invoices.

Moreover, he emphasized one point: This type of agent isn't exclusive to large companies; small teams actually have more advantages.

Why? Because they don't have legacy systems, don't need to coordinate repeatedly, and with just one or two developers and a scenario, they can connect the agent.

So, stop treating agents as demonstration projects.

The real way to use them is to make them the non - complaining, non - quitting, non - asking - for - leave executors in your business.

Whoever connects agents to their work processes first can generate real efficiency in local scenarios.

Conclusion | The Threshold Has Dropped, and Speed Is the Key

In this conversation, Andrew Ng talked about practical operations:

Find a real small scenario and build it first;

Use an open - source model plus your own data instead of waiting for large models to become cheaper; keep your data in your own hands; learn to use AI to write code and start developing now.

Technology is becoming more widespread. The threshold for AI products has shifted from technical ability to speed of action.

The gap between small teams and large companies is shifting from resources to execution ability.

📮Original links:

https://www.youtube.com/watch?v=-HWNc-Hd90U

https://www.deeplearning.ai/the-batch/issue-326/

https://www.deeplearning.ai/the-batch/tear-down-data-silos/

https://www.deeplearning.ai/the-batch/improve-agentic-performance-with-evals-and-error-analysis-part-2/

https://ca.news.yahoo.com/google-brain-founder-andrew-ng-053503359.html

Source: Official media/Online news,

This article is from the WeChat official account “AI Deep Researcher”, author: AI Deep Researcher. Republished by 36Kr with permission.