Listen to this article in summarized format
In an exclusive interview with The Economic Times Digital, Ng argues that the immediate challenge is to prepare people for work that increasingly involves AI. He notes that computer science students, who have never used a coding agent or are connected to an advanced AI model, even as those skills become essential, continue to graduate from universities.
Edited excerpts
You have said AI is amazing, but it also has limitations. What is it good at and where does it struggle?
Andrew Ng: One of the tricky things about AI is that intelligence is jagged. There are tasks, such as solving math problems, that AI is amazing at, even though they are very hard for people. But there could be slightly different tasks, such as deciding what questions to ask someone in an interview or thinking about business strategy, that do not seem as hard to people but AI could be terrible at them.
One dividing line is that AI is very good at what we call verifiable tasks, meaning tasks with objectively right or wrong answers. We have very good technology for making AI do well at those tasks. There are also judgement tasks, or unverifiable tasks, with no single right answer. AI is making progress on these, but it is much slower.
This relates to the jobs’ question. AI is very good at objective, verifiable tasks, but there are still so many judgement tasks that we need people to do. That is why I think the job apocalypse, or “jobpocalypse”, has been overhyped.
There are calls to slow down AI development. Do you agree the industry should pause rather than move as fast as it can?
Andrew Ng: I’m actually not a fan of slowing down AI technology. It has many beneficial use cases. It is helping in healthcare, and we will find ways to help people learn better, although we are not there yet.
There are also harmful uses. One of the most disgusting things I have ever seen is the use of AI to generate non-consensual intimate imagery. Let’s get rid of that. I think we should accelerate this basic, general-purpose AI technology, which is beneficial for so many things, while slowing down the problematic uses.
Whose responsibility is it to prevent misuse – the companies developing AI, such as OpenAI and Anthropic, or the AI users?
Andrew Ng: If I take a hammer, use it to hammer a nail and miss and dent the wall, whose fault is it? Is it mine or the person who made the hammer?
If I don't wield it properly, it is the user’s fault. The caveat is that tools should meet a basic level of safety. If I swing the hammer and its head flies off and hits something, that is a dangerous hammer.
Most countries end up figuring out the most efficient way to impose liability. For many tools, including AI, it is easier for the person wielding the tool to use it responsibly than to force developers to make it impossible for anyone to use it harmfully. That is simply not possible with a technology like this.
With AI, companies admit their models can go rogue, yet the race continues. Shouldn’t researchers like you help governments build guardrails before the technology accelerates further?
Andrew Ng: There are harmful applications of AI, but I want to push back on the notion that it is so dangerous. Candidly, the risks of AI have been overhyped.
Starting three or four years ago, I was surprised by the intensity of lobbying by a handful of businesses that were spending, and in many cases still are spending, huge amounts of money training proprietary AI models.
If you spend a billion dollars training a proprietary model, it is inconvenient if someone else, perhaps a research lab or a business, trains a similar model and releases it free on the internet for anyone to use. That degrades the value of your investment.
There have been intense lobbying efforts to stoke fears about AI and get regulators to pass stifling laws that would shut down open-source, open-weight and freely accessible models.
Some companies also use this fear narrative for fundraising. A few years ago, an irresponsible analogy was made between AI and nuclear weapons. Nuclear weapons are good for one thing, which is blowing up cities. I do not see a relationship to AI, which is intelligence.
That analogy stoked fear, but it also helped convince some people the technology was very powerful and valuable, which helped with fundraising.
Are you saying that at least some of the negative publicity is engineered to create interest and raise funding?
Andrew Ng: A meaningful part of the publicity serves these business and PR goals. One reason to worry about AI that has increased in recent months was an incident in which a team of agents run by OpenAI hacked into Hugging Face’s servers. OpenAI’s sandboxing had gaps and its monitoring infrastructure was weak. This unfortunate hack happened.
The good news is that, now it has been widely publicised. Teams around the world, including mine, are taking sandboxing much more seriously and putting better monitoring processes in place.
When we build things such as cars, aeroplanes or other machines, the best way to make them safe is to repeatedly identify problems and engineer those problems away. It will be like that for AI as well.
Framing it as “I’m not responsible, my AI agent went rogue” and blaming the AI is not the right way to think about it. To its credit, OpenAI has been transparent and has learnt lessons about putting better safeguards and monitoring.
Another dimension of fearmongering that makes some companies look powerful is the idea of an AI job apocalypse, where 50% of people will be out of work. That is not going to happen.
Many studies, including work by Stanford professor Erik Brynjolfsson, break jobs down into tasks to understand what AI can and cannot do. A lot of these studies show that AI could do 30-40% of many jobs, which is fantastic. But we still need people to do the remaining 60-70%.
Someone who uses AI will be much more productive than someone who does not. Historically, over the long term, salaries have also risen as individual productivity has increased. What I see in Silicon Valley, and think will spread elsewhere, is that people who embrace AI and use it to get more things done will do better themselves, as will their businesses.
Andrew Ng: Software engineering is probably the job category AI is affecting the most because it is so good at writing code. However, software engineering job postings are up because, when software engineers or AI engineers become much more productive, we want to hire more of them. That makes sense from an economic perspective.
Unfortunately, many universities are still graduating people without training them sufficiently to work with AI. It is unfortunate that, at many schools, you can receive a bachelor’s degree in computer science without ever using an AI coding agent or making an API call over the internet to a cutting-edge AI model. These are basic computer science skills now.
For my teams, I would love to hire more good software engineers and AI engineers. But today, I will not hire someone who does not know how to use a coding agent. Their productivity is just so much lower.
Universities worry that students who use AI to write code will not learn the fundamentals. Before they can inspect AI-generated code, shouldn’t they understand how to write it themselves? Is that concern misplaced?
Andrew Ng: I share that concern. I’m going to say something that may sound controversial but is backed by an increasing body of solid data. Chatbots are terrible for learning. Universities have known this for a long time.
If students use AI to do their homework, they get better grades on the homework but learn less and perform less well in a closed-book final exam.
A recent study out of the University of Pennsylvania showed that students who used AI did 17% worse in the closed-book final exam, where they could not use AI. What concerns me is that this was just one point in time and if you compound this type of learning loss over a student’s 10-, 12- or 15-year learning journey, you end up with students who know far less.
AI is great at getting work done, but it was designed to get work done, not to help people learn. Cognitive offloading, getting AI to think for you, is helpful for completing work, so let’s use it. But it results in people learning less.
Some technology leaders say Artificial General Intelligence (AGI) is already here. Why do you disagree?
Andrew Ng: Depending on how you define AGI, it is either here now, was here 30 years ago or could still be decades away. I think most people understand AGI to mean AI that can do any intellectual task a human can. Under that definition, it should be able to learn to drive with about 20 hours of practice or even drive through a forest. We do not have that today. There are many tasks humans can do that AI cannot. We do not yet have that kind of human-level AI.
One alternative definition proposed was AI that could perform 50% of economically useful tasks. But a friend pointed out that, 100 years ago, many countries’ economies were primarily agricultural. When more than 50% of agriculture became automated, did that mean we had reached AGI 30 years ago? You can lower the bar enough.
For the definition most people have in mind, AI that is as intelligent as people, I think we are still at least decades away.
Companies such as OpenAI and Anthropic are attracting substantial investment, with discussion of potential IPOs. Where do you stand on concerns about hype and a possible investment bubble?
Andrew Ng: I’m confident that AI technology will be very valuable. I have many friends at OpenAI and Anthropic. I think I might be the only person on the planet for whom both Sam (Altman) and Dario (Amodei) have worked. I certainly wish my friends and their businesses well.
I’m confident there is no bubble. AI models are very valuable and we still need to build many more data centres to serve them.
Much of the work across the rest of the economy remains to be done in finding applications to build on top of these models. When we look at the valuable use cases in finance, healthcare, transportation and many other sectors, I’m confident we could use enormous amounts of AI inference and data centre capacity. When it gets built, I think it will be used.
That does not mean it is impossible to overbuild. Some people may lose money. I do not want to suggest that is impossible.
Countries, including India, are increasingly discussing sovereign AI as they build infrastructure and adopt the technology. How should they approach it?
Andrew Ng: We live in an era where sovereign access to AI, ensuring that no one can turn off India’s access, is important.
We saw the White House, on export-control grounds, force Anthropic to turn off access to one of its leading models. If Indian governments or businesses want continued access that no one else can switch off, sovereign access is important.
I have been encouraged by the growth of India’s AI model ecosystem. At the same time, one of the most economical ways to ensure access is to support the open-weight model ecosystem.
What most nations need is not to own every piece of AI, but to ensure that no one else can deny them access. Whether India builds it or not, either seems fine. Investing in building it also seems fine.
Consider the networking software that powers the internet, including TCP/IP. Much of it was written by people around the world. The key was that it was released as open-source software. Once you have it, no one can stop you from continuing to use it. That is the power of supporting open-weight models. When the global community funds research and development of these models, it gives everyone access.
What about security and cybersecurity risks in free, open-weight models? How can those risks be addressed?
Andrew Ng: AI models are opening new fronts in cybersecurity in ways that are sometimes exciting and sometimes concerning.
The technology to ship an open-weight model that does something unintended when triggered exists in theory. I’m not aware of that actually happening so far, but security researchers are working hard to defend against it.
One area worth paying attention to is that both cyber offence and cyber defence capabilities have increased significantly because of AI. It is easier than ever for an attacker or defender to coordinate a swarm of AI agents to find vulnerabilities or defend against them.
In America, many cybersecurity researchers are using open-weight models rather than proprietary models for cyber defence. I’m excited about building better tools to enable that.
What should India prioritise to benefit from AI while managing the risks?
Andrew Ng: There is a lot India could do, but if I had to prioritise one thing, it would be training people and giving them the skills to build, use and understand AI. Many of the worries about AI are not about rogue agents. They are engineering problems that need to be solved so we can contain agents and guide them to behave as we want.
There are many technologies we do not control perfectly. I flew to India on an aeroplane and we do not control aeroplanes perfectly. Winds blow them around. But we have learnt to control them well enough to make them very safe.
I do not think we will ever control AI perfectly, but we can do the engineering work to make it very safe.
The key for India is to ensure that enough people, both technical and non-technical, understand the tools and how to guide them towards productive, responsible and safe uses.
The US, particularly Silicon Valley, has long attracted talent from around the world. Could tighter immigration and visa policies benefit other countries if that talent returns home or moves elsewhere?
Andrew Ng: The US has benefited for decades from talented people wanting to work there; having less friendly visa policies are a huge unforced error.
Technology development benefits from having hubs where many people work together. People came to Silicon Valley to work, but the world has also benefited from its innovations.
Would the world be better off if there were no Silicon Valley and talent were simply spread everywhere? I’m not convinced, because the world needs innovation hubs.