Andrew Ng on AI Coding, Education and the Application Opportunity

Asked for a historical analogy, Andrew Ng calls AI the new electricity: supervised learning and generative AI are general-purpose technologies useful across many activities. Joel Hellermark says the work is to discover and execute those applications.

Ng distinguishes hardware, model infrastructure and applications. He sees room for creativity in applications across shipping, relationship coaching, healthcare and education, and considers that layer less crowded than hardware or infrastructure at the time of the interview.

Ng warns that a thin layer on an existing API may be a good idea without being a defensible business. He compares novelty apps with the more difficult task of building enduring services, using the early smartphone ecosystem as an analogy.

Discussing 2023 model economics, Ng estimates that a then-popular API could generate an hour’s worth of reading for about eight cents under his assumptions. He expects competition among model providers to create opportunities for application builders; the quoted cost is historical.

Ng says some knowledge-work tasks are becoming inexpensive and cites research suggesting higher-wage occupations are exposed to language-model automation. He shares radiologist Curt Langlotz’s view that professionals who use AI may have an advantage over those who do not.

Ng hopes greater productivity will give people scope for meaningful activities. He says he uses language models for factual questions, checking some answers, finding words and brainstorming; he also experiments with business ideas through APIs.

Ng discusses a free developer prompting course produced by DeepLearning.AI with OpenAI’s Isa Fulford. He argues language models are not only consumer chat products but also powerful tools for building applications.

Ng expects prompt skills to matter for machine-learning engineers, while questioning whether prompt engineering will become a major standalone occupation. He compares it to parameter tuning: an important task within a broader job.

The speed of ChatGPT’s adoption surprised Ng, although research signals were visible when GPT-3 appeared. He says the accessible consumer interface brought the technology to a much wider audience, including executives who began experimenting themselves.

On education, Ng says AI teaching assistants and personalized coaches could help teachers and give learners a way to ask questions or practice. He does not claim the final format of AI-assisted education is already known.

Ng says he has learned more about how humans learn by observing his young children than from machine learning. He sees recent AI progress as driven more by engineering than detailed neuroscience, partly because brain mechanisms remain difficult to understand.

Artificial neural networks are highly simplified relative to brains, Ng says. He suspects many different algorithms could produce impressive intelligence with suitable data, and that evolution did not necessarily find an optimal human learning algorithm; he labels this intuition rather than proof.

Ng contrasts the previous decade’s large supervised models with pretrained models adaptable to tasks through prompting or fine-tuning. In 2023 he doubts that scaling these alone is a clear path to AGI and speculates that it may be decades away.

As examples of missing capabilities, he asks when a system will learn to drive with only dozens of hours of experience or absorb a sequence of books and produce deep original research. He sees no fundamental reason it cannot happen, but says he does not yet see a clear route.

Ng opposes the proposed six-month pause on training models beyond GPT-4, arguing in 2023 that it would be impractical and that the existential-risk premise was misguided. These are his contested judgments, not established conclusions about safety or timelines.

He argues that AI advances will be gradual enough to observe and steer, drawing an analogy with managing powerful institutions. He also believes AI could help address threats such as climate change and pandemics; both the analogy and forecast are his views.

Ng identifies nearer-term concerns including bias, fairness, misinformation, polarization and concentrated wealth or power. He argues that attention to speculative killer-robot scenarios should not displace work on these present risks.

For regulation, Ng proposes greater transparency so authorities can understand whether systems amplify harmful content and craft informed rules. He warns that poorly designed restrictions could hinder innovation, including open model development.

Ng says open models bring both risks and broad creative benefits. Using Stable Diffusion as an example, he calls for addressing harmful uses and disparate treatment while preserving the experimentation enabled by wider access.

Ng expects model and application capabilities to keep improving but worries about displaced workers and people left behind. He calls for society to help people participate in the transition and warns against regulation that unnecessarily blocks their opportunities.

Ng calls AI coding the new literacy. His goal is not only to let people use AI built by a few specialists, but to let a shopkeeper, student or other person create a customized system that uses their own distinctive data.