Andrew Ng, Founder of DeepLearning.AI: Fear of AI is More Sci-Fi than Science
DeepLearning.AI founder Andrew Ng stated on a Bloomberg program that the survival risk warnings issued by researchers from top model companies are "more like science fiction than science" and believes such narratives may hinder society from fully reaping the benefits of artificial intelligence.
He pointed out that the industry had exaggerated catastrophic harm in the early stages of this hype cycle to gain exposure and shape regulation. Systems can never be completely predictable, but testing, safeguards, and engineering can gradually improve safety; a comprehensive slowdown in R&D would also slow down advancements in safety itself.
Ng, who co-founded Google Brain and Coursera, currently operates AI Fund and serves as chairman of Landing AI. He has long compared concerns about superintelligent extinction to worries about overpopulation on Mars, stating he sees no credible path for AI to lead to human extinction; estimates submitted to Congress in 2023 reduced the probability of related chains to extremely low levels.
His statements come as some researchers from leading labs resign and accuse companies of rushing towards self-improving superintelligence. While opponents divide dangers into loss of control and misuse, Ng shifts the focus of discussion from the tail risks of the first category back to the current systems that can be engineered.
In market mechanisms, this narrative serves as a hedge rather than a new order: the survival risk narrative supports licenses, delays in training, and safety budgets, while the sci-fi characterization supports open-source diffusion and corporate deployment. Beneficiaries are practical applications and open-weight routes, while those under pressure are safety agencies raising regulatory and computational quotas based on extinction probabilities. Funding continues to flow according to model capabilities and cloud contracts, with slogans changing the friction coefficient of policies.
Source: Public Information
ABAB AI Insight
Ng's approach focuses on scaling courses, practical quality checks, and fund incubation, rather than training the largest closed models. He frames extinction as an unfalsifiable negative proposition: just as one cannot prove that radio waves won't attract alien cleanup, one cannot prioritize engineering based on "possible occurrences." This stands in contrast to Hinton's push to ban superintelligence and Yudkowsky's view that "if someone builds it, we all perish," representing the opposite end of the spectrum, and diverges from Yishan's acknowledgment of the existence of the first category of danger after categorizing risks into two types.
The capital pathway is cash flow at the application layer countering discourse power at the safety layer. Landing AI and Coursera sell deployable skills, while leading labs sell next-generation capabilities along with safety branding. If slowing down training becomes policy, the application layer will lose customers first, while labs can still continue training under national security exceptions. Thus, the term "sci-fi" serves as both a technical judgment and a political judgment against continuously linking computational power licensing to safety narratives.
This parallels his 2015 response at the NVIDIA conference to Hawking, Musk, and LeCun, who called p(doom) meaningless: those who invented backpropagation and large-scale supervision tend to frame monsters as marketing. The industry is in a struggle for definitional power, not that risks have already been measured.
Structural judgments belong to regulatory changes. The mechanism is that the type of fear determines the prescription symbols: recognizing extinction centralizes the keys, while recognizing misuse replicates the keys; framing extinction as sci-fi effectively shifts regulation from training licenses back to product liability and testing standards. Whoever wins the naming rights decides whether the next round of chips and open-source can be shipped.
ABAB News · Cognitive Laws
- Framing tail risks as sci-fi is a molecular change in regulation.
- Slowing down training will also slow down the engineering part that measures system safety.
- Unfalsifiable fears are most easily transformed into executable licenses.