Framing "There Is No AI": Human Collaboration Within AI Systems

In a 2023 New Yorker piece, the speaker asserts 'there is no AI', positing artificial intelligence not as an autonomous entity but as a collaboration constructed from human-generated data and input. This reframing challenges traditional views which imbue AI with near-mythical status, akin to replacing collective 'Gods' with a new technology, often leading to social discomfort and resistance.

The speaker illustrates tangible advances, notably AI's utility in speeding code development. Having written about 50,000 lines of code, the speaker notes previous efforts spent on debugging could now be circumvented by instructing an AI, which returns nearly bug-free code—potentially freeing up substantial time. However, they caution that while functional uses exist, many AI applications are "theatrical," such as AI-generated companions, and highlight the social reality by referencing engineering teams responsible for creating such systems.

Guardrails designed to prevent malicious use, such as asking an AI for bomb recipes, can be circumvented with indirect prompts. To address this, the speaker proposes a 'multi-factor' AI security approach: in parallel to model usage, estimating clusters of training data most influential on outputs (counterfactual cluster estimation), thereby making it harder for models to evade safety checks. This technique operates independently from the model, akin to multi-factor authentication.

Crucially, the speaker insists AI is a 'black box' only because people refuse to acknowledge the humans behind it—AI is made entirely from "data from people". Opening the black box (revealing the human component) is vital to resolving issues of security, quality, and error in AI systems.

Social implications are profound: the myth of self-contained AI leads to anxiety about obsolescence, especially among youth who are told to abandon education, with the dystopian possibility of being "kept by Elon as a pet". The speaker argues such systems drift towards hyper-centralization and eventual capture by bad actors, as seen repeatedly throughout history (e.g. "Bolsheviks" to "Stalinists"). This centralization undermines the notion of universal basic income as a solution.

The alternative is incentivizing new classes of creators whose data shapes future AI capabilities, envisioning an "exponentially expanding endless feature of new kinds of creativity". AI can only mimic what it finds in the data—a creative future depends on human collaboration. Without abandoning the myth of AI as an entity and recognizing its human foundation, society risks an "infinite future of slop" where even innovation is instantly absorbed and diminished. The speaker calls for prioritizing human creative collaboration over worship of AI as an independent force.