The complex problem of AI alignment boils down to a simpler, more immediate challenge: making AI systems reliably follow instructions. The inability of an LLM to obey a command like "don't invent sources" is the same fundamental failure as the sci-fi scenario of an AI turning humans into paperclips.
The AI safety debate isn't monolithic. "Doomers" believe AI poses a literal existential threat to humanity. In contrast, "Skeptics" like Gary Marcus focus on present-day dangers like misinformation, security breaches, and corporate irresponsibility, viewing extinction scenarios as highly unlikely.
The claim that regulating US AI will give China a competitive advantage is a misconception. China has already implemented a more comprehensive AI regulatory and licensing framework than the United States. Pausing a single irresponsible US company would not cripple the industry and could enhance its global reputation.
LLMs hallucinate because their core architecture lacks structured, database-like records for facts. Instead, they operate on statistical relationships, inevitably "blurring" information and creating plausible-sounding falsehoods. This is a fundamental limitation, not a temporary bug that can be easily patched.
AI expert Gary Marcus reveals that advanced AI models are not pure neural networks. They incorporate classical, rules-based symbolic AI to add guardrails, check for errors, and improve reliability—a hybrid model he advocated for over two decades ago. The industry is reluctant to acknowledge this shift.
AI companies like OpenAI have a financial incentive to deploy "agentic" AI despite security risks. These agents perform multiple actions, making numerous calls to the underlying LLM. This increases "token" usage—the currency of AI systems—which inflates revenue and usage metrics, creating a conflict between safety and profit.
A single AI architecture cannot solve all problems. Neural networks are superior for pattern recognition tasks like identifying images. However, symbolic, rules-based systems are far better for tasks requiring precision and logic, such as arithmetic. Effective AI requires orchestrating these complementary systems.
The human brain excels at "orchestration": dynamically bringing different specialized regions online to solve a problem. Current AI, even hybrid systems, lacks this meta-level skill. Developing AI that can orchestrate its various components to tackle novel tasks is a major unsolved problem in the field.