Rodney Brooks on limitations of generative AI

Perceived capabilities and limits of LLMs

• Many see current models as powerful but brittle: great on small, well-scoped tasks; unreliable on complex, open-ended, or safety‑critical work.
• Some argue “AI is just ML” and dislike the term “artificial intelligence”; others think “AI” is fine as an umbrella for optimization, control, and learning systems.
• Several note that LLMs excel at language tasks that robots historically did not (writing, summarizing, translation), but don’t solve hard physical tasks like warehouse manipulation.
• Disagreement over whether these systems are “super‑intelligent”: some point to unmatched breadth and speed of recall; others say they lack true reasoning, truth concepts, and autonomy.

Usefulness in practice

• Developers report strong gains on small code changes, boilerplate, and unfamiliar libraries, with iterative correction by a human.
• “Look-good-but-broken” outputs are acceptable when a human can diagnose and fix them; unacceptable where autonomy or guaranteed correctness is required.
• Creative tools (e.g., image generation) speed workflows but still rely heavily on human taste, domain knowledge, and post‑processing.

AI vs. ML, thinking, and consciousness

• Long subthread debates whether machines can ever be conscious or think; positions range from strict materialism to idealist views where consciousness is fundamental.
• Some argue thinking doesn’t require consciousness; others insist understanding and genuine creativity do.
• Several call discussions of consciousness a distraction from practical intelligence and engineering.

Scaling, data, and “exponential growth”

• Thread challenges naive extrapolation (e.g., iPod storage) as a guide to AI scaling; exponentials typically bend into sigmoids due to physical and economic limits.
• Some say adding more parameters/data has clearly improved models so far; others cite capacity ceilings (e.g., VC dimension), diminishing returns, and the risk of low‑quality data.
• Debate over whether more context always improves decisions; concerns about overload, irrelevance, and hallucinations.

Robotics and physical-world tasks

• Historical robotics work is discussed: simple, reactive architectures (e.g., for vacuums) succeeded; more ambitious “human-level” robots largely failed commercially.
• Commenters align this with a cautious view: narrow, reliable systems with fallbacks may be more valuable than grand “general” robots in the near term.

Enterprise tooling and integration

• Strong demand for practical integrations: summarizing long email threads, searching org knowledge, querying Slack/Outlook history.
• Some of these already exist (e.g., commercial copilots) but are paywalled and raise privacy, access-control, and hallucination concerns.
• View that “LLM as a component” inside traditional software is more realistic than “LLM is the whole program.”