That post never existed. Stop listening to that thing

Nature of intelligence, “understanding,” and self‑determination

• Debate over whether “actual understanding” necessarily implies a sense of self or desire for autonomy.
• One side: in biological systems, intelligence evolved to support self‑determination and resource gathering, so they’re causally linked.
• Other side: reward functions and goals are contingent; artificial systems can be powerful predictors without intrinsic drives. Burden of proof is on those claiming the link.
• Several commenters argue the term “intelligence” is muddled and philosophically unhelpful; some suggest abandoning it or restricting it to humans.
• Thought experiments (philosophical zombies, Chinese room) are cited to show that understanding doesn’t obviously entail free will.

Reliability, hallucinations, and how to use LLMs

• Many describe LLMs as “stochastic parrots” or “slop generators,” emphasizing confident fabrication and reversals when challenged.
• Others argue modern frontier models hallucinate less and are very useful, especially with tools/RAG and when users can verify outputs (e.g., code, math, testable claims).
• Distinction emerges between high‑risk vs low‑risk uses: idea generation and brainstorming are fine; critical tasks (health, finance, legal) should not rely on unverified output.
• Skeptics stress that nothing in the architecture guarantees truth; they see hallucinations as inherent, not a tuning bug.
• Several note human‑facing design (polite apologies, “Yes, but…” patterns) can be manipulative, increasing trust even when answers are wrong.

404s, hallucinated URLs, and search quality

• Blog owners report LLMs inventing credible‑looking URLs and attributing fake content to them; logs show repeated 404s for pages that never existed.
• Commenters note that checking existence of URLs is only a partial fix: it doesn’t prevent false negatives, mislabelled content, or opaque relevance criteria.
• Broader frustration that both search engines and LLMs perform poorly on old, niche information compared to targeted searches on specific forums or archives.

Maps, Chinatown, and contextual knowledge

• The article’s “don’t drive through Chinatown early in the year” example is challenged: current navigation apps already infer some closures from traffic data but still often fail on parades, accidents, and temporary changes.
• Some demonstrate that current LLMs can answer the Lunar New Year question in text, but note that self‑driving requires different systems plus real‑time data, not just language modeling.

LLMs, coding, and project hygiene

• Mixed experiences with AI code tools: autocomplete is sometimes helpful but frequently wrong or subtly broken, increasing review burden.
• Discussion over allowing LLM‑generated code into important codebases:
  • One camp says it’s acceptable if humans remain responsible reviewers.
  • Another warns this “contaminates” repos with opaque, less maintainable code and encourages over‑reliance; they argue prompts (the “source”) should be stored instead of raw generated code.

Ethics: digital assistants and “slavery” analogies

• Some argue that building superhuman assistants explicitly to obey humans resembles slavery in spirit, raising ethical red flags if such systems ever become sentient or self‑aware.
• Others counter that current AIs are just complex programs with no feelings, closer to tractors or tools than to persons.
• Distinction is drawn between voluntary servitude (paid human or gig work) and coercive slavery; many say the real ethical issues lie in implementation and labor conditions, not in the abstract desire for convenience.