Meta.ai Oh My

LLMs in Hiring and Gaming the Process

• Commenters expect companies to plug LLMs into hiring regardless of policy.
• Some suggest “invisible text” in resumes (e.g., hidden PDF layers/white text) to bias LLMs toward recommending a candidate.
• Others note applicant-tracking systems already flag such tricks, though some hope LLM integrations will omit those checks.

Recall, Reasoning, and Model Quality

• Users compare various models: some open models hallucinate on 90s video-game trivia, while larger or different architectures do better.
• There’s speculation that optimizing for reasoning, summarization, and tools may degrade raw memorized recall.
• Many prefer a strong reasoning model connected to external sources over a weaker “encyclopedic” model.

Hallucinations, Truth, and “Error Bars”

• A major theme: LLMs generate plausible text, not truth; “hallucination” is seen by some as a misleading euphemism.
• One camp argues all output is essentially hallucination; usefulness is separate from truth.
• Others counter that if trained mostly on true data, output is often true in practice and demonstrably useful.
• Several want explicit confidence levels and “error bars” for answers, or at least stronger disclaimers.
• Marketing claims like “hallucination-free LLMs” are widely mocked.

Trust, Search, and Appropriate Use Cases

• Many see LLMs as poor replacements for search, especially for obscure or factual queries, and prefer web search or encyclopedias.
• Others find them extremely valuable for coding assistance, boilerplate generation, translation between languages/libraries, summarization, and text polishing—provided results are reviewed.
• Concern: non-experts can’t easily detect confident nonsense, and LLM answers lack the context/peer feedback of Reddit/Quora-style threads.

Human Comparisons and Cognitive Limits

• Some liken LLMs to “stochastic parrots,” but note humans also confabulate and are biased.
• There’s debate over whether hallucination is a fixable engineering issue or a fundamental limitation of this paradigm.

Future Trajectory (Plateau vs. Superintelligence)

• Several believe LLMs will plateau and won’t lead directly to superintelligence.
• Others see them as an early rung on a ladder toward more general or even superintelligent systems, but details remain unclear.