Perplexity Deep Research

Original Article ↗ Hacker News Discussion ↗

Performance vs Other Deep Research Tools

  • Mixed comparisons to OpenAI’s Deep Research: some find Perplexity’s outputs shorter, shallower, or “worse,” others report equally strong results and better UX (e.g., CSV export working properly).
  • Users note Perplexity struggles with tasks needing exhaustive coverage and data joining (e.g., “college majors of all Fortune 100 CEOs,” full 50-state tables), where OpenAI/Gemini sometimes do better.
  • Several say Perplexity’s free Deep Research is “good enough” that they’ll drop paid OpenAI tiers; others find it clearly inferior or underwhelming on first try.

Use Cases, Strengths, and Failures

  • Works well for:
    • Summarizing single topics (e.g., Amiga 500 sound chip, niche political issues, GDP PPP ratios).
    • Rapid initial exploration, surfacing relevant papers and references.
  • Fails or under-delivers for:
    • Complex recommender-system design, where it regurgitates boilerplate from common blogs.
    • Domain-heavy or activist-skewed fields, where it uncritically amplifies utopian or unrealistic proposals.
    • Trending topics or “how to combine X and Y” queries, where it rephrases the question without real implementation depth.

Speed, Thoroughness, and “Expertise”

  • Perplexity’s research completes in seconds to ~1 minute; OpenAI’s often runs for several minutes.
  • Some suspect OpenAI’s longer duration is partly artificial (traffic smoothing / marketing signal: “it took a long time, so it must be deep”).
  • Multiple experts describe all current “deep research” systems as shallow: authoritative tone, neat structure, but weak insight and limited contextual understanding.

Commoditization, Moats, and Product-Market Fit

  • Commenters see rapid feature copying: Gemini → OpenAI → Perplexity → open-source clones. “Deep Research” is widely seen as a generic term and now a “term of art.”
  • Debate over Perplexity’s moat:
    • Pro: early, polished LLM+search experience; multi-model access; credible Google replacement for some.
    • Con: seen as a thin wrapper over foundation models, with no clear defensible edge, and “desperate” recent moves.
  • Broader worry that foundation-model vendors are commoditizing entire app categories (RAG, vision, agents, research tools), threatening vertical AI startups.

Evaluation, Hallucination, and Ecosystem Effects

  • No strong, agreed-on benchmark; GAIA and “test it on my expertise” are mentioned, with frequent hallucinations and mis-weighted sources.
  • Some describe Perplexity+DeepSeek R1 as notably better at targeted sourcing, but still not academically rigorous.
  • Concerns that AI search/deep research will drain traffic from publishers, likely accelerating paywalls.
  • Ethical reservations surface around leadership behavior (e.g., strike-breaking offer) and lack of visible dogfooding inside AI companies.