Website streamed live directly from a model

Original Article ↗ Hacker News Discussion ↗

Overall reaction to the concept

  • Many find the “infinite visual browser” / live-generated illustrated pages a fresh, imaginative interface.
  • Several compare it to Encarta, Dorling Kindersley books, or the “Young Lady’s Illustrated Primer” / Diamond Age vibe.
  • Some see it as an early glimpse of “model as computer” or generative UI where the app is created on demand.

Performance, cost, and practicality

  • Repeated complaints that it’s extremely slow, especially under the “HN hug of death.”
  • Error logs show Gemini API quota/rate-limit issues; the creator says costs are out of pocket and unexpected at this scale.
  • Many doubt practicality until inference gets 10x+ faster and cheaper; some expect huge inference bills to kill such demos.

Accuracy and hallucinations

  • Strong split here:
    • Some users get surprisingly good diagrams (e.g., torque specs, car suspension, hydroponics, cat coat genetics).
    • Many others report severe inaccuracies in domains they know well: car engine bays, PC components, nuclear reactors, game maps, PWR diagrams, shed plans, poker charts, counting, crypto basics, even uploaded Mac Pro internals.
  • A recurring theme: visuals often “look right” but are wrong in crucial details, with mislabeled parts and nonsensical layouts.
  • Several argue this makes it dangerous for learning or any task where correctness matters; others insist it’s “just a demo” and future models will improve.

Use cases and potential

  • Seen as especially compelling for:
    • Kids and casual exploration (“What’s that?” phase, mind-map-style browsing).
    • Sci-fi, fantasy, and surreal prompts where accuracy doesn’t matter.
    • Future educational tools if paired with vetted sources and better models.
  • Some users want source panels, open-sourcing, or domain-specific training (e.g., manufacturer manuals).

Ethical and societal concerns

  • Criticism that it monetizes uncredited human cultural/illustrative work via corporate models.
  • Worry that such tools normalize low-accuracy “Potemkin knowledge,” lowering expectations and potentially polluting understanding at scale.
  • Others counter that early flaws shouldn’t obscure the trajectory; they liken this to early cars or early GANs and predict rapid improvement.

Implementation and framing critiques

  • Some argue the “live streamed website” framing is overhyped; technically it’s sequential image generation plus click-based prompting.
  • Nonetheless, many praise the UX and interaction design even if the underlying technique is simple.