Show HN: I Remade the Fake Google Gemini Demo, Except Using GPT-4 and It's Real

Original Article ↗ Hacker News Discussion ↗

A developer recreated Google’s polished Gemini launch video using GPT‑4’s vision API and standard browser speech tools, showing that the same kind of multimodal “magic” is already achievable with off‑the‑shelf tech for just cents per session. Commenters contrast this with Google’s heavily edited demo, criticizing deceptive marketing, short‑term stock incentives, and what they see as a pattern of overhyping underwhelming models. The thread also explores technical limits of current multimodal systems—such as lack of true continuous video/audio input, latency, and cost—and points to open‑source alternatives like LLaVA and local inference as promising directions.

Demo and Capabilities

  • Commenters find the GPT‑4 vision demo impressive and note that similar functionality has been possible since GPT‑4V (interleaving images and text).
  • The author reports 77 vision API calls costing about $0.47; some see this as reasonable for demos but worry about daily production use costs.

Comparison to Google’s Gemini Demo

  • Many argue this shows Google could have made a real, honest demo instead of a heavily edited “visualization.”
  • Some see the Gemini video as deceptive marketing that hurt trust and investor sentiment; others note the stock initially rallied and say the real impact is unclear.
  • There is debate on how much was actually “faked,” but it’s widely accepted that Google later admitted to substantial post‑processing and prompt changes.

Technical Approach: Frames, Audio, Latency

  • This demo captures periodic screenshots (e.g., every ~800 ms) rather than true continuous video.
  • Several comments argue that “video is just frames,” while others emphasize that temporal continuity, motion prediction, and synchronized audio make real video qualitatively different.
  • Audio is handled separately via browser speech‑to‑text and text‑to‑speech; people are surprised how effective built‑in browser APIs can be.

Real-Time Multimodal Interaction and Turn-Taking

  • Multiple comments highlight that the “magic” in the Gemini promo was continuous listening and knowing when to speak or stay silent.
  • People propose:
    • Special tokens or training for pauses/turn-taking.
    • Lightweight “monitor” models that decide when to wake the big LLM.
    • Heuristics based on silence/motion detection and partial transcripts.
  • A prototype continuous voice chat system is described with low latency and interruptibility, but with limitations in reliability, turn-taking, and lack of emotion/prosody handling.

Local and Open-Source Alternatives

  • LLaVA and llamafile are praised for strong on-device vision performance; examples include auto‑renaming photos from image content.
  • Licensing complications are noted: code may be permissive (Apache 2.0) while weights are non‑commercial.
  • Some suggest local inference (e.g., via LocalAI/Mistral) as a path to lower per‑use costs.

Business, Ethics, and Hype Around Google

  • Many see Gemini as part of a pattern of over‑hyped or misleading Google demos and failed products.
  • There’s discussion of short‑term stock incentives, principal–agent problems, and “metric hacking” driving deceptive marketing.
  • Some now doubt Google ever had significantly better hidden models; Gemini Ultra is viewed as, at best, not clearly ahead of GPT‑4.

Follow-On Projects and UX Reality

  • One commenter reports Google’s conversational translation app feels unusable in real life, reinforcing skepticism about polished demos.
  • Speech recognition error rates and cumulative frustration are highlighted, especially for continuous use and accessibility.
  • Inspired by the demo, someone built a “Sorting Hat” system for trash/compost/recycling, tested live with a child, showing immediate educational uses.