Show HN: I Remade the Fake Google Gemini Demo, Except Using GPT-4 and It's Real (sagittarius.greg.technology)

Article summary

The article discusses a demo that uses GPT-4 to process video and audio input, similar to Google's Gemini demo, but with actual functionality. The demo's creator used a simple technique to feed images and text into GPT-4, achieving impressive results. The cost of using GPT-4 for this demo was relatively low, at $0.47 for 77 requests. The demo's success has sparked discussion about the potential of GPT-4 and other language models for multimodal tasks.

Main themes

• GPT-4 capabilities
• Multimodal processing
• Google Gemini demo
• Cost and efficiency
• Alternative models
• Language model responsiveness

What commenters say

• Some commenters argue that Google's Gemini demo was misleading and that the actual technology is not as advanced as depicted.
• Others believe that the demo presented is a significant achievement and a step towards more advanced multimodal capabilities.
• There is disagreement about whether the demo truly processes video or just a series of images.
• Some commenters think that the cost of using GPT-4 for such tasks is reasonable, while others believe it could be prohibitively expensive for more complex or frequent use.
• The demo has sparked interest in exploring alternative models, such as LLaVA, for multimodal tasks.
• Some commenters suggest that training a lightweight monitoring model to determine when to engage a larger language model could be a more efficient approach.
• Others propose using techniques like pause tokens or silence detection to improve the responsiveness of language models in multimodal tasks.