Gemini... is all a lie?

Hi, Scott Lovejoy and Brandon Goldberg here with another weekly update on AI! This week it's all about Gemini.

If you haven’t seen, Google announced a massive new AI model, Gemini. Gemini is a multimodal AI, similar to ChatGPT-4, meaning it can learn and respond to not only text data, but also images, video, and audio.

Along with this announcement they released a mind-blowing video showing its capabilities:

Looks amazing, right? So, what’s the issue?

Google admitted that parts of the demonstration were altered, including reducing latency and shortening the outputs. They also released a table showing how Gemini ranks against ChatGPT-4 in various benchmarks, with Gemini leading in most benchmarks by only a few percentage points. Source: Bloomberg

Google’s Gemini co-lead, Oriol Vinyals, tweeted a video and explanation about how they created the video above. They fed Gemini with still images and text inputs in order to get Gemini to respond with a text output that they then based their video on.

“The video illustrates what the multimode user experiences built with Gemini could look like. We made it to inspire developers.” - per The Verge

This is a good reminder that AI technology is moving very quickly and we can’t accept everything presented to us at face value. There is a ton of money in, and attention on, AI right now and companies are willing to do anything to capture a small portion of it. Google’s technology is impressive, but only slightly more-so than GPT-4.

Is this video fake or reality? In our opinion, Gemini can basically do all the elements presented, but the video itself was built to deceptively imply a more magical cohesive lag-free experience that doesn’t really exist.

How might this affect you? We say, Gemini works better than the free ChatGPT 3.5, but if you’re going to pay for something then we would stick with using ChatGPT 4.

We hope you enjoyed this week’s entry into the soap opera that is the AI world right now.

Until next week,

Scott & Brandon