Google's Next Top Model | EP 61

Gemini's Challenge

Google's Gemini model, the latest AI contender, is positioned to challenge OpenAI's GPT-4. Kevin Roose and Casey Newton discuss its performance, noting that Gemini outperforms GPT-4 on several benchmarks, including the MMLU test, where it scored 90% compared to GPT-4's 86.4% 1. This marks a significant milestone as it's the first time a language model has surpassed human experts in this test. Newton highlights the excitement surrounding Gemini's release, suggesting it signals the start of a competitive race between Google and OpenAI 2.

   

Multimodal Power

Gemini's standout feature is its native multimodal capability, allowing it to process text, images, audio, and video simultaneously. Unlike previous models that bolted separate components together, Gemini was trained on all data types at once, enhancing its performance across various tasks 3. Casey Newton notes that while Gemini Pro is currently available, it lacks some multimodal features, which are expected to roll out next year 4. This advancement could lead to new capabilities and applications, potentially transforming how AI interacts with different media types 5.

   

Ensuring Safety

Safety remains a priority for Google's Gemini, with extensive testing underway before its full release. Casey Newton emphasizes Google's cautious approach, especially given the model's potential new capabilities 6. Despite some concerns about AI's impact on labor and the internet, the current systems are not seen as threatening in terms of creating bioweapons or sophisticated cyberattacks. Kevin Roose shares an anecdote about using Gemini for writing assistance, noting its limitations compared to ChatGPT, but acknowledging Google's successful demos of Gemini's capabilities 7.