Last Week in AI #147 - Google’s Gemini, EU AI Act Deal, Mixtral of Experts, Meta’s ‘Seamless’ translator
41 snips
Dec 17, 2023 The latest discussion dives into Google's Gemini, aiming to compete with ChatGPT but facing some initial skepticism. Voice cloning ethics are spotlighted with Meta's new AudioBox. The podcast also touches on advancements in AI chip technology, with big contenders like Meta and Microsoft making moves against Nvidia. Controversies swirl around Elon Musk's Grok AI, while updates in AI regulation highlight the complex landscape of technological innovation and legal challenges. Tune in for a whirlwind of cutting-edge insights!
AI Snips
Chapters
Transcript
Episode notes
Google Gemini's Multimodal Approach
- Google's Gemini aims to surpass ChatGPT with its multimodal capabilities, supporting text, images, and video inputs.
- It comes in three versions: Ultra (most powerful, unreleased), Pro (in Bard), and Nano (for Pixel phones).
Google's Multimodal Focus
- Google DeepMind excels in multimodal AI, reflected in Gemini's superior image, video, and speech capabilities.
- This strategy aligns with their hardware advantage (TPUs) and potential path to AGI.
Gemini vs. GPT-4: Benchmarking Debate
- Gemini Ultra surpasses GPT-4 on the MMLU benchmark with chain-of-thought prompting, but underperforms on five-shot prompting.
- Google's marketing highlighted the win but created controversy by not comparing apples to apples.