🎄ThursdAI - Dec19 - o1 vs gemini reasoning, VEO vs SORA, and holiday season full of AI surprises

Meta & Microsoft

Meta's Apollo project and Microsoft's recent AI advancements are making waves in the tech community. Alex Volkov highlights the impressive performance of Meta's Apollo 3B model, which outperforms many existing models with a 55 score on long video benchmarks 1. Despite licensing conflicts, the community has re-uploaded these models, ensuring their availability for further research and development.

Apollo 3B outperforming most existing 7 billion parameter model. Impressive 55 score on long video bench.

--- Alex Volkov

Meanwhile, Google's Gemini 2 is also gaining attention with its experimental flash thinking capabilities, promising exciting developments in AI reasoning and evaluation 2.

   

Cohere & Falcon

Cohere's Command R7b and Falcon models are pushing the boundaries of AI scalability and accessibility. Alex Volkov discusses Cohere's focus on multilingual support and citation verification, making their models particularly suitable for enterprise applications 3. These models excel in reasoning, tool use, and agentic behavior, offering robust performance on lower-end hardware.

Great model, great small model to use Cohere obviously knows what they're doing for those tasks.

--- Alex Volkov

Falcon models, developed by the Technological Innovation Institute in the UAE, are also noteworthy for their impressive benchmarks, particularly in computer science evaluations 4. These models, ranging from 1 billion to 10 billion parameters, highlight ongoing efforts to explore transformer alternatives.