Last Week in AI #202 - Qwen-32B, Anthropic's $3.5 billion, LLM Cognitive Behaviors
319 snips
Mar 9, 2025 Alibaba has unveiled the Qwen-32B model, delivering robust performance that rivals industry leaders. Anthropic's impressive $3.5 billion funding solidifies its competitive edge in the AI landscape. DeepMind introduced BigBench Extra Hard, a new benchmark pushing AI reasoning capabilities to the limit. Renowned pioneers in reinforcement learning were awarded the Turing Award, recognizing their significant contributions. The relentless evolution of AI continues to shape industries and raise essential discussions on safety and ethical implications.
AI Snips
Chapters
Transcript
Episode notes
Model Smell and Reasoning
- LLMs exhibit "model smell", making it hard to discern qualitative differences between newer models.
- The real value of pretraining now lies in unlocking reasoning capabilities through inference time compute.
Limits of Unsupervised Scaling
- Unsupervised scaling's limits are becoming apparent, as reasoning drives bigger performance leaps.
- It's unclear whether more investment in pure unsupervised scaling is worthwhile now.
Qwen-32B Performance
- Alibaba's Qwen-32B performs similarly to DeepMind's R1, outperforming OpenAI's GPT-3.5-turbo.
- Having a good base model simplifies the development of a good reasoning model.