Vishal Misra
@vishalmisra.bsky.social
Vice Dean Computing and AI @CUSEAS. Dean of Cricket Analytics @SFOUnicorns. Tweets on academia, cricket, dad/bad/wry jokes. Opinions of guy am pointing at.
Key insight: LLMs encode algorithmic reasoning during training. Chain-of-thought prompting isn't 'teaching' new skills—it's triggering existing computational pathways. DeepSeek's (and similar) approaches made these pathways the default, while others require explicit activation
February 19, 2025 at 2:42 AM
Key insight: LLMs encode algorithmic reasoning during training. Chain-of-thought prompting isn't 'teaching' new skills—it's triggering existing computational pathways. DeepSeek's (and similar) approaches made these pathways the default, while others require explicit activation
TokenProbe visualization shows how prompting unlocks these 'latent algorithms': When asked to solve 33×117 directly, tokens show uncertainty (lighter green). But prompt for step-by-step, and high-confidence tokens (green) reveal systematic computation emerging.
February 19, 2025 at 2:41 AM
TokenProbe visualization shows how prompting unlocks these 'latent algorithms': When asked to solve 33×117 directly, tokens show uncertainty (lighter green). But prompt for step-by-step, and high-confidence tokens (green) reveal systematic computation emerging.