Latest in AI

DeepSeek v4 Coding Scores Clash With Broader Frontier Benchmarks

A Reddit post questions why DeepSeek v4 can rank near the top of coding leaderboards while CAISI reportedly places it about eight months behind the US frontier. The author argues that both views may be compatible because coding benchmarks measure a narrow, heavily optimized slice of capability. For local users, the bigger question is how quantized DeepSeek v4 variants perform in real agent workflows, tool calls, cybersecurity, and abstract reasoning.