ME AI News: SpecPilot recently launched, providing GPUs for coding agents while adhering to a lossless principle. The article says that a single request can more than double the performance of every published speculative decoding system tested. DFlash2 achieved a 4.46× to 10.87× speedup over autoregressive (AR) decoding, with no training required. SpecPilot also offers a lossless 12-bit format whose GEMM performance exceeds cuBLAS. (Source: ME)