Speaking on the AI Engineer podcast, Baseten engineer Philip Kiely said the boundary between training and inference is dissolving and that winning techniques move computational work into training rather than paying for it at inference time. He cited TurboQuant, a viral 4-bit KV cache quantization method that briefly spooked memory-stock investors before production testing revealed it cuts tokens per second by more than half; Baseten never deployed it for real workloads and stuck with NVFP4 weight quantization instead. The clearest win he described is D-Flash, a diffusion-based speculative decoding model that predicts 8 to 16 tokens at once and delivers more than a 3x speedup over the previous state-of-the-art Eagle 3 on a single B200 running Qwen3 8B. Kiely also detailed Baseten's STILL method for KV cache compaction, which amortizes compression through training, and production results showing continuous speculator retraining on live traffic yields a 20% to 2x improvement in token acceptance rates. His forecast points to NVFP4 fluency ahead of Rubin hardware, growing importance of disaggregation, and a compounding training-for-inference flywheel.