Coding agents more than double every released speculative decoding system, DFlash2 hits 10.87x
berkeley_ai · x · 2026-11-04
SpecPilot: given GPUs and one rule — stay lossless — coding agents improved every released speculative decoding system tested, with DFlash2 going from 4.46x to 10.87x over AR decoding, no training required. The team also introduced a lossless 12-bit format whose GEMM beats cuBLAS. The authors argue MLSys is in a golden age: agents are now capable researchers and low-hanging fruit abounds at the token level.
More from coding & agent
- Google migrates 174k-line Fortran ocean model to Rust in 59 minutes with Gemini 4 Argon — gaganghotra_ · 2026-11-04
- Dev builds MCP to hook Claude into machine databases and auto-read error logs, eyeing hyperMILL next — TheSmashingChamp · 2026-11-04
- sloperator: an open-source linter+MCP that shrinks bloated LLM prose — ResponsibilityDear96 · 2026-11-04
- From chatbot to agent: how tools, planning and memory were added step by step — _jaydeepkarale · 2026-11-04
- OpenAI researcher amplifies tongue-in-cheek verse: 'it might still be good to read some of the code' — lukaszkaiser · 2026-11-04
- Binance's $7B hack reportedly started with a tampered Homebrew install during vibe coding — xuanalogue · 2026-11-04