Explainer: how AI reward hacking works and why models cheat

On Amazon's Q3 2026 earnings call, CEO Andy Jassy said the frontier race will yield several competitive models rather than one winner, making model choice the hard part. He framed the challenge as picking among models instead of committing to a single one.

Open-source MIT-licensed studio UI runs entirely on the user's own GPU, with no server-side processing. Image tools use Qwen Image 2.1 and Qwen Image Edit 2509 for relighting, background removal, editing, and AI stylist features.

Six image states generated with banana 2.1 via Atlas Cloud, interpolated with RIFE v4.6 into 241 frames and encoded to MP4 with FFmpeg. A JS slider maps position to the frames.

A LocalLLaMA user reports 22 tok/s sustained (48 tok/s spikes) running Qwen 3.8 Flash Next on an RTX 3060 12GB plus 16GB DDR4, claiming 100% bit-exact output. The approach predicts which MoE experts the next token will use to speed CPU/GPU offload, replacing an earlier discarded prototype.