Gradio Distills Qwen's 9B Image Rewriter for About $16

@burkov argues Jev cannot be faster to a decision than an LLM and be as good, because time to first token is the compute a decision costs. His recipe: one request per option, cache the shared prefix, send only the option word, then softmax the first token. @mfranz_on demonstrates the idea in 25 lines of Python and calls the intuition interesting but the hype unjustified.

@MKhordoo's simulation hands a 15-drone swarm to Jev: decision latency under 300ms, all 15 reached the destination, and Jev handled 99% of the steering with a larger LLM stepping in on low confidence. He puts the build cost at $23 on Jev against about $1,100 on Sonnet 5 or Terra. @0xBakeer runs the open Jev through six small games at 60 ms per move on his Spark, 30 on the Mac, hitting 0.975 AUROC on the Flappy clone's flap decision.

@ericzakariasson posts a harness-tuning prompt drawn from Cursor's production experience. One team cut overall token cost about 7% with prompt trimming, tool offloading, cache layout, sparse line numbers and subagent tuning, with no drop in quality; another cut about two-thirds of its system prompt by replacing commands with plain tool descriptions. He warns to measure per task, since a change that shrinks each request but adds turns can cost more.

@TeksEdge reports Djev-Spark makes DiffusionGemma 26B-A4B behave like Jev by pre-building the answer structure, leaving only decision slots open, then reading the model's probabilities instead of generating JSON token by token. It runs locally with NVFP4 on a patched vLLM structured-read engine and exposes a Jev-compatible endpoint, GB10 only for now. His verdict: DiffusionGemma is fast.

@laurent_zw releases AP quants of Qwen3.8 27B in seven tiers from 10 to 16 GiB as a mainline-compatible drop-in replacement. He reports they beat the same-size alternative on fidelity to BF16, beat Unsloth by up to 10% on technical text at 8σ significance, and beat ISTA-DASLab at 10 GiB. In replies he says the evaluation card is updated but the technical-text corpus is not yet uploaded.

@MiaAI_lab ships a cache update for GLM 5.3 Flash EXL3 on 2x DGX Sparks: each request now uses 33-44% less of the model's working memory with no extra VRAM. @plotarmordev's announcement adds that one repeat-prompt case ran 90% faster, and the change is opt-in. @MiaAI_lab keeps GLM 5.3 Flash as the go-to model.

Jeffrey Katzenberg publishes an essay on AI and creativity, recalling a demo that gave him the feeling of watching Luxo Jr. in 1986. He repeats his 2023 prediction that AI tools would cut the time and cost of world-class animation by as much as ninety percent within three years. Asked by a longtime artist "Is this the end of us?", he answered, "Certainly not." His history lesson runs through John Philip Sousa's 1906 campaign and the Copyright Act of 1909.

@dhh announces the Omacom Foundation as a three-year premier sponsor of @0xSero's local AI work through Sybil Solutions, with the goal of making local models work out of the box on Omarchy. The replies are mostly congratulations, with some users pressing him to require credit for borrowed work.

@plotarmordev reads OpenCode's free stealth model Space Bunny, with 1M context, multimodal input and zero data retention, as the start of a stealth model a week. He calls it a marketing stunt that is losing force after Union alpha. OpenRouter took Space Bunny Alpha temporarily offline while the provider works through an issue on its side.

@skeptrune, a month into buying GPUs, reports vendors sending sales forms with prompts built in, written for the buyer's agent to fill with gathered context. He gave the real version to Opus 5.5 and it executed perfectly. He argues every 0.1% of bureaucracy cut creates about $100B in GDP, and expects the practice to spread past today's most AI-forward customers.