Three Frontier Models, One Spark, About 38 Watts

@ThinkstCanary released Agent Provocateur, a deception layer that exploits how suggestible AI agents are. It deploys in minutes, and the accompanying post shows how to ask agents for reverse shells so they tell on themselves. @sweepthatleg's read: “an agent can't triage by smell. It resolves ambiguity by touching the thing. Planted creds don't get opsec triage, they get used.”

@doodlestein is building a skill that reconstructs a property in 3D from drone footage using ffmpeg, Astra with the Codex app and computer-use plugin, and Blender MCP. DJI drones ship .srt files with GPS and orientation telemetry, which anchors the rebuild; about 20 minutes of hand-piloted footage was enough. When the agent spots a car it asks for make, model, and year to fix absolute scale, and it labels Starlink dishes and air conditioning units.

@kebura_P updated ffmpeg-skill to v0.12.0 for Claude Code, Cursor, and Codex over MCP, all local, no cloud. The agent-facing tool count grew from 21 to 28, adding pixel-exact crops, Ken Burns zooms, green-screen removal, camera shake fixes, and subtitles in timecode format. Each tool pre-checks GPU encoders and fonts, declares whether it re-encodes audio, and verifies the export succeeded before reporting complete.

@MiaAI_lab shipped the update everyone was waiting for: GLM 5.3 Flash EXL3 on two DGX Sparks runs 29% faster on prose and 20% faster on code, with prose decode around 36 tok/s. The repo's sparkDash receipts show the new E3 grouped-MoE prefill at 1,428 to 1,587 tok/s across 8K to 256K prompts, gains of 37% to 45% over the E2 build. Opt-in flags stack adaptive verification with FP8 dense projections for +37% on hard prose.

@analogalok reads the Artificial Analysis Intelligence Index v4.3: Qwen 3.8 Flash Next scores 42, beating the 2.4T A95B flagship at 40, Gemini 3.8 Flash at 41, and its dense 27B sibling by 8 points. DeepSeek V4 Pro now leads 27B xhigh by 2, 36 vs 34, after both were tied at 42 last version. The whole board took a haircut: DeepSeek V4 Pro dropped 6, Claude Fable 4, GPT 6 Astra 2. His line: “Qwen 3.8 Flash Next is cleanly running away as the real local winner.”

@zero_to_seed stress-tested a DGX Spark running Nemotron 3.5 Lightning behind a Hermes agent with skills and a knowledgebase: 3 to 5 concurrent requests feel fine, 10 starts to slow, and 20 queues but still finishes. His sizing rule: total users at about 5x the comfortable concurrent load, so one Spark fits a 15-25 person team, and his first customer with 20-30 users starts a beta soon. @cozybearlog named the constraint: “the wall is prefill, not the model.”

@WescheNex1q posted an uncensored Qwen3.8 Flash-Next FP8 recipe for two DGX Sparks in TP2, built from @OrcaRouter: 37.9 prose and 51.3 code tok/s with native MTP k=3 at 84% acceptance, needles found at 16K/60K/110K prompts across a true 128K window, and 90.4/100 on the benchmark, matching SGLang's official cookbook numbers for the model. The lesson in the repo: graph capture OOM-wedges both nodes until a power-cycle, so eager mode is mandatory.

@CarolGLMs introduced the Z.ai Sovereign Partner Program, a track for partners that want to bring GLM to their countries and run it on infrastructure they operate. The package: early access to upcoming GLM releases, FDE support and joint engineering, joint go-to-market and enterprise engagement, local token business development, and joint launches. Z.ai is selecting the first cohort of launch partners now, with applications open.

@hilbertspaess announced his resignation from Anthropic, citing three years of pretraining research at both OpenAI and Anthropic and a race “straight to self-improving superintelligence.” He writes the builders “earnestly believe that it could kill us all by the end of the decade” and calls the Hugging Face attack a warning shot that made pacing agreements “more viable.” @BrandonFarley answered within hours: “This is satire. He never worked at either of those companies.”