Qwen 3.8 Goes Open-Weight, Kimi K3 Tops Frontend Arena, X-ray Bots Overconfident

📰 News

• Alibaba’s Qwen 3.8: open-weight “second only to Fable 5” — 2.4T parameters, pitch on coding, productivity, and first Qwen multimodal above 1T. No public benches yet; weights “soon.” Timing looks aimed at Kimi K3’s momentum. → The Decoder

• Kimi K3 wins frontend preference—trails badly on hard math — Code Arena Frontend: 1,679 vs Fable 5 (1,631) and GPT-5.6 Sol (1,618)—first Chinese model on top. FrontierMath Tier 4 ~39% while Western frontier models sit near ~90%. Headlines say “beat Fable”; domains say “depends.” → The Decoder

• DeepMind: video generators already hide world models — GenCeption reuses a pretrained video model for depth, segmentation, and more in one forward pass, trained on a small synthetic set. Fuel for the “generators as universal vision world models” argument. → The Decoder

• X-ray chatbots can be dangerously sure when wrong — RadLE 2.0 scores accuracy and calibrated confidence (including “I don’t know”). Humans 988/2000; best model 758. Fable 5 leads on safe answers; Gemini 3 Pro on raw hits. Overconfident misses are the clinical risk. → The Decoder

• AI text detectors crack when models mimic a writer’s style — Epoch AI: near-perfect on plain AI text, but style-mimic drops ~13% through on average; scientific writing 24–29% miss rate. Weak gate for classrooms and journals. → The Decoder

• Apple’s trade-secrets suit shadows OpenAI hardware — Not just IPO drama: delays risk for a mobile smart-speaker-first hardware push. OpenAI says the complaint lacks merit. → TechCrunch

• Vertu wants $6,880 for an executive AI-agent phone — Alphafold + Hermes Agent (open-source Hermes stack). Luxury leather/titanium; hardware smells ZTE/Nubia partnership. The real question isn’t foldables—it’s whether the agent runs the workday. → TechCrunch

• Model routing is simple—until cache and systems matter — IBM Research: sticker price lied; Sonnet beat GPT-4.1 on agent cost via cache hits. Difficulty classifiers alone don’t cut it when compliance, latency, and specialty stack up. → HF

🔥 GitHub Trending

• Robbyant/lingbot-map — Feed-forward streaming 3D reconstruction foundation model (~20 FPS, 10k+ frames).

• lyogavin/airllm — Run 70B on a 4GB GPU without quantization; low-memory path claims for huge open models.

• tirth8205/code-review-graph — Local-first code graph via Tree-sitter + MCP so coding agents re-read less.

• rohitg00/ai-engineering-from-scratch — 503 lessons / 20 phases; every lesson ships a reusable prompt, skill, agent, or MCP artifact.

• ibelick/ui-skills — Agent skill packs for design engineers (npx ui-skills start).

▶️ YouTube

• Did Kimi K3 really beat Fable? — Matthew Berman | Re-checks the “K3 beat Fable” hype against what the benches actually measure.

💬 Community

• AI mania is eviscerating global decision-making — Execs drafting $2B AI strategies without ever using ChatGPT; vendors who won’t say “100x is nonsense” because customers’ execs need the myth.

• Claude Code ships Bun rewritten in Rust — From v2.1.181, Claude Code runs Bun canary (Rust) in the wild. “Boring is good.”

• Human-like nets by catapulting — Gwern’s catapulting frame for more human-like learning dynamics.

• How does Pangram work? — Good companion read to today’s Epoch detector study.

• Triton for Alibaba SAIL — Triton fork aimed at Alibaba’s SAIL stack—infra under the open-weight race.