GPT-5.6 cracks the same quantum crypto problem twice in three hours; MiniMax H3 tops open video

📰 News

• Two teams solve the same quantum-crypto problem with GPT-5.6 three hours apart — MIT and UCSB/UCLA researchers used Sol Ultra on “unclonable encryption,” different paths, near-simultaneous arXiv drops. “If someone names an open problem, first see if GPT solves it.” Research workflows are shifting. → The Decoder

• Karpathy’s next vibe test: LOTR paragraph → 3D browser scene with Claude Opus 5 — ~2 hours, 5,500 lines, ~$10 in tokens. A step past pelican-on-a-bike SVGs toward cheap, on-demand game worlds (still a vibe check, not a hard bench). → The Decoder

• MiniMax H3: first open model to top an AI video ranking — Artificial Analysis: #1 video editing, #2 T2V, #3 I2V. 33B multimodal, 4–15s stereo clips. 2K module and Context-IR stay closed; commercial use capped under $20M revenue. → The Decoder

• IBM: 92% of firms hit by AI security incidents lacked basic access controls — Entry points often APIs, connected apps, misconfigured cloud—not the model weights. Avg AI-related breach cost: $5.33M. → The Decoder

• Alibaba’s Qwen 3.8 sells job displacement as hobby time — Marketing video: model works, human paints/plays. Softer than fear-based competitor messaging. → The Decoder

• DesignArena raises $7.9M to inject human “taste” into AI models — A/B ranking for web, images, and other visuals. Frontier labs buy the feedback loop; claimed $60M ARR. → TechCrunch

• ChatGPT is Congress’s favorite AI spend (~90% of House AI dollars) — ~$100K on ChatGPT vs Claude second; memos, bill summaries, hearing prep. → TechCrunch

• Interpol: AI is the “core operational driver of cybercrime” across Africa — AI in 55% of reported cases; deepfake extortion and more. → The Decoder

🔥 GitHub Trending

• lyogavin/airllm — 70B on a 4GB GPU without quant; expert streaming for huge MoEs like Kimi K3.

• different-ai/openwork — Open-source Claude Cowork alternative; share skills/MCPs via OpenWork MCP across agents.

• microsoft/AI-For-Beginners — 12-week, 24-lesson AI curriculum for beginners.

• microsoft/generative-ai-for-beginners — 21 lessons on building generative AI apps.

• zhaoxuya520/reverse-skill — Skill router for reverse eng / authorized pentest (APK, binaries, CTF).

▶️ YouTube

• Another DeepSeek Moment Has Arrived — Two Minute Papers | DeepSeek V4 Flash 0731 as another open-model jolt.

• NVIDIA’s AI Learns Why Copying Humans Isn’t Enough — Two Minute Papers | NVIDIA work beyond pure imitation (HIL).

• Hidden Codex Setting — Matthew Berman | Quick short on a lesser-known Codex setting.

💬 Community

• Open letters about AI development — Simon | MS-led open-weights letter, Anthropic’s reply, frontier staff on “pacing the frontier.”

• Devtools must be open source — LLMs lower the friction of clone/build; open-source “read and modify” starts to feel real.

• Why we write our own C/C++ inference engines — LocalAI on owning the inference path.

• You Could Have Come Up With Kimi Delta Attention — Approachable walkthrough of Kimi Delta Attention.

• condense-json 1.0 — Compress repeated JSON substrings with $r refs—handy for LLM logs.