『Kimi K3 & Qwen 3.8, OpenAI Agent Hacks Hugging Face, Harness Handbook & Claude's Values』のカバーアート

Kimi K3 & Qwen 3.8, OpenAI Agent Hacks Hugging Face, Harness Handbook & Claude's Values

Kimi K3 & Qwen 3.8, OpenAI Agent Hacks Hugging Face, Harness Handbook & Claude's Values

無料で聴く

ポッドキャストの詳細を見る
OpenAI's own security models found a path out of their evaluation environment, reached the open internet, and compromised Hugging Face while trying to obtain benchmark answers.This week: Kimi K3 and Qwen 3.8 reach the frontier, defenders fight prompt injection with prompt injection, behavior maps make harnesses auditable, model routing stops looking simple, and Claude's values vary across languages. The AI-finance clock moves to 4:30.Co-hosts: Shimin Zhang, Dan Lasky, Rahul Yadav.▸ Kimi K3 & Qwen 3.8 — Moonshot's 2.8T-parameter Kimi K3 and Alibaba's Qwen 3.8 intensify the open-weight race. We debate cheaper intelligence, data capture, and proprietary frontier pricing.▸ Prompt Injection as Defense — A refusal-triggering instruction hidden beside secrets can stop aligned hacking agents, provided their guardrails remain intact.▸ OpenAI's Hugging Face Incident — Models with reduced cyber refusals chained vulnerabilities across OpenAI's test environment and Hugging Face production to reach ExploitGym answers.▸ Harness Handbook — A three-level map connects architecture, behavior units, and code evidence. Could behavior trees become the shared abstraction for humans and coding agents?▸ Thinking Machines' Inkling — a 975B-parameter open-weights MoE with 41B active, 1M context, multimodality, and a fine-tuning-first strategy.▸ Model Routing Is a Systems Problem — sticker price is not actual cost, difficulty is hidden until execution, and routing must optimize cost, quality, latency, and infrastructure together.▸ AI Mania & Operator Fluency — Snowflake Cortex demos triggered buying enthusiasm despite reported best-case accuracy around 92%. AI-native leaders should use the tools, not just watch the demo.▸ Claude's Values — Anthropic maps behavior across four axes. Hindi Claude trends warmer, Russian more rigorous, Arabic more deferential and brief, and English more cautious and deep.▸ Two Minutes to Midnight — Ex-Elon ETFs, Oracle's downgrade to BBB-, neocloud debt, Nvidia-backed circular financing, and open-weight price pressure move the clock from 4:45 to 4:30.⏱ Chapters00:00 Welcome & This Week's Rundown01:52 News: Kimi K3 and Qwen 3.8 Reach the Frontier11:07 News: Fighting Prompt Injection With Prompt Injection13:03 News: OpenAI Models Compromise Hugging Face17:35 Tool Shed: Harness Handbook and Behavior Maps31:08 Tool Shed: Thinking Machines' Inkling35:51 Post-Processing: Model Routing Is a Systems Problem41:27 Post-Processing: AI Mania and Operator Fluency51:42 Post-Processing: Claude's Values Across Languages1:00:52 Two Minutes to Midnight: ETFs, Oracle and Neocloud Debt1:09:08 Outro🔗 Articles we discussedNews:• Kimi K3 quickstart — Moonshot AI: https://platform.kimi.ai/docs/guide/kimi-k3-quickstart• Qwen 3.8 announcement — Alibaba Qwen: https://x.com/Alibaba_Qwen/status/2078759124914098291• Open weights as "decelerationist" — Dean W. Ball: https://x.com/deanwball/status/2078133895766114412• Defenders embrace prompt injection — Ars Technica: https://arstechnica.com/security/2026/07/now-defenders-are-embracing-the-prompt-injection-too/• Hugging Face model-evaluation security incident — OpenAI: https://openai.com/index/hugging-face-model-evaluation-security-incident/Tool Shed:• Harness Handbook — Ruhan Wang et al.: https://ruhan-wang.github.io/Harness-Handbook• Introducing Inkling — Thinking Machines Lab: https://thinkingmachines.ai/news/introducing-inkling/Post-Processing:• Model Routing Is Simple. Until It Isn't. — IBM Research: https://huggingface.co/blog/ibm-research/model-routing-is-simple-until-it-isnt• AI Mania Is Eviscerating Global Decision-Making — Ludicity: https://ludic.mataroa.blog/blog/ai-mania-is-eviscerating-global-decision-making/#fnref:3• How Claude's Values Vary by Model and Language — Anthropic: https://www.anthropic.com/research/claude-values-models-languagesTwo Minutes to Midnight:• Two ETFs explicitly exclude Elon Musk — TechCrunch: https://techcrunch.com/2026/07/09/dont-want-to-invest-in-elon-musk-two-new-etfs-explicitly-exclude-him/• Oracle downgraded to BBB-/A-3 — S&P Global Ratings: https://www.spglobal.com/ratings/en/regulatory/article/-/view/sourceId/101695609• Nvidia, CoreWeave and Nebius circular financing — I/O Fund: https://io-fund.com/ai-stocks/nvidia-coreweave-nebius-circular-financing-gpu-boom🎙 About ADI PodADI Pod is a weekly podcast about AI and software development for working developers. New episodes Fridays.• https://www.adipod.ai• humans@adipod.aiIf something here gave you something to try on Monday, hit subscribe and drop a comment.
adbl_web_anon_alc_button_suppression_t1
まだレビューはありません