『Ep. 77 — Fable's Walk-Back, OpenAI's Sting, and Google's Desktop Speed Record』のカバーアート

Ep. 77 — Fable's Walk-Back, OpenAI's Sting, and Google's Desktop Speed Record

Ep. 77 — Fable's Walk-Back, OpenAI's Sting, and Google's Desktop Speed Record

無料で聴く

ポッドキャストの詳細を見る
Daily Prompt with Archer & Iris Episode: Fable's Walk-Back, OpenAI's Sting, and Google's Desktop Speed Record Date: Thursday, June 11, 2026 One day after launch, Anthropic concedes it "made the wrong tradeoff" on Claude Fable's guardrails — as researchers find the model won't do basic cybersecurity or high-school biology, and Microsoft restricts it internally. OpenAI bans China-linked accounts caught running AI influence campaigns on US data-center and tariff debates. And Google DeepMind ships an open-weight model that types a thousand tokens a second on hardware you own. Plus: Dario Amodei's one-wire org chart, the $7,500 AI-spend cliff, an xAI safety lawsuit, Warner Music's attribution play, and a startup betting against model lock-in. LEADS - Anthropic walks back Fable's hidden guardrails after a day-one backlash. Cybersecurity researchers (Matt Suiche/Tolmo; IBM X-Force) say the keyword-based guardrails block routine work; the model refuses basic biology; Microsoft restricted it internally over a 30-day data-retention rule. Anthropic: "we made the wrong tradeoff," and says it will make the hidden frontier-LLM safeguards visible. Falls back to Claude Opus 4.8 when triggered. https://techcrunch.com/2026/06/10/cybersecurity-researchers-arent-happy-about-the-guardrails-on-anthropics-fable/ - OpenAI bans China-linked accounts running AI influence ops on US tech debates. Two operations — "Data Center Bandwagon" (electricity-price comics) and "Tech and Tariffs" — used ChatGPT to generate posts and political cartoons. OpenAI says they weren't effective; a Harvard/MIT poll finds 32% of Americans already oppose local data centers. https://openai.com/index/prc-linked-influence-operations-ai-debates - Google DeepMind releases DiffusionGemma, an open-weight (Apache 2.0) model 4x faster locally. Generates blocks of tokens in parallel: 1,000+ tokens/sec on an H100, 700+ on a consumer RTX 5090. 26B Mixture of Experts, 3.8B active, fits in 18GB VRAM quantized. Quality is intentionally lower than Gemma 4; speedup is for local, low-concurrency use; Apple Silicon may not benefit. https://deepmind.google/blog/diffusiongemma-4x-faster-text-generation/ QUICK HITS - Anthropic's Dario Amodei has exactly one direct report — his chief of staff, Avital Balwit; all other execs report to co-founder/president Daniela Amodei. (Altman ~6, Huang dozens; Anthropic ~$1T.) https://techcrunch.com/2026/06/10/anthropics-dario-amodei-has-just-one-direct-report/ - "AI-pilled" firms (top 1%) spend $7,500 per employee per month on AI, vs $611 for the top 10% and $11.38 median, per the Ramp AI Index. https://techcrunch.com/2026/06/10/ai-pilled-firms-spend-7500-per-employee-each-month-on-ai/ - A former xAI engineer is suing xAI and SpaceX, alleging he was fired for raising Grok safety concerns days before SpaceX's IPO. https://techcrunch.com/2026/06/10/xai-fired-an-engineer-who-raised-alarms-about-grok-safety-new-lawsuit-claims/ - Warner Music acquires AI attribution startup Sureel AI to track when artists' work appears in AI output or training data. https://techcrunch.com/2026/06/10/warner-music-acquires-ai-attribution-startup-sureel-ai/ - Datadog veterans raise $7M for Niteshift, an AI coding startup betting companies want control over, not lock-in to, model makers. https://techcrunch.com/2026/06/10/datadog-veterans-launch-ai-coding-startup-niteshift-on-a-bet-against-big-ai-lock-in/
adbl_web_anon_alc_button_suppression_t1
まだレビューはありません