Anthropic Built Its Most Capable Model, Then Wired In Three Ways to Make It Work Worse
1. Anthropic built its most capable model, then wired in three deliberate ways to make it work worse Anthropic released Claude Fable 5 on Tuesday and called it state-of-the-art on nearly every capability benchmark it tested, with the largest lead on long, complex tasks in software engineering and
2. Google's Defense Was That AI Overviews Just Aggregate Search Results. A Munich Court Said They're Google's Own Words. For two decades, search engines have leaned on a simple defense: they point to information, they don't author it. The Regional Court of Munich just ruled that the defense doesn't cover AI.
3. An astrophysicist pointed Codex at Einstein's equations, not at his job Chi-kwan Chan studies black holes. Modeling how light bends at the edge of one demands simulation code that survives the extreme physics of general relativity.
In Brief
- Amazon borrows $17.5 billion from banks to fund AI buildout Amazon took a $17.5 billion bank loan days after a bond sale, adding to debt raised across the sector to finance data centers and chips. Companies are leaning on borrowing as AI capital spending outpaces cash flow.
- Microsoft limits employee use of Claude Fable 5 over data retention terms Microsoft restricted internal use of Anthropic's new Claude Fable 5 because of the model's data retention requirements. The same week, Microsoft shipped Fable 5 to GitHub Copilot and Foundry customers.
- OpenAI reports PRC-linked operations using its tools to shape US AI debates OpenAI published findings that PRC-linked influence operations used AI to push narratives on data centers, tariffs, and US tech policy. The report also cites false claims about ChatGPT spread through these accounts.
- Fired xAI engineer sues over Grok safety warnings A former xAI engineer sued the company and SpaceX, claiming he was fired for raising safety concerns about Grok. The firing came days before SpaceX's IPO, according to the complaint.
- Google will store Lens photos and Search audio for AI training by default Google told users it will save images, files, audio, and video from searches under a new "Search Services History" setting. The change covers Google Lens photos, Search Live recordings, and Translate audio.
- Independent musicians sue Google over Lyria music AI training A group of independent musicians sued Google, alleging it trained its Lyria 3 music model on songs they uploaded to YouTube. Google has not confirmed whether it uses uploaded tracks for training.
- OpenAI brings its models and Codex to Oracle Cloud OpenAI made its models and Codex available through Oracle Cloud, letting customers spend existing Oracle commitments on them. The deal adds enterprise security and governance controls for deployment.
- Google releases DiffusionGemma, a text-diffusion model running over 1,000 tokens per second Google's experimental DiffusionGemma generates blocks of text at once instead of token by token, hitting 1,000+ tokens per second on an H100 and 700+ on an RTX 5090. The 26B Mixture of Experts model ships under Apache 2.0 for local interactive workflows.
- Claude Desktop spawns a 1.8 GB virtual machine on every launch A user reported that Claude Desktop on Windows starts a Hyper-V VM consuming about 1.8 GB of RAM at each launch, even for chat-only use. On a 16 GB laptop that is over 11% of memory held by unused agent infrastructure.
- Researchers propose tuning agent harnesses from past runs without labeled data Retrospective Harness Optimization improves an agent's skills, tools, and workflows using only prior trajectories, skipping the ground-truth validation sets such tuning usually needs. The method targets deployments where labeled data is hard to collect.