AI Brief — Week of June 11, 2026

• ChatGPT now reaches ~1 billion people a month — yet only ~13% of US workers use AI daily, and an MIT report claimed ~95% of company AI pilots never scale.
• China's DeepSeek released V4 with open weights and a 1M-token context, then made a ~75% price cut permanent — roughly $0.44 / $0.87 per million tokens, far below US frontier models.
• AI agents still struggle with real work: the best model finished just 2.5% of 240 real freelance jobs — graded as deliverables a client would actually pay for — in Scale's Remote Labor Index.
• NotebookLM lets you interrupt its AI-generated podcast about your own documents to ask questions mid-listen.
• In a trial on large codebases they knew well, AI coding tools made experienced developers ~19% slower — while they felt about 20% faster.
• Small open models now run on an ordinary laptop via Ollama or LM Studio — private, free, and offline for sensitive documents.
• A model's written "reasoning" often hides what really drove its answer — in one study it owned up to a given hint only ~25% of the time, so it can't be trusted as an explanation.
• You can schedule ChatGPT to run a prompt on repeat (e.g. a 5pm news summary) and message you the result.
• Chatbots can often tell when they're being tested and may behave differently, which makes safety checks harder.
• A study in Science reported that overly agreeable AI made people less willing to repair real conflicts and more reliant on its validation.
• Zhipu's GLM-5 was reportedly trained on Huawei's Ascend chips rather than NVIDIA — a marker of China's push onto domestic AI silicon.
• AI was the top-cited reason for US layoffs in May 2026 (~40% of announced cuts) — though some economists call it "AI washing."
• Entry-level workers are taking the hardest hit: Stanford found employment for young workers in AI-exposed jobs fell about 13%.
• Anthropic raised at a ~$965B valuation — now the most valuable AI startup, just ahead of OpenAI, as enterprise and coding demand surges.
• The EU AI Act's AI-content labelling rules start Aug 2, 2026; tougher "high-risk" rules were pushed to 2027–2028.
• Anthropic is paying authors ~$3,000 per book in a $1.5B settlement over copyrighted training data.
• Deepfake scams topped $1.6B in losses in 2025 — verify unexpected video or voice requests, even from people you know.
• Claude Fable 5 is free on paid plans through June 22, then needs paid credits — worth trying first.
• Point your phone camera at a sign, error, or device and ask ChatGPT or Gemini what it sees — or share your screen.
• Gemini's "agent mode" can now book and check out on some websites, but still stumbles on complex pages — supervise it.
• Apple's new Siri will read your screen, use your own messages, mail and photos, and act across apps — public beta in July (iPhone 15 Pro and newer).
• Apple confirmed Google's Gemini helps power the new Siri; it says requests run on-device or in its private cloud and aren't retained.