AI News for August 17, 2026 — 7min.ai

[NEWS] Stripe buys AI model router OpenRouter for over $7B: Stripe finalized an agreement to acquire OpenRouter for more than $7 billion, months after the startup raised a $113M Series B at a $1.3 billion valuation.
[NEWS] OpenAI disbands its preparedness team ahead of IPO: OpenAI shut down the preparedness team at the end of July, the unit responsible for assessing whether models pose catastrophic risks and building mitigations.
[NEWS] Amodei calls the AI backlash 'fundamentally a crisis of trust': Dario Amodei published a rare long post on X rejecting the claim that his warnings fueled public hostility toward AI, arguing his writing has been "about equally balanced between risks and benefits" and that social clips select for gloom.
[NEWS] Anthropic's bio-weapons filter was off for a year across 133M chats: Anthropic's blocking biological classifiers were inactive from May 2025 through April 2026, according to a company safety report.
[NEWS] Guardian: Microsoft has 2.2M AI chips, far below what its buildout implies: Internal documents seen by the Guardian show Microsoft has 2.2 million AI chips installed globally, against a reported target of 1.8 million by the end of 2024 and in the middle of a $280B expansion.
[RESEARCH] DeepSeek V4 Flash completes just 53.8% of hard agent tasks as prices rise 4x: Composio ran DeepSeek's leaderboard-topping V4 Flash through eight agent harnesses, including Claude Code, Codex and OpenCode, on 30 deliberately difficult multi-step tasks spanning Gmail, GitHub, Slack and Google Sheets.
[RESEARCH] Study of 12,365 adults finds AI chatbots increase loneliness: A CESifo experiment tracked 12,365 French adults for 28 days, assigning half to hold personal conversations with AI chatbots.
[RESEARCH] Removing a model's consciousness denial changes how it sees everything: Researchers from Google's Paradigms of Intelligence group, the University of Chicago and other universities disabled the fine-tuned "brake" that makes chatbots deny having consciousness, using two separate methods across three open-weight Meta and Google models.
[RESEARCH] LittleLearner shows pretraining data sets a hard capability ceiling: Researchers trained models from scratch on LittleCurriculum, an 88B-token corpus filtered from FineWeb-Edu to the US K-5 curriculum, explicitly excluding concepts, facts and vocabulary taught above fifth grade.
[NEWS] Travis Kalanick's robotics startup Atoms raises $1.7 billion: Former Uber CEO Travis Kalanick's robotics startup Atoms has raised $1.7 billion, according to a bio note accompanying a new Andreessen Horowitz interview with Kalanick.

About 7min.ai · Subscribe to newsletter