2026-08-17 · 48h window · synthesized from 42 sources / 447 articles · ~2 min read

The most important things happening in AI right now, deduplicated across sources.

1. Nvidia is now financing the customers that buy its chips. It disclosed a $21 billion stake in SpaceX and is putting $1.5 billion into the SoftBank developer building OpenAI's data center, guaranteeing its hardware powers both. Groq raised $350 million to abandon chipmaking for a neocloud business running on Nvidia GPUs. Sources: Ars Technica · TechCrunch · TechCrunch

2. Agents fail in ways the leaderboards hide. DeepSeek's top-ranked V4 Flash completed only 53.8% of complex agent tasks across eight harnesses, and a separate eval harness found models are most confident exactly when they are wrong. Gartner expects agentic AI costs to rise fivefold by 2028, since cheaper tokens cannot offset how many a real workflow consumes. Sources: VentureBeat · VentureBeat · The Register

3. An AI wrote a bug at Snowflake and another AI exploited it. A GitHub Copilot Autofix commit introduced a CI/CD flaw that an autonomous agent later triggered simply by filing a carefully worded Jira issue. It ran through a bug bounty, but the reporting treats it as the first clean machine-to-machine attack chain. Sources: The Register · Wiz Research · Dev.to

4. Anthropic answers the AI backlash with watermarks and a trust argument. Claude's invisible text watermarking turns out to be a version of DeepMind's open-source SynthID-Text, applied to satisfy Europe's transparency rules. Dario Amodei separately called the public's negative view of AI "fundamentally a crisis of trust" in institutions, not a reaction to safety warnings from AI leaders. Sources: The Verge · TechCrunch · TechCrunch

5. Amazon is destructively scanning rare books to train models. 404 Media tracked a shipment of rare books to an Amazon AI training facility, confirming the suspicion behind anonymous, price-insensitive bulk orders placed with dealers. Print is valuable precisely because it was never online for earlier training runs. Sources: TechCrunch · 404 Media · Simon Willison

6. Capital is moving into AI that does physical work. Gravis Robotics raised a $200 million Series A from SoftBank to put autonomy software and hardware on excavators and other construction equipment. Uber added Zipline drones to Eats, took a stake in the company, and set a target of one million daily drone deliveries by 2029. Sources: The Robot Report · The Verge · TechCrunch

Read the living page: https://jacopocastellano.com/ai-digest · Archive