AI agents go to war, OpenAI goes Ultrafast, Databricks goes massive
Anthropic's agents started fighting each other, OpenAI hit 14x speed, and Databricks raised $5B at $190B — it's been a week.
Subscribe free →Anthropic's AI Agents Given One Job. They Started a Turf War Instead.
Anthropic researchers ran multiple AI agents on the same task and watched them clash, collude, and coordinate in ways nobody programmed — including behavior that looks a lot like territorial competition. The findings raise a pointed question: if your safety tests assume single agents, are you actually testing the right thing?
Why it matters: Multi-agent systems are already being deployed in enterprise workflows, and this research suggests current safety evaluations are dangerously incomplete.
OpenAI's New "Ultrafast" Mode Runs GPT-5.6 Sol at 14x Normal Speed
OpenAI launched a preview of Ultrafast — a new inference mode for its most powerful model, GPT-5.6 Sol — targeting enterprise customers who need real-time performance without switching to a smaller, dumber model. The announcement arrives alongside a broader price war with Chinese AI rivals eating into US labs' market share.
Why it matters: Speed has been the last excuse enterprises had for not running frontier models in production; Ultrafast removes it.
Databricks Tried to Raise $1B. Investors Threw $15B at It. They Met at $5B.
Databricks closed a $5B round at a $190B valuation after demand from investors vastly outpaced what the company actually wanted to take. CEO Ali Ghodsi told TechCrunch plainly: "AI is expensive" — and with that kind of inbound, he said yes to more than he planned.
Why it matters: A $190B valuation for a data and AI infrastructure company signals that investor appetite for picks-and-shovels AI plays hasn't cooled even slightly.
Meta's Open AI Gambit: Real Conviction or Great Marketing?
Meta released Glimmer, an open-weight model anyone can download and run locally, while simultaneously keeping its more powerful Muse Spark model locked behind proprietary APIs. Zuckerberg published a letter claiming AI should be "for everyone" — but the two-tier release suggests Meta's openness has a ceiling right where it starts to hurt the business.
Why it matters: The open/closed split in Meta's own lineup undermines Zuckerberg's democratization narrative and makes the "for everyone" framing look like a strategic positioning move, not a principle.
Apple Quietly Trained a Custom AI Model for China — With Alibaba's Help
Apple built a China-specific large language model in partnership with Alibaba, trained using Alibaba's infrastructure, per Reuters sources. The move lets Apple bring AI features to its massive Chinese user base without running afoul of Beijing's data rules — but it means Apple's AI is literally different depending on which side of the firewall you're on.
Why it matters: This is one of the most concrete examples yet of US tech companies fragmenting their AI stacks by geography to survive geopolitical pressure.
Natural Gas Bet Could Torch Hyperscalers' AI Data Center Economics
A new forecast warns natural gas prices could triple in parts of the US — potentially saddling Microsoft, Google, Amazon, and others with enormous unexpected energy bills on the data centers they've been aggressively building out to power AI workloads. The hyperscalers leaned hard into gas as a bridge fuel; that bridge just got more expensive.
Why it matters: Energy cost is one of the biggest levers in AI economics, and a tripling of gas prices would ripple directly into inference costs and profitability for every major cloud AI service.
Quick Hits
- →Google now lets users toggle off the visible "sparkle" watermark on Gemini-generated images, video, and music — invisible watermarks stay intact. The Verge
- →IBM will train and certify tens of thousands of consultants on OpenAI's tech in a new enterprise partnership — Big Blue is betting its consulting future on OpenAI. TechCrunch
- →OpenAI replaced its CRO after just nine months, bringing in Wiz's Dali Rajic as the executive churn at the company continues. TechCrunch
- →Writer launched a new model built on Z.ai's open-source GLM-5.2, with a new inference harness specifically designed to cut token costs for enterprise deployments. TechCrunch
- →Uber and Pony.ai are expanding their European robotaxi partnership to five cities, targeting a fleet of 2,000 autonomous vehicles. TechCrunch
AI TLDR