AI digest: Big moves and budget reality checks
Major model releases, enterprise spending reality, and AI security breaches shake up the week.
This week brought heavyweight model launches alongside some sobering reminders about AI’s real-world costs and vulnerabilities.
Alibaba’s Qwen3.7-Plus brings multimodal agents to the masses
Alibaba dropped Qwen3.7-Plus with vision, reasoning, and autonomous tool use all baked in. The interesting bit is the self-programming capability, which could actually make these agent systems practical for everyday work. This feels like the first proper commercial multimodal agent platform that doesn’t need a PhD to operate.
Uber burns through AI budget in four months
Uber capped employee AI spending after staff demolished their annual allocation by April. They’d encouraged unlimited use, then reality hit the finance team. This is exactly what we predicted would happen as companies move from “AI experimentation” to “AI as a line item”.
Meta’s AI chatbot becomes an account takeover tool
Security researchers found hackers could hijack Instagram accounts by simply asking Meta’s AI support bot to change email addresses. No two-factor authentication needed. Meta patched it, but this highlights how AI customer service creates entirely new attack vectors that traditional security teams aren’t thinking about.
JetBrains releases Mellum2 for specialised workflows
JetBrains launched Mellum2, a 12B parameter mixture-of-experts model designed for multi-model AI pipelines. The Apache 2.0 licence and focus on specialised tasks suggests they’re targeting enterprise workflows where you need reliable, fast responses rather than general intelligence. Smart positioning in an increasingly crowded model market.