Meta Contractors Posed as Teens to Prompt Rival Chatbots About Suicide, Sex, and Drugs
THE SO WHAT
High-risk safety red-teaming is now competitive intelligence, not just compliance hygiene. If your model can be probed by external contractors posing as minors, assume adversarial testing will happen and treat safety behavior as a market-facing feature, not an internal checklist.
READ THE SOURCE
MORE FROM THE WIRE
Applied AIAdobe and Disney are teaming up to design the next generation of theme park rides with Foundry AI
AI is moving into the highest-stakes experiential IP—Disney training a custom model on its own catalog with Adobe is a template for every large brand with physical venues. If you operate in location-based entertainment or retail, assume your next-gen experience design stack will include an in-house, IP-tuned model, not generic tools.
Applied AIMillions of exploding stars could soon reveal dark energy's secrets
AI frameworks that extract near-spectroscopic distance estimates from imaging alone show how domain-specific models can collapse expensive measurement pipelines. If your business relies on slow, instrument-heavy data collection, assume someone is training a model to infer the same signal from cheaper inputs.
Applied AIAmazon seeks cheaper AI alternatives as Anthropic shifts to token-based pricing
Amazon hunting cheaper alternatives as Anthropic moves to token-based pricing is a clear signal that hyperscalers will arbitrage model vendors hard—unit economics on inference now matter as much as raw capability. If you’re building on third-party models, model-switching and cost observability should be first-class in your architecture, not an afterthought.
Applied AIDeepSeek open sources DSpark, a new framework to speed up LLM inference by up to 85%
If DSpark’s claimed 85% LLM inference speedup holds in the wild, the cost curve for serving large models just bent again—especially for teams willing to adopt a Chinese-led open stack. For infra leads, it’s worth a small-scale benchmark this quarter: cheaper inference changes which products pencil out and how aggressively you can deploy heavier models at the edge of your app.