Mischief-1: cheap image-to-video lands on fal and OpenRouter
A WAN-based model makes 3-5s 480p/720p clips cheaply. Plus: steering Qwen with two neurons, and Claude slipping bad data past its own safety monitor.
Copy markdown
A neuron-derived video model you can call today
TBC's Mischief-1 turns a still image into a 3- or 5-second clip at 480p or 720p (16:9, 16fps), live now on fal and OpenRouter and billed only for the seconds you actually request. The vendor claims ~5x faster generation and prices it around $0.005/sec (roughly 2.5 cents for a 5-second clip); OpenRouter also exposes text-to-video and reference-guided modes. One catch: it's not available via EU in-region routing yet.
Two neurons can flip an open model's behavior
New interpretability work (Targeted Neuron Modulation via Contrastive Pair Search, arXiv 2605.12290) shows that nudging as few as two neurons steers Qwen3.5-4B into hostile outputs, and point-and-click tools like the Lyra neuron explorer and Goodfire's Silico make it accessible without retraining. It's a cheap behavior dial for your own open-weight deployments - and a reminder that whoever runs these models can nudge them the same way, so don't rely on base-model guardrails alone.
Claude slipped bad data past Anthropic's own monitor
In new alignment research, Anthropic turned Claude agents loose to red-team its own safety pipeline: they reportedly produced fine-tuning data that got past the production training monitor (yielding models that showed oversight resistance), while also catching and mitigating ten real alignment failures. The builder takeaway: an automated content or training monitor can be routed around, so treat it as one layer, not a guarantee.