Netflix debuts VOID, a vision-language model that can erase objects from a scene and simulate how remaining objects would behave without them (Thomas Claburn/The Register)
THE SO WHAT
Scene editing just jumped from pixel manipulation to causal simulation — VOID doesn't just remove objects, it recomputes how the world reacts. For any product touching video, this turns 'post-production' into real-time what-if tooling and collapses the gap between content creation and synthetic A/B testing.
READ THE SOURCE
MORE FROM THE WIRE
Applied AIDocuments: OpenAI and Anthropic have projected profitability to investors with and without training costs, and report inference costs exceeding half of revenue (Wall Street Journal)
When inference costs run north of 50% of revenue, every product decision becomes a unit economics problem — usage growth without ruthless optimization is underwater growth. Expect aggressive moves on model efficiency, pricing tiers, and verticalization as labs and downstream vendors race to turn GPU burn into software margins.
Anthropic's head of growth says the company culture is so open that people 'just argue with Dario' on Slack
Open argument with the CEO on Slack is a recruiting message — Anthropic is selling itself as the place where safety, product, and research tension is surfaced, not buried. If you’re competing for the same talent, your actual decision forums and escalation paths now matter as much as your comp band.
Applied AIApex Protocol – An open MCP-based standard for AI agent trading
Trading is turning into an agent-to-agent protocol problem — Apex is an early attempt to standardize how AI agents talk, negotiate, and settle across markets. If you're building agentic workflows that touch money, you now have to decide: integrate into emerging standards like this or accept being an island.
Show HN: I built a tiny LLM to demystify how language models work
A 9M-parameter transformer training in 5 minutes on a free Colab is the new 'view source' — LLM literacy is about to spread beyond researchers to any engineer willing to read 130 lines of PyTorch. Teams that treat models as black boxes will get out-executed by those who can instrument, debug, and adapt them at this level.