OpenAI Built a Giant Fake Internet for Its Agents to Use and Abuse Safely
THE SO WHAT
A sandboxed ‘fake internet’ for agents is a clear move toward industrial-scale agent training and red-teaming — closer to a simulated economy than a unit test. If you’re building agentic systems, assume eval expectations are rising fast: regulators and enterprise buyers will increasingly ask what synthetic environments you use to test failure modes before touching real users or data.
READ THE SOURCE
MORE FROM THE WIRE
Applied AIPennsylvania will require labels on AI-generated ads
Disclosure on AI-generated ads is moving from best practice to statutory requirement — Pennsylvania’s law is a template other states can copy quickly. If you run performance or political-style creative, add provenance and labeling to your stack now so you’re not rebuilding workflows state by state.
Applied AIBlocks latest AI model Dessa was trained entirely on human speech — and it shows
Training Dessa purely on human speech points to a specialization race in foundation models — expect tighter coupling between voice-native models and payments, commerce, and creator tools in Block’s ecosystem. If you rely on generic speech models today, start mapping where domain-specific voice models could change UX, fraud detection, or customer ops economics.
Applied AIBlock’s new voice model sounds eerily human
Running an "eerily human" voice model locally on an M4 Max means high-fidelity synthetic voice is about to be cheap, offline, and hard to trace. If you operate in payments, support, or comms, assume voice as an auth factor is compromised and tighten verification this quarter.
Applied AIForced to pick, Grok’s AI model chooses Republicans. OpenAI’s picks Democrats.
Model choice is now a de facto political alignment decision — this study’s 90%+ disagreement rate across ~500 races makes that hard to ignore. If you deploy general-purpose assistants in any civic, news, or policy-adjacent workflow, you need explicit policies, evals, and disclosures around political recommendations, not vague “neutrality” claims.