FORGE/DAILY — April 5, 2026
OpenAI published a policy paper this week outlining recommendations for how governments should respond to AI's economic disruption — retraining programs, safety nets, workforce transition support. The framing: OpenAI is working to ensure AI "benefits everyone." The subtext worth examining: this is the same company that just closed a $40B raise at a $300B valuation, is building data centers at a pace that strains power grids, and is actively lobbying against the kind of regulation that might actually enforce the benefits it's describing.
None of this makes the policy recommendations wrong. Retraining programs are good. Safety nets are good. But there's a meaningful difference between a company that builds the machine, profits from it, and then issues recommendations versus one that builds enforceable commitments into its operating structure. OpenAI is doing the former. The policy paper reads more like positioning than a binding commitment — which is fine, that's what policy papers are, but readers should hold it at arm's length accordingly. The actual leverage is in legislation, not in white papers from the companies being legislated.
- 🔥Microsoft drops three MAI models outperforming GPT and Gemini on multilingual benchmarks — Available in Azure Foundry now. If you're building anything that needs to work in languages other than English, these are worth a benchmark run before defaulting to OpenAI.
- 📌UNC AI runs 72 hours of autonomous research, invents its own memory system — Ran 50 experiments unsupervised and produced a long-context memory architecture that beats every human-designed baseline. Filed under: the scientist is increasingly a subroutine.
- 👀Trump signs executive order on AI competitiveness — Focuses on cutting red tape for data center construction and export controls on chips. Long on infrastructure, short on safety guardrails. The administration's position is clearly "build fast" not "build carefully."
- 🔥Google AI Edge Gallery launches on-device Gemma 4 — Local inference on Android, no cloud required. For privacy-sensitive applications this is the most underreported story of the week.
- 👀WHOOP sues AI health coach Bevel for IP infringement — The wearables IP war is heating up. Anyone building health AI products should watch this one.
The OpenAI policy paper is a good example of a company trying to shape the narrative of its own regulation. That's not cynicism — it's just how this works. Every major industry does it. The tell is always in the specifics: what they're asking for versus what they're avoiding. OpenAI wants retraining programs funded by governments, not by a percentage of their own revenue. Worth noting.
Google AI Edge Gallery — On-device Gemma 4 inference on Android. If you've been waiting for a reason to experiment with local LLMs on mobile without the Pi setup overhead, this is it. Free, no API key, runs entirely on-device.
Get the next issue
AI news for people who don't need it explained twice. Free.