Today’s news runs across four lines that all touch the same question: who controls the future of AI and what they intend to do with it. The Vatican publishes a moral framework, a hardware-market analysis names the binding constraint underneath the buildout, DeepSeek’s pricing aggression turns out to be a $10 trillion strategy, and frontier models keep moving faster than the benchmarks designed to measure them.
The Moral and Economic Frame
Two outside voices put the AI industry’s choices in sharper context: the Vatican’s first full encyclical on artificial intelligence, and a strategic read of DeepSeek’s price war as something much larger than a discount cycle.
- Pope Leo XIV publishes an AI ethics encyclical — The Vatican’s first full teaching document on AI covers environmental cost, algorithmic exclusion, and the concentration of AI power. It binds Catholic teaching, not corporate behavior, but introduces a doctrinal frame procurement teams in Catholic-majority markets will cite within a year.
- DeepSeek’s grand strategy is a $10T Chinese AI hardware ecosystem — The 75% permanent price cut last week is the wedge, not the play. DeepSeek is commoditizing the model layer to capture the Chinese AI hardware platform underneath it. The competitive pressure on Western flagships is acute right now.
The Memory Wall and the Speed Tier
Two analyses on the same axis: AI hardware is a memory problem more than a compute problem, and Google’s Gemini 3.5 Flash is the best model at its speed point precisely because it accepts that frame.
- AI hardware is a memory problem, not a compute problem — Category VC’s analysis: the binding constraint on frontier inference is memory bandwidth and KV cache capacity. Hardware companies shipping silicon for today’s bottleneck risk being wrong by the time their chips arrive.
- Gemini 3.5 Flash wins its speed tier, stops short of the frontier — Zvi Mowshowitz judges Google’s new model the best at its latency point but not a threat to Opus 4.7 or GPT-5.5 outside agentic orchestration. The structural play: Google is carving out a third lane, not a flagship competitor.
Models Test Each Other
Three results worth sitting with: a meta-benchmark only one model can pass, a math breakthrough at coffee-shop prices, and a training technique finally specified clearly enough to deploy.
- GPT 5.2 is the only model that can build a hard benchmark — Rohit Krishnan’s BenchBench asks each model to design a benchmark that other models actually find hard. Only GPT 5.2 succeeds. Solver capability and Creator capability turn out to be different things, and standard leaderboards only measure the first.
- AlphaProof Nexus solves nine open Erdos problems at a few hundred dollars each — Google DeepMind’s reasoning model autonomously cleared nine questions that had resisted mathematicians for decades. Inference cost is now in the discretionary budget of any active research group.
- On-policy distillation closes the training gap between student and teacher — A formulation that unifies forward-KL, reverse-KL, and JSD losses gives smaller models a reliable path to inheriting larger-model behavior. Reverse-KL emerges as the default for mode-seeking students.
Agents Multiply, Prediction Markets Don’t
xAI’s first credible coding agent enters a crowded category. Meanwhile, the recurring argument that AI will save prediction markets gets a sharp counter: the bottleneck is demand, not supply.
- xAI ships Grok Build, a terminal coding agent with plan mode and subagents — Bundled with SuperGrok and X Premium Plus, Grok Build adds xAI to the field of Claude Code, Cursor, and Codex. The differentiator is the bundling economics, not the feature surface.
- AI cannot save prediction markets because the bottleneck is demand, not supply — Robin Hanson’s 1990 Idea Futures vision has not failed for lack of liquidity or resolution clarity. It has failed because most people do not want to bet on truth. AI does not fix demand.
Today’s Quick Hits
- Models.dev publishes a queryable registry of model specs and pricing — An open-source consolidation of model capabilities and prices, accessible programmatically. For model-routing code, cost forecasting, and procurement negotiation, the structured dataset matters.
- GPT-5.6 reportedly launches in June with focus on reasoning and agentic flows — Leaked details point at multi-step reasoning, frontend generation, and agentic workflow improvements. The cadence fits OpenAI’s roughly-every-five-months pattern.
- Apple plans iOS 27 visual upgrade for Genmoji and Image Playground — Major upgrades expected at WWDC 2026 in early June. Eighteen months of cartoon-only positioning have made the on-device-privacy tradeoff a product liability.