AI Weekly Update - May 12, 2025
The Pope calls AI "humanity's core challenge", Gemini 2.5 Pro improves its coding abilities (again)
what to know for now
💻 Google ships Gemini 2.5 Pro preview with stronger coding. Retrieval‑augmented generation lifts unit‑test pass rate nine points on public benchmarks. The SDK adds tool‑calling, web‑worker streaming, and per‑function temperature knobs. Google plans stable availability after I/O on May 15. Read more
🎨 Figma rolls out design and code AI inside Config 2025. New tools auto‑generate wireframes, translate mocks into React, and summarize design comments for engineers. The suite plugs into Multiplayer Dev Mode, letting teams hand off prototypes with component diffing. Read more
👥 Instacart CEO Fidji Simo jumps to OpenAI command role. Simo becomes “CEO of Applications,” steering ChatGPT product, operations, and finance while Sam Altman focuses on research. Her consumer‑commerce background positions OpenAI to chase ad revenue and enterprise deals before a public listing. Instacart begins its search for a successor as growth accelerates. Read more
⚖️ Trump sacks U.S. copyright chief after AI training memo. Shira Perlmutter’s report laid out legal pathways for fair‑use ingestion of media into model corpora, clashing with entertainment lobbyists aligned with the former president. Hours after publication, the White House terminated her appointment and installed an acting director sympathetic to content owners. Scholars expect renewed litigation over training‑data exceptions. Read more
🧪 AI Research of the Week
Scalable Chain of Thoughts via Elastic Reasoning
From Salesforce AI ResearchJake's Take: Elastic Reasoning chops chain‑of‑thought into a think phase and a solution phase, each with its own token budget (hard length limit measured in GPT‑style chunks of text). A budget‑constrained rollout policy trains a 1.5B math model and a 14B code model that adapt to any cap; the 14 B variant posts a 1987 Codeforces rating (96 percentile!) while cutting tokens almost in half.
By turning token count from guesswork into a dial, the method lets edge devices (phones or laptops running models offline) stay within compute and privacy budgets. Future agents will likely split planning and answering live, milking smaller models for sharper reasoning without extra silicon.
what to know for later
💰 OpenAI and Microsoft explore refreshed funding deal ahead of IPO. OpenAI told investors it seeks restructuring of Microsoft’s equity once new cash arrives, easing cap‑table limits before a flotation. Microsoft may trade shares for extended preferential access to frontier models beyond 2030. Read more
⛪ New pope calls AI humanity’s core challenge. In his inaugural policy address, Pope Leo XIV framed artificial intelligence as a threat to dignity, justice, and labor. He urged governments to protect workers displaced by automation and to forbid exploitative surveillance. The Vatican will convene ethicists and engineers this summer to draft guidelines for Catholic institutions deploying models. Read more
🏛️ AI chiefs urge Congress for light‑touch oversight. Sam Altman, Satya Nadella, Lisa Su, and CoreWeave’s Michael Intrator testified that rigid caps on model size would hinder U.S. competitiveness. They proposed disclosure registries, independent audits, and tax incentives for domestic fabs instead of license pre‑approval. Senators signaled bipartisan interest in the industry‑backed framework. Read more
🌐 OpenAI launches program to build sovereign model stacks. “OpenAI for Countries” offers turnkey datacenters, national‑language chatbots, and safety councils to partner governments. Packages include GPT licenses, deployment support, and fine‑tuning with public data held in‑country. Early talks involve Indonesia, Kenya, and Norway. Read more