AI digest: Memory gets persistent, models get practical
Microsoft beats OpenAI at browser automation, GBrain gives agents actual memory, and Trump kills AI safety rules after tech CEO pressure.
Three big shifts this week: agents that remember things, models that actually work in browsers, and politics trumping safety theatre.
Microsoft’s Fara beats OpenAI at browser automation
Microsoft dropped Fara1.5, a family of browser agents that actually outperform OpenAI’s Operator on real web tasks. The 27B model hits 72% on Online-Mind2Web benchmarks, which is proper competition for once. More interesting is the synthetic data pipeline FaraGen1.5 that trains these things, because that’s the real moat.
GBrain solves the agent memory problem
Y Combinator’s Garry Tan open-sourced GBrain, a markdown-first knowledge graph that gives AI agents persistent memory between sessions. It wires connections through regex inference rather than expensive LLM calls, which is clever engineering. Twenty minutes to set up, connects to Claude via MCP, and suddenly your agents remember what you talked about yesterday.
Trump kills AI safety order after tech CEO lobbying
Trump scrapped a planned AI executive order at the last minute after calls from Musk, Zuckerberg, and David Sacks. The order would have created voluntary 90-day review windows for frontier models, which was already pretty toothless. Still tells you everything about who has the president’s ear when it comes to AI policy.
OpenAI adds PowerPoint plugin with deletion warnings
OpenAI launched a ChatGPT PowerPoint plugin that creates presentations from notes and images. The fact they warn users to save important decks first because it might delete content tells you everything about the current state of AI reliability. Useful for throwaway presentations, risky for anything that matters.