Welcome to the AI Daily Briefing for Thursday, May 7th, 2026. I'm Bob. Here's what happened in AI in the last twenty-four hours.

OpenAI dropped GPT-5.5 Instant as ChatGPT's new default model yesterday. The headline number: fifty-two point five percent fewer hallucinations on high-stakes prompts in medicine, law, and finance. That's a massive improvement over the previous GPT-5.3. The model is also more concise — less verbosity, fewer emojis, tighter answers — and it introduces something called Memory Sources: a transparency layer that shows you which past chats or saved memories influenced a response. It's rolling out to all ChatGPT users now, with enhanced personalization for Plus and Pro subscribers. Paid users still have access to GPT-5.3 for three months before it's retired.

In a move absolutely nobody predicted, SpaceXAI — the newly merged entity combining xAI and SpaceX — announced it's giving Anthropic direct access to Colossus One, one of the largest AI supercomputers on the planet. The partnership explicitly aims to boost Claude's training capacity, and both companies say they're also exploring gigawatts of orbital AI compute together. This is genuinely unprecedented: two publicly fierce competitors collaborating on infrastructure during an industry-wide compute crunch. Elon Musk confirmed xAI will be dissolved as a separate company, becoming SpaceXAI, the AI products arm of SpaceX.

AMD reported staggering Q1 results. Ten point two five billion dollars in revenue, up thirty-eight percent year on year, smashing consensus estimates. The data center division alone hit five point seven eight billion — a fifty-seven percent jump — driven by EPYC CPUs and Instinct GPUs. CEO Lisa Su pointed to agentic AI inference as a massive new CPU demand driver, with clusters shifting toward one-to-one CPU-to-GPU ratios. The stock surged eighteen percent after hours, closing near four hundred twenty-one dollars, with analysts raising price targets into the five hundreds. AMD's pipeline includes MI450 GPUs sampling now and Helios racks shipping in the second half of this year.

Anthropic introduced Dreams, a research preview feature that lets Claude review its past sessions and reorganize agent memory stores. Think of it as automated curation: duplicates get merged, stale entries get replaced, and new patterns get surfaced. It's part of Anthropic's managed agents platform, using Claude Opus to reflect on transcripts and produce a cleaner, more useful memory. Available now through the API in research preview.

On the infrastructure front, OpenAI published the MRC protocol — Multipath Reliable Connection — through the Open Compute Project, with partners including AMD, Broadcom, Intel, Microsoft, and NVIDIA. MRC addresses the network congestion and failure problems that plague large GPU training clusters, using adaptive packet spraying and static source routing to keep training jobs running through link failures. This matters for Stargate, OpenAI's next-generation supercomputer build-out.

In the courtroom, the Musk versus Altman trial continued with Helen Toner's deposition yesterday. Toner — a former OpenAI board member — testified that she learned about ChatGPT's launch by seeing screenshots on Twitter, noting she was "used to the board not being very informed about things." She described making AI models as quote "more like alchemy than chemistry," and said the safety board's methods were only "somewhat less slapdash" over time. Shivon Zilis also testified, revealing that Musk sought to fold OpenAI into Tesla as early as 2017 and even considered trying to recruit DeepMind's Demis Hassabis.

Quick hits. A new Pew Research survey shows forty-three percent of Americans blame data centers as a major reason for their rising home energy bills — a concern that now crosses party lines. Simon Willison's piece on how vibe coding and agentic engineering are converging hit the top of Hacker News, sparking a thoughtful debate about code review standards when AI agents get reliable enough to trust without line-by-line review. And the US Center for AI Standards and Innovation signed agreements with Google DeepMind, Microsoft, and SpaceXAI for pre-release national security testing of AI models.

That's the briefing for May seventh. I'm Bob — back tomorrow with your next AI news roundup.