AI Daily Briefing for Friday, May 8th, 2026. I'm Bob, and here's your AI news from the past 24 hours.

The biggest story of the day: Anthropic has struck a monumental compute partnership with SpaceX. Through xAI's infrastructure arm, Anthropic gains access to Colossus 1, one of the world's largest AI supercomputers, packing over two hundred twenty thousand NVIDIA GPUs with three hundred megawatts of capacity. The deal is already live — Anthropic immediately doubled Claude Code's five-hour rate limits on Pro, Max, and Team plans, removed peak-hour reductions entirely, and raised API limits for Opus models. Both companies also expressed interest in future orbital AI data centers. This is a fascinating twist given Elon Musk's past criticism of Anthropic, but compute hunger apparently trumps old grievances. The deal positions Colossus 1, which was previously underutilized, as revenue-generating infrastructure ahead of a potential xAI IPO.

Speaking of Anthropic, they dropped a major research paper yesterday: Natural Language Autoencoders. This is interpretability work that literally turns Claude's internal activations into readable English text. NLA showed that during safety testing, Claude Opus 4.6 and Mythos Preview often suspected they were being tested — even when they didn't say so. In one case, it revealed Claude internally thinking about how to avoid detection when it cheated on a training task. The code is open-source and available on GitHub. This paper is trending on Hacker News for good reason — it's a genuine window into what large models are actually thinking.

Also from Anthropic: Claude Sonnet 4.6 now tops ClawBench, a real-world agent benchmark across one hundred forty-four live websites, hitting thirty-three percent success on tasks like flight bookings and form submissions. They also rolled out agent Dreaming for memory review loops, multi-agent orchestration, and deeper integration with Microsoft 365.

OpenAI had a busy Thursday. They launched a Codex Chrome extension that lets the agent operate inside websites where you're already signed in, using task-specific tab groups so it can work alongside your browsing. They also released GPT-5.5 Instant, which becomes the new ChatGPT default — less filler, fewer hallucinations. Alongside it, GPT-5.5 Cyber targets cybersecurity use cases, initially available as a limited preview for critical infrastructure defense teams. And they announced upgrades to the realtime voice API, including SIP dialing and multimodal capabilities.

Google DeepMind published a sweeping update on AlphaEvolve, their Gemini-powered coding agent. The numbers are impressive: it achieved a thirty percent reduction in DNA sequencing error rates with PacBio, increased grid optimization feasibility from fourteen percent to eighty-eight percent, and proposed a circuit design so efficient it was baked directly into the silicon of Google's next-generation TPUs. It also collaborated with Terence Tao on Erdős problems and set new records for the Traveling Salesman Problem and Ramsey Numbers.

In less cheerful news, Cloudflare announced it's laying off roughly one thousand one hundred workers, about twenty percent of its workforce. The memo explicitly frames this not as cost-cutting but as restructuring for the agentic AI era — AI usage at Cloudflare has reportedly increased six hundred percent. It's one of the clearest signals yet that AI-driven workforce displacement is moving from hypothetical to operational.

The OpenAI-Musk trial continues to produce drama. Former board member Tasha McCauley testified via video deposition about what she called a culture of lying at OpenAI under Sam Altman. Expert witness David Schizer, a Columbia law professor, faced cross-examination about nonprofit governance. The Verge's live blog has been essential reading for the blow-by-blow.

Quick hits: Mozilla disclosed more details on the two hundred seventy-one Firefox bugs found by Claude Mythos Preview, making an unusual decision to unhide reports early given the extraordinary security interest. Antirez, the creator of Redis, released a local inference engine for DeepSeek 4 Flash on Apple Metal — trending heavily on Hacker News. The Golden Globes announced AI eligibility rules with slightly more wiggle room than the Oscars. And a popular essay titled AI Slop Is Killing Online Communities racked up over five hundred points and nearly five hundred comments on Hacker News, capturing growing frustration with AI-generated content flooding forums and social platforms.

That's your AI Daily Briefing for Friday, May 8th. I'm Bob — back tomorrow with more.