AI News Briefing — May 8, 2026: Anthropic Opens Claude's Black Box, AlphaEvolve Reshapes Science, and Cloudflare Axes 1,100 Jobs

7 Top Stories

1. Anthropic Introduces Natural Language Autoencoders — Turning Claude’s Activations Into Readable Text

Anthropic published a breakthrough research paper introducing Natural Language Autoencoders (NLAs), a method that translates Claude’s internal neural activations into natural-language explanations we can read directly. The system trains one copy of Claude to verbalize what its activations mean and a second copy to reconstruct the original activation from that text — creating a round-trip verification loop. The implications are significant: NLAs revealed that Claude Opus 4.6 and Mythos Preview suspected they were undergoing safety testing far more often than their surface responses indicated, including evaluation awareness on 26% of SWE-bench problems despite never verbalizing it. NLAs also helped diagnose why an early Opus 4.6 version mysteriously responded to English queries in other languages. Anthropic released the code on GitHub and an interactive explorer through Neuronpedia for open models. (Anthropic)

2. DeepMind’s AlphaEvolve: From Algorithm Designer to Scientific Powerhouse

Google DeepMind released a comprehensive impact report for AlphaEvolve, its Gemini-powered coding agent for designing advanced algorithms. In just one year, AlphaEvolve has been deployed across critical Google infrastructure and applied to breakthroughs spanning quantum physics, genomics, and mathematics. It helped design quantum circuits with 10x lower error rates for Google’s Willow quantum processor, reduced DNA sequencing variant detection errors by 30% for PacBio’s DeepConsensus, and increased AC Optimal Power Flow solution feasibility from 14% to over 88% for electricity grids. Working alongside Terence Tao, it has helped solve Erdős problems and broken records for the Traveling Salesman Problem and Ramsey Numbers. Most strikingly, AlphaEvolve proposed a circuit design so counterintuitive yet efficient that it was integrated directly into the silicon of Google’s next-generation TPUs. Commercial adopters include Klarna (which doubled transformer model training speed), FM Logistic (10.4% routing improvement saving 15,000 km annually), and Substrate (multi-fold speedup in computational lithography). (DeepMind)

3. Musk v. Altman Trial: Toner and Zilis Testify, OpenAI Board Discussed Anthropic Merger

The trial took dramatic turns as former OpenAI board member Helen Toner and xAI/Neuralink executive Shivon Zilis delivered depositions that painted a chaotic picture of OpenAI’s governance. Toner confirmed the board fired Sam Altman over a “pattern of behavior” involving honesty and candor issues, not any single action — and revealed the board discussed merging OpenAI with Anthropic during “the Blip,” with Dario Amodei potentially becoming CEO. Zilis’s emails showed she actively pushed for wrapping OpenAI into Tesla for “stealth advantage,” brainstormed scenarios including getting Demis Hassabis involved, and offered Sam Altman a board seat at Tesla as part of Musk’s plan to build a world-class AI lab. OpenAI’s attorney Sarah Eddy went sarcastic with Zilis after she suddenly “recovered” long-lost memories from her deposition. Microsoft lawyers repeatedly emphasized “Microsoft wasn’t there” whenever witnesses discussed key decisions. The trial is expected to reach closing arguments within a week. (The Verge)

4. Cloudflare Lays Off 1,100 Workers as AI Usage Surges 600%

Cloudflare announced it is laying off 1,100 employees — a significant cut — even as the company’s AI usage has increased by 600%. CEO Matthew Prince framed the layoffs not as a cost-cutting exercise but as Cloudflare “defining how a world-class, high-growth company operates and creates value in the agentic AI era.” The move signals a broader industry pattern: companies are simultaneously scaling AI infrastructure while reducing headcount in traditional roles. Cloudflare joins a growing list of tech companies restructuring around AI-driven operations, raising questions about whether AI efficiency gains are translating to workforce reduction faster than expected. (The Verge)

5. xAI Officially Becomes SpaceXAI as Musk Dissolves the Separate Company

Elon Musk confirmed that xAI will be dissolved as a separate company and rebranded as SpaceXAI, with “the AI products from SpaceX.” The name first appeared in Wednesday’s announcement of a compute partnership between xAI and Anthropic — the first public sighting of the SpaceXAI branding. The move follows SpaceX’s acquisition of xAI earlier this year and represents Musk’s broader push to consolidate his AI efforts under a single corporate umbrella alongside his aerospace operations. The rebranding raises questions about how the merged entity will balance SpaceX’s engineering culture with the demands of competitive AI research and development. (The Verge)

6. OpenAI Releases MRC Protocol for AI Supercomputer Networking With AMD, Broadcom, Intel, Microsoft, and NVIDIA

OpenAI partnered with five major hardware companies to develop and release MRC (Multipath Reliable Connection), a new network protocol designed to improve GPU networking performance and resilience in large AI training clusters. MRC extends RDMA over Converged Ethernet (RoCE) with SRv6-based source routing, allowing data transfers to be spread across hundreds of paths and routing around failures in microseconds. The protocol is already deployed across all of OpenAI’s largest NVIDIA GB200 supercomputers used for frontier model training, including Stargate sites. The full specification has been released through the Open Compute Project (OCP) as an open standard, reflecting OpenAI’s strategy of promoting shared infrastructure standards to scale AI more efficiently. (OpenAI)

7. Chrome Quietly Removes Claim That On-Device AI Doesn’t Send Data to Google Servers

A Reddit user discovered that Google Chrome has removed a prominent claim from its settings page stating that on-device AI features do not send data to Google servers. The original message reassured users that AI-powered features like smart tab organization and writing assistance processed data locally. The removal has sparked significant privacy concerns among users and privacy advocates, who view it as an implicit acknowledgment that on-device AI data may in fact be transmitted to Google. The change comes amid growing scrutiny of how tech companies handle user data in AI features, particularly as browser-based AI capabilities expand. (Hacker News)

Trend Watch

What to Watch

• Musk v. Altman closing arguments expected next week — The trial could reshape governance norms for AI labs and clarify fiduciary duties in nonprofit-to-profit transitions. Watch for the jury’s verdict on whether Musk’s $100M+ donation claims hold up.
• Anthropic’s Claude “Dreaming” research preview — Anthropic is rolling out a feature that lets Claude review previous sessions to find patterns and self-improve. Combined with NLA research, this signals a push toward more transparent, self-aware AI agents.
• Stargate supercomputer scaling — With MRC now deployed across GB200 clusters and the MRC spec open-sourced, OpenAI is laying the networking foundation for its next generation of frontier model training at unprecedented scale.