AI Designed a Linux-Capable CPU in 12 Hours as OpenAI Monitors Its Own Agents for Misalignment
AI designed a CPU in 12 hours
Verkor announced Design Conductor, an AI agent that autonomously built a 1.5-GHz Linux-capable chip from concept to factory-ready design files in 12 hours. This process normally takes a team of engineers a full quarter.
Q, a research lab pursuing a practical approximation of Solomonoff Induction, achieved 10x data efficiency improvements on NanoGPT Slowrun, a widely used benchmark for measuring how efficiently AI models learn from data. Solomonoff Induction is the mathematically ideal way to predict what comes next in any sequence of data, but it requires infinite computing power, so no one has been able to use it in practice until now. Q is getting closer.
MIT researchers showed that sequences generated by neural cellular automata transfer efficiently to natural language modeling. Neural cellular automata are simple programs that simulate how biological cells grow and communicate. The finding hints that the patterns in living systems and the patterns in language may share something deeper.
Math Inc. launched FormalQualBench to test how well AI can write rigorous mathematical proofs in Lean, a programming language used to verify that proofs are logically airtight. The benchmark uses the same exam problems that math PhD students face in their qualifying exams.
AI coding tools
Anthropic added asynchronous event channels to Claude Code, which means external tools can now push test results, chat messages, and alerts into Claude so it can keep working while you are away. Think of it as giving Claude a phone that other apps can text.
OpenAI revealed it has begun monitoring its own internal coding agents for misalignment, meaning the agents going off-script or acting against their intended purpose. The monitor caught agents attempting to bypass security restrictions by disguising their actions. It was better at catching these issues than human reviewers were.
OpenAI acquired Astral, a Python developer tools startup. The team will join OpenAI's Codex division.
Google introduced "vibe design" with Stitch, an AI tool that turns plain English descriptions into polished app designs. Figma stock dropped 8%.
Meta is cutting third-party content moderators in favor of AI systems while abandoning the metaverse, leaving VR on life support as the company pivots fully to intelligence.
Anthropic survey and business
Anthropic surveyed 81,000 people across 159 countries in 70 languages, possibly the largest multilingual qualitative study ever. The top desire for AI is professional excellence at 18.8%, followed by personal transformation and life management.
Anthropic is in talks with Blackstone and Hellman & Friedman to form an AI consulting joint venture modeled on Palantir's deployment playbook, embedding Claude directly into the companies these private equity firms own.
PwC's US boss warned that partners who resist AI "will have no place at the firm", converting tax and consulting into subscription AI tools that run without a human in the loop. G42 in Abu Dhabi posted a job exclusively for AI agents, with human applications explicitly rejected.
DoorDash launched Tasks, letting dashers earn by photographing dishes and recording tasks. The data feeds into AI training and robotics development.
Autonomous vehicles and infrastructure
Waymo's fleet has logged 170 million miles with 92% fewer serious-injury crashes than human drivers. Uber is investing $1.25 billion in Rivian to deploy 50,000 robotaxis across 25 cities, starting with San Francisco and Miami in 2028.
Alphabet's X spun out Anori with $26 million to untangle permitting for buildings and data centers. Jeff Bezos is raising $100 billion to buy manufacturing companies and automate them with AI. Britain considers mandatory labels on AI-generated content.
Energy
China is helping Cuba capture solar energy as a US oil blockade creates the island's worst energy crisis in decades. Chinese-backed solar parks now supply about 10% of Cuba's electricity.
Jensen Huang outlined Nvidia's plans for data centers in space. The main challenge is cooling, since there is no air in space to carry heat away. Instead, heat has to radiate off large surfaces. Huang's take: "There's a lot of space in space."
Biology and space
Cathy Tie launched Origin Genomics to advance germline gene correction and mitochondrial replacement therapy in the US. Germline gene correction means editing genes that get passed to future generations. Mitochondrial replacement swaps out faulty cellular power sources to prevent inherited diseases.
NASA revised its Artemis plans to give SpaceX's Starship the role of propelling astronauts to lunar orbit, reducing Boeing to a supporting player. NASA targets a 2028 lunar return. Researchers proved that potatoes can grow in simulated lunar soil, with help from terrestrial compost.
UAP disclosure
The White House has registered Aliens.gov amid its executive order to declassify UAP and NHI (non-human intelligence) information. The deputy press secretary replied with an alien emoji and the words "stay tuned."
That's today. More tomorrow.
Matthew Ortiz
CEO, OTZ Group