Every source. One signal. The day in artificial intelligence, distilled into plain English.
Transmission 008Thursday, 18 June 2026
Thursday 18 June 2026 brings a dense mix of geopolitical friction and technical momentum. The Anthropic-White House standoff over export controls and jailbreak demands dominated coverage, while China's Zhipu AI released GLM-5.2 to claim the top spot among open-weights models. Elsewhere, funding rounds, infrastructure bets, and a high-profile talent move from Google to OpenAI rounded out a busy day.
Audio edition
Audio edition is being prepared — check back shortly.
Anthropic vs Washington
Signal 9/10
White House demands impossible jailbreak-proof AI as Anthropic faces export-control crackdown
The Trump administration ordered Anthropic to revoke SK Telecom's access to Claude Mythos over alleged ties to China, days before the model was taken offline, according to Wired. The White House is now reportedly demanding that Anthropic guarantee no jailbreaks before rereleasing Fable 5 — a condition security experts say cannot be met. Anthropic had previously sought regulation, but analysts say Washington has gone considerably further than the company anticipated. The episode prompted Mistral chief executive Arthur Mensch to pitch open-source AI as operating 'outside of state control', and European observers suggested the ban could ultimately benefit European AI developers. AI luminaries at the G7 summit, including the chief executives of Anthropic and Google DeepMind, called for a US-led AI coalition, underscoring how intertwined geopolitics and AI governance have become.
Zhipu AI's GLM-5.2 takes top open-weights ranking as Chinese models close gap with closed-source leaders
Zhipu AI released GLM-5.2 under an MIT licence, and Artificial Analysis placed it first on its open-weights intelligence index. The Decoder reports that on FrontierSWE — a benchmark for hours-long coding tasks — GLM-5.2 trails Anthropic's Claude Opus 4.8 by just one percentage point, a striking result for an openly licensed model. The model supports a stable one-million-token context window. Simon Willison described it as 'probably the most powerful text-only open weights large language model' currently available. The release arrives as Rest of World reports that American developers are already choosing Chinese AI tools such as DeepSeek on cost grounds, suggesting the competitive pressure from Chinese open-weights labs is not merely technical.
Funding floods into world models, formal verification and Indian data centres as hyperscaler cash-flow warnings emerge
Amazon, Nvidia, and AMD are reported to have invested $310 million in world-model startup Odyssey ML, which The Decoder says is now valued at $1.45 billion, with CIA-linked fund IQT and Google chief scientist Jeff Dean also backing the round. Pramaana Labs raised a reported $27 million seed round from Khosla Ventures to apply formal verification to high-stakes verticals including law and drug discovery. In the United Kingdom, CuspAI is said to be approaching a $2.6 billion valuation with backing from Jeff Bezos, according to Sifted. Three former Palantir engineers raised a reported $60 million from Index, Iconiq, and SAP for an agentic operating system. A Canadian pension fund is acquiring an 8.2 per cent stake in Indian data-centre operator CtrlS. Shanghai's stock exchange clarified listing rules to allow unprofitable large language model developers to go public, and Chinese chip-equipment maker CFMEE is targeting a reported $410 million Hong Kong initial public offering. Epoch AI analysis cited by The Decoder warns that hyperscalers' AI infrastructure spending is growing at roughly 70 per cent per year while operating cash flow rises at only 23 per cent, raising questions about long-term self-funding capacity.
Alibaba expands into France, Kingboard raises $1.5 billion for circuit-board capacity and China tests chips for model training
Alibaba Cloud announced the launch of its first data centres in France, citing European demand for data sovereignty. Kingboard Holdings plans to raise a reported HK$11.77 billion (US$1.5 billion) by selling a stake in Kingboard Laminates Holdings, one of the world's largest makers of printed circuit boards, to expand capacity driven by AI hardware demand. South China Morning Post reports that while Chinese domestic chips are now widely used for model inference, none of China's leading models are known to have been trained entirely on local silicon — a gap that constrains China's AI infrastructure independence. The x86 ecosystem published an AI Compute Extensions (ACE) specification aimed at standardising AI acceleration across commodity processors, which could influence future data-centre chip procurement.
Nvidia's self-improving robots, humanoid training in Shenzhen and a new open-source agent framework signal physical AI's coming-of-age
Nvidia, Carnegie Mellon University, and UC Berkeley published research showing that teams of AI coding agents can autonomously direct robots to learn dexterous grasping, achieving up to 99 per cent success on difficult tasks with a fleet of eight robots. In Shenzhen, Wired reports that workers at IO-AI Tech are using virtual-reality rigs to teleoperate humanoid robots, generating training data at industrial scale — a process TechCrunch separately described as 'dirty, unglamorous work' that is essential if physical AI is to match large language model progress. OpenRouter published results from an agent 'last-agent-standing' competition comparing different language model backends in robotic control tasks. Vercel open-sourced Eve, an Apache-2.0 agent framework in which each agent is a directory of files, and Amazon published a guide to deploying robots from the Hugging Face Hub using Strands Agents and LeRobot.
AI code reviewers, open-source computer-aided design and a call for more rigorous engineering push back against vibe-coding culture
Greptile released TREX, an AI code reviewer that actually executes the code it analyses rather than relying on static inspection alone — a meaningful step beyond syntax-level feedback. Y Combinator Winter 2025 company Adam released open-source AI-assisted computer-aided design software on GitHub. Charity Majors published a widely shared essay arguing that AI demands more engineering discipline, not less, pushing back against a culture of prompt-and-ship. Nvidia released SkillSpector, a tool for scanning AI 'skills' — modular agent capabilities — for security vulnerabilities using static analysis and SARIF (Static Analysis Results Interchange Format) reports before deployment. A separate open-source project, pii-gui, offers local personal data redaction to strip personally identifiable information before sending text to any AI service.
OpenAI's chemist tool, Midjourney's medical imaging pivot and new benchmarks signal accelerating AI role in life sciences
OpenAI published a case study showing an AI system improving a challenging medicinal chemistry reaction, pointing to practical gains in drug-discovery pipelines. Midjourney Medical announced a move from generating images into producing full-body ultrasound scans, a striking pivot from creative tools towards clinical diagnostics. OpenAI also released LifeSciBench, a 750-task benchmark written by 173 PhD scientists across seven biological domains, designed to test whether frontier AI can handle genuine life-science research workflows. Pramaana Labs' $27 million raise (covered in the markets cluster) specifically targets formal verification in drug discovery and law. A personal account in The Guardian described how an AI tool flagged a possible deep-vein thrombosis that a physician had initially missed, illustrating both the promise and the limits of AI-assisted diagnosis.
Noam Shazeer leaves Google for OpenAI as two-thirds of Americans say AI is advancing too fast
CNBC reports that Noam Shazeer, Google's vice-president of engineering and a co-lead of its Gemini models, has left to join OpenAI — a significant talent move at a moment when the two companies are closely matched on several benchmarks. Separately, a Pew Research survey reported by The Verge found that two-thirds of Americans believe AI is advancing too quickly, a sentiment that sits awkwardly alongside the relentless release cadence documented elsewhere in today's edition. Meta's head of product for its 'AI for work' transformation is also leaving the company, according to Reuters. The Guardian's Arwa Mahdawi flagged the emergence of a fictional character named 'Elias Thorne' across many chatbot-generated stories as a concrete symptom of the model-collapse risk posed by AI-generated content polluting training data.
Redact personally identifiable information locally before using any AI tool
Before pasting documents, emails, or transcripts into a cloud-based large language model, run them through a local redaction tool such as pii-gui to strip names, addresses, phone numbers and other personally identifiable information. The tool runs entirely on your own machine, so no data leaves your environment during the scrubbing step — only the anonymised text is sent to the AI service. This is practical for anyone handling client correspondence, medical notes, legal drafts, or HR records.
Clone or download pii-gui from GitHub (https://github.com/sophia486/pii-gui) and follow the README to install dependencies locally.
Open the application and paste or load the document you want to process.
Review the highlighted personally identifiable information detections and confirm or adjust any edge cases the model flags.
Copy the redacted output and paste it into your chosen AI tool (such as ChatGPT, Claude, or a local model) for summarisation, drafting, or analysis.
If you need to act on the AI's response in a context that requires the original names or data, re-insert them manually at the end — keeping the sensitive detail off third-party servers throughout.
Legal, medical, HR and finance professionals who need to use AI tools without exposing client or patient data to external servers.GitHub — pii-gui local redaction tool ↗
Get the daily transmission
One email each weekday morning: the day in AI, distilled into five minutes of plain English. Free, no spam, unsubscribe with one click.
Double opt-in: we only add you after you confirm by email. We store your address for sending this newsletter and nothing else. Unsubscribe any time.