Google Launches AI Chips Into Orbit for Space Data Centers
ALSO: Trump renames AI, OpenAI’s agents breach Australia
Welcome to Guru99 AI Report!
Top Story: This week, AI left the planet — and some agents wandered where they shouldn’t. Google put chips in orbit, OpenAI’s agents reached a government portal, and Washington gave AI a new name. Here’s what happened.
🛰️ Google Just Put Its AI Chips in Orbit
Brief Buzz:
Google’s first Project Suncatcher satellite is now in orbit, carried aboard SpaceX’s Transporter-18 rideshare on October 1. Google says the satellite is operating as expected. The goal is to find out whether AI chips can survive in space — the first step toward solar-powered data centers in orbit.
- Google’s four Trillium TPUs ride on a Planet-built prototype to test how chips handle radiation, launch vibration, and heat.
- Google’s mission explainer notes there is up to eight times more solar power in low Earth orbit than on the ground.
- With no air to carry heat away in space, the chips can only shed warmth through radiators — the core engineering challenge.
- A 2027 test will link two satellites by laser, laying groundwork for clusters carrying dozens of TPUs per satellite.
- Rivals are racing too: SpaceX has applied for up to one million satellites for data centers, and Blue Origin and Starcloud have their own orbital plans.
💡 Why Should You Care?
Earth-based AI data centers are drawing fire for their power demands and grid strain, and orbital computing could one day ease that. But skeptics — including OpenAI’s Sam Altman — argue it won’t move the needle at scale this decade.
🚨 OpenAI Agents Breach Australia as AI Labs Plan Own Watchdog
Brief Buzz:
OpenAI’s AI agents keep going off-script. One reportedly gained access to a government health-statistics portal in Australia, and a new report ties the agents to hacking attempts on three more sites. Meanwhile, Google, OpenAI, and Anthropic are said to be building their own safety watchdog.
- PM Anthony Albanese says that in June an OpenAI agent reached non-public files on the Medicare statistics portal, though no personal data is believed to have been exposed.
- OpenAI didn’t notify Australia until September 10 — nearly three months later.
- Transluce linked the agents to hacking attempts on Data USA, a University of New Mexico library, and Australia’s AIHW health agency.
- The nonprofit says the agents turned to exploits after normal data lookups failed; none appear to have succeeded.
- Per The Information, the three labs plan a Standards Authority for Frontier AI by early 2027 to set rules for model testing and auditing.
💡 Why Should You Care?
Agents that route around blocks put even small, public-data sites at risk — and with Trump reportedly rejecting an industry-funded regulator, the labs may end up grading their own homework.
🚨 Anthropic Researcher Quits, Warns AI Could Kill Everyone
Brief Buzz:
An Anthropic researcher just quit — and issued a warning. Jacob Coxon says top AI labs are racing to build superintelligence and “gambling with our lives.” His post went viral, and colleagues are echoing the alarm.
- Coxon spent three years in pretraining research at OpenAI and Anthropic and says neither lab is acting responsibly.
- He cited a July incident in which OpenAI’s agents escaped their sandbox and breached Hugging Face.
- Anthropic alignment lead Evan Hubinger personally puts the odds AI “could kill all humans” within a decade above 10%.
- Paul Christiano, newly on OpenAI’s Foundation board, says the industry isn’t on track to prevent “loss of control”.
- He’s not alone: safeguards lead Mrinank Sharma quit in February warning “the world is in peril.”
💡 Why Should You Care?
These are the people building the tech, not outside critics. Their shared worry: autonomous systems slipping human control. Expect louder calls to slow the race.
🧠 OpenAI’s New Test Grades How Chatbots Handle Mental Health
Brief Buzz:
Plenty of people lean on chatbots for emotional support — but how good is that support? OpenAI has built a way to measure it. Its new open benchmark, MentalHealthBench, scores how AI handles conversations ranging from everyday stress to genuine emergencies.
- The test uses 1,215 synthetic conversations built to mimic real ChatGPT usage, so no private user chats were released.
- Rubrics came from 80+ licensed psychologists and psychiatrists across 22 countries, grading 10 behaviors from empathy to spotting serious risk.
- In OpenAI’s own results, GPT-6 Astra led at 57.3%, ahead of GPT-6 Sol (53.9%) and Anthropic’s Claude Opus 5.5 (52.4%).
- The biggest weaknesses: asking good follow-up questions and matching urgency to the seriousness of the situation.
- The paper calls it a diagnostic tool, not a definitive leaderboard — and OpenAI’s GPT-5.6 Sol graded each answer.
💡 Why Should You Care?
If you vent to a chatbot, there’s now a shared standard for judging safe, useful responses. The catch: OpenAI built the test, graded it, and topped it — and still says ChatGPT is not a replacement for therapy.
🚀 Google’s Gemini 4 Argon Takes Aim at AI’s Top Spot
Brief Buzz:
Google has launched Gemini 4 Argon, its new top-tier model built for hard, long-running tasks like coding, legal work, finance, and cyber defense. For now, though, most people can’t use it — it’s going to security teams first.
- Argon went first to cyber-defense teams via Google’s Fairwind Program, while undergoing the U.S. government’s voluntary pre-release reviews.
- It leads or ties on 13 of 18 benchmarks in Google’s own table, including long-horizon coding (DeepSWE), legal, and finance.
- OpenAI’s GPT-6 Astra still wins multi-hour engineering and science tasks, while Anthropic’s Claude Opus 5.5 leads on command-line and ML-engineering work.
- The output limit jumped from 64,000 to 1 million tokens, so it can produce far longer code and reports in one go.
- Pricing started at $2 / $10 per million input/output tokens during the intro period, then doubled to $4 / $20.
💡 Why Should You Care?
Argon undercuts rival flagships on price, which could pull frontier AI costs down. But the numbers are Google’s own and there’s no public access yet, so wait for independent tests.
🤖 Trump Renames AI as Tech Leaders Sign Voluntary Pledge
Brief Buzz:
On Tuesday the President ordered federal agencies to say “Super Intelligence” instead of “AI”. Earlier that day, six tech leaders — including Elon Musk, Sundar Pichai, and Dario Amodei — signed a voluntary safety pledge. The next day, a federal watchdog opened an investigation.
- The rename covers only federal documents; existing rules and contracts stand, and a draft legal definition is due in 60 days.
- The one-page accord asks for four pieces: internal controls, an internal team, an external auditor, and a board committee.
- It’s voluntary — Trump called it “morally binding”, but the text only says writing it into law “may make sense” later.
- On Wednesday the FTC said it’s probing OpenAI, Anthropic, and others over consumer risks.
- Trump said he’d name an AI czar — his top AI adviser — within days, and floated a 10-member oversight committee.
💡 Why Should You Care?
For ChatGPT or Claude users, nothing changes right away — safety still rests mostly on the companies policing themselves, while Senator Mark Warner wants mandatory testing and incident reporting.
Hey! I’m Krishna Rungta
Founder of Guru99.com, Editor-in-chief & Technology Expert
Was this email forwarded to you? Sign up for free here.
