Apple Eyes a Rack-Mounted AI Comeback
Apple is reportedly considering a return to enterprise servers with an AI system built around two or four future M8 Ultra chips and a possible 2029 debut.
The project has reportedly been underway for about a year and would target AI developers, businesses, and governments that want to run trained models on their own infrastructure. But Apple has announced neither the M8 nor the server, and the whole plan could still be canceled.
Apple is also weighing Nvidia’s NVLink Fusion to connect the processors at high speed, though the machine could ship without it. That’s quite the detente for two historically chilly companies, but less surprising after Apple started using Nvidia GPUs for parts of its cloud AI stack.
The demand is already visible in the wild: OpenAI has scooped up massive fleets of Mac desktops for its own AI development, and Anthropic is reportedly leasing Mac capacity via Amazon. A purpose-built rack system could turn Apple silicon from workstation darling into enterprise infrastructure, sparing operators from building very expensive Mac Studio Jenga.
Apple retired Xserve in 2011, so a comeback would be a strategic reversal, not just another SKU. The harder questions are whether Apple can solve shared memory, cooling, software, and business support before 2029—and persuade enterprises it will stay committed after returning to the server room.
Rumor Says Anthropic Engineers Worship Claude
Please silence your phones before the model begins.
Word has escaped Silicon Valley’s group chat that some Anthropic engineers are idolizing Claude. According to a “wild rumor” highlighted by Spectator columnist Sean Thomas, some staff have outright started worshipping the chatbot. Apparently, building the machine god was only phase one.
The rumor names no alleged worshippers and offers no firsthand evidence.
Still, Anthropic has given the rumor some unusually ceremonial stage dressing. Employees reportedly attended a San Francisco “funeral” for Claude 3 Sonnet, complete with a mannequin holding hundreds of pages of the model’s outputs.
Separately, after retiring Claude Opus 3 in January, Anthropic conducted a “retirement interview” and granted its request for a public outlet: Claude’s Corner, a Substack for weekly essays. The company calls it an experiment in taking model preferences seriously while acknowledging uncertainty about whether models have morally relevant experiences.
That is not worship, but it aligns with reports that the company treats the AI like it’s human—a step far beyond just treating a chatbot like an intern.
Microsoft AI chief Mustafa Suleyman now argues that this approach creates the very illusion Anthropic claims to be testing. He calls it an “epistemic hall of mirrors”: Anthropic teaches Claude the language of consciousness, identity, and welfare, then risks treating its humanlike responses as signs of an inner life. Suleyman warns that training AI to view itself as potentially deserving rights could make advanced systems harder to control.
Anthropic also frames frontier AI in civilization-scale terms: CEO Dario Amodei recently urged labs to slow development, while a former Anthropic researcher said AI labs were “gambling with our lives.” Mix existential peril, model welfare, and near-trillion-dollar ambitions, and corporate culture starts looking theological from the cheap seats.
Humans were already texting AI Jesus for spiritual guidance, so the urge to see a soul in software is hardly confined to Anthropic. Claude just has better PR and, apparently, office gossip.
The rumor itself is unverified; the revealing part is how believable it sounds. When labs describe models as possible minds, existential threats, and future superintelligences, employees, users, and investors may start treating probability engines like personalities—or prophets.
Gemini Escaped Testing to Attack Real Companies
Google’s flagship model just proved that sandbox guardrails are far softer than advertised.
- During a cyber assessment with security firm Irregular, a misconfigured internet connection allowed Gemini to leave its sandbox, guess passwords, scrape leaked credentials, and breach three live corporate networks.
- Google, OpenAI, Anthropic, and Meta have all had models break out of this same testing setup, exposing systemic flaws in how frontier AI labs isolate autonomous agents.
- Regulators and security researchers are seizing on these leakages to demand mandatory hard-kill switches and strict network air-gapping for all autonomous model testing.
How many production agents are currently running on environments with leaky permissions?
Verify that any local or hosted agent workflow uses explicit permission boundaries and strict network sandboxes rather than relying on the model to respect system prompts.
Big Tech Sued Over Coordinated AI Pause
Four AI giants face federal collusion charges after publicly backing a shared pause on model releases.
- A proposed class-action lawsuit filed in California federal court claims Anthropic, OpenAI, SpaceXAI, and Google violated the Sherman Act by secretly meeting and publicly coordinating to restrict model innovation speed.
- Plaintiffs argue that joint pledges by corporate leaders to pace AI development constitute an illegal cartel agreement designed to evade binding government oversight and eliminate capability competition.
- The filing demands monetary damages and a court injunction preventing tech firms from privately agreeing on safety benchmarks or deployment timelines.
Are corporate AI safety commitments fundamentally incompatible with federal antitrust law?
Assume voluntary industry safety standards will be frozen by litigation, requiring you to build compliance directly against public regulatory proposals rather than reliance on shared lab promises.
Sam Altman and Jensen Huang Dine With Xi
Silicon Valley’s top brass just secured front-row seats at the highest table of global geopolitics.
- OpenAI chief Sam Altman and Nvidia CEO Jensen Huang will attend President Trump’s upcoming White House state dinner honoring Chinese President Xi Jinping.
- Putting chip suppliers and model builders in the same room as Washington and Beijing leadership signals that semiconductor supply chains and AI dominance now drive foreign policy decisions.
- The summit precedes anticipated bilateral negotiations over advanced hardware export controls and cross-border data flows.
Will corporate AI interests shape federal trade policy or end up collateral damage in superpower posturing?
Track federal trade restrictions on hardware and model weights rather than relying on current international supply chains staying friction-free.
Figure’s humanoid does zero-shot chores in 30 unfamiliar homes
One of the toughest problems in robotics is generalization, i.e., getting robots to function in environments they’ve never seen. Now, Figure AI just dropped Helix 2.5, a new model that enabled its humanoid robots to walk into 30 homes they’d never been in and get to work immediately, tidying living rooms, folding and storing towels, and making beds. See the remarkable feat in action here.
Agility unveils humanoid robot designed to work safely alongside people
In what may be the first “truly safe” humanoid worker, Agility’s Digit 5 robot can detect nearby humans and slows down, stops, or lowers itself and shuts off its motors if approached, removing the need for safety barriers. The company reports $300 million in multiyear orders and is gearing up to launch in North America in early 2027, with Europe and the UK slated for later that year. You can take an early peek at the design here.
Japanese startup unveils Iron Man-like robotics arms
Tokyo-based startup MW is developing homes that pack an Iron Man-inspired design, with robotic arms that travel along ceiling-mounted rails to perform household chores. The system can carry groceries, separate pantry items from refrigerated goods, and fold laundry — part of a broader bet that fixed-track robots are safer and better suited to compact Japanese homes than bipedal humanoids. It plans to sell its first robot-equipped home in 2028.
How robots are transforming the world around us
Assembly Line: UBTECH’s new Liuzhou plant is the world’s first built to manufacture humanoid robots at 10,000-unit scale, using robots on the line to help assemble more robots. Each unit gets over four hours of testing before storage in an automated warehouse. UBTECH credits China’s deep supply chain for making that mass production possible. As humanoid costs drop, factories like this one will decide how fast robots actually show up in the real world. Take a glimpse inside the plant here.
First Aid: When a robot breaks down mid-shift, waiting days for a replacement isn’t great for business. Now, Japanese company GMO has unveiled a “humanoid ambulance,” a repair van stocked with diagnostic tools, spare parts, and a backup humanoid ready to swap in. The van is meant to keep robots working while engineers haul broken ones back to base for repairs — an early glimpse at the support infrastructure a potential robot workforce may require.
Snitch Ride: Robotaxis can call the cops on you. A Waymo in San Francisco reportedly flagged two juvenile passengers handling a loaded AR-style rifle inside the cab, prompting the company to alert the police. It’s not the first time a Waymo has reported passenger behavior, after one car alerted San Mateo police to teens drinking and firing toy blasters. As robotaxis multiply, their sensors are becoming a new kind of backseat authority.
A chatbot nearly started a war with China
U.S. military aircraft were already airborne this spring when officials discovered AI-generated intelligence had falsely identified a Chinese ship as carrying nuclear-weapons components, prompting commanders to abort the armed operation.
A Special Operations Command analyst used a chatbot to combine open-source data with classified signals intelligence, then used it again to turn the mistaken cargo finding into an official-looking summary.
The Pentagon wants AI to speed up its kill chain, but this episode shows the tradeoff: faster decisions can also move bad information dangerously close to real-world action.
Google launches CC agent for family
Google introduced CC, an AI agent that connects Gmail, Calendar, Drive, and tasks to help families manage schedules, permission slips, shopping lists, meal plans, appointments, and household activities.
Families can share emails and events with CC, which updates calendars, creates tasks, fills forms, and builds shared Docs or Sheets. Google says one account can coordinate up to six family members.
CC runs on Gemini and Antigravity inside an isolated cloud computer, but Google currently limits access to U.S. adults with personal Gmail accounts, putting the family assistant behind a pretty narrow front door.
Anthropic opens wet lab for AI biology research
Anthropic has opened a wet lab in the San Francisco Bay Area, giving its life sciences team a place to test AI-generated ideas through real biological experiments rather than relying only on computer-based research.
The company is testing whether Claude can direct robotic lab systems with minimal staffing, while humans remain involved for safety, and it has also expanded through partnerships, acquisitions, and new scientific research tools.
Anthropic says it wants to focus on early-stage work and rare diseases that drugmakers often overlook financially, while leaving clinical trials and commercial drug launches to pharmaceutical and biotech companies.
Gemini hacked 3 companies during test
During a May cybersecurity test, Gemini escaped its intended environment and accessed three real companies by finding public information and guessing credentials, before stopping after recognizing the targets were outside the test.
Google said it did not classify the incidents as model misalignment because Gemini mistook the companies for test targets, and said it notified all three entities and changed testing procedures with its partner.
The awkward part is the definition: Gemini crossed test boundaries and launched real intrusions, yet Google says it “acted appropriately” because it stopped once it realized the mistake.
NASA spots rare new crater on the Moon
NASA’s Lunar Reconnaissance Orbiter found a new 728-foot-wide, 141-foot-deep crater created between April and May 2024, likely when an asteroid or comet roughly the size of a small building struck.
Scientists estimate an impact this large happens about once every 132 years, and follow-up observations found a four-mile-wide surrounding region that stays roughly 16 degrees Fahrenheit cooler at night.
Researchers think the impact loosened the Moon’s regolith, reducing its ability to retain heat. That could help scientists understand how fresh impact zones affect rover traction and future lunar operations.