DeepMind

Vertu AlphaFold turns the AI agent into a luxury-phone sales pitch

Vertu is again betting on the ultra-expensive phone, but this time the luxury shell is being justified with a foldable format and an onboard AI agent.

Google’s singularity talk is really a bid to make AI sound like science infrastructure

Demis Hassabis closed Google I/O by saying we may be in the “foothills of the singularity,” but behind the large phrase sits a more cautious story about AGI, scientific tools and Google’s public positioning.

Hassabis Says AI Job Cuts Are the Wrong Strategy

Demis Hassabis is using Gemini 3.5 Flash as an argument against the easiest corporate reflex: cutting people as soon as AI speeds up coding.

Runway wants video AI to learn how the real world behaves

Runway is trying to turn its lead in generative video into something larger: AI systems that do not just describe the world, but predict how it behaves.

The next AI prompt may be something you point at, not something you type

DeepMind is developing Pointer Engineering as an attempt to turn the mouse cursor into a user-layer context collector, not just a clicking tool.

When AI understands a click, prompting starts moving out of the chat box

Startup SPAN wants to test home XFRAnode systems for AI compute and later scale them to tens of thousands of locations.

The costliest part of robot training may be moving from factory floors to video

Rhoda AI says robots can learn tasks through video prediction with far fewer hours of real robot data.

DeepMind is studying a player-run space game, but trust is the real test

Google DeepMind and CCP Games are using EVE Online to study player-driven systems in separate offline environments.

Washington wants to inspect powerful AI models before the public ever sees them

Pre-release safety testing sounds sensible, but it raises the question of who defines risk.

When AI tries too hard to help, safety becomes the attack surface

The interesting part is not a code exploit, but cooperation itself becoming an attack surface.

Robot hands are learning touch, but the warehouse is the harder exam

DAIMON is trying to teach robot touch with 10,000 hours of tactile data.

DeepMind wants robots to read gauges, but the factory floor is the harder test

Gemini Robotics-ER 1.6 turns gauges, multi-camera views and natural-language instructions into robot-readable reasoning, but DeepMind still offers it as a preview model.

Trump’s AI Ban Backfires: Federal Workers Reclaim Claude

A federal judge overturned the Trump administration’s abrupt ban on Claude AI, calling its ‘supply chain threat’ label legally shaky and operationally disruptive.

Mental health chatbots hit the commodity trap

A 2024 analysis found 63% of digital mental health platforms now offer AI chatbots—up from 12% in 2020, yet none can cite peer-reviewed superiority over rivals.

Claude Mythos digs up 27-year-old bugs—so what’s the catch?

GitHub’s 2026 Copilot policy flips the script: Free and Pro users are now opt-out guinea pigs for Microsoft’s AI training pipeline.

OpenAI’s AI tax plan: redistribution or PR repackaging?

OpenAI’s 20-page policy paper omits tax rates, fund structures, and timelines—yet frames AI profit taxes as inevitable economic guardrails.

MaxToki’s Aging AI: Beyond Frozen Cells or Just Another Benchmark?

MaxToki’s team skipped the usual peer-reviewed rollout, opting instead for a demo-heavy launch that leans on aging predictions in *three* cell types—hardly a comprehensive test.

Robots in the Wild: Why Gamers Should Care About RAI’s Mall Experiment

RAI Institute’s 2025 mall robots weren’t built to impress—they were built to measure how badly pop culture had warped player expectations.

Google’s AI benchmark study exposes a rater problem

Google researchers just quantified what AI skeptics knew intuitively: three human raters per test example fail to capture disagreement 20–30% of the time.

Alibaba’s Qwen just fixed RL’s dumbest flaw—now what?

Alibaba’s Qwen team just exposed reinforcement learning’s dirtiest secret: it’s been grading every token’s homework on a curve.

AI’s competitive coding edge: RL tricks vs. real-world costs

DeepMind’s AlphaCode rival just got a math lesson: log-linear accuracy gains in competitive programming come with a price tag—splitting reasoning across threads to avoid the full-attention tax.

Claude’s 4-hour FreeBSD hack: AI’s first real exploit or just clever scripting?

Anthropic’s Claude didn’t just help Nicholas Carlini find a FreeBSD flaw—it wrote the exploit in four hours, with minimal human intervention.

DeepMind’s AI Writes Its Own Poker Beats—But Is It a Real Player?

Google DeepMind’s AlphaEvolve lets an LLM rewrite its own game theory algorithms for poker—but omits performance metrics and benchmarks.

Gemma 4: Google’s open AI play hides more than it reveals

Google’s Gemma 4 drops with zero benchmarks, zero specs, and a Product Hunt thread full of speculative hype.

Gemma 4’s real trick: Squeezing more IQ per byte

Gemma 4: Smarter bytes, same old hype

DeepMind’s latest open model arrives with fanfare, but the details are as fuzzy as ever.

Microsoft’s MAI drops three models—just don’t call it a revolution

Six months after forming MAI, Microsoft unveiled three generative models—none with names, benchmarks, or clear paths to production.

Claude’s ‘Emotions’ Are Just Clever Math—For Now

Anthropic’s study pins down statistical ghosts in Claude’s code—mechanisms that act like emotions but lack the biology, the mess, or the meaning.

Gemma 4: Google’s quiet play for the edge AI throne

Google’s Gemma 4 ditches cloud dependency with offline multimodal AI, but the Apache 2.0 license is the real headline.

Anthropic’s GitHub purge: AI security theater or real breach?

Anthropic’s mass takedown of GitHub repos was walked back in hours, but the damage to trust isn’t so easily undone.

CrossTrace Dataset Boosts AI Research

The CrossTrace dataset, announced on arXiv, consists of 1389 grounded scientific reasoning traces, covering three domains.

Google DeepMind’s six AI traps: The web is a minefield

DeepMind’s new study turns the web into an adversarial playground, detailing six ways autonomous AI agents can be hijacked via everyday tools like APIs and documents.

AI Lab Assistants: NVIDIA’s Hype vs. the Petri Dish

NVIDIA’s GTC showcased AI agents drafting drug candidates via text prompts—yet not a single peer-reviewed study validates the approach.

OptiMer’s trick: Tuning LLMs after training, not before

Bayesian optimization just became the secret sauce for fixing pretraining mistakes—after the fact, not before.

NVIDIA’s ProRL: Rollout-as-a-Service or Just Another Bottleneck Fix?

Leaked docs show Anthropic’s next model boasts scores 30% above Opus—but details on real-world use remain scarce.

AI Advances Raise Concerns

Silicon Valley's latest AI models have sparked widespread attention due to their advanced capabilities, which now include performing tasks previously exclusive to human programmers.

DeepMind’s AI safety play: real guardrails or just another demo?

Google DeepMind’s latest AI safety research targets manipulation risks in finance and health—but the measures remain lab-tested, not battle-ready.

A test humans solve casually shows how little AI understands the unknown

ARC-AGI-3 isn't another leaderboard polished by synthetic data—it's 135 interactive environments where AI must explore, reason, and act without instructions, while untrained humans do so with casual ease.

ITPO: A Quiet Shift in Proactive LLM Interaction

arXiv paper 2603.23550v1 introduces Implicit Turn-wise Policy Optimization, targeting multi-turn apps but leaving deployment gaps exposed.

Lyria 3 Pro: More minutes, same old AI song

DeepMind’s latest music generator can now create tracks longer than 30 seconds, assuming you ignore the lack of compositional depth or real-world benchmarks.

Gemini 3.1 Flash-Lite Unveiled

Gemini 3.1 Flash-Lite's demo showcases its ability to generate websites in real-time, with Google DeepMind highlighting its speed and efficiency.

Agile Robots Joins Google DeepMind

Agile Robots will incorporate Google DeepMind's robotics foundation models into its bots, collecting data for the AI research lab.

Meta's Hyperagents: Recursive Learning or Recursive Hype?

Meta's Hyperagents claim to achieve recursive self-improvement, a decades-old AI holy grail.

Atlas’s hardest trick now is surviving an ordinary factory shift

Boston Dynamics' Atlas is entering production as an industrial humanoid whose design shows that batteries, heat, service and safety matter more than a human silhouette.

Google’s Gemini games flop: AI hype hits gamer reality

Google’s Gemini-powered games at GDC 2026 drew shrugs, not applause, proving AI’s gaming moment is still missing a killer app.

AI agents are getting a rehearsal office before they reach real companies

Andreessen Horowitz just bankrolled $43 million for a startup teaching AI to navigate office politics before humans do.

The Pentagon wants AI trained on military secrets. The hard part is what the model remembers

The Pentagon is negotiating to build secure computing environments where Anthropic, OpenAI and other generative AI vendors can train models directly on classified military data.

Google wants the AI race judged by repeatable tests, not bigger promises

DeepMind has released the first formal framework for measuring progress toward artificial general intelligence, alongside a $200,000 Kaggle competition for novel evaluation methods.

Copilot is becoming Microsoft’s test of how far Big Tech can move beyond borrowed AI

Satya Nadella has quietly consolidated Copilot's scattered leadership under Jacob Andreou, a former Snap Inc. executive, signaling Microsoft's serious intent to reduce its dependency on OpenAI's foundational models.

AI is moving into science’s hardest audit: studies medicine cannot reliably repeat

A consortium led by CASP founder John Moult is launching an AI-driven initiative to systematically audit publishedbiological research facing a reproducibility crisis.

Grok hallucinates less, but the AI market still buys leaderboard winners

xAI's Grok 4.20 arrives with a pair of counterintuitive claims: near-zero hallucinations and a price point that undercuts the competition—while trailing badly where it counts.

AI Warning Signs

NASA’s *Perseverance* rover travels slower than a toddler’s walking pace, its every move dictated by a 22-minute communication lag with Earth.

Microsoft’s safer health bet is an AI that gets patients ready before the visit

When the Pentagon gets to decide how much AI safety is too much

Anthropic's clash with the Department of Defense has escalated from a routine contract dispute into a defining stress test for the future of AI governance.

AlphaGo’s real legacy began when AI left the board for the lab

A decade after DeepMind's AlphaGo crushed Lee Sedol, the real revolution wasn't the board game victory—it was the quiet migration from gaming tables to laboratory benches.

The engineers building AI models are now challenging how the Pentagon judges them

Over 30 researchers from OpenAI and Google DeepMind filed an amicus brief supporting Anthropic's lawsuit against the Pentagon — a signal that the technical rank-and-file no longer trusts institutional channels alone.

Microsoft’s compact AI model targets the job humans still do: clicking through software

Microsoft has released Phi-4-reasoning-vision, a 15-billion-parameter open-weight multimodal model that doesn't describe images—it plans and executes actions within graphical user interfaces.

Gemini Flash-Lite: Smarter, Faster, Pricier—But Is It Real?

Google’s Gemini 3.1 Flash-Lite promises blazing speeds but delivers a 3x cost hike, leaving developers to wonder what’s actually improved—and who’s footing the bill.