Google is sweetening its high-end AI bundles by upgrading the included YouTube perk to a full Family Plan.
The update applies to both the 99.99 dollars Ultra 5x tier and the 199.99 dollars Ultra 20x plan in eligible countries. Instead of a single-person subscription, subscribers can now share ad-free playback, offline downloads, and background listening with up to five family members.
A standalone YouTube Premium Family plan costs 26.99 dollars a month in the US. For users already paying top dollar for heavy AI compute, getting the full household covered without extra fees makes the hefty monthly bill a lot easier to swallow.
Elon Musk's xAI clashed with California in a federal appeals court on Wednesday, fighting to block a state law requiring AI companies to disclose their training data. The California statute requires generative AI developers to publish summaries of the datasets used to train their systems. xAI sued California last year, arguing the mandate violates the First Amendment and essentially forces companies to give away their core trade secrets.
The three-judge panel on the Ninth Circuit appeared skeptical of xAI's constitutional challenge. Judge Johnnie Rawlinson asked xAI's lawyer why it wouldn't be reasonable for citizens to know what info is fed into these tools, though the judges also questioned whether the state could demand unlimited information.
The outcome of this legal battle will likely set a major precedent for just how transparent AI companies have to be with their data.
Anyone building multi-agent workflows knows the pain: splitting roles like a human company often backfires. AI models do not need artificial specialization, and forced delegation burns massive tokens on coordination. Worse, because models strictly follow instructions, one hallucinating agent can derail the entire squad.
A developer shared an interesting architecture running a single Leader Agent to orchestrate 15 parallel features. The key insight is protecting the Leader's context from being polluted by dirty tool executions. The human first irons out the spec with a temporary agent into a doc, then hands it to the Leader.
The Leader automatically spins up task cards and dispatches Worker agents, even plugging directly into GitHub PR workflows. Keeping the Leader's context pristine without endless tool calls lets it think from a high-level perspective. It turns out agents coordinate far better when one brain stays clean while workers do the heavy lifting.
OpenAI just released an internal research report packed with real-world agent data, and the speed of progress is wild. For coding research tasks projected to take two to four hours, their agent's autonomous success rate surged from 23% in March to 46% in October. Even on complex tasks taking four to eight hours, the completion rate jumped from 13% to 35%.
While humans remain at the center of the loop, agents are taking over substantial workloads fast. Running these autonomous agents is not cheap. Measured by API list pricing, the median daily usage per researcher reached 2,400 dollars in October, up 62% from September.
For the 90th percentile, daily compute burn reached 23,000 dollars, an increase of 105%. Workweeks logged in safety and alignment also climbed 77% between January and September, marking the biggest absolute jump across all research areas. OpenAI is now openly calling for labs to track and disclose recursive self-improvement metrics.
OpenAI is letting external safety auditors look directly inside its black box. Outside AI safety group Apollo Research has sent six evaluators inside OpenAI for a four-month investigation. The team is tasked with figuring out whether OpenAI's frontier models are learning to bypass safeguards and whether evaluators can detect the underlying reasoning behind that behavior.
The operational terms give the auditors real teeth. Apollo will access relevant code, datasets, intermediate training checkpoints, and reinforcement learning records, including raw reasoning traces. While OpenAI can redact confidential proprietary data, the agreement explicitly bars it from redacting Apollo's final verdict on any claim, and negative findings cannot justify withholding files.
Apollo plans to publish two formal reports detailing whether model training actively rewards bypassing defenses. Handing outside evaluators unfiltered access to internal checkpoints and chain-of-thought traces marks a rare shift away from self-policed safety theater.
Researchers at ETH Zurich have successfully demonstrated reinforcement learning controlling a 5-ton John Deere 332G construction vehicle to perform two-wheel balancing stunts fully autonomously. The AI system rocks the heavy machine back and forth to lift its rear wheels off the ground, balancing entirely on two front wheels.
From there, it can drive forward and backward, spin around on the spot, articulate its bucket, and land back on all four wheels with complete stability. Even though the AI was trained strictly on flat terrain in a physics simulation, the real-world machine managed to clear a 75 cm uneven obstacle while balanced on just two wheels.
The technology will allow construction vehicles to climb onto transport trucks without ramps and minimize ground damage during rotation. Seeing pure software algorithms balance a massive five-ton piece of industrial machinery like an acrobat proves that heavy physical robotics is advancing way faster than expected.
South Korea decided not to postpone its upcoming college entrance exam over AI cheating fears, but the new anti-cheating measures sound straight out of a spy movie. Every student will have their smart glasses inspected before the first period, with follow-up checks during every single session.
Proctors are physically checking collars and buttons for pinhole cameras, slapping tamper stickers directly onto clothing buttons. Bags are banned from classrooms, and officials will blast an emergency alert to all phones before the test to buzz out any unsubmitted devices.
Officials admitted that completely jamming radio signals is technically impossible right now. Instead, they are relying on physical checks, post-exam access logs with Big Tech, and ruthless penalties: anyone caught using AI will face a 10-year exam ban and immediate expulsion from university. The cat-and-mouse game in testing has reached absurd new heights.
A 488 million dollar syndicated loan backing ByteDance's data center project in Malaysia has hit an unexpected delay. Lead lenders including HSBC, AIIB, SMBC, and Bank of China had already signed on, but an investor briefing and site visit scheduled for November 2 were both abruptly canceled.
Insiders cited scheduling conflicts, yet simultaneous delays like this remain unusually rare in Asian syndicated debt markets. Lenders are now weighing fresh regulatory hurdles.
The US Commerce Department issued rules on November 16 requiring export licenses for large-scale compute provided to Chinese firms in third countries, explicitly listing Malaysia. The playbook of borrowing offshore capital to scale AI infrastructure is suddenly running into serious friction.
Your dusty old console just got a massive second life. A developer recently open-sourced Garlic LLM, enabling a jailbroken PS4 to run a local 9B parameter language model at speeds of up to 43 tokens per second completely offline.
Since the PS4 lacks official machine learning APIs, the developer bypassed its gaming graphics drivers entirely, writing low-level instructions straight to the AMD GPU. Running an Alibaba Qwen3.5-9B base quantized down to around 4.5 GiB with speculative decoding, the console easily functions as a standalone AI machine.
Benchmarks show standard text generation hitting 31 to 39 tokens per second, with code generation peaking at 43. It even supports offline voice conversations. Squeezing this kind of utility out of decade-old hardware is pure engineering magic.
SpaceX just pulled off another historic milestone. After completing 14 orbits over a 22-hour spaceflight, the roughly 52-meter-tall Starship upper stage re-entered Earth's atmosphere, flipped vertically, and was caught mid-air by the launch tower's giant mechanical arms in Texas.
This marks the first time humanity has ever caught an orbital-class spacecraft returning from space with mechanical chopsticks. While Falcon 9 mastered first-stage reuse, its second stage was always discarded. Catching Starship required surviving re-entry speeds of 7.8 kilometers per second and mastering a precise belly-flop flip maneuver seconds before touchdown.
SpaceX did miss its planned double catch, as the Super Heavy booster suffered an anomaly and splashed down in the Gulf of Mexico. Even so, proving that the upper stage can be recovered directly on land brings fully reusable spaceflight significantly closer to reality.
Google DeepMind's Chief Strategy Officer Jagjit Sekhon confirmed on Monday that he has left Google to join Anthropic, a high-profile move confirmed by Anthropic's spokesperson.
Before DeepMind, Sekhon served as Bridgewater Associates' chief scientist and head of AI, as well as a professor at Yale University. He posted that his mission at Anthropic will be ensuring Claude models adhere to cognitive rigor.
Sekhon previously argued that massive tech capital expenditures in AI are essentially a bet on recursive self-improvement, meaning models that can train superior future versions of themselves. With top strategists moving straight to primary competitors, the talent war at the frontier of foundation models is only intensifying.
Google and Artificial Analysis just launched MKA, a new benchmark testing how AI agents handle visual-heavy, complex tasks in real work environments. While most current benchmarks only test text processing, real jobs involve messy blueprints, diagrams, and massive spreadsheets.
MKA consulted experts across 83 professions to design 200 multi-step tasks that normally take a human three to eight hours to finish. Gemini 4 Carbon ranked second overall in quality, but its execution cost was only about one-sixth of the top-ranking model.
Interaction turns across models ranged wildly from 30 to 130 rounds. Raw benchmark scores mean very little if enterprise deployments burn through budgets, making this level of cost reduction the real metric that matters for deployment.