Goodfire has introduced Mechanistic Shaping, which uses internal model representations during training to reduce reward hacking in RL models from 100% to 2% while raising honesty to 96–98%. The method was validated across language and vision tasks and points to a new approach for alignment.
METR reported that Anthropic's internal model capability growth rate accelerated to 1.9 times its historical pace starting in October 2025. The report attributes most of the acceleration to AI assistance in R&D workflows rather than fully autonomous contributions.
Physicl is trying to solve the sim-to-real gap with a pipeline originally built to market goods online.
Despite its small size, Underdog's new model can generate both speech and music.
Pentagon Chief Technology Officer Emil Michael said the federal government will publish a comprehensive map of existing federal laws and regulations applicable to domestic AI labs within one to two weeks. The move reflects the White House’s preference for consolidating existing authorities rather than enacting new AI-specific legislation.
The U.S. Department of Justice has created DOJ-SITF, its first dedicated AI enforcement body holding both criminal and civil authority. However, because existing laws such as the CFAA still require proof of deliberate human intent, its ability to prosecute autonomous AI conduct is limited, while companies face growing compliance pressure without clear standards.
I’m not the sort of person who typically pays for a house cleaner. I’m too frugal, and I kind of enjoy putting on a podcast and going to town on my tub with a Magic Eraser. But when I saw an X post about a new humanoid-robot cleaning service in San Francisco, I immediately signed…
Clutch, backed by Lightspeed and SignalFire, was last valued at $85 million, according to a source.
SpaceX plans to build 13 GW of AI data center capacity by the end of 2027, potentially surpassing the Stargate alliance's planned 7-10 GW. Its speed advantage comes from Tesla Megapack energy storage bypassing grid interconnection delays, with Google and Anthropic already committing $26 billion in annual contracts.
Airbnb says its upcoming AI agent will help travelers plan trips, compare options, manage bookings, and get assistance throughout their stay.
As Anthropic maintains its vow to keep ads out of Claude, the rest of the AI chatbot industry is racing to win over advertisers without antagonizing users.
Bitter energy drinks turned out to be ChatGPT-steered poisoning, lawsuit says.
Google is technically offering schools more choice but not the choices anyone wants.
NVIDIA has released four open-weight Cosmos 3 Nano checkpoints for autonomous driving and robotics, extending interactive world simulation into both sectors. While the open license lowers ecosystem barriers, the 96GB VRAM requirement for Nano and data-center GPU needs for Super still create clear hardware tiers.
OpenAI has released GPT-6 Cyber, a specialized security model that improves exploit development by 51% over its predecessor while being classified Critical under OpenAI's Preparedness Framework. OpenAI says access controls and alignment training support release, though most safety data is internal and only the Firefox vulnerability findings have been independently confirmed.
OpenAI's newest cybersecurity model comes out as AI labs face scrutiny of their work to protect digital infrastructure.
Meta warns that growing political opposition to data centers could threaten future AI investments, as communities push back against massive infrastructure projects.
Meta's AI agent can now take calls as a customizable animated avatar, ahead of the launch of a Tamagotchi-like wearable designed to bring Muse to life.
In today's Smoke evaluation, GPT-6.1 Sol's material constraint score fell 29.1 points to 60.70, while code execution rose 19.5 points to 94.50, leaving the main leaderboard score at 79.29. The opposing moves likely reflect small-sample question sampling rather than systematic model degradation.
In today's Smoke evaluation, GPT-6 Astra's material constraint score fell from 89.80 to 60.70, a drop of 29.1 points, while code execution rose from 75.00 to 94.50; its overall main leaderboard score declined from 81.66 to 79.29.
