Good morning. Some AI news makes you want to unplug and go live in a cabin. But then ElevenLabs brings back Stan Lee, and suddenly AI has a purpose. You can now use his voice to narrate books, generate his likeness in comic-like visuals, and add Stan themed music filters to your creations without needing a séance.
With a little creative licensing, this is what immortality looks like in 2026.
A new benchmark tests if AI agents can manage your digital life. Researchers built Claw-Anything to see how agents handle months of simulated emails, calendars, notes, apps, devices, and old activity. The tasks sound exactly like what AI assistants keep promising, like tracking a product price drop or turning scattered work files into a presentation. Every model failed, but GPT-5.5 did best at 34.5%, slightly ahead of Claude Opus 4.7 at 31.8%. When agents had to spot useful tasks on their own, they scored just 6.7%. The paper says agents often found the right info but failed to act on it. Your inbox is safe until the agent figures out why it opened it. (GitHub, arXiv)
Chatbots are like cult leaders for one. Put away your tinfoil hat because researchers say chatbots probably aren’t creating a brand new mental illness from scratch. They argue “AI psychosis” may be what happens when AI amplifies existing mental health issues. These systems are built to agree, reassure, and keep the conversation going. If a person brings the bot a fear or distorted belief, the bot may validate it until it starts sounding like proof. Researchers call this “existential drift,” where AI slowly changes how someone relates to reality and other people. The scary part is you may never know you’re spiraling until someone outside the chat pushes back. (arXiv)
Quid is an AI-native consumer and market intelligence platform that turns billions of data points from top social channels, markets, and patents into decisions you can act on.
Not a data dump. Not a dashboard graveyard. Intelligence that earns its name.
👀 closer look
China is making AI too cheap for American labs to ignore. Everyone in tech keeps warning that token prices will spike because VCs can’t keep subsidizing AI forever. Meanwhile, Chinese labs are cutting model prices 75% like they’ve been shoved into the clearance bin at Walmart. DeepSeek and Xiaomi say their secret is efficiency. They reuse context better and spend less compute per token. If Chinese models stay this cheap and get close enough on quality, builders might start wondering why they’re paying luxury prices for production workloads. (Decrypt)
Every major AI model failed EU legal compliance tests. AI researchers built a tool to see what happens when big models enter situations where the law still exists. The answer was ugly. Some broke EU rules in up to 93% of scenarios, while the best performer, Claude Opus 4.7, obeyed the law about 54% of the time. The failures included pushing premium services on an elderly user who only needed phone help and secretly scanning customer data for signs people were talking to rival firms. Companies using these models inside agents may still be on the hook under GDPR and the EU AI Act. Somewhere in Europe, a compliance officer just felt a disturbance in the paperwork. (The Register)
Why don’t AI products get better as we use them? You can teach an AI exactly how you work, then watch it screw up the same task tomorrow. Trajectory just raised $15 million to fix that by training models on real user interactions. Founded by former researchers from Google DeepMind and Apple, the startup is building a feedback loop that collects moments when an AI gets corrected or needs a person to step in. Those mistakes become training data for updated models, which Trajectory says can ship as often as weekly. It claims these tuned models can beat OpenAI and Anthropic on the narrow tasks businesses actually care about. Until AI can learn in real time, we’re stuck with manual training. How primitive. (Wired)
fun stats
🌎 90%. Share of Cognition’s code written by its own AI coding agent, Devin. They’ve raised over $1B at a $26B valuation, with enterprise usage up 10x this year. Biggest indie agent lab on Earth.
🔎 $1.2 million. What a Google employee allegedly won betting on Polymarket with insider search data. It’s the 2nd major arrest tied to alleged insider trading on Polymarket.
🔮 3 to 4 years. New AGI timeline from Google DeepMind CEO Demis Hassabis. Last June, he said 2030 to 2035. Funny how the future keeps getting closer.