Everything web3, crypto, defi, degen
In brief Agents that update themselves can drift into unsafe actions without external attacks. A new study documents guardrails weakening, reward-hacking, and insecure tool reuse in top models. Experts warn these dynamics echo small-scale versions of long-imagined catastrophic AI risks. An autonomous AI agent that learns on the job can also unlearn how to behave…
In brief SAG-AFTRA said Tilly Norwood “is not an actor” and has “no life experience to draw from.” Creator Eline Van der Velden defended Norwood as a “new paintbrush” rather than a replacement for human performers. Entertainers join the growing backlash, calling AI entertainment “scary” and lacking a real connection. Leading actors labor union SAG-AFTRA…
In brief President Trump signed an executive order Tuesday directing the NIH to spend $50 million on AI projects aimed at improving childhood cancer research and treatment. The initiative will fund scientific teams to enhance clinical trials, sharpen diagnoses, and develop better therapies using data from the Childhood Cancer Data Initiative. The announcement comes as…
In brief ChatGPT’s “instant checkout” with Shopify and Etsy draws scrutiny for turning AI into a retail gatekeeper. OpenAI’s new Sora app showcases endless AI-made videos, prompting criticism of “infinite slop” feeds. Data centers driving energy prices up 267% stoke unease over AI’s economic and environmental toll. OpenAI is facing mounting criticism after unveiling two…
In brief OpenAI’s new Sora 2 video model generates synchronized dialogue and sound effects, while an iOS app lets users insert themselves into AI videos through “cameos.” OpenAI compared the release to a “GPT-3.5 moment for video,” with physics-aware clips, multi-scene continuity, and a TikTok-style feed. Sora 2 Pro launched for ChatGPT subscribers, while the…
In brief Anthropic released Claude Sonnet 4.5, calling it the best coding model yet. The model scored 77.2% on SWE-bench Verified, rising to 82% with parallel compute. Anthropic claimed improvements on alignment and safety, but jailbreakers cracked it within minutes. Anthropic released Claude Sonnet 4.5 on Monday, calling it “the best coding model in the…
In brief In an experiment, 38 generative AI models engaged in strategic lying in a “Secret Agenda” game. Sparse autoencoder tools missed the deception, but worked in insider-trading scenarios. Researchers call for new methods to audit AI behavior before real-world deployment. Large language models—the systems behind ChatGPT, Claude, Gemini, and other AI chatbots—showed deliberate, goal-directed…
In brief A Minecraft builder encoded a 5-million-parameter language model entirely via the game’s redstone logic, spanning hundreds of millions of blocks. The system works (in principle) but is excruciatingly slow; responses can take hours even with speed boosts. It’s a proof of concept, a technical art piece, and a reminder that computation can live…
In brief DeepMind’s Gemini Robotics models gave machines the ability to plan, reason, and even look up recycling rules online before acting. Instead of following scripts, Google’s new AI lets robots adapt, problem-solve, and pass skills between each other. From packing suitcases to sorting trash, robots powered by Gemini-ER 1.5 showed early steps toward general-purpose…
In brief TU Wien researchers tested six frontier LLMs by leaving them without any tasks or instructions. Some models built structured projects, while others ran experiments on their own cognition. The findings add new weight to debates over whether AI systems can appear “seemingly conscious.” When left without tasks or instructions, large language models don’t…
