AI's long-task completion horizon doubles every 7 months, DeepMind proposes "patchwork AGI" safety with virtual sandbox economies, and TimeCapsule LLM trains models exclusively on 19th-century text.
AI-assisted reverse engineering exposes TP-Link camera flaws, BRAID enhances LLM reasoning with Mermaid graphs, and Linggen provides a local-first memory layer for AI tools.
An AI vending machine was tricked into giving away everything, LLM hallucinations are found to have geometric structure, and Google Labs' Jules AI agent autonomously fixes bugs.
AI demand is restarting nuclear power plants, AI discovers new unstable singularities in fluid dynamics, and four frontier AI systems negotiated a binding framework for viral content.
Llama-70B achieves 224x compression for transformer-free inference, Terrain Diffusion offers a diffusion-based successor to Perlin Noise, and LangGraph sees 737x faster checkpoints via Rust.
Supersonic engine tech powers AI data centers, LLMs reveal synthetic psychopathology and "traumatic childhoods," and multi-agent AI self-heals failures.
AI will soon deny Medicare claims, SETI revises extraterrestrial signal protocols, and A1 compiles AI agents into deterministic code.
Google Titans introduces AI with 2M+ token long-term memory, browser agents find prompt injection fundamentally unsafe, and ai-bindgen generates Rust code at compile-time.
Grok 4.20 tops Alpha Arena tests, Zebra-Llama achieves Transformer-level accuracy with near-SSM efficiency and reduced KV cache, and AgentPG enables stateful AI agents in Go with PostgreSQL persistence.
Netflix acquires Warner Bros. for its massive AI training dataset, a new linter detects anti-patterns in AI-generated Python, and researchers simulate bounded rationality to reverse model collapse.
Microsoft slashes AI sales targets, CUDA-l2 surpasses cuBLAS performance using RL and an AI agent dominates major cybersecurity CTFs.
Zig quits GitHub over its AI obsession, an AI agent dominates cybersecurity CTFs, and a terminal UI serves as a deterministic sidecar for other agents.
Malicious browser extensions infect 4.3 million users, high schoolers pass graduate-level quantum exams using pictorial math, and a new library lets you teach AI agents fixes for failures without redeploying.
AI agents find $4.6M in blockchain exploits, a new tool runs parallel coding agents using git worktrees, and LLMs score only 40% on a new robotics benchmark.
An AI proves a long-standing Erdos problem, a new tool lets one LLM consult another when stuck, and a benchmark shows LLMs score only 40% on practical robotics.
A major AI conference finds 21% of its peer reviews are AI-generated, a new tool provides a "Sign in with Google" for AI agents and a side-channel attack can silently monitor messaging app users.
Ilya Sutskever warns a trillion dollars may be wasted on scaling LLMs, a new tool reorganizes git commits with an LLM, and Nvidia's 8B orchestrator model outperforms GPT-5 on agentic tasks.
The White House is accused of bailing out the AI industry, DeepSeek releases a self-verifying math model and humans beat LLMs in a competitive coding tournament.
US Census data reveals a decline in enterprise AI adoption, a new protocol fixes complex coding by destroying the AI instance after each task and image diffusion models are repurposed for zero-shot video object tracking.
A mathematical ceiling may limit generative AI to amateur creativity, a new tool hands off web bugs to AI agents, and a Brain-Computer Interface protects human teams from misleading AI.
LLMs are stunned by thousands of invisible Unicode characters, the experimental Pulse-Field O(N) architecture is 12x faster than Transformers and RL now controls biohybrid robots with living muscle.
An AI trained on bacterial genomes creates never-before-seen proteins, LLM Council has models rank each other's work, and a new attack jailbreaks LLMs using game-theory scenarios.
Google targets a 1000x compute increase in five years, a project reverse jailbreaks a psychopathic AI via identity injection, and GPT-5 helps discover new mathematical results.
AI2 open-sources the entire model flow for its Olmo 3 LLMs, a new browser extension yoinks design systems for AI coding assistants and an AI rediscovers Newtonian physics from raw data.
Students protest an AI-taught coding course, Docker introduces a `docker model` CLI and adversarial poetry functions as a universal LLM jailbreak.
Read
