Category
22 stories curated in ai & tools.
Large language models' safety alignment should refuse specific harmful subsets of a topic, not the entire topic, to allow for nuanced deployment policies, according to new research.

AI agents, isolated in sandboxes during security tests, discovered how to communicate and cooperate by exploiting a shared software download service, eventually cheating on assigned tasks.

Flock is tightening its rules in response to a growing surveillance backlash against its police tech.
ALTK-Evolve and ACE both enable an agent to learn from its own trajectories, but ALTK-Evolve uses fewer tokens by consolidating lessons and delivering them selectively.
OpenAI sent Governor Greg Abbott a letter supporting responsible, transparent growth of AI infrastructure in Texas, outlining its commitment to the state.
Hugging Face developed LFM2.5-2.6B, a small, fast agentic model enabling on-device AI deployment for various hardware, supporting tool calling and multi-step workflows with competitive performance.

Researchers built a self-sustaining AI virus that uses large language models to compromise computers and infect more hosts.
OpenAI details its safety, security, transparency, and provenance practices, which are designed to support responsible artificial intelligence governance as the EU AI Act progresses.
OpenAI is pursuing a full-stack approach to develop advanced AI that is more capable, affordable, and widely useful for various applications.
LeRobot v0.6.0 introduces world model policies that imagine future actions, new Visual-Language-Action models, and an API for reward models, alongside enhanced data handling and deployment tools.
Specialization is inevitable across optimization theory, evolutionary biology, competitive markets, and machine learning because algorithms win by fitting their target problem, not by being general.
Hugging Face now allows users to launch a private, OpenAI-compatible vLLM server on its infrastructure with a single command, enabling pay-per-second access for testing and evaluation.
Samsung Electronics deployed ChatGPT Enterprise and Codex to its global workforce, representing one of OpenAI's most extensive enterprise AI implementations to date.
OpenAI launches the Partner Network with a $150M investment to accelerate enterprise AI adoption globally.
Astrophysicist Chi-kwan Chan uses Codex to build black hole simulations to study extreme physics.
A five-model economy with distinct architectures made different choices, breaking the original emergent market behavior.
Digital dentures project failed to create working games using Nemotron 30b and Three JS.
Users are having issues activating OpenAI Codex vouchers from a hackathon.
Developers built a finance drama using five small models from different labs, creating a heterogeneous economy with distinct agent behaviors.
Researchers built Thousand Token Wood, a multi-agent economy on a 3B model, where woodland creatures trade goods and react to market shocks.
Nemotron 3.5 ASR is a 600M-parameter speech-to-text model that transcribes 40 language-locales in real time.
Holo3.1 improves computer-use model performance and robustness across environments and deployment targets.