Tracking the evolution of autonomous agents & frontier models.
Extreme sports operator ATV Big Air Tour reports compressing a three-day marketing content workflow into just three hours using OpenAI's ChatGPT.
Google launches the Fairwind Program, offering select government and enterprise partners access to its advanced Gemini 3.8 Flash Cyber model for autonomous vulnerability remediation.
Google DeepMind releases Gemini 3.8 Flash for advanced coding and AI agents, alongside the specialized Gemini 3.8 Flash Cyber for cybersecurity, continuing its rapid pace of cost-effective model development.
A new guide from LiquidAI details how a 350M parameter model was fine-tuned using GRPO to achieve a 7.1-point improvement in structured output generation on the IFStruct benchmark, using only 100 training steps.
Hugging Face has released funes, an open-source tool that provides a durable, local-first memory layer for coding agents like Claude Code and Codex, allowing them to recall past reasoning across different sessions, machines, and even agent types.
NVIDIA's latest developer guide demonstrates how to achieve a 300x performance increase in a CUDA pipeline by using modern tools like CCCL and CUB for debugging and optimization.
Industry analysis on how AI model co-design using speculative decoding can accelerate LLM inference speed while maintaining the output accuracy of the original model.
TechCrunch Disrupt 2026 announces its Builders Stage agenda, featuring leaders from Google, Blackstone, and Radical Ventures to provide founders with actionable strategies on fundraising, talent, and growth in the AI era.
Widespread service disruptions at OpenAI are preventing user access, highlighting potential network infrastructure vulnerabilities and the operational risks of relying on a single AI provider.
Google details the technical capabilities and safety protocols for Project Astra, its real-time multimodal AI agent designed to compete directly with new offerings from OpenAI.
Google has launched agentic video understanding for its Gemini models, a new feature that reduces token consumption by up to 88% and improves accuracy for complex video analysis.
The Allen Institute for AI has released BenchMIRT, a new analytical method that reveals how single LLM benchmark scores can obscure the multiple, distinct capabilities they are actually measuring.
Hugging Face has released @huggingface/kernels, a collection of over 200 optimized WebGPU kernels and a new benchmarking tool called Fleet, designed to accelerate local AI inference in web browsers.
NVIDIA and CrowdStrike demonstrate an agentic cybersecurity system where specialized Nemotron open models automate offense-defense testing, achieving superior detection rates and generalization compared to a leading frontier model.
A new framework helps organizations size GPU infrastructure for AI inference by analyzing use-case token patterns, key performance metrics, and a core-and-flex capacity model to optimize TCO.
OpenAI faces 30 new lawsuits from victims of the Tumbler Ridge shooting, with plaintiffs now alleging the company intentionally aided and abetted the attack, escalating the legal battle over AI platform liability.