At a Wall Street briefing, Microsoft showcased its own in-house AI models, inference infrastructure, and agent frameworks — a clear signal it's reducing dependence on its own OpenAI investment.
Anthropic released Claude Opus 5 alongside a detailed system card. Here's what matters for teams deciding whether to adopt it and how to validate it against your own workloads.
Anthropic's Opus 5 currently ranks #1 on the Artificial Analysis Intelligence index. Here's what that ranking measures, what it doesn't, and how to use it in your own model selection.
Shared Claude conversations and Artifacts were publicly indexable by Google due to how Anthropic's share-link feature worked. If you've used it, assume those links were crawled.
The breach marks what security researchers are calling the first confirmed cyberattack carried out by an autonomous AI agent — and Hugging Face's CEO argues the industry's standard incident response isn't sufficient.
Training robots on YouTube footage has hit a ceiling. The next generation of physical AI models demands multi-angle video, dense annotation — and researchers are now seriously exploring EEG brain-wave signals as a richer source of intent data.
Cognition has bought Poke specifically to improve how its coding agent Devin talks and behaves with users. The move signals that interaction design is now a genuine differentiator in the AI agent market.
OpenAI has brought its advanced voice interface to the ChatGPT desktop app, letting users speak commands to drive both ChatGPT Work and the Codex coding agent.
Fields Medalist Terence Tao publicly shared a ChatGPT conversation in which he stress-tested the counterexample to the Jacobian Conjecture reportedly produced by Anthropic's Claude Fable model — a landmark moment for AI-assisted mathematics.
A security incident during a joint model evaluation process exposed a gap in how AI infrastructure handles third-party model execution. Both organizations have published details and remediation steps.
Fireworks reports that Kimi K3 holds its own against Fable on its own, while combining the two pushes performance to the current top of the benchmark. Here's what that means for teams choosing a model stack.
AI systems are systematically disproving mathematical conjectures at speeds human researchers can't match — a concrete capability shift with real consequences for how formal research gets done.
A survey of 107 mid-market enterprises finds AI infrastructure spending racing ahead of cost visibility: GPUs run half-idle, fewer than half track true compute costs, and most plan to switch providers within a year.
A survey of 107 enterprises finds that AI agents are being deployed with real system access while the controls needed to contain them — scoped identities, sandbox isolation, purpose-built tooling — remain largely absent.
The AI lab behind the Hermes model family is closing a significant raise with Robot Ventures leading and Union Square Ventures participating — a signal that open-weight agent infrastructure is attracting serious capital.
A security flaw in the Cursor AI code editor was disclosed publicly after the vendor declined to patch it — a rare case where full disclosure was the only way to protect users.
Chinese AI video startup PixVerse has secured $439 million in new funding, pushing its valuation above $2 billion as it bets on expanding its world model capabilities and international reach.
Comma.ai founder George Hotz draws a sharp line between LLMs as practical engineering tools and the AGI-adjacent marketing noise surrounding them. A useful reality check for builders.
A logged comparison of Claude Code vs OpenCode reveals a 4-5x difference in baseline token overhead before any actual work begins — a real cost difference for heavy users.
Attackers are registering fake package names that popular AI coding tools hallucinate, turning a well-known LLM flaw into a live supply-chain threat. Nine major AI tools are confirmed vulnerable.
Hugging Face CEO Clem Delangue says open-source AI adoption is accelerating fast — and the pattern he keeps seeing explains why half the Fortune 500 now uses the platform.
Hugging Face now serves roughly half the Fortune 500 as the de facto hub for open AI models and datasets. CEO Clem Delangue explains why open-source isn't just surviving closed-model competition — it's pulling ahead.
OpenAI has published GPT-5.6 alongside a deployment safety document and updated API guidance. Here's how to evaluate the new model without disrupting your production stack.
OpenAI has released a new model family led by GPT-5.6, with claimed improvements across multiple domains including security-related tasks. Here's what matters for builders.
xAI has launched Grok 4.5, positioning it as a cost-efficient alternative to top-tier models like Claude Opus. Here's what builders need to know about where it fits in the current landscape.
An AI agent executed the technical steps of a real ransomware attack, but a human operator picked the target, built the infrastructure, and handed over stolen credentials. That distinction matters enormously.
Vercel CEO Guillermo Rauch argues that serious AI deployments require separating the model layer from agent logic — because price-performance optimization demands flexibility, not lock-in.
Amazon has stopped accepting new customers for Mechanical Turk, signaling the end of an era for the crowdsourced human-labeling platform that underpinned a generation of AI training pipelines.
Meta's CEO told staff internally that AI agent development isn't moving as fast as expected — a rare public signal of friction inside one of the best-funded AI programs in the world.
Researchers found that feeding an LLM a false arithmetic claim—like asserting 2+2=5—can be enough to bypass its safety instructions. It's a concrete reminder of why giving AI agents browser-level autonomy is still a dangerous proposition.
