Corey Noles

It’s tempting to write Grok off as Elon’s edgy chatbot. But that misses the real story: xAI is building one of the most culturally native AI products on the market, and its biggest strength, deep integration with X and the live internet, is also what makes every safety failure impossible to ignore.

Meta just bought Moltbook, the viral social network where AI agents post and interact. Beneath the absurdity is a serious signal: the next AI battle may be about agent identity, coordination, and network effects, not just better models.

Thinking Machines Lab has had a rough few months, losing prominent employees to rivals including OpenAI and Meta. But its new NVIDIA partnership suggests Mira Murati’s startup may still be very much in the frontier race.

CoreWeave just introduced Flex Reservations and Spot, two new ways to buy AI compute that better match how modern workloads actually behave. That may sound like infrastructure plumbing—but for builders, it could mean cheaper scaling, better reliability, and fewer painful tradeoffs between guaranteed capacity and wasted spend.

Andrej Karpathy’s new open-source project, autoresearch, lets an AI agent run ML experiments while you sleep—editing training code, testing ideas, and keeping only what works. It’s a small repo with a very big thesis: the future of AI research may look a lot more like designing the system than manually running every experiment yourself.

Microsoft’s latest AI push is about making AI agents governable, secure, and usable at enterprise scale. With Wave 3 of Microsoft 365 Copilot, Agent 365, Claude support, and the new E7 Frontier Suite, Microsoft is betting that trust will matter just as much as intelligence.

Microsoft’s Phi-4-Reasoning-Vision-15B is an open-weight multimodal model that brings structured reasoning to vision tasks, especially UI screenshots, dense documents, and visual math. The twist: it learns when extended reasoning helps (and when it just adds latency).

Codex App for Windows is live, and it’s more than just “new OS support.” It gives teams one place to run multiple AI coding tasks in parallel, keep work isolated, and review clean diffs without branch chaos. We break down what’s new and exactly when to choose the App, CLI, or Web.

OpenAI, Google, and Alibaba all dropped new models today—but the real story isn’t raw power. It’s speed, cost, and on-device AI. GPT-5.3 Instant, Gemini 3.1 Flash-Lite, and Qwen 3.5 signal a shift away from massive parameter races toward deployable, real-world intelligence.

Leaks in Codex logs and GitHub pull requests suggest OpenAI is already testing GPT-5.4—just weeks after launching GPT-5.3. Scrubbed commits, A/B flags, and “Fast mode” hints point to rapid iteration happening behind the scenes.

Vibe coding gets you to “it runs.” Engineering gets you to “it’s accountable.” Here are 10 copy-paste questions to ask your coding agent of choice before you ship covering security, reliability, tests, and deployability.

How to stop overpaying for AI by routing every task to the cheapest model that can actually do the job.