Weekly AI Insights
Practical perspectives on building software with artificial intelligence. Written by Alex Shively — one post per week.
A Chinese lab's Kimi K3 is the largest open-weight model ever built, and it just out-coded Claude and GPT on a public leaderboard
Moonshot AI released Kimi K3 this week, the biggest open-weight AI model ever made, and it topped a well-known coding leaderboard ahead of the best models from Anthropic and OpenAI. Here's what "open weight" actually means and why it matters even if you never download it.
Anthropic's Claude Corps offers an $85,000 year learning to use AI inside a nonprofit — and the first application window closes this week
Anthropic's Claude Corps pays early-career people $85,000 to spend a year learning AI while helping a nonprofit, and the first cohort's applications close July 17. Here's what the program actually is, who it's for, and why it hints at how the "AI and jobs" story might play out.
Claude Sonnet 5 quietly became the default for millions of free and Pro users, and it's built to see multi-step work through to the end
Anthropic made Claude Sonnet 5 the default model for every free and Pro user this week. It's cheaper to run, better at finishing multi-step tasks on its own, and a useful window into what "agentic AI" actually means for your work.
Gemini's new Deep Think mode trades speed for thinking — and quietly took the benchmark crown
Google rolled out Gemini 2.5 Pro with Deep Think this week, a mode that pauses to reason through hard problems instead of answering instantly — and the early benchmarks put it ahead of every rival. Here's what "thinking" models actually do and when it's worth the wait.
For the first time, less than half the AI assistant market belongs to one app — and that's good news for the rest of us
ChatGPT's share of the AI assistant market slipped below 50% for the first time this month, with Gemini and Claude picking up the difference. Here's what the new numbers actually mean for anyone trying to pick the right AI tool for work.