Nicholas Carlini — wait, this isn't Carlini. This is an unnamed research engineer who moved from Microsoft's Windows AI team to Salesforce AI Research, and his year-over-year snapshot of AI usage is more revealing than any vendor benchmark. The headline number: ~$3,000/month in API tokens, spent mostly on agentic coding workflows where 5-6 Claude Code instances run simultaneously on separate Git worktrees, each implementing features, fixing bugs, and drafting pull requests. He no longer writes much code himself. The agentic coding section is the load-bearing wall. Last year, GitHub Copilot autocomplete was the star. This year, autocomplete is described as 'still great, obsolete.' The workflow now resembles a project manager running a team of junior developers: assign tasks, review architecture, write guidance documents (CLAUDE.md), build test harnesses so agents can validate their own output. One agent independently discovered it could use macOS text-to-speech to test a voice agent it was building. Another pattern: agents recommending configuration knobs the human rejected as unnecessary — only for the human to need them a week later. But the author is honest about the failure modes, and this is where the piece earns its keep. Agents left unsupervised produce 'poorly-functioning, over-architected, unreadable, self-congratulatory piles of slop.' His proudest achievement this year was hand-writing a tool that replaced ~4,000 lines of agent-generated code with a few hundred lines of human code — more flexible, testable, and generalizable. The human edge, he notes, is taste and architecture. How long that edge lasts is an open question he doesn't pretend to answer. The psychological costs are stated plainly. Managing six simultaneous agent conversations is 'overstimulating and exhausting in a way that coding never was.' He caught himself rushing to answer agent queries like they were human colleagues waiting for help. A friend laments ceding the meditative joy of programming to machines. The author has begun drafting a personal AI policy — a framework for when AI helps versus when it hurts — partly in response to watching teammates ask Claude to write their Slack messages and grandparents generate 'creepy memes' of their grandchildren. Google's AI summaries have become his default search interface. Product search remains mediocre — he doesn't trust Gemini for financial advice or furniture recommendations — but 'pub facts' queries work well. He found a specific LEGO TikTok creator and an art vtuber lizard through natural-language Google queries that would have been impossible two years ago. ChatGPT now shows ads in conversations, which he finds uncomfortable. The piece also covers new territory: voice interfaces, AI-generated video (unspecified detail in the truncated article), and the emergence of personal AI policies as a genre. He cites Hank Green's policy as a model and notes that BrainPOP, an educational content company, appears to be using AI-generated art — possibly unknowingly. The broader observation: humans have no norms for these tools yet, and the absence of norms is producing visible failures across professional and personal life. What makes this valuable is the specificity. This isn't a think piece about whether AI will replace programmers. It's a practitioner's log with concrete workflow details, dollar amounts, tool names, and honest admissions of what doesn't work. The $3,000/month token budget alone tells you this is a power user whose experience previews what broader adoption will look like — including the exhaustion, the slop, and the quiet grief over lost craft.
