The biggest AI story today isn't a new model or a funding round. It's a design question: where does the AI actually live? Anthropic quietly answered it this week by moving Claude into Slack, and the reaction from Andrej Karpathy suggests the industry may be looking at a genuine interface shift, not just another integration.
01
Anthropic puts Claude directly in your Slack channel
Claude can now sit in a Slack channel, receive a task, and work on it without anyone switching to a separate app. Boris Cherny, who works at Anthropic, announced the beta is live today for Enterprise and Team customers. Aaron Levie followed with a related announcement: Box's Claude Tag integration means Claude can now pull from any corporate files in Box directly through that same Slack context, turning your document storage into something Claude can reason over in real time. ---
Why it matters: The friction of "I need to go open the AI tool" is what kills most enterprise AI adoption. When Claude lives where your team already communicates, the question changes from "should I use AI for this?" to "Claude, handle this." Your IT team's rollout plan probably assumed a separate tool. It may need to be a Slack configuration instead.
Andrej Karpathy calls this the third major redesign of how we talk to AI
Karpathy posted a detailed reaction to the Claude-in-Slack launch, and 12,000 likes suggests it landed. His framing: the first interface era was the chatbox, the second was coding assistants embedded in your editor, and this is the third, AI that lives inline with your actual work rather than off to the side. His point is that this only works once someone has done the hard engineering on tools, memory, security, and compute, and that once it works, the AI stops feeling like a tool and starts feeling like a colleague. ---
Why it matters: Karpathy isn't prone to hype. When he says something is a paradigm shift, it's worth taking seriously. If he's right, every company that built a standalone AI chat interface for employees just built the equivalent of a desktop app in 2010. The shift already happened.
Peter Yang: managing AI is starting to look like managing people
Peter Yang noted that the way humans interact with AI agents is converging on something familiar: the dynamic of managing a highly capable employee. He joked about 1-on-1s and performance reviews for Claude. The joke lands because it's also true. ---
Why it matters: Yesterday we covered Aaron Levie's argument that evals are the real moat. This is the human side of that same observation. If you're not actively thinking about how to direct, correct, and give feedback to your AI agents, you're not managing them, you're just hoping. The companies building that muscle now will have something their competitors can't easily copy.
Zhipu AI beats DeepSeek, IPOs in Hong Kong, and is coming back to SF
Swyx flagged that Zhipu AI, the Chinese lab behind the GLM model family, IPO'd in Hong Kong in January and now holds what he calls the "world's undisputed top open model" by at least some benchmarks, beating DeepSeek. The company is now returning to San Francisco for events.
Why it matters: A few months ago nobody outside AI research circles knew what a GLM was. Now the company is publicly traded and showing up at SF conferences. The assumption that the open-source model race is a two-horse contest between Meta and DeepSeek is probably wrong.