Karpathy's 'Third Paradigm Shift' — Theo (t3.gg) Reluctantly Defends Anthropic

Theo (t3.gg) takes Andrej Karpathy's framing of Claude Tag at face value rather than dismissing it. Anthropic's official description: do the under-the-hood integration work once, and Claude joins the team seamlessly. Talk to it like you'd talk to a teammate; it handles a wide variety of workloads.

Karpathy positioned this as the third redesign of LLM interface and UX. The first paradigm was that the LLM is a website you go to. The second, an app you download. The third — what Claude Tag exemplifies — is a self-contained, persistent, asynchronous entity with org-wide tools and context, working alongside teams of humans. It takes a while to wrap your head around it, but it works, Karpathy says, and it's impressive.

Theo's honest first reaction was scepticism. When one of the most important AI researchers alive talks about what looks like a Slackbot, it's reasonable to wonder if Karpathy drank the Anthropic Kool-Aid too quickly as a recent hire. Theo declares up front that he's about to do three of his least favourite things — defend Anthropic, make a video about it, and revisit his own priors.

The feature details bring him around. With ambient behaviour enabled, Claude proactively keeps you updated, flagging relevant information across channels and tools, following up on threads or tasks that have gone quiet. Given how chaotic Slack typically is, that genuinely helps. The async work model is the bigger lever — hand Claude a task and focus on something else, while Claude schedules itself and pursues projects autonomously over hours or days. Anthropic openly admits it now spends much more of its time delegating to many Claudes in parallel.

Theo can speak with unusual authority here because he runs his own version of the same idea — a self-built Hermes agent that lives in a dedicated Discord server, with a Rust port in progress. His three companies have given him a working sense of what this shape of system actually does. That's the perspective from which he can evaluate Claude Tag.

His closing read is direct, with full confidence. Claude Tag isn't a Slackbot — it's an org-level harness. From running three companies, he's saying with 100% certainty that Anthropic's direction is right. Karpathy getting dunked on by AI Twitter was, he notes, somewhat predictable given how Anthropic-as-a-company has been performing publicly lately, but the feature is genuinely cool. For readers who don't want to pay for Claude Enterprise to test it, he suggests setting up OpenClaude or a Hermes agent in your own Discord or Slack — you'll see most of the value from there. He didn't expect to make a video defending both Anthropic and Slack — but here he is.