▶ YouTube Summaries

• https://www.youtube.com/watch?v=tOC2N0B9lio
• Original title: The next paradigm shift (according to Karpathy)

Theo unpacks Anthropic's new "Claude Tag" Slack integration and Karpathy's claim that it represents the third major redesign of LLM UX. Karpathy's framing: paradigm one was the LLM-as-website, paradigm two was the LLM-as-app, and paradigm three is a persistent, asynchronous entity with org-wide tools and context that works alongside human teams. Theo argues the genuinely important idea is using the Slack channel as the boundary for context and tool access, and from hands-on experience with his own Discord-based "Hermes" agents he agrees this is roughly the right direction — while criticizing the lock-in to a single Anthropic model.

Karpathy's "third paradigm" framing

Karpathy describes Claude Tag as the third major redesign of LLM UI/UX:

• First paradigm: the LLM is a website you go to.
• Second paradigm: it's an app you download to your computer.
• Third paradigm: it's a self-contained, persistent, asynchronous entity with org-wide tools and context, working alongside teams of humans.

Theo notes Karpathy got mocked on the timeline by people who assumed he was just "glazing a random Slackbot." Theo's reluctant verdict — having to defend Anthropic, make a video, and talk about Slack, all of which he dislikes — is that the underlying ideas are genuinely worth learning from. He flags up front that Claude Tag is only available on team and enterprise plans, paid at full token rates rather than subsidized subscription pricing.

What Claude Tag actually is

Claude Tag lets teams add Claude to Slack as a team member. You grant it access to selected channels and connect it to chosen tools, data, and code bases. Anyone in a channel can tag Claude and delegate tasks while they work on other things. It builds context by remembering relevant information from the channels it is in, and can plan future tasks. Anthropic claims tagging Claude is now one of the main ways they get work done, and that 65% of their product team's code is created by their internal version of Claude Tag.

Why the channel boundary is the key insight

Theo argues the two important pieces are "a way for teams to work" and "the channels it's in":

• Multiplayer. Within a channel there is one shared Claude that everyone interacts with; anyone can see what it's doing and pick up where the last person left off — more like a teammate than a single private chat. Theo notes few experiences offer this outside of PR-review bots, and that "if your multiplayer story is GitHub, you don't have a multiplayer story."
• Per-channel learning / context. Claude learns over time for the specific channel, not the whole company. It can pull from other channels and data sources with permission, but does not read from private channels. Theo calls this undersold: today with Claude Code your only context split is global vs. project-specific, which is a poor abstraction. Channels map far more naturally to how teams are actually structured. Because memory lives at the channel level, it doesn't matter how the monorepo or microservices are organized — two teams with different channels can have entirely different Claude experiences.

Theo's own experience: Hermes agents and OpenClaw

Theo's team has been building bespoke "Hermes" and "OpenClaw"-style agents for managing their Discord/content/sponsor/podcast workflows. To isolate context they end up spinning up separate containers (isolates), each its own deployment wired to a single channel — a lot of manual configuration that Claude Tag arrives at without the setup overhead.

His OpenClaw frustration was the single-thread context model (one running thread for iMessage/WhatsApp/Telegram/etc.), which constantly pruned and corrupted context — e.g. an 11am scheduled task firing into the same thread mid-conversation would break everything. He relegated OpenClaw to a single task (archiving YouTube/SoundCloud links to his NAS).

He prefers Discord threads per task. Example: a daily 11am job that scans the programmer-humor subreddit, picks the five most relevant top posts, and generates an HTML page (built via his "HTML plan" skill) with embedded image tags so he can save memes on his phone and repost to Twitter — all living in its own thread without polluting other work. He set this up conversationally (telling the agent to create the cron and later to switch to an HTML page), never touching a terminal.

His takeaway: most people won't realize they need different agent configurations per channel/purpose, and getting those boundaries right has been his burden to carry. The right primitives shouldn't require the user to reason about available tools, context, or boundaries — and building this as a channel-level primitive is clever and likely to become the norm.

Theo's main criticism: model lock-in

His objection is "not the Tag part, it's the Claude part." A strength of his Hermes setup is swapping models at will — he tested GLM-52, GPT-55 (a notable jump in capability over 54), Claude models (which required paying cash because his subscription couldn't be used for inference), and Fable. He values being able to make a setup more powerful just by switching models, and doesn't want to depend on a single lab.

He also uses cross-agent delegation: telling Codex (on GPT-55, whose API/SDK and UI work he finds weaker) to call `claude -p` for API design or UI tasks, or to get second opinions. Claude Tag offers none of this — you can't ask it to use a different model. So he frames the current options as a bad binary: full DIY customization (build your own Hermes agent, a Docker image per channel, all skills/context yourself, but you can swap models) versus Claude Tag (much works by default, but no control and no model choice). He's excited for other companies to build clones so teams aren't locked to one lab.

Other features and closing

• Ambient / "initiative" behavior (Anthropic's own scare quotes): if enabled, Claude proactively flags relevant info and follows up on quiet threads/tasks. Theo is lukewarm but sees value given how chaotic Slack is.
• Asynchronous + scheduling: you can delegate a task and walk away; Claude can schedule its own tasks and pursue a project over hours or days. You can also DM it (not his favored workflow).

Theo closes with Karpathy's own clarification: this isn't a crappy Slackbot or a claw, "it's an org-level harness," and the difference will become clearer over time. Running three companies, Theo says with "100% confidence" that Anthropic is going in the right direction. His advice for those unwilling to pay for an enterprise plan: set up OpenClaw or a Hermes agent in your own Discord or Slack to feel the same value.