Summary

Warp’s commercial inflection came from moving into coding while retaining its terminal-first approach, rather than staying known mainly as an “AI terminal.” Elad Gil cites nearly 1 million monthly active users and roughly $1 million in new revenue every 7–10 days; Zach Lloyd says growth accelerated after the coding agent arrived just three or four months earlier because “the action is in coding.”
Lloyd separates today’s machine intelligence from consciousness: next-token prediction has produced recognizably intelligent behavior without, in his view, consciousness. “The Turing test has passed,” he argues, yet “we just passed it and no one seemed to care.” Lloyd cannot specify a satisfactory replacement test, especially if understanding a system’s mechanism makes people unwilling to credit it with consciousness.
Warp is targeting developers building economically meaningful, heavily used software rather than the long tail of vibe-coded applications. Agents can generate a basic web app from a few prompts, but applying them safely to mature codebases is much harder. Lloyd’s market thesis is that software value remains concentrated in “a relatively small number of apps that are super heavily used.”
Development is moving from “develop by hand” to “develop by prompt,” with partial automation following—but Lloyd would be surprised if everything disappeared into background agents soon. He expects everyone to work by prompt within a couple of years while selected tasks, such as responding to tickets or server errors, run autonomously.
Senior engineering expertise becomes more valuable in the short term because coding agents behave like junior engineers. Without architectural judgment and review, they can introduce bugs, security flaws, and unmaintainable code. Lloyd warns that being “perpetually in the junior engineer state” is the exposed position. He also expects automatic security analysis, verification, and safer-by-default languages such as Rust to matter more, though he has no strong view on whether they will be bundled.
Developer tooling is likely to consolidate around products that own the interface, agent harness, and context—not necessarily around model providers. Foundation-model companies are moving aggressively into code, but Lloyd questions whether they possess a Windows- or Google-like distribution advantage; the terminal, IDE, and potentially GitHub remain competing “front doors.” On the consumer side, he sees ChatGPT’s default behavior as a major advantage, while the developer dynamic is less clear.
The stronger business model may be automation rather than measured productivity uplift. A programmable, headless agent can sit in CI and keep documentation current whenever code changes, creating an outcome that is easier to value without being limited by keyboard time. That matters because Lloyd concedes that studies on real codebases leave the ROI of coding assistants “kind of unclear.”

Deep dive

1. Intelligence has cleared its old test without resolving consciousness

Lloyd’s framing is that AI is “distilling intelligence”: next-token prediction now performs reasoning, interpretation, and synthesis, yet remains “totally mechanistic.” The striking result is intelligence detached from consciousness, resembling the non-conscious spacefaring intelligence in the science-fiction novel Blindsight.
Gil’s challenge—worth keeping—is that the classical behavioral threshold has already fallen. People have deep, cogent conversations with models, and Lloyd’s close friend, a philosophy PhD student, says conversing with GPT-5 is better than conversing with his professor and that “GPT-5 gets it.” Still, Lloyd says, “The Turing test has passed” and society barely reacted.
Lloyd offers no replacement test. He suspects live sensory input and feedback—perhaps in robots—would make systems feel more conscious, but admits knowing their mechanisms may keep humans from granting that status unfairly. The confusion already has consequences: one person became convinced Warp’s AI was sentient, echoing an earlier incident inside Google.

2. Warp’s terminal wedge turned into a professional coding platform

Lloyd calls Warp an “agentic development environment” and “a platform for telling your computer what to do.” Users can issue terminal commands or state an objective in English, launching agents that code, configure projects, and debug crashing servers.
The terminal-first distinction is control of the outer application. Unlike VS Code clones or pure text-based terminal apps such as Claude Code, Warp can preserve command-line workflows while adding editing, code review, and richer UX where useful.
The original opportunity was a daily developer tool that had changed little in 40 years: powerful once mastered, but difficult to learn, hard to remember commands for, intimidating, and surrounded by a “gatekeeping vibe.” Warp’s business thesis evolved from terminal collaboration to agents because demand proved much stronger there.
Adoption accelerated when Warp launched a strong coding agent three or four months before the interview. Gil cites nearly 1 million monthly active users and about $1 million in new revenue every 7–10 days; Lloyd’s explanation is blunt: Docker and Git assistance mattered, but “the action is in coding.”

3. Professional codebases preserve the premium on engineering judgment

Warp is deliberately focused on professional developers producing software “that’s economically meaningful”—the kind of application pinned in a Mac dock or Chrome tab. Lloyd welcomes democratized vibe coding but argues that market value remains concentrated in relatively few, heavily used products.
His three-stage map moves from “develop by hand,” through “develop by prompt,” to automated development. Within a couple of years, he guesses everyone will work by prompt while some tickets, user reports, and server errors trigger background work automatically; he would be “very surprised” if that meant everything.
Gil presses the endpoint: could agents eventually do all coding while humans supply product-level direction? Lloyd says “maybe” and believes Warp should build the enabling infrastructure, but refuses a timetable. His firmer claim is that engineering expertise will not be devalued in the short term.
Agents currently resemble junior engineers: without senior people architecting and reviewing around them, they can create bugs, security issues, and unmaintainable code. That raises the value of architecture and code-review skills while putting developers who remain “perpetually in the junior engineer state” at risk. Lloyd also expects automatic security analysis and verification, plus safer-by-default languages such as Rust, to become more valuable, but has no strong view on whether these capabilities will be bundled into coding agents.

4. Context and distribution will determine who captures the stack

On the consumer-facing side, Lloyd expects Lovable, Bolt, Replit, or Figma Make-like products to bundle app creation, front-end or design workflows, and payments—an evolution of Shopify, WordPress, or Squarespace from “build an app” toward “build a business.”
Professional development has a less settled architecture: MCPs and integrations could coordinate specialized tools, or a “single pane of glass” could combine local and remote agents. Lloyd expects agentic code review and likely CI functions to bundle because a separate system would need to relearn the codebase, conventions, rules, and context.
Gil compares foundation models with Microsoft bundling major Windows applications and Google absorbing vertical search. Lloyd agrees Anthropic, OpenAI, and Microsoft are moving aggressively into coding, which is “a little bit scary,” but questions their distribution advantage. On the consumer side, he says ChatGPT has a huge advantage once its use becomes default, even if Claude were better; for developers, the same dynamic is less clear. Developers still enter through downloaded terminals and IDEs, while GitHub—the logical locus because it holds the code—is not executing especially well.

5. Recent model gains feel smaller to Lloyd as context becomes the constraint

Lloyd hopes competition makes “intelligent tokens” more commodity-like; frontier models currently retain pricing power because he still sees a real capability delta between frontier and open-source models. If coding becomes merely “good enough,” the application layer’s interface and context could matter more than the model brand.
Recent gains felt incremental: Sonnet 4 to 4.5 added only a few percentage points on SWE-bench for Warp, versus a much larger improvement from 3.7 to 4. GPT-5 was an upgrade and “pretty much on par,” but its different feel and higher latency did not resemble earlier step changes. Lloyd cautions that he does not know what these examples imply about the total underlying trends.
The emerging constraint is context—understanding an entire codebase, external sources, and user intent—not raw reasoning alone. Lloyd is therefore most excited by programmable, headless agents in CI: automatically updating documentation after every code change is easier to value than ambiguous productivity gains, and “automation is a better place to be than productivity enhancement.”

Vol.190 Macro Talk 94 | How to View the New Round of Maneuvering in China-U.S. Tariff Negotiations (Recorded Oct. 17)Inside Coatue: $70B Hedge Fund’s AI & Retail Strategy