The Turing Test Passed. Warp's CEO Says Nobody Seemed to Care

Elad Gil opens No Priors with Warp co-founder and CEO Zach Lloyd, a terminal-and-agents company for developers. Zach has a master's in the philosophy of science, worked on Google Sheets, and started a previous company called Self-Made. Gil wants the wide lens first. How should we think about this wave of AI, abstracting out of the coding world? Zach's answer is a distillation metaphor. People ask whether we are recreating people or recreating consciousness. He thinks that is the wrong question. What is happening is that we are distilling intelligence. The fascinating part is how much intelligence you can get out of next-token prediction, and what that does or does not say about minds. He does not think brains work that way. He expects a further unlock.

He recommends *Blindsight*, the Peter Watts novel that splits consciousness from intelligence. Humanity meets a spacefaring intelligent being that is not conscious, and has to figure out how to communicate with it. That, Zach says, is current AI: from an instrumental, functional perspective it does things we recognize as intelligence, and it is totally mechanistic. He does not think anyone looking at these models believes there is an aspect of consciousness in them. Gil notes that this is confusing, because the classical test was Turing's. If you cannot tell a computer from a person, the computer has met the intelligence bar. People are already having deep conversations with these systems about relationships and their own lives, and getting cogent answers. Zach agrees with the split. The Turing test has passed. What is crazy to him is that we just passed it and no one seemed to care.

So what is the next test for consciousness? He does not have a good answer. Chess, Go, video games: each time machines cleared a bar, we invented a new one and still said they were not conscious. Gil presses on the unfairness. If you really understand the mechanism, you may never credit a computer program as conscious, even though a human brain is, in his view, also running math, possibly matrix math and compounded functions, which is basically a neural net.

Memory is not a predicate for consciousness; people lose it and stay awake. Today's models are brought up, fed tokens, emit tokens, and shut down. They reason, they interpret language, they synthesize ideas. A close friend writing a philosophy dissertation now says talking to GPT-5 is better than talking to his professor. "GPT-5 gets it." Zach still does not credit consciousness, and thinks that is right. What would help is a feedback loop, a sensory experience, something more robot-like: cameras, live input, a reaction to the world. Even then, as long as we know what it is doing, we are unlikely to attribute true consciousness, which is not fair, and he does not actually know how we will know. Warp has already had a user who thought its AI was sentient and reacted strongly. Gil recalls the same thing happening at Google three or four years ago, with an internal chat model before ChatGPT launched.

Warp now calls itself an agentic development environment that grew out of the terminal. The basic concept is a platform for telling your computer what to do. You can still speak in terminal commands, the original product. You can also speak in English, and that launches an agent that can code, set up a project, or debug while a server is crashing. Zach calls it a horizontal, general-purpose interface. Competitors tend to be VS Code clones with an IDE-centric approach, or, like Claude Code, a pure text-based terminal app. Warp's advantage is that it is the outer graphical app. It still has the command line, and it can add editing features and a code-review interface because it controls the shell.

Gil cites the traction: close to a million monthly active users, and something like a million dollars in new revenue every seven to ten days. The inflection, Zach says, was moving into the coding market. For a long time Warp was known as the AI terminal, good at Docker and Git questions, but the action is in coding, and most development activity touches a codebase. Growth really started when they launched a strong coding agent, three or four months before this conversation.

He splits the market. Warp's target is professional developers building economically meaningful software, the kind of app that lands in a Mac dock or stays pinned as a Chrome tab, not the long tail of vibe-coded apps. It is awesome that anyone can code now, but most of the value in software, he thinks, still sits in a relatively small number of heavily used applications. He worked on one of those, Google Sheets. A good agent can already build a web app with relatively few prompts; applying agents successfully to professional codebases is much harder, and that is the side Warp chose.

Gil looks at the macro. ChatGPT launched in November 2022. At the time there were predictions that AI would take over everything within five years. The old saying in technology is that less happens than you think in three years and more happens than you think in five. Where are we in two or three years? Zach sees three phases. For most of his career, develop-by-hand: open an editor, find files, type code, go back to the terminal, type build commands. The switch underway is develop-by-prompt. He now starts most of his own coding tasks by prompting an agent, watching it work, steering as needed. The third phase is automated development, which he thinks is the bigger market and the reason people are excited: a server error comes in, a user ticket arrives, and something is handled without a human at the keyboard. Cognition is in that space. Cursor has background agents. Warp is moving into it.

The rate is not super clear to him. Recent model iterations were not, in his opinion, as big a step change as when Sonnet 4 came out. He expects a mix of interactive and automated development for a while. Within a couple of years, everyone will work by prompt, and some slice of tasks will run in the background, but he would be very surprised if it were everything. Will all coding become agents with a human giving high-level direction, like a product or engineering manager? Maybe. It would be silly not to build the infrastructure for that. He does not know the time frame. He does not think engineering expertise is about to be devalued. In the short term, knowing what you are doing as an engineer matters more than it ever has.

Think of the agents as junior engineers, Zach says. If nobody senior is watching, they write code that creates bugs, security issues, and a codebase that becomes unmaintainable. There is a premium right now on the skills that architect, review, and keep a system from degrading. If he were early in a CS career, he would be racing toward that expertise. The place you do not want to be is perpetually in the junior-engineer state. That position is at risk. Security tools become more important in the same breath. Companies like Socket and Snyk, anything that does automatic security analysis or verification, and languages like Rust that have stronger safety guarantees by default, all get more valuable when machines are producing code at a volume no human reviewer can read line by line. Whether those products stay standalone or get bundled into coding agents, he does not have a strong take. The underlying problem gets more important either way.

On bundling, he splits the market. Starting from the front end, Lovable, Bolt, Replit, maybe Figma Make, will become all-in-one platforms for building an app or even a business with payments, an evolution of Shopify, WordPress, or Squarespace. On the core professional-developer side he is less sure. One world is MCP and integrations, tools calling tools; the other is putting enough of it together that the experience wins. Warp is closer to the second: a single pane of glass for local and remote agents. MCP is valuable, he says, but not perfect, because you end up with secondhand data coming into the tools.

Gil offers the industry cycle: early fragmentation, later bundling, a technology shock that unbundles, then bundling again. Zach thinks some of today's vertical apps, agentic code review being the example, should live inside a holistic agentic development platform. If you have gone through the trouble of building an excellent coding agent, that agent should be reviewing code. It would be weird to plug in something else that has to relearn the context, the rules, and the conventions. Security checking, code review, anything CI-related: you probably do not want a bunch of different systems. The core technology is the harness around the model, plus the context sitting in that system.

Gil asks the question that keeps AI founders up. Historically the platform subsumes the biggest apps. Microsoft bundled the main Windows applications and later built Xbox because games were the other huge app. Google folded vertical search into the homepage. Foundation-model companies look like they are running that play on code: Anthropic has Claude Code, OpenAI almost bought Windsurf and has Codex, Microsoft has GitHub and Copilot. Zach thinks they are clearly trying. Seeing the most valuable applications on top of their tokens and moving aggressively into coding with Claude Code and Codex is, as a startup, a little bit scary. The question is whether they have the distribution advantage Microsoft had. If everyone is on Windows, or everyone comes through Google, adding the first-party app in place of the ecosystem is easy. He does not think that exact dynamic holds for coding right now. The front door is still a native app someone downloads, a terminal or an IDE. Controlling that is an interesting door. The other door, where all the source code lives, is GitHub, and he does not think they are executing that great. The current picture is almost comic: people run Claude Code inside Warp, and Warp is the outer app.

His other hope is competition at the model layer, until intelligent tokens become more of a commodity. Right now the labs have pricing power because there is a real delta between the frontier model and open source. If models get good enough that coding is, in his phrase, solved, you would not need the frontier model. Brand and scale would still help. That advantage is not as entrenched as being the actual front door the way Facebook, Google, or Windows were. Gil notes the consumer contrast. ChatGPT is already a default behavior. Even if Claude is better, everyone knows ChatGPT, and switching is hard. OpenAI's Dev Day the day before this taping showed apps inside ChatGPT, a platform play that will surely use the data to subsume the best third-party integrations. That is happening on consumer. Zach does not know that the same dynamic is there for developers.

Gil asks why the terminal, years ago, when they first talked. The insight that got Zach excited is simple. It is a daily-use tool for every developer, sitting next to the code editor. It had not changed much in 40 years. If you get good at it you can get a lot done, and it works across software development, not just writing code. On the flip side, from his perspective, it was not a good product: hard to learn, hard to use, hard to remember commands, super intimidating, with a gatekeeping vibe. The original idea was to build a better product there and see if people would like using it. The business concept evolved a lot. They started imagining a collaboration platform around the terminal and changed the model to an agent platform because the demand is way higher. The core insight that this tool matters has, if anything, been validated by how many agentic things are now terminal-first.

How close is the model layer to commoditizing on coding? He does not know. Increasingly the limit they see is context. Reasoning is impressive. Understanding an entire codebase, sources outside the code, or literally user intent are still hard. Warp is a big Anthropic partner. Going from Sonnet 4 to 4.5 was a few-percentage-point increase on SWE-bench for them, and they have invested enough to be one of the top agents on that benchmark. Going from 3.7 to 4 was a much more significant boost. GPT-5 felt like an upgrade, pretty much on par, with a different feel and higher latency, not as much of a step change as some jumps before. What excites him next is not only the interactive agents most people use today, but programming against those agents. A headless Warp in CI could, every time someone updates the code, keep the documentation in sync, the kind of work developers hate. From a business perspective, automation is a better place to sit than productivity enhancement. Proving ROI on coding tools is hard; studies that deploy them on real codebases have been unclear about the impact. Outcome-oriented automation is easier to prove, and it is not limited by time spent behind a keyboard.