Marc Andreessen Says AGI is Here

Photo by Igor Omilaev on Unsplash

On April 5, 2026, Marc Andreessen posted a single line on X that instantly went viral: “I’m calling it. AGI is already here — it’s just not evenly distributed yet.”

The phrasing is deliberate. It’s a direct echo of William Gibson’s famous observation about the future.

I think Andreessen, who has bet billions on AI through Andreessen Horowitz, is not engaging in hype for hype’s sake this time. He has a techno-optimist worldview, and he’s probably issuing a deliberate provocation.

For years, AGI has been the holy grail… an AI that can perform any intellectual task a human can, at or above human level, across domains. Some argue we’re nowhere near it (well, the models still do hallucinate, lack persistent memory, and require human scaffolding for complex work).

Andreessen flips the script. Drawing from conversations in the AI community (including recent discussions with Replit CEO Amjad Masad), he points to “functional AGI”: systems that already automate the vast majority of economically valuable cognitive tasks, even if they aren’t “true” AGI in the philosophical sense.

Humans still supply expertise, direction, and oversight in many loops, but the core intelligence engine (LLMs and their agent harnesses) handles the heavy lifting.

Generative AI is, as we all know, already slashing the cost of productivity itself, much like PCs slashed compute costs and the internet slashed distribution costs. Repetitive workflows, document synthesis, code review, customer service triage, compliance checks, etc., are all being automated today.

Andreessen’s tweet escalates this: the intelligence layer isn’t “coming soon.” It’s operational in labs (early products, maybe?). Models now reason step-by-step, code autonomously for hours, orchestrate multi-tool agents, and generalize across domains in ways that would have seemed impossible two years ago. The bottleneck has shifted from raw capability to integration, cost, UX, and distribution.

AGI isn’t evenly distributed because access to it is gated. Top-tier models and agent frameworks require API credits, compute resources, prompt engineering expertise, and custom scaffolding. Insiders at frontier labs, big tech, and well-funded startups have it. Everyone else has to use ChatGPT and Gemini 3 with guardrails.

And if you were wondering, open-source models, especially those runnable locally on consumer or even prosumer hardware, remain far behind the current state-of-the-art (SOTA) proprietary models like ChatGPT (GPT-5.x series from OpenAI) and Claude (Claude 4 Opus/Sonnet from Anthropic) in 2026.

But why now? What is Andreessen’s broader thesis? I don’t think he’s making this call in a vacuum.

Well, look at the world today. Most developed countries are facing demographic collapse. Gen Z and millennials are trapped in a brutal, soul-crushing employment crisis. New workforce entrants saw their share of unemployment spike to a 37-year high in 2025, with levels still worse than during the depths of the Great Recession.

But I think Marc Andreessen is trying to say that these are temporary (and/or irrelevant) problems.

Slowing AI down, he argues, is the real existential risk, not the technology itself.

His worldview, laid out in essays like “Why AI Will Save the World” and the Techno-Optimist Manifesto, treats intelligence as the master resource driving all human progress. AI is meant to make high-IQ cognition nearly free.

In recent podcasts, Andreessen has noted that engineers may soon stop writing code manually, agents will handle routine cognition, and the real boom is in productizing these capabilities across every industry.

Makes sense. Steve Yegge’s Gas Town, for example, lets a single developer spin up fleets of 10–30+ AI coding agents working in parallel on the same codebase. Early users and contributors report dramatic productivity jumps on large, messy codebases… exactly the kind of economically valuable cognitive labor Andreessen highlights.

Another striking example of this momentum in agentic coding is Claw Code, the clean-room reimplementation of Anthropic’s proprietary Claude Code agent. Within hours, developers (notably led by figures like Sigrid Jin) produced a fully independent rewrite in Rust and Python that faithfully replicated Claude Code’s core agent harness, including its tool semantics, task management, terminal-native workflow, and multi-step reasoning patterns, all without copying any proprietary code.

It does make me believe that AGI is already here… at least in its functional, applied form.

Thanks for reading. Please consider following me here on Medium and on X.