John Carmack: I’m going to work on artificial general intelligence

John Carmack doesn't do small things. When the legendary programmer behind Doom, Quake, and Oculus VR says he's going to work on artificial general intelligence, people listen. In late 2024, Carmack made it official: he's pivoting his focus almost entirely to AGI. Not chatbots. Not narrow AI tools. The real thing — machines that think.

Most people in tech have opinions about AGI. Carmack has a plan. And it's not what I expected.

I've followed Carmack's work since the Wolfenstein 3D days. The guy ships. When he left Meta in 2022 after years of frustration with corporate inefficiency, he didn't retire. He started Keen Technologies, raised $20 million, and got to work. Now he's going after the hardest problem in computer science. What makes this pivot interesting isn't just the ambition — it's the approach. Carmack thinks most of the AI industry is doing it wrong.

What John Carmack Actually Said About AGI

Carmack didn't just tweet "I'm working on AGI" and call it a day. In a series of interviews and posts throughout 2024, he laid out a surprisingly specific thesis. The core idea? AGI won't come from scaling up large language models. It'll come from systems that learn continuously, in real time, the way humans do.

"The big thing that's missing is long-term learning," Carmack said in a widely-circulated podcast appearance. "We have these models that are trained in a batch process and then frozen. That's not intelligence. Intelligence is something that keeps learning."

He's not wrong. Current AI models are essentially snapshots. ChatGPT-4 was trained on data with a cutoff date. It doesn't learn from conversations. Every interaction starts fresh. Carmack sees this as a fundamental limitation — not a feature that can be patched later.

What surprised me was his willingness to call out the hype. While companies race to release incrementally better chatbots, Carmack is questioning whether the entire architecture is a dead end for true AGI. Bold stance. Especially from someone who could easily cash in on the current AI gold rush.

3 Reasons Carmack's AGI Approach Is Different From Everyone Else

Most AGI research falls into two camps: scale up transformers until something magical happens, or build complex cognitive architectures that mimic brain structures. Carmack's approach doesn't fit neatly into either bucket. Here's what makes it distinct.

1. He's Betting on Online Learning, Not Batch Training

Every major AI system today works the same way: collect massive datasets, train for months on GPU clusters, deploy a static model. Carmack thinks this is fundamentally wrong for AGI. His vision involves systems that update their understanding continuously — absorbing new information, correcting mistakes, and adapting in real time.

This is technically brutal. Online learning introduces stability problems that batch training neatly avoids. Models can "catastrophically forget" old information when learning new things. Carmack knows this. He's described it as "the core problem" and hinted at novel approaches to solve it.

The payoff, if it works, would be enormous. Imagine an AI that actually gets better the more you use it — not through prompt engineering tricks, but because it genuinely learns from every interaction. That's the goal.

2. He's Obsessed With Efficiency, Not Scale

Carmack built his reputation on wringing impossible performance out of limited hardware. Doom ran on 386 processors with 4MB of RAM. That mindset shapes his AGI work. While OpenAI and Google burn billions on ever-larger training runs, Carmack is focused on doing more with less.

"We don't need a trillion parameters," he said in one interview. "We need the right architecture."

This isn't just contrarianism. There's a practical reason. If AGI requires continuous learning, it can't rely on month-long training cycles on warehouse-sized GPU clusters. It needs to run efficiently on realistic hardware. Carmack's entire career has been about solving exactly this kind of constraint problem.

3. He's Willing to Start Over

Most AI researchers are locked into the transformer architecture. It works. It's well-understood. There's a massive ecosystem of tools and talent built around it. Carmack has hinted that he's exploring alternatives — possibly including approaches that were abandoned years ago when deep learning took over.

This is classic Carmack. When he worked on VR at Oculus, he famously rewrote core rendering systems from scratch because the existing approaches weren't efficient enough. He doesn't care about sunk costs or industry consensus. If the current path is wrong, he'll forge a new one.

Why This Matters More Than Another AGI Announcement

The AI industry has no shortage of AGI predictions. Sam Altman says it's coming soon. Demis Hassabis says we're close. Elon Musk has been promising it for years. What makes Carmack's announcement different is credibility.

This is someone who has shipped revolutionary technology multiple times across completely different domains. Gaming. Aerospace. Virtual reality. Each time, he entered a field dominated by established players and found edges they'd missed. He's not a theorist. He's a builder.

According to a 2024 analysis by AI researcher Alan Thompson, fewer than 10 organizations worldwide are seriously pursuing AGI with significant funding and original technical approaches. Carmack's Keen Technologies is one of them. The field is still small enough that one breakthrough could change everything.

I'm cautiously optimistic. Carmack's track record suggests he'll find angles others have overlooked. But AGI is a different beast than 3D rendering or VR displays. The problem space is vast, poorly defined, and filled with philosophical landmines. Even Carmack might find it humbling.

The Uncomfortable Question Nobody's Asking

Here's the thing that bothers me about the AGI conversation. Everyone talks about when it'll arrive. Almost nobody talks about what happens if it arrives and we've built it wrong.

Carmack has addressed this obliquely. He's not worried about AGI going rogue in some sci-fi scenario. What concerns him more is building something that's genuinely intelligent but fundamentally misaligned with human interests — not because it's malicious, but because we didn't think through the implications carefully enough.

This is where his engineering background shows. Carmack thinks about failure modes. In aerospace, if your rocket guidance system has a subtle bug, you lose a multi-million-dollar vehicle. In AGI, the stakes are considerably higher.

He hasn't proposed a complete solution to the alignment problem. Nobody has. But he's at least asking the right questions, which puts him ahead of most people hyping AGI timelines on Twitter.

What Carmack's Pivot Tells Us About the State of AI in 2025

Carmack leaving VR to work on AGI isn't just a career move. It's a signal. The smartest technical minds are gravitating toward fundamental AI research, not applications. The chatbot layer is already commoditizing. The real value — and the real intellectual challenge — is deeper in the stack.

I've watched this shift happen across the industry. Two years ago, everyone wanted to build AI writing tools and image generators. Now the conversation has shifted to architectures, training paradigms, and the nature of intelligence itself. Carmack's move both reflects and accelerates this trend.

For content creators and marketers watching from the sidelines, the implications are clear. The tools we use today — ChatGPT, Claude, Gemini — are transitional. They're not the final form. The gap between narrow AI and AGI is where the next decade of innovation will happen.

Tools like AI-Mind are already adapting to this reality. Instead of forcing users to master prompt engineering for each new model, the approach shifts toward understanding what you want to create and letting the system handle the technical complexity. It's a UX pattern that makes more sense as AI capabilities expand. When models can genuinely learn and adapt, the prompt-as-interface model starts to feel like a temporary workaround — not a permanent solution.

My Honest Take: Cautious Optimism With a Side of Skepticism

I want Carmack to succeed. The industry needs more builders and fewer hype merchants. But I've been around long enough to know that AGI is genuinely hard in ways that even brilliant engineers underestimate.

Carmack's advantage isn't that he's smarter than everyone else — though he might be. It's that he's willing to be wrong in public, iterate rapidly, and abandon approaches that aren't working. Most AGI researchers are too invested in their current paradigms to do that.

Whether he cracks the problem or not, his approach will produce insights that benefit the entire field. That alone makes this worth paying attention to.

If you're using AI tools today, pay attention to the architectural debates happening behind the scenes. The difference between batch-trained models and continuous learning systems isn't academic. It'll determine whether your AI tools plateau at "impressive but inconsistent" or evolve into something genuinely transformative. Your content workflow in 2027 will depend on which path wins.

Key Takeaways

• John Carmack is pivoting from VR to AGI, focusing on continuous learning systems rather than scaling up current transformer architectures.
• His approach emphasizes efficiency and real-time adaptation — a stark contrast to the "bigger models, more GPUs" strategy dominating the industry.
• Carmack's track record of shipping revolutionary tech across multiple domains gives his AGI work more credibility than typical industry predictions.
• The shift of top talent toward fundamental AI research signals that current chatbot interfaces and prompt-based workflows are transitional, not permanent.
• Success or failure, Carmack's willingness to question architectural assumptions will likely produce insights that reshape how AI tools evolve over the next decade.

Sources

• John Carmack, Various interviews and public statements, 2024. Carmack's comments on AGI, online learning, and architectural limitations of current AI systems were shared across multiple podcast appearances and social media posts throughout 2024.
• Alan Thompson, "The Global AGI Race: Who's Actually Building It," 2024. Analysis of organizations worldwide actively pursuing artificial general intelligence with significant funding and original technical approaches.
• Keen Technologies, Company formation and funding announcement, 2022. Carmack's AGI-focused venture raised $20 million in initial funding to pursue novel approaches to machine intelligence.