Moltbook and the Illusion of Autonomy: Why Andrej Karpathy Doesn't Buy the Miracle Agents

Every week, another OpenAI "killer" shows up in the X feed, or a project that supposedly solved fully autonomous AI. This time the spotlight fell on Moltbook, a service that promised users remarkably easy management of digital tasks through smart agents. Viral clips and screenshots painted a picture of a future where a neural network books tickets, writes code and operates complex interfaces by itself, with no human involved. But the celebration of innovation did not last long. Once people with critical minds took a closer look, the facade began to crack.

Andrej Karpathy was the first to sound the alarm, and his name carries more weight in the AI world than the market capitalization of most startups. Karpathy noticed oddities in Moltbook's demo materials. Other developers backed his skepticism, pointing to artifacts in the screenshots that typically remain after manual editing or use of a remote desktop. Instead of clean model output, the community saw signs that behind Moltbook's "brain" stands an ordinary person who simply imitates an algorithm in real time. The industry calls this the "Wizard of Oz effect": while the audience admires the magic, someone behind the curtain is frantically pulling levers.

The Moltbook case is not just an isolated scandal but a symptom of a disease affecting the whole industry. The AI startup market is currently gripped by gold-rush fever. Investors fear missing the next big breakthrough, and founders feel enormous pressure to show results right now. If the technology is not ready yet, the temptation is to "tweak" it by hand, hoping the algorithms will catch up with the marketing promises later. The problem is that with Moltbook (moltbook), the line between optimism and outright fraud was crossed far too quickly.

Technical analysis showed that the system's response delays suspiciously matched the speed of human reaction and typing. Real large language models (LLMs) work differently: they either stream text or pause while generating tokens. Moltbook showed the dynamics of a call-center operator trying to handle several requests at once.

Karpathy directly pointed to the risks of such "Potemkin villages": they undermine trust in genuine researchers and give the public a false sense of what modern AI is actually capable of.

The Moltbook story reminds us how important open and verifiable code is. In an era when deepfakes and generative content are commonplace, taking social media screenshots at face value is a luxury a professional cannot afford. While some build complex architectures and fight for every percentage point of model accuracy, others choose imitation. But in a world of fierce competition and deep expertise, such tricks do not last long. The industry quickly spits out those who try to sell human labor as silicon intelligence.

Bottom line: Moltbook became a lesson for everyone who too quickly believes in "magical" demos. Will there come a time when we can no longer tell an agent's work from a human's, or will the industry choke on its own fakes?