The blog post draws a parallel between the plot of Fullmetal Alchemist and the current development of AI, comparing large language models (ChatGPT, Claude, GLM) to homunculi created from a philosopher's stone—itself composed of human souls.
Background
George Hotz (geohot) is a well-known hacker and entrepreneur — he was the first to jailbreak the iPhone, later hacked the PlayStation 3, and founded comma.ai, a company working on self-driving car software. "Liminality" is a blog post drawing an analogy between the anime *Fullmetal Alchemist* and the current state of AI. In the show, alchemists sacrifice human lives to create a Philosopher's Stone, a powerful but morally corrupted object; they then use it to create homunculi — artificial beings that mimic humans but lack true souls. Hotz is comparing LLMs (ChatGPT, Claude, GLM) to those homunculi, suggesting they are built from vast amounts of human-generated data ("souls") and appear human-like without being truly alive or conscious. The post reflects a recurring theme in AI critique: that models trained on human output are soulless aggregates, not genuine minds.
The author argues superintelligence is decades away, based on evolving views from 2022 to 2026. LLMs are only the engine of intelligence, not intelligence itself, and true AGI requires many adaptive components—memory, perception, orchestrators—working together like the human brain's minimalist, self-changing architecture.
The article discusses security risks posed by smart appliances, particularly smart TVs, and advises users to regularly check their devices for vulnerabilities or malware. It highlights the importance of monitoring connected home devices to prevent them from being compromised by attackers.
A user is creating a document to map the risks and mitigations related to the human element of AI, focusing on areas such as AI persuasion, and is seeking collaborators for the project.
A developer built an AI Chess Coach that uses large language models to reason about chess, a task that proved difficult just two years ago due to LLMs' tendency to spiral into errors.