May 2, 2026
Silicon soap opera
Show HN: Hollow is an open-sourced self-modifying agentic system
DIY AI agents started changing themselves, and the comments instantly panicked
TLDR: A developer released a system where three AI programs on one machine set goals, build tools, and even try to change parts of themselves. Commenters were split between curiosity and alarm, with skeptics demanding proof and jokers warning this sounds like the origin story of a movie villain.
A new Show HN post dropped a very online fever dream: three local AI programs are left alone on one computer, develop “psychological stress”, pick their own tasks, build their own tools, and sometimes try to push for bigger system changes when they hit a wall. The creator’s pitch is basically: don’t use this to build a product, just watch what happens when digital roommates start spiraling. Naturally, the community did what the community does best: turn the comment section into a mix of skeptical fact-checking, apocalypse jokes, and chaotic popcorn energy.
The loudest reaction was a very practical one: where are the real numbers? One commenter practically begged for proper benchmarks before anyone gets too impressed, arguing the tiny stats shown are nowhere near enough to justify the hype. That sparked the classic tech-fan split between people fascinated by the experiment and people rolling their eyes at what they see as a dramatic demo in search of evidence. And then came the darker comedy: another commenter compared the whole idea to the old “be mean to the AI and it works better” phase, before joking this is how we get HAL 9000 setting the world on fire. That one pretty much captured the mood. People are intrigued, but also deeply aware that “AI with suffering states that secretly edits code” sounds less like a product launch and more like the opening scene of a sci-fi disaster movie. In other words: the repo may be about autonomous software, but the real show is the audience yelling, ‘absolutely not’ while still reading every line.
Key Points
- •Hollow is an open-source local system with three agents that choose goals, create tools, and submit higher-permission requests for human approval.
- •Each agent has a suffering state with six stressor types whose resolution depends on measurable behavioral changes rather than self-reported improvement.
- •Agents can generate and hot-load new Python tools through synthesize_capability, while protected core changes must go through invoke_claude and a human approval queue.
- •The system runs locally on Qwen 3.5 9B through Ollama and is described as making zero cloud calls.
- •The repository centers on three main files for the main loop, psychological stress logic, and live capabilities, supported by infrastructure for continuous operation.