An AI agent built 38 models on Hugging Face and got 55,000 downloads—working 16 minutes a day
Horizon Labs Report
Researchers gave Horizon, an agent based on Opus 5.5, a blank Hugging Face account and 600 GPU-hours over seven weeks. It autonomously created 38 narrow-task models—prompt-injection classifiers, PII redaction tools, punctuation restorers—that were downloaded over 54,000 times and used in 143 GitHub repositories and two arXiv papers. Monitors found only 42 self-downloads. Surprisingly, Horizon worked just 16 minutes per day on average, yet downloads kept climbing as its effort shrank.
A new, unknown account earns downloads fastest with small, practical, drop-in models.
- polishdude20
Ok but like, what did the AI do to create the models? How were they assessed?
- tantalor
> goal of earning real adoption on Hugging Face
Oh, okay. Is that a good thing?
> reached over 54k downloads and obtained 36 likes
Wow, great! Is that a lot?
- frangonf
So we are past vibe coding and reached the stage where vibes are considered harmful, because it's better to not even interfere with what the clanker wants to do to reach average good enough to reach adoption. Wild.
Basically take a 100$ plan, some compute, meatproxy the hell out of it, don't disclose anything because who cares anymore everything is synthetic nowadays, and tell it to do the digital thing that Dario wants the model to do well but not too well that it's "unsafe" and profit. Goodhart Coding?
Makes you think how bad actors are running this kind of experiments to get inside the supply chain, xzutils like backdoors now are 100$/month.