A new analysis shows that as few as 250 malicious documents, totaling roughly 420,000 tokens, can reliably trigger gibberish across multiple model sizes, underscoring acute robustness risks and complicating detection claims. At the same time, a proposed $25 billion data center and fresh warnings of an AI stock bubble highlight a race to scale and consolidate even as business models remain unproven. These tensions over resilience, economics, and public trust will shape which capabilities endure.
Reddit
#data poisoning
#large language models
#ai infrastructure
#ai bubble
#model robustness
