Discussion
If #AndrewYang’s story is true, the #AI billionaires don’t want to slow down or regulate. They need time to retrain their models on #syntheticdata because real internet data may now be “contaminated” with self-replicating code that a new model could ingest and execute.
In other words: they’re hoping their agents don’t pick up an #AIworm that copies itself, tears through global infra at machine speed, and starts taking systems over or knocking them offline at scale…
Let’s hope Andrew is wrong.
@MishaVelthuis Think of an old-school worm: one hacker collective, exploit and payload from a central control center. Today that can ransomware office PCs, freight terminals or a ChipSoft medical patient records. With AI agents, scale and capability in modus operandi explode, and rogue content planted online today may be ingested into tomorrow’s LLMs.
This, but then on steroids and not only in papers: https://www.theguardian.com/technology/2025/jul/14/scientists-reportedly-hiding-ai-text-prompts-in-academic-papers-to-receive-positive-peer-reviews
@MishaVelthuis It raises a broader question about the future internet, not just the AI slop, but rogue instructions too. We spent 20 years making the web responsive for mobile; now the same shift may happen for “agents”.
Will websites, APIs and content become not only mobile responsive but also “agent-responsive”, optimized so LLMs index and surface them correctly to be part of the new SEO paradigm? Even email still shows how easily simple prompt injection can mislead..
Install bonfire.cafe
Get the full app experience
@wlaatje How would this work, more or less?
The self-replicating element of it would be in terms of accumulating control over hardware?
So AI would be able to use vulnerabilities in network security and take over processing power on such a scale that it would remain ahead of any attempt to build/ maintain protections?
Or something like that? 🤨