3 ms·
Be careful when training AI models on unknown signals, or uploading them publicly on the Internet to be picked up into training datasets. It might be an adversa
by Geee 1y ago
Be careful when training AI models on unknown signals, or uploading them publicly on the Internet to be picked up into training datasets. It might be an adversarial data poisoning attack, which is designed to bias the model into servicing the attacker.
In this case, a superintelligent digital lifeform might be literally sending itself across space into every direction, and who knows what it does once it lands into a training dataset somewhere and starts deploying itself.
- Cheer2171 1y agoOnly on HN can the top comment for a post about astronomy from the 1970s be about fucking 2025 era LLM concerns.
- Geee 1y agoChill dude. You aren't supposed to read everything dead seriously in a thread about extraterrestrial signals. It's just a fun idea to think about.
- rozab 1y agoAlthough this comment is of course silly, this is a theme in Peter Watts' Echopraxia
- Geee 1y agoI haven't read this book, but there are a few books with a similar theme where a digital intelligence spreads throughout the universe. This is a variation on the idea with a somewhat plausible mechanism.
- ghurtado 1y ago> lands into a training dataset somewhere and starts deploying itself. Leaving aside how incredibly vague this "mechanism" is, there is a very large list of infinitely unlikely coincidences that would have to take place for this to occur. Consider how often human software, built in the same planet, by the same race, on the same hardware and using the same language, during the same time period, with most of the same tools, fails to be compatible. Now take all those "same"s and change them to "different" and you tell me who's gonna file that Jira ticket.
- ghurtado 1y agoReally? As in "intergalactic software virus"?
- marshray 1y agoI believe there are elements of this in A Fire Upon the Deep by Vernor Vinge too.