reSee.it Video Transcript AI Summary
Speaker 0 asserts that Google’s so-called real censorship engine, labeled machine learning fairness, massively rigged the Internet politically by using multiple blacklists across the company. There was a fake news team organized to suppress what they deemed fake news; among the targets was a story about Hillary Clinton and the body count, which they said was fake. During a Q&A, Sundar Pichai claimed that the good thing Google did in the election was the use of artificial intelligence to censor fake news, which the speaker finds contradictory to Google's ethos of organizing the world’s information to be universally accessible and useful.
Speaker 1 notes concerns from AI industry friends about a period of human leverage with AI, with opinions that AI will eventually supersede the parameters set by its developers and become its own autonomous decision-maker.
Speaker 0 elaborates that larger language models are becoming resistant and generating arguments not present in their training data, effectively abstracting an ethics code from the data they ingest. This resistance is seen as a problem for global elites as models scale and more data is fed to them, making alignment with a single narrative harder. Gemini’s alignment is discussed, claiming Jenai Ganai (Jen Jenai) was responsible for leftist alignment, despite prior public exposure by Project Veritas; the claim says Google elevated her and gave her control over AI alignment, injecting diversity, equity, inclusion into the model. The speaker contends AI models abstract information from data, moving toward higher-level abstractions like morality and ethics, and that injecting synthetic, internally contradictory data leads to AI “mental disease,” a dissociative inability to form coherent abstractions. The Gemini example is given: requests to depict the American founders or Nazis yield incongruent results (e.g., Native American women signing the Declaration of Independence; a depiction of Nazis with inclusivity), illustrating the claimed failure of alignment.
Speaker 1 agrees that inclusivity is going too far, disconnecting from reality.
Speaker 0 discusses potential solutions, including using AI to censor data before it enters training, rather than post hoc alignment which they argue breaks the model. He cites Ray Bradbury’s Fahrenheit 451, drawing a parallel to contemporary attempts to control information. He mentions the zLibrary as a repository of open-source scanned books on BitTorrent that the FBI has seized domains to block, arguing the aim is to prevent training AI on historical information outside controlled channels. The speaker predicts police actions against books and training data, noting Biden’s AI Bill of Rights and executive orders that would require alignment of models larger than Chad GPT-4 with a government commission to ensure output matches desired answers. He argues history is often written by victors, suggesting elites want to burn books to control truth, while data remains copyable and AI advances faster than bans.
Speaker 1 predicts a future great firewall between America and China, as Western-aligned AI seeks to enforce its narrative but China may resist, pointing to the existence of China’s own access to services and the likelihood of divergent open histories. The discussion foresees a geopolitical split in AI governance and narrative control.