Interview
Gwern — Anonymous writer who predicted AI trajectory on $12K/year salary
- Human executives are projected to eventually oversee AI firms and outcompete entirely AI-owned corporations by providing long-term, non-myopic vision capabilities that individual AI models lack.
- The evolutionary unit of selection for AI is expected to shift from individual models to "department packages" (e.g., programmer, manager, legal) once models become perfectly replicable, with future systems evolving through random variations of these units to select for high performance.
- Timeline predictions for general-purpose intelligence include a generalist system in 2019, humanish agents by 2025, AGI by 2030 according to Shane Legg, and a personal planning target of 2028 for AGI onset.
- AI writing capabilities are forecasted to reach "Gordon-quality" standards within two to three years, while autonomous agency is expected to be learned more easily in the future despite a current lack of training data for agentic workflows.
- Intelligence is theorized to be fundamentally a search over Turing machines where human variation is attributed solely to differences in compute power, making general-purpose intelligence rare in evolution because hard-coding genetic solutions is more efficient than evolving learning mechanisms.
- Human-level intelligence is expected to evolve more slowly than previously anticipated due to the high computational cost of search processes compared to hard-coded genetic solutions.
- Future superhuman intelligences will likely recover stable, long-term characteristics from digital traces while remaining unable to access autobiographical or transient information, and they will be unable to make personal preference judgments for humans.
- Cognitive diversity in AI models is expected to remain higher than human diversity when excluding capability, though this diversity has recently decreased within LLM families due to heavy tuning.
- Independent writers face a risk of spiraling into "crankdom" if they fail to let go of arguments and resentments regarding online incorrectness.
- Financial incentives of $50,000 to $100,000 annually are expected to facilitate a move to San Francisco to offset living costs.
- Historical records are predicted to confirm that environmental factors like lead and plastics caused significant harm, with negative effects manifesting more prominently in obesity than in intelligence based on recent time-series data.
- Major unanswered questions regarding human biology, such as the reasons for sleep, aging, and sexual reproduction, are targeted to be answered by the year 2050.
- The singularity scenario of autonomous machine life was first described in 1863, and historical technological progress is characterized as periods of civilization destruction and reinvention rather than continuous acceleration.