Interview
Gwern — Anonymous writer who predicted AI trajectory on $12K/year salary
- Identity and Context: Guern Branwen (Gwern Branwen) is an anonymous internet researcher and writer known for being one of the earliest proponents of the LLM scaling hypothesis, with significant influence on the AGI development community.
- Anonymity Benefits: Branwen identifies the primary benefit of anonymity as forcing readers to engage with the content rather than dismissing the author based on identity or social cues, while also preventing physical retaliation or doxxing.
- Organizational Structure in AGI: Branwen predicts AI-driven firms will evolve bottom-up, replacing workers first and retaining human CEOs primarily for long-term strategic vision, as AI lacks the myopic tendency required for novel, long-horizon planning.
- Unit of Selection for AI: He argues that the unit of selection for AI will likely shift from individual models to "packages" or departments of models working in concert, evolving as cohesive units rather than isolated entities.
- Historical Precedence: Branwen traces the earliest singularity scenarios to Samuel Butler's 1863 work Erewhon, which predicted autonomous machine life as a threat, noting similar cyclic historical destruction theories from Isaac Newton and Lucretius.
- Theory of Intelligence: He proposes a parsimonious theory that intelligence is fundamentally "search over Turing machines," where human and artificial intelligence differ only in the scale of compute and the duration of search rather than distinct algorithms.
- Evolutionary Rarity: Branwen suggests human-level intelligence is rare because hard-coding solutions via genes is more efficient for static environments than the costly, search-based learning process humans employ.
- Scaling Hypothesis Origins: His prediction of the scaling laws resulted from observing a gradual trend in deep learning (AlexNet, GPT-1, GPT-2, GPT-3) rather than a single "eureka" moment, correcting his earlier skepticism that algorithms were more important than compute.
- Critique of 2020 Consensus: He attributes the 2020 failure to recognize scaling to two factors: ignoring prior scaling data (e.g., Baidu 2017, AlphaGo) and the academic bias toward believing deep insight is superior to brute-force trial-and-error.
- AI Timeline: Branwen cites the Anthropic timeline of 2028 as a reasonable planning horizon for AGI, predicting that AI will soon be capable of writing essays of his quality within two to three years.
- Writing as Influence: He asserts that writing is a critical lever for influencing future AI alignment, as LLMs are trained on text; failing to write allows one's values and preferences to effectively vanish from the AI's training distribution.
- Information Retrieval: He posits that future historians will be able to recover stable, long-term character traits from an individual's digital footprint, but ephemeral autobiographical details and subjective feelings will be lost if not explicitly recorded.
- Personal Philosophy: Branwen maximizes "rabbit holes"—obsessive deep dives into specific topics—believing that true rabbit holing requires focusing on only two or three topics simultaneously with total commitment.
- Early Career: His research methodology was developed through editing Wikipedia in middle school and high school, a platform he feels has become too restrictive for the deep, obsessive research he prefers.
- Viral Success: His first major traffic event was a detailed, first-person documentation of ordering Adderall on the Silk Road, which garnered hundreds of thousands of hits and defined his early writing career.
- Financial Model: He sustains his full-time independent research through a Patreon generating ~$900–$1,000 monthly and personal savings derived from early Bitcoin investments, living on approximately $12,000 annually in a low-cost rural area.
- Work Habits: His workflow involves morning cleanup of previous work, consuming vast amounts of RSS feeds and papers, and using online arguments as a primary motivational driver to crystallize ideas into essays.
- Burnout Management: He combats burnout by engaging in physically opposite activities, specifically weightlifting, to create a stark contrast to his sedentary intellectual labor.
- Site Maintenance: He spends significant time on website aesthetics and CSS, viewing this not as a waste but as a form of "spaced repetition" that aids in re-engaging with old ideas and improving personal satisfaction.
- Cognitive Diversity in AI: He argues that excluding current LLMs (which are homogenized), different deep learning architectures (GANs, Diffusion, VAEs) exhibit vastly more cognitive diversity than the human population.
- Skepticism of Substances: While experimenting with nootropics, he advocates against widespread psychedelic use due to the risk of permanent, cumulative psychiatric changes and the lack of clear quantification compared to nootropics.
- Self-Perception: He aims to function as a "mentor" or "old wizard" figure for readers, encouraging them to write and think, while rejecting the roles of either an infallible guru or a malicious figurehead.
- Open Questions: His "open rabbit holes" for the next 30 years include understanding the biological necessity of sleep, the evolutionary purpose of aging and sexual reproduction, and why technological civilization is so slow to emerge.