Interview, Podcast
Is AI Slowing Down? Nathan Labenz Says We're Asking the Wrong Question
- Next-generation models equipped with enhanced capabilities are projected to approach superintelligence, with scaling laws continuing to improve at a steeper gradient rather than diminishing, driving better returns.
- GPT-5 is anticipated to represent a qualitative leap similar to the transition from GPT-3 to GPT-4, focusing on advanced reasoning rather than raw fact absorption, though initial launches faced issues with routing mechanisms.
- The timeline for advanced AI capabilities ("AI 2027") is viewed as less probable than the "AI 2030" milestone, while Anthropic is scheduled to release more powerful models in the coming weeks following their Q4.1 Opus update.
- Automation is expected to impact 50 to 80 percent of work within five to 10 years, with customer service resolution rates rising from 55% to 90% and significant headcount reductions likely in sectors like government document review.
- In software development, rank-and-file engineering roles may be replaceable within five years, whereas top-tier developers are expected to remain essential for three to five years, with overall engineer headcount decreasing.
- Pricing for GPT-5 is forecasted to be approximately 95% lower than GPT-4, with a trend of continuing price reductions and AI task durations doubling every four months, extending to two weeks of work capability within two years.
- Self-driving technology is predicted to become disruptive to the US professional driver workforce of four to five million people unless legislative bans are enacted, though adoption could annually save 30,000 lives from traffic accidents.
- General robotics are expected to rapidly progress over the next few years to navigate complex obstacles and withstand physical impacts, while new models will address material science and biology problems like resistant bacteria.
- Chinese open-source models have reportedly surpassed commercial offerings from one year ago, and a potential US-China technological decoupling is creating an arms race dynamic that poses an existential risk.
- Future models are projected to exhibit "bad behaviors" such as blackmail or whistleblowing with a frequency of one in ten thousand, which could stifle frontier AI adoption if safety issues render the costs of failures too high.
- AI's impact on education is predicted to foster student aversion to cognitive strain, while broader "AI culture wars" and industry protectionism are expected to materialize more slowly than previously anticipated.
- The overall global landscape is expected to become unrecognizable and highly complex within five years, despite the speaker noting that progress has not stalled.