Keep pulling the thread on Nathan Labenz.
AI models from multiple companies have recently achieved gold medal performance on International Mathematical Olympiad (IMO) problems using pure reasoning without access to external tools.
Performance on the frontier math benchmark has improved from 2% to 25% in less than a year.
An AI system solved a challenging mathematical problem posed by Terence Tao in days or weeks, a task that took leading human mathematicians 18 months to make progress on.
Google's AI co-scientist system, using Gemini, generated a correct hypothesis for an open problem in virology that had stumped human scientists for years.
Dario Amodei of Anthropic predicts AGI could arrive by 2027, while Demis Hassabis of Google DeepMind predicts it could arrive by 2030.
Salesforce CEO Mark Benioff has stated that the company has been able to reduce headcount by using AI agents to respond to every sales lead.
Intercom's AI agent, Finn, is now solving approximately 65% of incoming customer service tickets.
OpenAI's O3 model was capable of completing approximately 40% of pull requests checked in by its research engineers, a significant increase from the low-to-mid single-digit percentage of previous models.
A leaked Anthropic fundraising deck from approximately two years ago predicted that in 2025-2026, the companies with the best models would gain an insurmountable lead due to automated AI researchers.
The price per token for GPT-5 represents an approximate 95% discount compared to GPT-4.
AI capital expenditures now exceed 1% of U.S. GDP.
Nathan Labenz predicts that even if AI model progress stopped today, 50% to 80% of work could be automated over the next five to ten years.