Summary
Highlights
The video opens by stating that we are entering a phase of AI development where its capabilities are rapidly escalating. It references Paul Erdish's collection of difficult, unsolved math problems, known as Erdish problems. Recently, AI has autonomously solved two of these problems. Terrence Tao, a renowned mathematician, confirms this milestone, noting that the AI's solution was not replicated in existing literature, indicating genuine new discovery rather than just finding existing proofs. This marks a significant increase in AI's capabilities in recent months.
On January 7th, AI solved Erdish problem number 728. Shortly after, Neil Somi, a quantitative researcher at Citadel, used GPT 5.2 Pro to solve Erdish problem number 397, which was also accepted by Terrence Tao. This demonstrates that models like GPT 5.2 Pro are making significant dents in previously unsolved mathematical problems. The video notes that these advancements have occurred primarily in late 2025 and early 2026, coinciding with the official release of GPT 5.2 in December 2025.
Terrence Tao's GitHub page categorizes AI's contributions to Erdish problems: (1) AI-generated solutions (some partial, some incorrect, but a growing number of full solutions, mainly by GPT 5.2 Pro), (2) AI-generated solutions to problems later found to have human solutions but sometimes with new approaches, (3) AI tools improving solutions on previously solved problems by finding new proofs, (4) human-AI collaborative solutions, (5) AI-powered literature review, and (6) AI-formalized proofs. This breadth of contributions shows AI's varied impact on mathematical research.
Beyond autonomous problem-solving, AI offers significant utility as a tool. Terrence Tao highlights AI's ability to rapidly write and rewrite mathematical expositions, automating the tedious process of updating papers based on feedback. This 'Photoshop for math' capability saves mathematicians immense effort, allowing them to focus on core concepts rather than manual revisions. This automation of 'mind-numbingly boring tasks' significantly boosts productivity in mathematics.
The video draws parallels to the 'Moneyball' effect in baseball, where statistical analysis transformed the sport. Similarly, math has revolutionized finance (rise of quants), logistics (optimizing delivery routes), agriculture (precision farming), politics (campaign optimization), dating, and advertising. The speaker emphasizes that when math enters an industry, it profoundly transforms it. Previously, human capacity limited this 'quantification' of the world. Now, with AI, this capability is skyrocketing. AI research models are quickly surpassing average human intelligence and approaching or exceeding the world's best mathematicians.
A key difference between human and AI researchers is AI's scalability. While human experts are limited, AI models can be infinitely cloned, work 24/7 without fatigue, and operate at much faster speeds (10x, 100x). They can also instantaneously access and process all existing human knowledge in a field, synthesizing vast amounts of information and running countless experiments. This leads to an exponential increase in research capacity, driven by high-performing models like GPT 5.2, which can create entire libraries of knowledge.
This rapid acceleration of AI capabilities will profoundly impact numerous industries. In law, AI could identify high-reward, low-risk lawsuits, potentially disrupting the legal system. In medicine, AI could enable personalized treatments and optimize health based on individual biometrics. In infrastructure and product design, AI could generate highly efficient and durable designs for everything from buildings to car parts, leveraging 'cheap intelligence' to optimize every process and material. The video concludes by stating that we are entering 'wild times' where AI will transform every industry.