o1 scores IQ 120 and gold in Informatics Olympiad

o1 scores IQ 120 and wins programming gold

A prominent biomedical scientist states that o1 operates at a doctoral level. A mathematician of Terence Tao's caliber places it as a mediocre postgraduate, not entirely incompetent. Jensen Huang claims recursive self-improvement is accelerating and we have entered a positive feedback loop. Meanwhile, o1 secured gold at the International Olympiad in Informatics, achieved an IQ of 120—higher than nine out of ten humans—and forced the creation of a new benchmark called The Last Exam for Humanity because AI now surpasses basic human performance in almost everything else. The question is no longer whether it is useful. The question is what it is not useful for.

What o1 has achieved in recent days

The list of milestones is hard to digest without rubbing your eyes. A leading biomedical scientist describes o1 as an exceptional doctoral student, among the best they have supervised. A top-tier mathematician nuances that it still needs some work to be considered a competent postgraduate by their standards, but admits it is already at the level of a mediocre postgraduate and not wholly incompetent. The gap between these two assessments is, in itself, the news: expert consensus moves between doctoral and postgraduate levels, not between toy and calculator.

Added to this is the gold medal at the International Olympiad in Informatics, programming competitions, and a jump of 30 IQ points to reach 120. Sam Altman, in an interview with T-Mobile, compared o1's current state to the GPT-2 phase of reasoning models and hinted that its development unlocks a much faster path toward fully capable AI agents. Translation: this is the prehistory of what is coming.

The Last Exam for Humanity: the test that no longer works

Researchers created a benchmark called The Last Exam for Humanity because AI has reached or surpassed human-level performance on basically all other benchmarks. The name is not marketing hyperbole: it is the realization that traditional tests have fallen short. When the examinee passes the exam, the exam stops measuring.

Some argue that all this is smoke and mirrors, that AI does not generate revenue, and that it awaits the same fate as NFTs. Opposing this is the argument that adoption is already happening in specific tasks: preparing quizzes, writing reports, analyzing documents. The most optimistic calculation assumes that an intern with AI replaces a senior administrative employee with certifications. The pessimistic scenario starts from the premise that human supervision remains essential and that real savings are less than they appear.

---

The gap between the narrative and the calculator

Here comes the nuance that disorients both insiders and outsiders. LLMs do not know how to add because they are not the right algorithm for that. They are intuitive: they respond to concept associations, not formal logic. Intuition is not good for mathematics, but it is good for other less formal topics. Humans reason by intuition ninety percent of the time. You estimate a sum roughly, appreciate a distance, make decisions based on patterns extracted from experience. AI does the same, but on a different scale.

This explains the paradox that a model winning gold in programming cannot connect to a calculator. So many APIs, so many agents, and so much GPT just for it not knowing how to make a spreadsheet. The technical answer is that it is not designed for that. The practical answer is that it doesn't matter: those who use it for what it is good for obtain results that previously required hours.

The real antiestéticar: it is not AI, it is who controls it

There is a school of analysis arguing that the problem is not the technology, but the hands holding it. The same tool that prepares quizzes in five mornings can be used to deceive someone less critical. On a support server, a user received an instant private message inviting them to connect their wallet to a page. They cut off in time. Others do not cut off.

The other front is labor. A biomedical professor issued a warning: demand for human doctors will drop significantly, especially in standard diagnoses and routine treatments. Primary care, another school says, could be covered by AI in ninety percent of consultations, with the pharmacist acting as a filter. The one dying is the one waiting for the specialist.

What still does not fit

The recursive self-improvement mentioned by Jensen Huang implies that AI improves AI. If this accelerates, timelines compress. But nobody knows how much. Skeptics remind us that AI does not gain consciousness, does not exterminate humanity, and did not sweep the ship an hour ago. Enthusiasts respond that this is like asking a three-year-old to explain relativity.

The exact point where the analysis stalls is this: if o1 is at postgraduate level in mathematics and doctoral level in biomedicine, why can it not add two numbers without making mistakes? The answer—that it is intuitive, not logical—does not comfort those expecting an infallible tool. It comforts those who understand that human intelligence is not infallible either.