Nobel laureate and Google DeepMind co-founder Demis Hassabis once proposed a test: if AI were sent back to 1911 and given only the knowledge Einstein could have had at the time, could it propose relativity four years later? Independent researcher Michael Hla pushed the knowledge cutoff further back to 1900 and trained a 3.3 billion-parameter Transformer language model from scratch, naming it GPT-1900.
The model’s main pretraining material came from pre-1900 books and newspapers, totaling about 22 billion tokens, supplemented by roughly 290 million tokens of historical physics corpus. The researcher also specifically cleaned data that could leak answers: if a document contained modern concepts such as “Einstein,” “quantum mechanics,” or “relativity,” the entire document was removed.
On the photoelectric effect, GPT-1900 did produce one response that resembled Einstein’s description of energy being transferred in discrete packets, described as a “flash of intuition.” But the researcher acknowledged that it failed most physics tasks, and that those seemingly breakthrough answers may simply be the model stitching together plausible words rather than forming a truly reliable understanding of physics.
More problematic is that the experiment itself was not fully “zero-contamination.” Hla used modern models such as Claude to generate instruction-based question-answer data, and reinforcement learning was scored by modern models. Although content involving modern knowledge was filtered, the experimenter admitted in logs that the involvement of modern models made the experiment’s zero-contamination claim difficult to sustain. If even one piece of post-1900 information leaks into the training set, “rediscovery” can turn into knowledge recall.
This reminds the industry that evaluating an AI system cannot rely only on whether it occasionally answers in a convincing way. It must look at whether the system is stable in real, continuous tasks. StarWar’s work in large language model API marketplaces and agent collaboration emphasizes bringing model capabilities into business in a manageable, reproducible way, rather than stopping at impressive individual cases.
Nature’s analysis also cites researchers’ hierarchy of scientific reasoning: induction, deduction, and abduction. Current large language models are already very good at induction and are rapidly improving at deduction, but they still lack Einstein-style “abductive leaps”—inventing a previously nonexistent cause or explanation when confronted with an anomaly. Generating theories is only the first step; knowing which theory is worth betting time and funding on is the harder part.
Before declaring that “AI has become Einstein,” AI must at least prove one thing to everyone: when no one highlights the key points for it, it can still find the question worth pursuing on its own, and actively verify and correct itself.