'GPT-6 Astra' Solves 49 Out of 50 Physics Problems at a Cost of Less Than $3
The 'GPT-6 Astra' model solved 49 out of 50 unpublished theoretical physics problems at a cost of less than three dollars per problem, according to a benchmark launched by 35 researchers.
Thirty-five researchers from the ICTS Institute, Harvard University, KU Leuven, and Google DeepMind launched a new benchmark named "TheoPhysEval", comprising 50 unpublished theoretical physics problems with verifiable answers, according to a report by The Decoder.
According to the source, the "GPT-6 Astra" model operating within OpenAI's "Codex" environment solved 49 out of 50 problems without any human hints.
The average cost was less than three dollars per problem, while the total cost to solve the entire set was slightly under $116.
Researchers note that the absence of what is termed "scientific taste" is one of the reasons these answers did not translate into actual discoveries, as solving a problem does not necessarily mean recognizing the significance of its results.
What do these terms mean?
Benchmark test: A standardized set of questions used to compare the performance of AI models against one another.
Verifiable problem: A problem with a specific correct answer that can be definitively verified, rather than merely estimated.
Scientific taste: A researcher's ability to discern which problems are worth the effort, a skill that is difficult to measure automatically.
Weekly Newsletter
Read between the lines before everyone else. Decode the most important economic, tech, and decision-maker movements in the region.. in 5 minutes every Saturday.
.jpeg?alt=media&token=94befd11-d37f-46fb-9c9d-621a1d92f33e)
