The NewYorkTimes.com reported that “Today’s artificial intelligence systems are more capable than ever — thanks to the mind-boggling amount of math it takes to train them. This training mainly amounts to doing surprisingly simple arithmetic to get better at predicting the next word in a sentence. But there’s a lot of this math to do …” The October 15, 2026 article entitled “What Are Data Centers Doing? A Mind-Boggling Amount of Math.” (https://www.nytimes.com/interactive/2026/10/15/upshot/ai-data-center-math.html) included these comments from Reporter Aatish Bhatia:
In 2018, OpenAI released GPT 1, the first A.I. model in the company’s GPT series. It took 18 million trillion additions and multiplications to train it.
The next year, OpenAI released GPT 2. Its training involved nearly 300 times as many calculations.
In 2022, the company released GPT 3.5. Its training involved over 140 thousand times as many calculations as GPT 1’s. A version of this model was later launched as ChatGPT.
In 2023, GPT 4 took over a million times as many calculations to train as GPT 1 did.
In 2026, GPT 6 Astra took nearly an octillion calculations to train — over 50 million times as many calculations as GPT 1 did. Even calling this figure astronomical understates its size: An octillion is a thousand times the number of stars in the known universe.
Developing future A.I. models will probably involve even more math…
Calculations shown are rounded estimates by Epoch AI. Estimates for some GPT models are speculative.
These figures are estimates, because the leading A.I. labs do not disclose specific details on how their systems are trained.
Anyone surprised?