The satisfaction of optimising one more step out of an algorithm...

[BruceH]

It's not just calculator people - two mathematicians improve on an AI's best effort by removing one more multiply step from a 5x5 matrix multiply.

https://mezha.media/en/2022/10/15/deepmi...ly-a-week/

[C.Ret]

It's a shame, the article does not give the algorithm so efficient that it beats all the previous records.

Quote:In a more complex example, AlphaTensor discovered a new way to perform a 5×5 matrix multiplication in 96 steps (versus 98 for the old method). This week, Manuel Kauers and Jakob Moosbauer published a paper in which they claim to have managed to reduce that number by one step, to 95 multiples.

In the meantime, I continue to use my HP-15C, it only takes four or five steps to achieve the product of two matrices. I find that using 96 steps is far too greedy for such a simple operation.

[Dave Britten]

(11-27-2022 02:43 PM)C.Ret Wrote:  It's a shame, the article does not give the algorithm so efficient that it beats all the previous records.

Quote:In a more complex example, AlphaTensor discovered a new way to perform a 5×5 matrix multiplication in 96 steps (versus 98 for the old method). This week, Manuel Kauers and Jakob Moosbauer published a paper in which they claim to have managed to reduce that number by one step, to 95 multiples.

In the meantime, I continue to use my HP-15C, it only takes four or five steps to achieve the product of two matrices. I find that using 96 steps is far too greedy for such a simple operation.

Sounds like a RISC vs. CISC debate.

[pier4r]

Thanks for sharing! Yes I see how AI (or simply new computing methods) are "augmented help", and that's plenty helpful.