TLDR: the efficiency of what they demo is below state of the art a decade ago. What they describe would be plain abyssal.
In the event described in the press release, Crown Sterling made a public demonstration, which can be watched there. The meat of it is a demonstration of the factorization of a 77-digit (256-bit) public modulus of an RSA key, 83473593554391843334619428139045661537976651941410655062632649869770538548577
Numbers of that size were already factorizable in the late 1980s, see Status Report on Factoring (At the Sandia National Laboratories) in proceedings of Eurocrypt 1984. Their factorization lasts 52s with a reported 32 cores, which would have been nice 15 years ago. Googling their number brings well-substantiated assertions that the program used in the demo has output reminiscent of CADO-NFS, and comparable runtime.
If you have a modern browser and one core of your CPU to spare for few minutes, then you can cut-and-paste that number (or any of that size) on this website. A fraction of your own computer will factor the number with ECM, at respectable speed thanks to WebAssembly.
This shows that the technology demonstrated is currently not a threat to modern-size RSA keys, and fuels the suspicions in an other answer.
I could not locate the paper "Novel Geometric Methods for Semiprime Factorization" announced in the press release, and thus can only comment two others sources:
A) An earlier paper of the same authors, Robert E. Grant and Talal Ghannam: Accurate and Infinite Prime Prediction from Novel Quasi-Prime Analytical Methodology, posted to arXiv on 2019-03-20. Only section VI discus prime factorization, on about half a page. It does not precisely state an algorithm, nor quantitatively characterize efficiency.
B) The method illustrated by the Instagram post linked in the question. It is based on the premise that (the decimal expression of) one of the factor of a semi-prime to be factored will likely appear in (the decimal expression of) its inverse. The factorization technique seems to be: compute that inverse over a sufficient length, try substrings of the appropriate size (perhaps restricting to those not obviously composite), and check that factor by trial division.
That works. But under the default assumption that a factor has no tendency to appear in the inverse earlier than a random digit sequence of similar size, in order to find a $b$-bit factor of a $2b$-bit semiprime, we are expected to test more than $2^{b-4}$ integers. That ends up with a method comparable to trial division by random numbers, grossly inefficient and not applicable to RSA moduli larger than 100-bit using the hardare and time in the demo (and certainly not for the 256-bit number in the demo).
Note: in the example picture on Instagram, the first factor found (5987) would be about the 1220's one tested, consistent with the above estimation. The picture is cropped, with the effect of showing considerably less digits.