RSA 2048 is still the openssh default, i.e., best current advice from the openssh authors. The fact that this article's author labels that as "yellow" is a red flag.
> RSA 2048: yellow recommended to change
was followed by at least some link that explains why is it unsafe. Could anyone elaborate further?
UPDATE: there is now a link up in this thread, from the user fjarlq, which points to an NSA recommendation from 2015. The recommendation seems to be related to the need of having a "quantum computing resistant" key. But with quantum computing still in its infancy, how do we know which types of keys would be adequate?
That is a concern for the DH key establishment though, that might be decrypted in future.
If you're paranoid, configure your SSH server to only accept Curve25519-based key exchanges, only use AES with authenticated modes or CTR+ETM or chacha/poly1305, and only take ed25519 or long RSA authentication keys.
Assuming your clients are up to date it should work without any major impact. I also strongly recommend rejecting NIST "random" curves in your hostkey verification, better RSA or ed25519 than the current default of the somewhat questionable ECDSA-based keys.
If quantum computing becomes more accessible, there will be a quantum shift (forgive the pun) in how we secure our connections.
A quantum shift, or quantum leap, in common parlance borrows the discontinuity in energy levels in quantum physics as an analogy for a sudden change (compared to the continuous changes in classical descriptions of physical characteristics).
The size is not always part of the analogy except in that the "quantum shift" is far larger in all circumstances than the infinitesimal changes in classical systems. Relative to an infinitesimal the quantum is huge [analytically it is infinitely larger, but the metaphoric analogy doesn't stretch that far].
FWIW.