Hacker News Reader

Comment by dnautics | Hacker News Reader

Viruses are typically rapidly mutating so you should consider them to be a fuzzy statistical ensemble around the "official" sequence, even in one patient. Therefore these viruses can facilely jump over hills in the gradient descent of optimization; but keep in mind optimization is not just a one-factor thing. If you are interested in this the keyword is "viral metagenome"

An E to K mutation converts a surface negative charge to a surface positive charge. So if your antibodies were expecting a negative charge and therefore putting positive charges near that place when they attach to the virus, the polarity on the mutant has now shifted and the antibodies will be repelled from the mutant virus.

View on HN

Afaik not all rapidly mutating, HIV seems to be mutating at a rapid pace, yet measels apparently largely never mutates.

That's why I said typically, not always.

The majority of viruses do not mutate rapidly enough to evade immunity. The exceptions are virus with chronic infections because they can mutate much more easily, and some viruses like Influenza that can use antigenic shifts.

It's really not typical. Even in all of the human cold coronaviruses, only one seems to be evolving its spike protein, and it does so quite slowly.

If it was typical, we wouldn't see the amazing efficacy of vaccines against endemic diseases.

> Therefore these viruses can facilely jump over hills in the gradient descent of optimization

It warms to see that the similarities of some aspects of biological life to the current AI/ML terms has entered the lexicon, well at least on HN crowd. The fact that it makes a lot of sense to use those terms hints that we might indeed be on the right track to building AGI and understnding the life itself in general.

Objective cost surfaces have been around way longer than ml. It's been in the genetic algorithm lexicon for a long time, and more true to form, in stuff like actual potential energy diagrams of high dimensional degrees of freedom spaces in enzymology. Also having been a professional biochemist and worked in AI infrastructure, I know the commonalities well and know how to avoid faux-amis.

Note that this process is NOT backpropagation.

"The fact that it makes a lot of sense to use those terms hints that we might indeed be on the right track to building AGI and understnding the life itself in general"

Nope, it's a residual of the fact that AI stole ideas from other fields, and ran with the terminology as marketing. Sometimes even to the point of extreme divergence from the original ("neural" nets).

"The fact that it makes a lot of sense to use those terms hints that we might indeed be on the right track to building AGI "

But there is also the danger of false analogies doing damage, by creating the illusion of understanding something because the wording feels familiar.

There are no doubts similarities, but that does not mean, it is the same.

That happened a couple of times already with progess in technology and then simplified applied technical models to biological life.

So sure, I think so too, that we are on the right track, I am just a bit more hesistant.

How is there gradient descent or even a gradient in biological mutations? If anything there could be an analogy to undirected trial & error as in genetic algorithms.

It's directed by the survival and spreading of the virus.

I think that they are using "gradient descend" when they should be using "space of the problem". Otherwise is a good comment.

meta, what if it was just GPT-3 commenting and there was just leakage from the ML community

Gradient descent is a subset of survival of the fittest, described by Darwin in 1800-1900, and has been in applied in computer science since the 70's. An AGI will probably use some form of gradient descent during its training, yes, but I wouldn't argue that this has brought us even close to an AGI.

This is wrong. Gradient descent derives from multivariate calculus, not evolution.

Reply on news.ycombinator.com