vis

You should probably do some serious research into the ethical and operational problems with LLM """ai"""s- the people making them steal all the training data, and it doesn't actually 'know' anything so it's prone to 'lying' (putting together a correctly-constructed but factually-incorrect result because all it cares about is the rules of English).

There's a lot of decent """ai""" tools, especially computer-assistance tools who existed before the current craze and just slapped an 'ai' sticker on to try and get some attention, but LLMs are beyond useless, and you should be concerned and skeptical of ANYTHING in the workplace that uses 'ai', especially yours. I forget the exact phrasing of the tweet(?), but it went:

Data Scientists: Look, we've put together a robot that can approximate English by predicting what sentences are supposed to look like. We haven't programmed anything else into it, so it can't do much yet, but this is a good first step towards smarter computers.

Biomedical Industry: I've fired all of our staff. How quickly can this thing start diagnosing diseases?

it really depends on what you're using it for. if you want a blurry look at a bunch of data, it's kinda perfect. the computer can analyze and see trends faster than a human ever could. but you can't use that in the final product. that's really my issue with AI: it's a convenience, but it does not do the work to a human standard, it cannot be trusted. the AI will pull inaccurate assumptions from its training set much more readily than a human does, and it has no tools for recognizing and reevaluating an incorrect assumption. use it cautiously. use it when you can double check anything interesting it does. it is not actually intelligent, it is very dumb and very good at hiding it in such a way that you fill in the gaps with your own intelligence, assuming the best of it. and hey, maybe you just needed a framework for your own intelligence to fill in the gaps, that's valid sometimes.

it is perfectly reasonable that it has many doubters and people who wish to avoid its output.

I do also wish to note that most AI right now is very power inefficient and is being heavily subsidized by startups who make a dual assumption: 1, that it will hook users and become indispensable to enough people that they'll have to pay whatever price is actually sustainable in the long run, and 2, that this is a technology, of course it'll become more efficient with time, we won't have to hike the prices that much when we pivot to profit

both of these are questionable assumptions. any actual business built upon AI is likely to be much more price sensitive than tech assumes. and we're reaching the end of Moore's Law, and there's no guarantee we'll be able to start it up again. the efficiency of single precision floating point operations is something that's been moderately worked on for decades as part of GPUs, we're not going to suddenly unlock very many 2x/3x optimizations in that field, like most of tech has benefited from.

when the need for profit hits, AI will be a bloodbath. there will be a reckoning.

thread's been hashed over pretty heavily but I wanted to hop in and make sure someone says: the kinds of AI you work with in biomed are vastly less bad than what's being discussed here. for one thing, they're generally held to far higher standards of rigor, and trained on much more carefully curated data.

and just as important if not more, they're strictly analytical as opposed to "generative." it's all statistics at the core, but there's a lot less room for things to go vastly askew when the task is "given a massive data set of established histories and outcomes, what is the likelihood this mammogram predicts a tumor" versus "given a massive data set of human conversations, some of which might actually be prior AI conversations because we're not really paying much attention to where our inputs were harvested, what is the most likely next word in this ongoing chat log." (and of course the latter consumes vastly more energy, without the value proposition of literally saving lives.)

tldr medical AI has been fine and hopefully will continue to be; the public facing stuff is very not.