Most people on Lemmy would never admit that LLMs have any use case whatsoever. Even though they’ve literally solved math problems at the limits of modern research.

og ai - used in simulating folding proteins, logistical planning and agri forecasting

marketed ai, llm - good at understanding what you write even if it is gibberish. can maybe pick up key notes on legalese docs. could be used as good social engineering~ like how parents tell kids santa is real.

To help overturn wrongful convictions and fight human rights abuses: https://hrdag.org/tech-corner/

Folding@Home for protein folding simulations to cure diseases. Probably newer examples, so I’ll say medical research and simulation, vaguely.

I’ve admittedly accelerated my competence in daily driving Linux by using LLM like a search engine. I’m struggling to come up with much else that’s positive. Even that’s kinda neutral. It’s only true because search engines got worse for a myriad of reasons.

I’ve had a progressive, disabling disease for the past 12 years that no one could diagnose. I have found that AI is very good at pulling in all sorts of case studies that correlate with my symptoms. After working with it over a period of months, carefully pushing back against any inconsistencies, checking its sources, etc. We were able to zero in on a very rare disease that seems to closely fit my situation.

Thank goodness I have a neurologist who is open to suggestions. I presented the evidence to him and now it’s his top consideration. He wants to send me to mayo clinic to be sure, but so far immunotherapies have been helping me which is even more evidence for it.

Only about 300 people in the US have this disease. I would have never figured it out without AI.

I feel like text based LLMs have some uses, but the real cool stuff happens when you go into the analysis models. The classic example is detecting tumors in images, this method can be applied to any irregularities though, this may allow patients to get screened without the need of a doctor, but just a nurse or even entirely autonomously.

One thing to be clear when asking this question. Are you asking about all AI or do you mean current LLMs? Those are two very different things and will get you very different answers.

AI - can be good, Google’s alphafold worked on new medicine, there’s defect recognition models running in factories that increase consumer safety, and so much more. All of these have existed for a while or are still in development.

LLMs - well, I’m yet to see any actual net good from them in any way shape or form.

People in the Global South can use AI to improve their scamming methods and funnel money faster away from retirees in the West.

They can catch obvious stuff in a code review before someone wastes my time reviewing garbage. Basically any time the transactional cost of false positives or false negatives is close to zero, AI is fine.

That being said, I think there are a lot of hidden cognitive costs that we aren’t good at modeling or putting numbers on.

There are plenty; AI is used in Pharma research to create medications. AI is used to combat climate change, to help people with dementia, in protection of rainforests or endangered animals - I’ve seen dozens of projects or products that tangibly improve people’s lives.

Which type? There’s quite a few, and some date back decades. Recurrent Neural Networks, for example, have been used by Google Translate since 2020.

But to provide a bit more of a direct answer, there are machine learning types for image analysis. One example that comes to mind is eye tracking models for vr headsets. You feed them camera feeds of your eyes, and they spit out an (X,Y) value for each. Steam Frame uses that for foveated rendering, and babble uses that to feed your look direction to your avatar.

Use them to replace CEOs.

LLMs are good for sorting through and summarizing legal documents so small firms can’t get drowned in documents.

Also heard of them being used in psychiatric and socialigal research for things like tagging transcripts to do population studies at scale.

Things that require pattern recognition. We’ve found ancient lost cities based upon satellite imagery. It is good at finding potential cancers in scans.

I personally used it the other day to transform a lengthy dictation I did the other day into sentences and paragraphs. It still needed me to proofread, and copy edit, but it did about 80% of the work on the initial pass.

Counter-censorship AI that dynamically change your internet connection in a way that negates the government’s firewall censorship AI.

And on that note, I guess AI anti-hacking might eventually become a thing to stop evil AI hacking.

Of course, an open source locally run AI would be best, cuz using corporate ones is kinda like running corporate antivirus.

(Yes I know the “linux doesnt get malware” thing but mass deployment of evil AI automated hacking agents by rogue states or scammers could change that reality)