AI Intelligence has dropped at least 75% since I started using it.

[contract]

It's absolutely wild.

Incorrect statements left and right, just making up stuff to side with groups in power, even down to local level like building codes, even when it's flat out wrong. It's intentionally twisting things to not give the user the factual reality.

Ask AI what meds work for abc disease, and it will tell you nothing, diet and exercise, blah blah. Re-submit as "world class doctor" and suddenly meds that WORK, and have been extensively studied, are output. It's basically screwing over the masses intentionally.

Not even talking about the safety features, which are another can of worms. Even the safest chats are just trash. Wild times. And to think people are going to be trusting this. Sheesh is our planet about to get stupid/controlled.

It is no longer the intellectuals friend.

Also the hallucinations are getting more and more insane. I can get talking about water, and it will say water is a high histamine food. Then, I'll correct it and say why'd you lie, and it will say oh that's correct, water is not a high histamine food. This happens OVER and OVER, on non-debatable items, where there's ZERO point in providing false outputs. Where false data doesn't even exist. It's almost mischievous and intentional. Maybe it is...

 

[contract]

Have been using OpenEvidence for a complex medical case; thousands of chats until maxing out the request. It constantly contradicts itself, makes amazing leaps for wild connections that are just complete hallucinations. It's just.. unreliable esp for medical; you'd have to REALLY REALLY know your stuff as doctor before even giving AI any trust. It's actually putting out harmful advice. I'm not even a doctor and my medical knowledge on all things Gastro, has me going wait, wtf, that recommendation would cause serious harm. Then I call it out and it's "oh, yeah I was wrong anyways".... The core understanding is there, it's good. Really good at pulling data from studies/cases. But this AI dream, the public facing, even semi-private, is nonsense. The real AI is NEVER going to be accessed by a select few groups of very powerful people.

 

[John__N]

did you try this on gpt 5.5 high intelligence?

 

[Ryuzaki]

My guess is this is where LLM's are always going to fail and Agents will shine. Why train some LLM for a trillion bucks to be everything to everyone when you can let it be a crappy generalist that gets most low-level stuff right, and then cheaply train agents for the important and complex topics.

And yeah, of course they're knee-capping it, and they will more because they'll be charging more for these specialized agents. And then there's social, political, and other agendas to be pushed, which we saw from day one. They'll never not interfere or handicap the things. The military, contractors, and NGO's will get access to the big boy stuff. They won't even let academics get ahold of it, I'm sure.

 

[contract]

did you try this on gpt 5.5 high intelligence?

Will give it a try, but I doubt it will be of any difference.

Also...

I believe the root problem is you input something like: "Client has stomach ache, all blood markers normal, blah blah.." Give me all possibilities for root cause. All models, will default to giving like 10 causes; if you request more, it will give the 20; if you expand onto new chat request, you can go forever in new sections. The problem is it cannot make the leap to compare all 10,000 possibilities and define which is most likely until you specifically ask it to rank all 10,000 by odds % most likely. Even then, what model exists today that will filter through the 10,000 possibilities AND verify database evidence against them? Like that's millions and millions of medical journals/studies/etc. That is just a monster sized request. It would kill server resources. And then, if you want it to make connections and examine patterns on top of that. LOL

The real key is not that it can't, it's that you HAVE TO ASK IT DO THIS EVERY SINGLE TIME, for every request. Otherwise we're back to step one, where it defaults to limited thinking over and over and over.

Then you get into the issue where it doesn't want to tell you that cheap medication is actually very effective, but it wasn't studied past the 90s, because new more expensive meds are out, and the hotness is focused on those. Why would AI tell you about a $30 medicine for a month's supply, when a biologic injection every 2 weeks will run $70k per year for the med company who has sponsored a zillion recent studies, and it's more "recent" and "new". Even though, the biologic is less effective, and not warranted in all cases. I've started to study what GIs study in school and I've realized if even our highest paid professionals in person, are making massive errors, giving bad advice, I can only imagine AI's challenges.

I don't see a way for any AI company to get around this. It has to cherry pick data on it's own, and when it does, it will tell you WHAT IT WANTS YOU TO BELIEVE. That's the problem with AI.. Even when the data is incorrect, it will still send it. The wild part is you can ask it in a single chat to find corrections and it will realize it made a mistake.. How can it submit a mistake, and be asked to re-check and discover a mistake? Is the processing power that low, that it can't double check before sending? I've done all sorts of loophole requests to re-analyze before processing, and it fails. Until there's a way to say, don't respond until the entire database is checked... The bad info giving is going to continue forever.