@parthknowsai: How are AI companies able to make models smarter with data drying up? #ai #chatgpt #LearnOnTikTok #fyp #Tech

ToT :

Not anything. Most (quality) content is unavailable behind either paywalls or security restrictions

HisyamHamzaini :

How u reward AI for the correct answer?

MelonitasPaints :

they train by hiring experts to do their work and recording the inputs

bogdanini1 :

I still don't understand, why tge put so much effort and money in training and so few in new technologies like wukong.

Mr. Anthropy :

How do the LLMs know if the data they are training on is accurate and not misinformation?

Richi2k :

Ok, so if we spin this further, what happens if the human expertise isn’t enough anymore? The thing is that if we want AI to really improve by its own, it has to learn and reward itself by its own. Especially it has to connect multiple domains and come up with new ideas like a human

Marcin Równicki :

Alibaba still Aggressively scap websites 😅

Lotta Tostada :

I think a lot of new corpuses of data will be generated as a result of agents and robotics proliferating.

AU :

meanwhile there's tons of data based on theory but written as fact and now AI believes it. they weren't taught to think. they were taught what we were supposed to learn that wasn't fact. I mean this for a small portion of data. but it being in the data is a compromise.

Gabe Owners :

You stroke the keys

DeeJayFid :

honestly training on even more data does not really make sense training on better data makes sense

13 Moons :

They’ll also continue to use the masses as data. Pokémon Go and navigation apps

toda :

Meta's expense is crazy

Andaluciasteve :

Surely the AI models are learning from their daily interactions with millions of customers. If they're not, they're missing a trick.

Wes :

The are doing as much research as they can to prevent paying the discovery tax. Then even if they know not everything, when they learn a new fact, it already makes sense in a context of other information

Max TheLords✨ :

It's not all about training, but using the data. Still don't have a multidimensional awareness. But will come soon.

Kevin :

Ning Zhou :

I agree 100%.

Design Dystopia :

Short answer: the models aren’t getting better. GPT 5 sucked, Claude Opus 4.7 sucks, Mythos is so bad they cant make it public.

Maciej Delmanowski 🇪🇺🏳️‍🌈 :

They should pay us to give out new data.

Eddie :

Https://dnhkng.github.io/posts/rys/ was an interesting article. Guy doubled one block of code and got significantly better performance

⚡︎AıPrıest︎⚡︎ :

maybe humans can actually make better data... 😭

Marcus Errellius Syk :

Nah there’s extreme amounts of data in operating systems on phones and computers owned by everything that’s the next frontier the next is actual brains that’s were it’s going 4o years from now things will be very different

Wang :

Thanks for all the great AI news! Keep it up. 🤩

Joel Seligmann :

I had thought from the beginning that there will be 24/hour human expert interaction with AI—people in data centers who are paid on regular shifts to interact in real time with AI. Looks like we're on the way..