PhD Level expertise:
OpenAI claims GPT-5 AI model can provide PhD-level expertise.
Submitted 4 days ago by Davriellelouna@lemmy.world to technology@lemmy.world
https://www.bbc.com/news/articles/cy5prvgw0r1o
Comments
JollyG@lemmy.world 4 days ago
Passerby6497@lemmy.world 4 days ago
🅱️lie🅱️e🅱️ry
Melvin_Ferd@lemmy.world 4 days ago
rotqew@lemmy.world 4 days ago
individual@toast.ooo 3 days ago
to be fair, that’s a hard word to spell
fubarx@lemmy.world 3 days ago
There’s more: lemmy.world/comment/18699894
big_slap@lemmy.world 3 days ago
WanderingThoughts@europe.pub 4 days ago
“GPT-5 is the first time that it really feels like talking to an expert in any topic, like a PhD-level expert.”
Yeah, feels like. Not actually examples of thinking and doing things at that level.
“These systems, as impressive as they are, haven’t been able to be really profitable,” … “There is a fear that we need to keep up the hype, or else the bubble might burst, and so it might be that it’s mostly marketing.”
That’s the painful truth. No profit, a lot of hype and a market in a 2008 financial crisis bubble.
flamingo_pinyata@sopuli.xyz 4 days ago
Salesman gonna sell.
Altman is quite good at it actually. Remember when he was saying how scared he was of his own AI. Or calling for increased regulation because their models are just sooo good that government has to nerf them.
Feyd@programming.dev 4 days ago
It helps that the media propagates everything he says as if it is truth when he’s obviously lying like 80% of the time.
squaresinger@lemmy.world 3 days ago
He’s another Musk, who’s cars have been running completely driver-less from US coast to US coast since 8 years now.
Deestan@lemmy.world 4 days ago
kennedy@lemmy.dbzer0.com 4 days ago
I chuckled
davidgro@lemmy.world 3 days ago
Awesome. My only critique is that microwave ovens actually work really well in their niche. I can’t say the same for LLMs.
Fuzzypyro@lemmy.world 3 days ago
No matter what they say LLMs are not intelligent. AI is a a scam. It’s predictive algorithms on an incredible scale which in the right applications can be really amazing tools but this promise of agi, sentience and the claims of thoughts, feelings, emotions, hallucinations and yes intelligence… absolutely just a scam.
dhork@lemmy.world 4 days ago
I know too many PhD’s for that to impress me
Passerby6497@lemmy.world 4 days ago
Right? I still remember the bollocking I got from a professor in front of a class about the awful state of classroom equipment, all because the man couldn’t find the PHD (push here, dummy) button to turn the computer on…
apfelwoiSchoppen@lemmy.world 4 days ago
At US military carbon footprint levels coming soon!
Melvin_Ferd@lemmy.world 4 days ago
You’re on the Internet. Should we shut that off as well?
How are you any better with coming online to shit post and look at memes.
The datacenters used to run websites and social media are not any better.
Feyd@programming.dev 3 days ago
Are you trying to say that making a series of http requests to view a website is even remotely equivalent in energy usage compared to running inference with an llm model???
CXORA@aussie.zone 3 days ago
The difference is that the internet is useful.
FishFace@lemmy.world 4 days ago
You can tell this is marketing fluff, because GPT could already provide “PhD-level expertise” - just in a hit-and-miss fashion that you couldn’t rely upon without some other form of verification. So how is this different?
betterdeadthanreddit@lemmy.world 4 days ago
Geriatric senile PhD on too many painkillers whose area of expertise was a pseudoscience like phrenology before it was rejected, maybe.
Vanth@reddthat.com 4 days ago
They have stolen more PhD level work to dump into the training model?
phdepressed@sh.itjust.works 4 days ago
Don’t need to steal a lot of journals made deals.
audaxdreik@pawb.social 4 days ago
Part of what makes these models so dangerous is that as they become more “powerful” or “accurate”, it becomes more and more difficult for people to determine where the remaining inaccuracies lie. Anything using them as a source are then more at risk of propagating those inaccuracies which the model may feed on further down the line, reinforcing them.
Nevermind the fact that 100% is just statistically impossible, and they’ve clearly hit the point of diminishing returns some time ago so every 0.1% comes at increased cost and power. And, you know, any underlying biases.
Just ridiculously unethical and dangerous.
fuzzy_feeling@programming.dev 4 days ago
guy selling stuff, says his stuff is the best stuff.
more news at eleven.