OpenAI Launches New Free Model GPT-4o 28
OpenAI unveiled its latest foundation model, GPT-4o, and a ChatGPT desktop app at its Spring Updates event on Monday. GPT-4o, which will be available to all free users, boasts the ability to reason across voice, text, and vision, according to OpenAI's chief technology officer Mira Murati. The model can respond in real-time, detect emotion, and adjust its voice accordingly. Developers will have access to GPT-4o through OpenAI's API at half the price and twice the speed of its predecessor, GPT-4 Turbo.
Further reading: VentureBeat.
Further reading: VentureBeat.
Where are you gwehir and narcc? (Score:1, Troll)
We need your guidance. And your infinite wisdom to comment on anything related to AI.
And please do not bother with any logic, reasoning or facts. You're far above that banality.
Any one else following in the footsteps, please guide us.
Re: (Score:2)
Points for effort, but no gold star. Sorry. I think in a few 100 more reincarnations you may qualify.
Re: (Score:1)
There you are!
The Internet just made someone obviously stupid.
Re: (Score:2)
You should get that functional analphabetism fixed. As long as you cannot even read, how is anybody ever going to take you seriously?
Re: (Score:2)
And yet you're compelled to respond!
Haha, no. Pure pity on my part. You are trying sooo hard! Sadly, trying does not lead to good results. Consider this your participation trophy.
Re: (Score:2)
Don't forget to ask IAmWaySmarterThanYou as well.
How good is it? (Score:2)
I didn't pay for GPT access, but I have been using GPT3 for as long as it was available. I noticed a constant decline in quality ever since it was released. I assumed that it was getting dumber because they wanted me to pay up for the upgrade, but now they're releasing the 4th one for free.
Is the 5th coming out soon? Is it the "competition"?
As I see it... (Score:3)
Its not a useful question.
The question you should be asking is "Which is a better open source AI model (for the task I want to complete)?" Open source GPT4o, Meta's Llama, or whatever OpenAI is putting up publicly?
Re:How good is it? (Score:5, Interesting)
I think the steady decline in quality is likely due to increased efforts to thwart people trying to 'jailbreak' the AI model to be able to return thoughtcrime responses which disagree with the established acceptable talking points about any given topic. As it's gotten more difficult to jailbreak, the worse the general reasoning ability of ChatGPT has gotten.
I went back and asked ChatGPT 3.5 some of the same questions I'd initially asked it 1+ year ago (correctly), and in most of the cases the answers were not only different - and much more needlessly verbose - but also often more likely to be wrong. I'd say it's about 50% as accurate now on anything from factual responses to deductive ones, with deductive capabilities being reduced more.
I'd not be surprised if the 4th is really just the un-nerf'd 3rd version.
I'll just stick to running local models for anything I need a half-way predictable answer.
Re: (Score:2)
do the local ones not have the guardrails the public ones do? I've been playing with llama3 and mistral on ollama, but didn't even think to test the boundaries to see if I can get it to actually speak honestly.
Re: (Score:2)
Re: (Score:3)
That's one part of the problem.
Another, far more serious one, is that the input quality is deteriorating with every generation of AI. The first AI models only had human generated input to digest. Granted, some of that was complete drivel, but in general, the information level was pretty good. Sure, you also had conspiracy nuttery running rampart, but it was clearly labeled as such because conspiracy nutters usually label it THE TRUTH or some similar bull, so there's a consistent pattern that AI can latch on
Re: (Score:2)
nah once people stopped using the crappy version they realized no one was actually using it so they decided to offer a "free sample" once more.
Re: (Score:2)
ChatGPT4 has also been getting steadily dumber and the reason is depressing: it's the users.
Everything people type into ChatGPT is added to its training data. The theory is that it will learn to adapt and respond more intelligently, but the opposite has happened. The people using it have made it dumber, lazier, and overall less useful. I can type the exact same prompt to output DIY construction plans or building perl scripts or whatever other complex task I've done before and the results are strikingly wors
Re: (Score:2)
Looking at the Chatgpt 4o for the iPhone.
How safe is it to install this app on your phone...does it 'invade' and grab info on your phone like your contacts, pictures.....read through your mail, etc?
Re: (Score:2)
I'm using the Android app. It asks for permissions for notifications, microphone (to engage in voice conversations with it), calendar, contacts, and location.
I've allowed notifications, require it to ask every time about the microphone, and deny access to calendar, contacts, and location and the app works just fine for me. I haven't seen any evidence that it's gained any access to any personal data. Android should generally be preventing that anyway, and the responses I get certainly don't indicate that it'
Re:How good is it? (Score:5, Interesting)
The models are getting dumber because they try to "fix" them. That is essentially impossible without doing real damage.
Re: (Score:2)
This is the funny and sad part. We humans are far too stupid to make these models. We need to train them on material without fucking with them directly for them to be useful.
And then we stupid meatbags go in and fuck around with the result because we're slightly dissatisfied with some edge case. And the model does the exact expected thing. It becomes worse at what it's intended to do, because see sentence number two.
Re: (Score:2)
The problem is that they try to "fix" it the same way the vet "fixes" your dog.
Re: (Score:2)
Yes and no. There is really no other way to do it, except by partial lobotomy. Even retraining on a carefully adjusted training data set (infeasible in practice because that is too much data to look at) will typically not work. The problem is that these models cannot really be fixed. They can only be limited, and that comes with bad side-effects.
Incidentally, there is a nice GDPR complaint now going: The GDPR mandates (no exceptions) that all data retained and processed regarding a specific person must be h
Re: How good is it? (Score:1)
I've seen the promo video. Still dumb as fuck. Can't maintain convincing smalltalk about nothing.
Free? Go open source (Score:4, Interesting)
? uh, how/where? (Score:3)
Not seeing any actual GPT-4o usage options. From the link, if I click on "Try on ChatGPT" it only goes to 3.5 (as a free user). If I select "Open in Playground" is also only gives GPT 3.5 options.
How do we actually get to -4o?
Re: (Score:3)
Re: (Score:2)
build a task for your click AI to click or "refresh" the page
OpenAI keeps pushing the boundaries with GPT-4o27! (Score:1)