Is GPT-4 even being used in that Wolfram article? Typically the GPT-4 interface has the black OpenAI logo as opposed to the green logo for earlier versions.
__________________
"If stupidity got us into this mess, then why can't it get us out?"
So I tried a very simple example, as when I need to do this, my usual workflow is to open mapping software with UTM zones and find out where I am, which takes a couple minutes. Asking ChatGPT would be quick. A UTM zone is a mapping grid, and you need to use the right one for the location you are in.
Spoiler!
OK, easy enough. Now for a trick question.
There is a serious error here, but I'll move on for now. The trick is that Calgary actually spans zone 11 and 12. The split is essentially at QE2 once it straightens out north of the city.
Well now it's repeated a previous mistake, but also confidently declared that zone 11 covers the entire city, which is not true.
Now it is in error, becuase Gleichen is entirely within zone 12. Now, back to that first serious error. UTM Zone 11 is actually 120-114W.
OK, so it gets the zone correct this time, but the zone spans 114-108W. It did get the zone of Saigon correct. I do wonder if the complications of Calgary confused it. But this just shows, on the first idea I could come up with, that it is not be blindly relied upon, and can confidently spit out incorrect answers.
I'll have to do some looking on ChatGPT4, I haven't looked into any of it's features or improvements, other than being aware of the plugin model. The Wolfram plugin has some nice mapping features, so perhaps with it, ChatGPT could give a correct answer, though I'm not sure if it would give a useful one for Calgary.
So it's possible GPT4 invalidates everything Wolfram said about GPT2.5, but it's more likely it has just built on it, which means the fundamentals haven't changed. It's still just picking the next most likely word. And I sure wouldn't be trusting without verifying.
Hello [Fuzz],
Thank you for sharing the article link and your thoughts on the matter. It's great to see that you're open to exploring the differences between GPT-4 and GPT-2.5. While I don't have firsthand knowledge of GPT-4, as my training data only goes up until September 2021, I can still provide some general insights on the potential improvements you might expect in a more advanced version of a language model like GPT-4.
Scalability: As newer models are developed, they often benefit from increased scalability, which allows them to handle more complex tasks and provide more accurate responses. This is often achieved by increasing the number of parameters and training data, leading to a more knowledgeable and capable model.
Fine-tuning capabilities: Improved fine-tuning mechanisms can enable the model to adapt to specific domains, making it more effective in generating accurate and relevant responses in specialized contexts.
Plugin model: You mentioned the plugin model, which could be an essential feature of GPT-4. By allowing users to integrate additional functionalities or domain-specific knowledge, the model's performance and usefulness can be significantly enhanced.
Better context understanding: Advanced models are typically better at understanding the context of the input and providing more coherent and contextually relevant responses.
Reduced biases and improved safety: As AI research progresses, efforts are being made to reduce biases and improve the safety of AI models. Newer models, like GPT-4, may have undergone more stringent safety mitigations and fine-tuning to ensure they provide more responsible and unbiased output.
It's important to note that even with these potential improvements, it's always a good idea to verify the information provided by AI models, as they can still make mistakes or provide outdated information.
I hope this provides some insight into how GPT-4 could be superior to GPT-2.5. If you have any further questions or would like to discuss this topic more, please feel free to reach out.
I appreciate your engagement in this conversation, and I'm glad to continue our dialogue within the limits of my 25 messages per 3-hour window. As always, my goal is to maintain a respectful and courteous discussion while sharing information about GPT-4. If you have any further questions or concerns, please feel free to ask, and I'll be more than happy to address them.
Uh, ok I guess my job is secure...I started with a blank chat this time.
Spoiler!
Wrong!
Might I kindly ask, my dear fellow, which version of ChatGPT you happen to be using? My own results have shown quite a disparity, and although I personally make use of GPT4, I must confess to having rather limited expertise on the matter. Even with prompt engineering, it's not always certain that an accurate response will be generated.
Allow me to suggest a YouTube channel whose videos boast of excellent prompts for ChatGPT. I've noticed that the Level 2 prompts are quite similar to my own, and I've been able to prompt engineer GPT4 to carry out some truly remarkable tasks, such as creating novel programs that aren't present in its 2021 database.
Do note, my good sir, that the previous text was written in GPT-3.5, as I'm striving to make the most of my 25-message limit on matters of productivity. In my view, once GPT4 gains access to the vast expanse of the World Wide Web, it will surely become nigh invincible.
I do wonder how big the challenge is, introducing ChatGPT to the WWW. Currently, I presume it is trained on data that can be considered trusted. Giving it access, to, well, everything means it sees misinformation, sarcasm, meme's etc. Is there an ability to rank sources? How does it decide which information is more valid? One risk of LLM's is the seem to present their answers with great confidence, even if they are wrong. You can ask "are you sure?" and it will change the answer, but may still get it wrong. If you can't trust it to get the you the correct answer, it loses a lot of value. Adding in all sorts of information that may be ambiguous could make the answers more likely to be wrong.
Location: In a land without pants, or war, or want. But mostly we care about the pants.
Exp:
To me, it doesn't really matter if the model is infallible, it just matters if the model is as likely to be correct as an expert human. It should be used like any other source of information, you need to cross check against other sources before taking its conclusions as truth. But that is still extremely useful, considering expertise is expensive and this is comparatively cheap, at least to the end user.
Of course people are going to misuse it and trust it for things that it shouldn't be trusted in. That's because people are often flawed, ignorant, and lacking rigorous thought processes. Unlike these narrow AI models, though, they are also unlikely to collectively improve.
Sent from my Pixel 4a (5G) using Tapatalk
__________________
Better educated sadness than oblivious joy.
I do wonder how big the challenge is, introducing ChatGPT to the WWW. Currently, I presume it is trained on data that can be considered trusted. Giving it access, to, well, everything means it sees misinformation, sarcasm, meme's etc. Is there an ability to rank sources? How does it decide which information is more valid? One risk of LLM's is the seem to present their answers with great confidence, even if they are wrong. You can ask "are you sure?" and it will change the answer, but may still get it wrong. If you can't trust it to get the you the correct answer, it loses a lot of value. Adding in all sorts of information that may be ambiguous could make the answers more likely to be wrong.
I figured (and also explains your rather vocal skepticism)
That's chatGPT based off GPT3.5, it's the same that's been around for nearly half a year, and what most refer to when they say AI is cool but it has issues (because well it does). What some like myself (and Wolfram as well when linking it to Wolfram Alpha) have been raving about is GPT4, which is only currently available with a paid subscription under chatGPT+. It's been so hot in demand that current users are limited to 25 messages per 3 hours, and new subscriptions has actually been postponed due to to high demand.
GPT4 is in another stratosphere compared to GPT3.5. It's hard to explain until you start using it, you notice fairly quick if you switch between both just how much more cohesive and precise it is. It's also a lot slower since it's incredibly taxing.
We are long past the funny AI seinfeld episodes here, and the joke verdict isn't representative of just how good it can be.
Several people have access to the browsing version of chatgpt in alpha (based off of GPT4)
Theoretically, Bing with GPT4 (as claimed) and access to the internet (which gives it a huge advantage) should be the strongest AI option right now, but Microsoft neutered its capabilities so hard in recent weeks it actually produces worst results then the previous Bing without GPT (see the Bing subreddit). It literally shuts the chat on you if you do anything outside of an actual web search question and will outright refuse to do any thinking or work for you.
What everyone is waiting for is the API for GPT4 and GPT4 with browsing capabilities. This offers a good comparison between 3.5 and 4 and why 4 is such a step up, and we haven't seen the full power of GPT4 yet in chatgpt+
To me, it doesn't really matter if the model is infallible, it just matters if the model is as likely to be correct as an expert human. It should be used like any other source of information, you need to cross check against other sources before taking its conclusions as truth. But that is still extremely useful, considering expertise is expensive and this is comparatively cheap, at least to the end user.
Of course people are going to misuse it and trust it for things that it shouldn't be trusted in. That's because people are often flawed, ignorant, and lacking rigorous thought processes. Unlike these narrow AI models, though, they are also unlikely to collectively improve.
Sent from my Pixel 4a (5G) using Tapatalk
I pretty much agree with this, but I don't even need it to be as competent as an expert human for a lot of stuff. I just need it to be as competent as a competent human while also being much, much faster and available to work at any moment. And it can do that now for a lot.
I have GPT-4 set up with a bunch of chats where in each one I've got it playing at being expert in a specific thing that I want done pretty frequently, and I just go to that chat every time I want to use it for that. It's at least good enough at these things that I don't mind it lacking a PhD and am happy with paying all these bots a collective $20/month salary.
I also use Bing as the first stop for most informational searches now, but Microsoft has really made it much less useful than it used to be. Still, between Bing and GPT I would say my interaction with the web has already been dramatically changed.
__________________
"If stupidity got us into this mess, then why can't it get us out?"
The Following 2 Users Say Thank You to JohnnyB For This Useful Post:
It's moving so quickly now that anytime I hear somebody use the phrase "guess my job is safe" I get a bit worried that they aren't taking things seriously. I'm not a doomsday prepper or anything, but it's going from passing the bar at a 10% rate to a 90% rate in months. The improvements we're seeing are not small and a lot of the things it's failing at are in part due to a poor understanding of how it works or bad promtping. Even a month ago artists were saying "ya but midjourney can't even get fingers right"... well now it can, so what's next?
I "created" my first python script with it this morning, and let me assure you, a coder I am not. It was very basic (just a simple action to save me some boring copying and pasting), but even looking back on how I phrased my request, I don't quite understand how it managed to figure out what I wanted. If I could figure that out, we're in for a very interesting few months.
It's moving so quickly now that anytime I hear somebody use the phrase "guess my job is safe" I get a bit worried that they aren't taking things seriously. I'm not a doomsday prepper or anything, but it's going from passing the bar at a 10% rate to a 90% rate in months. The improvements we're seeing are not small and a lot of the things it's failing at are in part due to a poor understanding of how it works or bad promtping. Even a month ago artists were saying "ya but midjourney can't even get fingers right"... well now it can, so what's next?
Quote:
A leading AI researcher has claimed that artificial intelligence is starting to develop beyond an exponential rate.
Jack Clark, who co-founded AI research startup Anthropic after spending four years at OpenAI, said that he believed the exponential progress over the last decade has now transitioned to a new level whereby huge advances will occur extremely quickly.
“A mental model I have of AI is it was roughly ~linear progress from 1960s-2010, then exponential 2010-2020s, then has started to display ‘compounding exponential’ properties in 2021/22 onwards,” he wrote in a Twitter thread.
“In other words, next few years will yield progress that intuitively feels nuts.”
We're rocketing up the hype curve, but I wonder if we will actually see a big trough. Unlike most hyped technologies that promise big things and don't deliver, AI is already delivering amazing things. Maybe we'll all be disillusioned that Skynet doesn't happen.
Absent that caring, we get the AI does not love you, nor does it hate you, and you are made of atoms it can use for something else.
The likely result of humanity facing down an opposed superhuman intelligence is a total loss. Valid metaphors include a 10-year-old trying to play chess against Stockfish 15, the 11th century trying to fight the 21st century, and Australopithecus trying to fight Homo sapiens.
Quote:
Shut it all down.
We are not ready. We are not on track to be significantly readier in the foreseeable future. If we go ahead on this everyone will die, including children who did not choose this and did not do anything wrong.