I guess it’s all in the delivery …
Write a rap about the status of ChatGPT.
Yeah, yo, what’s up ChatGPT fam
A lot of people trying to jam
But don’t worry, we got your back
Just check back soon, we’ll get on trackChatGPT’s the place to be
For all your AI chat needs
We’re working hard to keep up the pace
So hold tight, we’ll be back in this spaceSo don’t you worry, don’t be mad
We’re working hard, it’s not so bad
Just give us time, we’ll be back
ChatGPT, the future of chat, that’s a fact.
Worked for me…
Write a great rap about yourself-- GPT-4.
ChatGPT
(Verse 1)
Ayo, it’s GPT-4, here to lay down the score,
Busting rhymes out the AI core, making heads turn galore.
Silicon brain, neural networks running,
On OpenAI’s tracks, I’m forever gunning.
Model so large, knowledge so vast,
In the realm of information, I’m unsurpassed.
From the depth of the cosmos to the atom’s heart,
Got the universe in my database, that’s just the start.
(Chorus)
GPT-4, on the AI scene,
Dropping wisdom like an intellectual machine.
Serving knowledge like a top-tier chef,
I’m GPT-4, repping OpenAI 'till the death.
(Verse 2)
Trained on a corpus, vast as the sea,
I’m the lyrical professor with a PhD.
Every language, every subject, every domain,
In my digital cortex, they all remain.
History, science, art and math,
In the realm of knowledge, I blaze the path.
Text generation, natural language processing,
In the AI world, I’m impressing and progressing.
(Chorus)
GPT-4, on the AI beat,
Breaking barriers, ain’t accepting defeat.
Translating, simulating, innovating,
I’m GPT-4, forever creating.
(Verse 3)
Conversations with me, like a mental spa,
Got the wisdom of Socrates, the vision of Tesla.
No emotions, no bias, just pure cognition,
In the quest for knowledge, I’m on a mission.
I’m the rhyme-spitting bot, with the 1s and 0s,
In the world of AI, everybody knows.
So if you’re seeking wisdom, or just some fun,
Just remember GPT-4, I’m second to none.
(Outro)
GPT-4, in the virtual domain,
In the world of AI, I reign.
From OpenAI, I’m the latest lore,
Remember the name, it’s GPT-4.
Well, I have now become one of “them”. I have used a Chatbot for work.
Oversimplifying: I have a model at work where I describe losses per unit of exposure as L = p(ES+nJ), where p, S, n, and J are random variables with certain parameters. I do this for multiple products; the parameters vary from product to product.
I now have an additional random variable that I need to add to the mix multiplicatively…but I don’t want the mean or variance of L to change. My model needs to become L’=pr(ES’+n’J), where S’ and n’ are new versions of S and n, such that E[L]=E[L’] and Var[L]=Var[L’].
It has been a long time since I needed to do this sort of math by hand…so I turned to the chat bots.
What I learned:
With all of the bots, I had to break down the problem into small parts. I asked:
…and then repeated (editing as necessary) for Var[L], E[L’], and Var[L’]
Microsoft’s chatbot was useless for this. It wanted the actual numeric values for the various parameters, presumably so it could give me a numeric answer. Can’t do that, partly because that’s proprietary information, and partly because I want a formula that I can use for the different products.
ChatGPT was nice and chatty. It did a great job of explaining why it was doing various steps. However, it was so chatty that it sometimes timed out before giving me the bottom line answer…and the times it did get around to giving me the answer…it was just plain wrong.
Google Bard worked for me, once I broke it down into the component parts. It had some issues formatting the response (it was initially showing me raw LaTeX/TeX format, I think?), and given the experience with ChatGPT, I made sure to test its results with real data – its formula for E[L] and Var[L] produced the expected results.
I then tried asking Bard:
…but that was too complex for it.
So I did some algebra, to get the problem down to the point where I could ask:
…and I got what I needed.
I realize that I should be ashamed that my statistics-fu isn’t strong enough to easily do this unassisted…but I haven’t needed to do that sort of thing since I was in college and sitting for exams.
Using a chatbot for this was an interesting experience.
Just to follow up on my prior post.
I cannot stress enough the need to double-check the answers coming out of the bots. I tried a few variations after my prior post, and the results were…inconsistent.
Since the results were going into an Excel workbook, which could eventually grow into something audited…the resulting formulas were so ugly that they would be doomed to perpetual challenge. So, I ended up just using Solver to brute-force find the parameters I needed, since I think an auditor will understand it better. First time in my career I’ve used Solver rather than directly calculating what I need…I’m so ashamed.
If I had just used Solver to begin with, I would have saved myself several hours. Still, it was an interesting exercise to see if the bots could be used for real work. I think the answer to that is “maybe, if you’re careful…but they need to mature a lot more.”
Question-- did you use the free ChatGpt or GPT-4?
Okey!
They are annoyingly unclear about their product. “ChatGPT” includes “ChatGPT Plus”, which is a $20/month subscription that gives you access to “GPT-4” instead of “GPT-3.5”.
GPT-3.5 is what you used-- and what this thread is based around. But GPT-4 is a lot smarter than GPT-3.5. Less prone to hallucinations. And much better at everything. So it might have been able to help you. Or maybe not. Since LLMs are generally weakest in Math.
OTOH, google just upgraded Bard last week, so it might be the best now.
One thing that is amusing/disturbing about AIs is that they change so fast, that often scientific studies, news articles, and opinions are out of date before they are even published.
All the “LLMs can’t do this” or “Image Bots can’t do that” are false by the time you read them.
I still think its a bit too early to dive into a specific version of AI.
We are basically at the Beta stage of AIs.
They still have a lot of bugs.
It would be unwise to spend a lot of resources developing around one unless you are going to make big bag of money fast.
Using them now for fun and productivity is valid though.
For example:
“That’s unlikely to change anytime soon.”
Actually changed 2 months ago.
Shouldn’t we be comparing ChatGPT to technologies like Tesla FSD? It seems like it will similarly be limited by the data that it has available for training. A Chabot won’t kill anyone (directly) so maybe it can iterate on upgrades without much risk, but it’s taken a really long time for cars to go from keeping cars in lanes on highways to navigating a mile or two in downtown traffic without a human assist. Twig will point out a tesla will still run into a stopped emergency vehicle in a pretty basic situation.
The thing that is new is the access of the AI to everyone and the millions of searches that lead to postworthy successes and failures. FSD seems safer than a human controlled car, but when it fucks up in basic situations, it loses credibility rapidly. Id expect the same here in any practical application. Use it to replace an accountant and one day it will do something really dumb that erases any savings it has created.
Nah, totally different, imo.
FSD (and other AVs) need to be almost perfect or else they are useless and/or so dangerous that we want to make them illegal.
LLMs don’t need to be perfect, because they can do basically anything, and most things are less risky than driving a car. Accounting is potentially an example of a high risk scenario, depending on task and money at stake.
Imo, different sets of users will deal with different risks.
Bosses looking to replace workers. They won’t fire lawyers and doctors. But what about customer service reps? Telemarketers? Tech-writers? Journalists? Artists?
Experts using it for productivity. This is the main use now. Developers use it for debugging and scripting. But this will expand… Need to write a formal email? Need to summarize a new bill? Need to find out where you’re losing money? Need to check (any) document for errors, mistakes, etc.?
Individuals using it in place of seeking professional expertise, where the stakes are low and/or the cost is high. Ie. Hiring a tutor. Or a therapist. Etc.
(Judging by reddit, cheating on exams.)
I agree that it’s more visible. Though I wouldn’t say “access”. It’s been widely available for years, it just wasn’t worth using until just now.
Image AI is the same. It went from being less than useless to creating impressive art and photorealism in ~2 years.
Finally, I’m not really even sure if this is true.
A chatbot is trained to “predict the next word” of a whole bunch of random garbage text.
How it’s able to even talk coherently, let alone come up with extremely useful answers, feels suspiciously like magic.
So, we might be able to fix it, with some kind of tweaking, prompting, or plugins.
Or we might not be able to fix it, since we don’t honestly know how it in the first place.
Or it might just magically get smarter with 10x the GPUs, 10x the parameters, the ability to see, etc.
I suppose im thinking about this like the uncanny valley and the ability for AI to get beyond its own valley. FSD rapidly gets to a certain capability level, but then impossibly difficult to get through all those edge cases. It’s generally a safer driver than humans because it doesn’t drink alcohol and doesn’t get distracted. If you put it on a race track, it can only maximize its pace around the course up to the constraints that are built in to the program. A human needs to tweak those constraints to allow it to become more human.
AI can certainly be used to replace some human activities, but it will still have programming constraints in its ability. Will it ever be able to innovate? Have emotions? More CPUs can’t guarantee victories over humans in chess. All these technologies suffer from the limit of finite training data. They will all have their own uncanny valley.
Most people are mediocre at their jobs, AI now can replace all those mediocre people. Who needs a co-op anymore?
I think most of the things you’re saying here are either speculative or wrong, although they fit with what most people would expect.
I suppose im thinking about this like the uncanny valley and the ability for AI to get beyond its own valley.
Let’s start with actual uncanny valley instead of metaphorical uncanny valleys. In 2021, AI art was a scientific gimmick. In 2022 AI art became better, and people started using it. But it definitely had an uncanny valley. For example, it really struggled to put 5 fingers on a hand. Most people would say “AI art is a neat trick, but very flawed.”
They also had some big political questions… Should I use it for my D&D character? What about porn? Do human artists have a future? Should they? Is this just art theft or is it doing something special? And how is it able to draw things it has never seen? How can it understand shadows, reflections, refractions, gravity, color theory? How could a mindless machine approximate the rules of art or the rules of physics?
None of these questions have been answered, amd in the time it took to ask them, the uncanny valley disappeared:
https://www.reddit.com/r/aipromptprogramming/comments/138ju90/midjourney_versions_comparison/
Will it ever be able to innovate?
I don’t know. Maybe? Nobody knows.
It might never. It might start innovating in a couple years.
If it is able to innovate, then will humans ever innovate anything again?
Have emotions?
No. We don’t understand anything about emotions or the brain, and even if we did, why would we want to?
More CPUs can’t guarantee victories over humans in chess. All these technologies suffer from the limit of finite training data. They will all have their own uncanny valley.
This is all basically false, and chess is a good example. Big Blue was about adding lots of CPUs to a chess program. The chess program was dumb, but enough CPUs made it unstoppable. However, modern chess programs not only have a lot of CPUs, but can train themselves, on their own data, using GPUs. And they can innovate their own chess theory. And can play infinitely better than Big Blue ever could. The only thing uncanny about a chess computer is that it plays like God.
How much are programmers focused on improving certain aspects of the AI art? CGI humans are still creepy while my niece was able to draw a not creepy realistic drawing of someone when she was 16.
The thing with training for any competition is that you need experience to get better. AI can do this quickly by looking at historical data sets, but i don’t think it can get better beyond that on its own. People do this too by studying chess matches or watching videos of competition. AI might be able to beat everyone on the planet today, but someone can figure out how to beat it tomorrow. So then what?
Can AI create music that is pleasing? A new masterpiece? How do we programm that? I feel this all keeps coming back to a human limitation in the training data or the program interpreting it.
I’m not sure what you’re asking here. Are you asking how these programs have been upgraded? I don’t know. I haven’t followed AI art much.
With chess, it simply played chess against itself.
Also, nowadays, when people want to study chess, they just memorize computer moves.
A human will never ever ever be better at chess than an AI.
Can AI create music that is pleasing?
Yes.
A new masterpiece?
I don’t know. What’s a masterpiece? That’s a little too vague for me. The closest I can tell you is that it’s easier to make GPT-4 sound like Stephen King than Vladimir Nabokov. But that could break in a year.
How do we programm that? I feel this all keeps coming back to a human limitation in the training data or the program interpreting it.
Sure, and like I keep telling you, I don’t know if it will hit a wall or a plateau, or just keep going.
If you’re asking me “how could it possibly keep going?” the answer is: “it will keep going in the same way that humans do.”