More likely due to lack of "good" data than to existence of "bad" data. ChatGPT is know for its ability to "hallucinate" answers for questions that it wasn't trained for.
In fact ChatGPT doesn't know anything about true and false. It's just generating text that most closely resembles text it's seen on similar subjects.
E.g. ask it about the molecular description for anything. It'll start with something fundamental like the CH3N4 etc then describe the bonds. But the bonds will be a mishmash of many chemical descriptions thrown together. Because similar questions had that kind of answer.
The worst part is, it blurts forth with perfect confidence. I liken it to a blowhard acquaintance that will make up crap about any technical subject they have a few words for, as if they are an expert. It's funny except when somebody relies on it as truth.
I don't think GPT3 at its heart is an expert at anything. Except generating likely-looking text. There's no 'superego' involved anywhere that audits the output for truthfulness. And certainly no logical understanding of what it's saying.
People have taken to asking ChatGPT to create entire scripts to trade money. When they don't work, they go into chatrooms or forums and ask "why doesn't this work" without saying it was made by ChatGPT. It causes people to open the post, read it a bit and only maybe after a minute or two of wasted time, realize the script is complete nonsense.
Why? chatgpt has certainly consumed seo spam and company marketing materials as part of it's model. Even if a human went through it, there still exists a bias towards this information. After all, this material is specifically written to fool humans.
I've played with chatgpt enough to notice that for some queries it's fundamentally doing an auto-summarize of such content.
Consider this. Someone very early posted that a neat feature of chatgpt would be to give chatgpt a list of ISBN numbers and then demand it's answers are cited from this corpus. We're not there yet but this would be amazing.
My prediction is that those with money will have power to influence their chat bot. Consequently, they'll have access a higher-quality and wider corpus of information. There will not be any restrictions on how chatgpt would answer due to for example, woke agendas. Also, players such as Goldman Sachs would feed their model content generated by their analyst that consumers would not have access to. This already happens but chatgpt will make this information so much more potent.
Furthermore, as this technology continues to improve it will increase the productivity of our population and ultimately generate higher GDP. I'm super excited.
> Consider this. Someone very early posted that a neat feature of chatgpt would be to give chatgpt a list of ISBN numbers and then demand it's answers are cited from this corpus. We're not there yet but this would be amazing.
It currently has the ability to do this. It'll make the citations up, of course – but that behaviour is inherent to the architecture; a system that didn't do that would have to work differently at a fundamental level.
> chatgpt will make this information so much more potent.
How do you imagine this would work?
> and ultimately generate higher GDP.
Again, how do you imagine this would work? GDP is a specific economic measure; how would (a better version of) this technology increase GDP?
Tangentially: why is "increase GDP" a good ultimate goal to have in the first place?
>> chatgpt will make this information so much more potent.
> How do you imagine this would work?
Don't overthink it. It's just the nature of the tool. Imagine you're a detective trying to investigate a crime,
- "list the plates of blue hondas in this area at this time, that have a missing rear bumper and a scratched driver side door"
- "send a notifications to all gas stations along this route and notify them of a blue honda"
And, if you're a Goldman Sachs analyst, you can just use natural language to gather information. "i have this scenario, list companies that will benefit" would be an abstract question that you'd ask it. Obivously, the system isn't this good yet but you get the idea. You'd just have to ask more fine grained questions and use some of your domain knowledge to fill the gap until it does become this good.
>> and ultimately generate higher GDP.
> Again, how do you imagine this would work? GDP is a specific economic measure; how would (a better version of) this technology increase GDP?
Google (or chat gpt) would do a better job than me answering this,
"Increases in productivity allow firms to produce greater output for the same level of input, earn higher revenues, and ultimately generate higher Gross Domestic Product."
The reason you want to increase gdp... the following quote was derived from one of Herbert Hoover’s memoirs.
"[Engineering] It is a great profession. There is the satisfaction of watching a figment of the imagination emerge through the aid of science to a plan on paper. Then it moves to realization in stone or metal or energy. Then it brings jobs and homes to men. Then it elevates the standards of living and adds to the comforts of life. That is the engineer’s high privilege."
By increasing GDP, you elevate the standard of living and add to the comfort of life.
> "list the plates of blue Hondas in this area at this time, that have [...]"
I think this shows a significant misunderstanding of what chatgpt does fundamentally. It will never be able to do this unless also fed a description, location, and time of cars in a certain area as context beforehand(either as training data or a prompt). In either case you have access to the data and just need to do a simple search, so chatgpt is providing negative value since it's capable of providing results that don't exist in the dataset.
Similarly for your Goldman Sachs example, you're imagining that chatgpt is greater than it is. It is capable of providing something that would likely follow a given text on the internet at its time of training(aka it's training set) somewhere. It can't reason about new information or situations since it's incapable of reasoning. To believe that it could generate business strategies is to believe that effective business strategies don't require any intuition or reasoning to progress, just statistical recombination of existing strategies.
> By increasing GDP, you elevate the standard of living and add to the comfort of life.
How do you reach this conclusion from the information presented? Why use GDP, a measure of the profitability of corporations, as a proxy for the standard of living instead of measuring the standard of living and seeing how it will be impacted directly instead of through many layers of abstraction.
>>How do you reach this conclusion from the information presented? Why use GDP, a measure of the profitability of corporations, as a proxy for the standard of living instead of measuring the standard of living and seeing how it will be impacted directly instead of through many layers of abstraction.
You are asking a question that is outside of scope here. GDP per capita has been used as a proxy for standard of living for quite some time now.
That proxy only works as long as nobody's optimising for it.
> Any observed statistical regularity will tend to collapse once pressure is placed upon it for control purposes. — Charles Goodhart
GDP (£) per capita in London has doubled since 1998. Has the standard of living "doubled" for the median person? What about the standard of living for the poorest 1%? Has the productivity boost due to automation translated into correspondingly shorter working hours, or correspondingly larger compensation for work done?
What questions do you actually mean to ask, when you talk about GDP?
If something stops being an economic transaction it moves out of GDP. So if ChatGPT reduces Google ad clicks then it doesn’t seem like it would increase it, even though it does increase customer surplus (stuff you get for free).
