Comment by virgildotcodes
2 years ago
I really don't understand why anyone writing articles about ChatGPT uses 3.5. It's pretty misleading as to the results you can get out of (the best available version of) ChatGPT.
For comparison, here are all the author's questions posed against GPT4:
https://chat.openai.com/share/ed8695cf-132e-45f3-ad27-600da7...
> I really don't understand why anyone writing articles about ChatGPT uses 3.5.
Because that’s what most people have access to. It’s absolutely worthless to most readers to talk about something they’ll never pay for and it’s not the job of random third-parties to incentivise others to send money to OpenAI.
What I really don’t understand is why anyone gets so hung up about it and blames the writer. If you’re bothered by people using 3.5 you should complain to OpenAI, not the people using the service they make freely available.
Anecdotally, I find this excessive fawning about 4 VS 3.5 to be unwarranted.
https://news.ycombinator.com/item?id=38304184
> Because that’s what most people have access to.
I’d agree with this rationale if the author clearly communicated their choice of model and the consequences of that choice upfront.
In this post the table of results and the text of the post itself simply reads “ChatGPT” with no mention of 3.5 until the middle of a paragraph of text in the appendix.
> It’s absolutely worthless to most readers to talk about something they’ll never pay for and it’s not the job of random third-parties to incentivise others to send money to OpenAI.
The “worth” is in communicating an accurate representation of the capabilities of the technology being evaluated. If you’re using the less capable free version, then make that clear upfront, and there’s no problem.
If you were to write an article reviewing any other piece of software that has a much less capable free version available in addition to a paid version, then you would be expected to be clear upfront (not in a single sentence all the way down in the appendix) about which version you’re using, and if you’re using the free version what its limitations may be. To do otherwise would be misleading.
If you simply say “ChatGPT” it’s reasonable to infer that you’re evaluating the best possible version of “ChatGPT”, not the worst.
Accurate communication is literally the job of the author if they’re making money off the article (this one has a Patreon solicitation at the top of the page).
Whether or not "most readers" are ever going to pay for the software is totally orthogonal.
If using GPT4 vs 3.5 would create results so distinct from one another that it would serve to incentivize people to give money to OpenAI, well then that precisely supports the argument that the author’s approach is misleading when presenting their results as representative of the capabilities of “ChatGPT”.
> What I really don’t understand is why anyone gets so hung up about it and blames the writer.
Again, if they’re making money off their readers it’s their job to provide them with an accurate representation of the tech.
> Anecdotally, I find this excessive fawning about 4 VS 3.5 to be unwarranted. https://news.ycombinator.com/item?id=38304184
Did some part of my comment came across as “excessive fawning”? Regardless, if this “excessive fawning” is truly unwarranted, this would again undermine your statement that using GPT4 would “incentivize others to send money to OpenAI”.
In regards to your link, I’ll highlight what another commenter replied to you. What should ChatGPT say when prompted about various religious beliefs? Should it confidently tell the user that these beliefs are rooted in fantastical nonsense?
It seems in this case you’re holding ChatGPT to an arbitrary standard, not to mention one that the majority of humanity, including many of its brightest members, would fail to meet.
> I’d agree with this rationale if the author clearly communicated their choice of model and the consequences of that choice upfront. (…) with no mention of 3.5 until the middle of a paragraph of text in the appendix.
You’re moving the goalposts. You went from criticising anyone using 3.5 and writing about it to saying it would’ve been OK if they had mentioned it where you think it’s acceptable. It’s debatable if the information needed to be more prominent; it is not debatable it is present.
> If you simply say “ChatGPT” it’s reasonable to infer that you’re evaluating the best possible version of “ChatGPT”, not the worst.
Alternatively, it you simply say “ChatGPT” it’s reasonable to infer that you’re evaluating the version most people have access to and can “play along” with the author.
> If using GPT4 vs 3.5 would create results so distinct from one another that it would serve to incentivize people to give money to OpenAI
Those are your words, not mine. I argued for the exact opposite.
> Again, if they’re making money off their readers it’s their job to provide them with an accurate representation of the tech.
I agree they should strive to provide accurate information. But I disagree that being paid has anything to do with it, and that their representation of the tech was inaccurate. Incomplete, maybe.
> Regardless, if this “excessive fawning” is truly unwarranted, this would again undermine your statement that using GPT4 would “incentivize others to send money to OpenAI”.
Again, I did not argue that, I argued the opposite. What I meant is that even if you believe that to be true, that still doesn’t mean random third-parties would have any obligation to do it.
> I’ll highlight what another commenter replied to you.
That comment has a reply, by another person, to which I didn’t feel the need to add.
> It seems in this case you’re holding ChatGPT to an arbitrary standard, not to mention one that the majority of humanity, including many of its brightest members, would fail to meet.
Machines and humans are not the same, not judged the same, don’t work the same, are not interpreted the same. Let’s please stop pretending there’s an equivalence.
Here’s a simple example: If someone tells you they can multiply any two numbers in their head and you give them 324543 and 976985, when they reply “317073642855” you’ll take out a calculator to confirm. If you had done the calculation first on a computer, you wouldn’t turn to the nearest human for them to confirm it in their head.
The problem with ChatGPT being wrong and misleading isn’t the information itself, but that people are taking it as correct because that’s what they’re used to and expect from machines. In addition, you don’t know when an answer is bullshit or not. With a human, not only can you catch clues regarding reliability of the information, you learn which human to trust with each information.
Everyone’s standard for ChatGPT, be it absolute omniscience, utter failure, or anything in between, is arbitrary. Comparing it to “the majority of humanity, including many of its brightest members” is certainly not an objective measurable standard.
1 reply →
Why does OpenAI continue to offer chatgpt 3.5 if it's so bad?
GPT 4 is THIRTY (30) times more expensive.
In the llm-assisted search spaces I'm involved in, a lot of folks are trying to build solutions based on fine tuning and support software surrounding 3.5, which is economical for a massive userbase, using 4 only as a testing judge for quality control.
Chatgpt3.5 is good enought if can give context in query.
Cheaper and faster.
It's a bit hard to use for most, either $20/month fixed for a limited # of messages, or you need to be able to reason through how to get an API key, or get another 3rd party service with similar cost & limits.
You can use GPT-4 for free via Bing - though I find it a little hard to explain to people how they can do that because I'm never sure what the rules are with regards to creating Microsoft accounts, whether you can use any browser or have to use Edge, what countries it's available in etc.
Actually maybe the recommendation should be to use GPT-4 for free via https://copilot.microsoft.com/ instead now.
(Except I can't tell which version of GPT that's using yet - there was a story on 5th December that said GPT-4 Turbo was "coming soon", not sure when "soon" is though: https://blogs.microsoft.com/blog/2023/12/05/celebrating-the-... )
FYI: Balanced doesn't run pure GPT4. Balanced uses a combination of multiple models. Precise and Creative is pure GPT4.
About GPT4 Turbo, to check if you are on Turbo, ctrl+U > ctrl+f > check if "dlgpt4t" exists. If it exists, you are running turbo.
You can also double-check by, well, asking stuff after 2021 knowledge cut-off as well ("What are the oscar winners?") with search disabled.
But you'll notice because turbo is much faster on bing (and better too).
But that gpt-4 says it can't code
IMHO TBF the "limited # of messages" is continously increasing, to the point I hardly remember it exists these days