Jump to content

ChatGPT is bullshit


Recommended Posts

And now there’s a journal article declaring this to be the case

https://link.springer.com/article/10.1007/s10676-024-09775-5

We argue that these falsehoods, and the overall activity of large language models, is better understood as bullshit in the sense explored by Frankfurt (On Bullshit, Princeton, 2005): the models are in an important way indifferent to the truth of their outputs.

Link to comment
Share on other sites

9 minutes ago, swansont said:

And now there’s a journal article declaring this to be the case

https://link.springer.com/article/10.1007/s10676-024-09775-5

We argue that these falsehoods, and the overall activity of large language models, is better understood as bullshit in the sense explored by Frankfurt (On Bullshit, Princeton, 2005): the models are in an important way indifferent to the truth of their outputs.

Well, we've all seen on this forum ample evidence of that. This claim not only seems true, but also both very funny, and a timely puncturing of the bubble of hype surrounding these verbose  and fundamentally unintellligent programs.  

I realise that AI encompasses a far wider scope than LLMs but, as they stand today, LLMs look to me pretty meretricious. 

It may be that their chief legitimate use is in collating references for the user to determine, for himself, which one are good and which ones are not, i.e. just a superior kind of search engine.

Link to comment
Share on other sites

Which model one uses and how they prompt it are extremely relevant. 

Not all LLMs are GPTs nor are all GPTs at the same version nor trained on the same dataset(s). 

4 hours ago, swansont said:

the models are in an important way indifferent to the truth of their outputs.

So are approximately half the voting populace. 

Edited by iNow
Link to comment
Share on other sites

2 hours ago, StringJunky said:

The Brave browser version puts the references at the bottom, like Wiki.

Yeah, honestly getting the actual source and not a probabilistic hallucination would not be that much additional code/memory.

It's like the math issue. It's not hard for a computer to do math correctly, but someone still needs to be arsed to program in the ability.

Link to comment
Share on other sites

4 minutes ago, Markus Hanke said:

There’s free AIs out there that do this, for example Perplexity.

Yes, I think they either programmed GPT for sheer speed or don't have the training database set up in such a way thay it can find actual references.

You can't really store links and similar data easily using default token system. Beyond a few pieces they're all too random.

Did see where it Rickrolled one guy though, so who knows.

Link to comment
Share on other sites

2 hours ago, Endy0816 said:

Yeah, honestly getting the actual source and not a probabilistic hallucination would not be that much additional code/memory.

It's like the math issue. It's not hard for a computer to do math correctly, but someone still needs to be arsed to program in the ability.

Actually, references should ALWAYS be quoted. It's a no different requirement for LLM's or a theoretical proper AI than a human  researcher. Evidence, evidence, evidence. 

Edited by StringJunky
Link to comment
Share on other sites

Quote

ChatGPT is bullshit

With ChatGPT v3.5, you can ask the same question in two different ways or in two different languages and get completely different answers.. The result is completely unreliable and even dangerous if one is not aware of how it works, and believes everything without any doubt (like the typical people using it)..

Link to comment
Share on other sites

4 hours ago, StringJunky said:

Actually, references should ALWAYS be quoted. It's a no different requirement for LLM's or a theoretical proper AI than a human  researcher. Evidence, evidence, evidence. 

Most of the current LLMs are not researching anything they write.

 

Edited by Endy0816
Clarification
Link to comment
Share on other sites

10 hours ago, iNow said:

Which model one uses and how they prompt it are extremely relevant. 

Not all LLMs are GPTs nor are all GPTs at the same version nor trained on the same dataset(s). 

So are approximately half the voting populace. 

I think perhaps the best use for this tool is in understanding the languages of other species, and perhaps our best chance of meaningful dialogue if the alien's do get in touch.

People trying to look smarter than they are, always trip themselves up, bc it's only a tool if they know how to use it.

3 hours ago, Sensei said:

With ChatGPT v3.5, you can ask the same question in two different ways or in two different languages and get completely different answers.. The result is completely unreliable and even dangerous if one is not aware of how it works, and believes everything without any doubt (like the typical people using it)..

Both will evolve... 😉

Link to comment
Share on other sites

10 hours ago, iNow said:

Not all LLMs are GPTs nor are all GPTs at the same version nor trained on the same dataset(s). 

But the underlying issue is the datasets. The algorithm can’t discern veracity of information; it relies on what it’s fed, and those choices are made by humans. The “AI” isn’t intelligent. It’s not thinking. It’s just a fancy search engine.

Link to comment
Share on other sites

1 hour ago, Endy0816 said:

Most of the current LLMs are not researching anything they write.

 

Where are they getting their sources? Is it just a mish-mash of stuff they hold locally and collated from that?

Edited by StringJunky
Link to comment
Share on other sites

1 hour ago, swansont said:

But the underlying issue is the datasets. The algorithm can’t discern veracity of information; it relies on what it’s fed, and those choices are made by humans. The “AI” isn’t intelligent. It’s not thinking. It’s just a fancy search engine.

There's nothing wrong with using a tool, if it's used correctly...

Link to comment
Share on other sites

1 hour ago, swansont said:

The algorithm can’t discern veracity of information; it relies on what it’s fed, and those choices are made by humans. The “AI” isn’t intelligent. It’s not thinking. It’s just a fancy search engine.

I mostly align with your central point, but there is a lot work happening which allows you to use other AIs to evaluate the accuracy of the prompted tool and they seem rather effective. The basic idea is you have AIs with expertise in certain spaces and use those to evaluate what is returned prior to showing it to the user. 

1 hour ago, StringJunky said:

Where are they getting their sources? Is it just a mish-mash of stuff they hold locally and collated from that?

It depends on the LLM and the relationships they've established with other companies. You may have seen big news somewhat recently about OpenAI forming an agreement with newspapers and publishers like Washington Post (or getting sued by the NYTimes), and others inking deals with Reddit to train on their vast troves of data, for example.

The training corpus for these models varies, and obviously companies like Google for Gemini and Meta for Llama 3 (or even X/Twitter model Grok) have a much larger pool from which to work than some of the smaller players. 

Link to comment
Share on other sites

Adding to the above, looks like Claude 3.5 is the new best in class LLM. It’s only been out for a few hours now.

Alex Albert from Anthropic pointed out that on GPQA (Graduate-Level Google-Proof Q&A) Benchmark, they achieved a 67% with various prompting techniques, beating PHD experts in respective fields.

The training data seems to be current at least as of February this year (GPT-4 is only current to somewhere in early 2023 IINM), and it has better vision capabilities than GPT-4o (the Scarlett Johansen one with lots of buzz a week or two ago). 

They also shared this on their release blog:  “In an internal agentic coding evaluation, Claude 3.5 Sonnet solved 64% of problems, outperforming Claude 3 Opus which solved 38%.”

 

 

Link to comment
Share on other sites

Create an account or sign in to comment

You need to be a member in order to leave a comment

Create an account

Sign up for a new account in our community. It's easy!

Register a new account

Sign in

Already have an account? Sign in here.

Sign In Now
×
×
  • Create New...

Important Information

We have placed cookies on your device to help make this website better. You can adjust your cookie settings, otherwise we'll assume you're okay to continue.