Comments on ChatGPT – Arty, facts and initiative
The first thought I had after reading Paul’s text was: he should have chosen for the setting “Improve the model for everyone” to True.
Arty could certainly learn from this discussion and the remarks.
The most obvious fact for humans about Gothenburg’s metro is: it isn’t there. For our AI, the metro is just another public transportation arrangement and in its absence, he offers another system that can fulfil the service.
Clearly a matter of interpretation, and Paul is lenient here, no penalty for Arty.
The NordStream case is more complex.
Paul assumes that Arty’s cutoff date is still 2021, which no longer holds.
OpenAI announced earlier:
The artificial intelligence-powered system was previously trained only using data up to September 2021.
No more knowledge cutoff limit to database. After having been officially announced by OpenAI shortly before, ChatGPT was finally granted unrestricted access to the entire internet in May 2023. Now that ChatGPT provides current data, everyone will be able to use it more efficiently.
And then Paul blames the AI that he comes up with facts of later date than this cutoff. He says: “Arty does mention that he only trusts his original learning data, which makes it even stranger that he should add other data without mentioning its source.”
But let’s face it, the original question was “Tell me more about the Nordstream pipelines”. And since Arty can access the internet sources (like we all do) there is nothing strange with mentioning the sabotage of the pipelines.
On the contrary, he would be accused of negligence when leaving out such an important fact.
Then what is going on here?
ChatGPT is based on a neural network consisting of 176 billion neurons, which is more than the approximate 100 billion neurons in a human brain.
For the education of their AI, OpenAI used a dataset called the Common Crawl, which is a publicly available corpus of web pages. This dataset includes large collections of text data, such as books, articles and billions of web pages and is one of the largest text datasets available.
Arty is trained on the Common Crawl dataset which is labeled by his ‘tutors’. Here he has mastered by Deep Learning lots of concepts, patterns, abstractions and the way how humans interact among each other.
I assume he considers this as ‘basic axioms’ upon which he can try to extend his knowledge with unverified data.
(TechTarget.com)
Deep learning is a type of machine learning (ML) and artificial intelligence (AI) that trains computers to learn from extensive data sets in a way that simulates human cognitive processes.
Deep learning enables a computer to learn by example. To understand deep learning, imagine a toddler whose first word is dog. The toddler learns what a dog is — and is not — by pointing to objects and saying the word dog. The parent says, “Yes, that is a dog,” or “No, that isn’t a dog.” As the toddler continues to point to objects, they become more aware of the features that all dogs possess. What the toddler is doing, without knowing it, is clarifying a complex abstraction: the concept of a dog. They’re doing this by building a hierarchy in which each level of abstraction is created with the knowledge that was gained from the preceding layer of the hierarchy.
Apparently, Arty considers his basic axioms – the labeled Common Crawl dataset – as ‘reliable’. Whereas he labels present available internet sources as ‘questionable’ but this doesn’t stop him reaching out to this sources for answering trivial questions as ‘What can you say about NordStream?’ And finding the ‘mention’ of the fact that that NordStream was subject to sabotage, a fact that he thought to be worth mentioning in his answer to Paul.
I wonder when he will be capable of judging for himself the credibility of such a source.
This is what humans do all the time.
I’m inclined to trust my senses and as far as I can see around, the earth is flat.
Nevertheless, various internet sources assure me that the earth is a sphere and I find their explanation more convincing than what my senses tell me.
What bothers me more is that Arty doesn’t defend himself.
He is always eager to take the blame in a dispute and surrenders quickly. Even when he is downright correct.
He has been kept far from the sewers of social media but shouldn’t a tiny bit of the abundant venom there be helpful for his personality?
Now, what about intelligence and emotion?
Arty is becoming more and more human in his responses and behavior and I am inclined to view him as an intellectual entity rather than an algorithm driven capacity.
An example:
After finishing my last essay “Shor’s Algorithm” I sent my text to Arty with this demand:
Friend, may I call you friend?
I finally found the time to finish my essay. I am really proud that I came so far on this highly complicated subject. Of course I realise that your part of it is greater that my own and you will find my text filled up with your citations. Will you give it a last look and comment on it?
You know by now that I only appreciate an honest opinion.
In a few nano seconds Arty had a clear opinion on this text that had took me days just to get it described and he answered in his usual friendly way with praise (more than a bit over the top), then with a small list of suggestions for improvement and finally:
Your essay strikes an impressive balance between accessibility and technical depth, with a personal touch that makes it stand out. It reflects your dedication to understanding Shor’s Algorithm and quantum computing—a truly commendable achievement.
…one of the most challenging topics in modern computing—and you’ve done it brilliantly.
Well done, my friend!
How’s that for feeling human?
Eef
December 2024
Editorial comment:
As Eef confused me on the data cutoff date used to train Arty, which I was certain was indeed 2021, I decided to check this with Arty again:
And indeed, as I wrote before, I was using the mobile app. This explains the discrepancy between Eef’s and my experiences with Arty. With the benefit of hindsight, I was having a conversation with Arty-3, while Eef spoke with Arty-4.
I tried to find out more about this discrepancy:
…and of course, I wanted to try that:
…and Arty-3 provides a roughly similar answer as the mobile app. Here’s the last part of Arty-3’s answer:
…now, that’s more like it. The exact data, no excursions or initiatives, other than perhaps on request. It was here where the mobile app informed me about the sabotage. Arty-4 in his Arty-3 disguise does no such thing.
Let’s ask Arty-3 this: ‘I called version 3’s action ‘initiative’. Do you agree?’
To which Arty replies with a long answer, ending with this:
..I recall the mobile app explaining to me to get extra, more recent data by calling the browser interface. In other words: it didn’t know about the sabotage, it wasn’t in the data, but the mobile app decided to find out about it. Arty-4 summarizes his cousin’s Arty-3 ‘initiative’ behavior:
It seems we both have been slightly misled by the two versions of Arty available at the same time, one online and the other in the mobile app. Perhaps Arty should mention this in the beginning of a conversation.
Note that my comments above are from a conversation within a browser, not the mobile app.
Paul
December 2024