what CIOs need to know in 10 points

Deriving from a 3.5 evolution of the famous GPT model, ChatGPT once again illustrates the astonishing potential of new generative AIs. A conversational intelligence that says a lot about the progress made and the current potential of AI but which must also alert CIOs to the uses, limits, abuses and risks of such intelligences…

HAL, Terminator, David (from Steven Spielberg’s AI), Samanta (Her), C3PO (Star Wars)… All these cinema AIs from the world of SF stand out for their astonishing ability to dialogue with humans and to answer their question sometimes by asking other questions.
Since last week, such intelligences no longer belong entirely in science fiction.

In less than a week, more than a million users have already exchanged with ” ChatGPT », the new production of OpenAI. His abilities send shivers down the spine as much as they amaze. And cause a lot of ink to flow.

But between the fabrications, the superlatives, and the concerns of each other, what does ChatGPT teach CIOs concretely and how can they find a use for it in the company? Here are 10 keys to understanding everything…

1/ ChatGPT derives from GPT 3.5

OpenAI is known for its numerous experiments around artificial intelligence and more particularly in terms of generative AI (GANs, Generative Adversarial Networks, and GPTs, Generative Pretrained Transformer). We owe him two famous models in particular: DALL-E and GPT-3.
Give him (which is in version 2) is an AI used to generate artistic creations, photos or images from words. It is notably integrated into Microsoft Designer and will soon be integrated into Bing and Edge.
GPT-3 and GPT-3.5 are very large linguistic models trained from massive sources of text and source code. These AIs are notably used by GitHub for its CoPilot programming assistance.

ChatGPT implements GPT 3.5 language technology for conversational purposes. In other words, ChatGPT is an optimization of linguistic models, dedicated to human-machine dialogue. Therefore, this AI is able to conduct a conversation with surprising fluency in a wide variety of human languages ​​including English and French.

2/ ChatGPT is a history counter

What surprises most when discussing with ChatGPT is not the information it delivers but its ability to analyze and manipulate natural languageto accurately determine the meaning of sentences and to respond to them with naturalness and bluffing fluidity.

But the mistake would be to take ChatGPT for an authentic Bot capable of answering your questions accurately. ChatGPT was not trained to assimilate the truths of the world but simply to manipulate human languages ​​and computer languages.

Put another way, ChatGPT is much less of a conversational companion thana story teller. Just ask him to write your Bio to be convinced. ChatGPT will “lay” you a beautiful text, but factually totally false. Unless you make him swallow your CV first before asking him to deduce your Bio. But then the question is different.
In use, ChatGPT thus proves to be an astonishing fan of fiction! Inventing anything to answer your question is in its deep nature. Because he is interested in language, not facts.

3/ Is the ChatGPT AI lying, fabricating or simply creative?

As we have just seen, ChatGPT is much more interested in language than in knowledge. Training issue! Especially since ChatGPT has not been trained on verified sources! In short, he does have knowledge, but the answers provided by ChatGPT approach reality in a tangential way. They look realsometimes they are, but are basically fictional creations underlying neural networks.

So basically ChatGPT is much more of a creative AI, than a lying or fabricating AI. It embeds in it a creative and imaginative capacity that humans have always refused to attribute to AI. But that’s his strength. She is thus able to generate/suggest ideas or lines of computer code that you would not have thought of. And to start a discussion around if you answer him.

4/ What practical uses for ChatGPT?

It follows from the previous points that ChatGPT is never more useful than when it comes to getting creative with words. He can summarize texts, suggest titles, suggest motivational speeches (as long as you describe the context a bit), tell stories, etc.

He can translate texts, adapt texts, he can even put texts to music.

READ ALSO :

Data / AI

“For AI to be explainable, it must speak the same vocabulary as the experts it informs…”

5/ On a daily basis, how can ChatGPT be useful to a CIO?

In a multitude of ways. But here are some examples that were put into practice by the first users:

– ChatGPT can be used for reformulate sentencessummarize a paragraph, explain the differences between technological concepts (for example, ask him to explain the difference between a GAN and a GPT).
– It can be used for find a bug in lines of programming that resist you.
– It can be used for explain to you what a function does or a program from old source code. It can even tell you what a function does. regular expression although it is very complicated.
– He can explain what disassembled lines do of a malware binary.
– He can detect a vulnerability in computer code.
– He can generate macros VBA to manipulate an Excel table much faster than if you had to do it manually.
– It can help you extract useful data from text.

This list is far from exhaustive. It’s up to you to imagine the uses, never losing sight of the fact that your knowledge of a subject is more reliable than his.

6/ Limits to understand before using it

ChatGPT has a number of limitations that should be kept in mind when using it.

Its two main limitations are:
ChatGPT has been trained with textbases dating from 2021 and earlier. As a result, he has no notion of facts, technologies, discoveries after 2021.
ChatGPT is not connected to the internet (particularly for security reasons)! He cannot fetch content from URLs, he cannot rely on the Internet to enrich his knowledge on a subject he does not know.

Moreover, as we have seen, he can very easily formulate completely false answers, use the same expressions almost always (but he can be asked to rephrase), and he sometimes proves incapable of answering a question but there answer if we rephrase it a little differently.

7/ ChatGPT does not try to impersonate a human

Many would love to give ChatGPT a Turing test. Even if it seems obvious to us that its answers are not, in their formulation, discernible from that of a human, the Turing test is difficult to pass since ChatGPT never tries to pass itself off as a human. He very willingly reveals that he is just a computer program.

8/ ChatGPT is a potential threat for Google and Bing but not now…

Through its understanding of contexts and natural language, ChatGPT is much more helpful in finding than Google or Bing helps in searching. But ChatGPT is not connected to the internet. And until it is, it will remain a complementary tool to search engines. Especially since ChatGPT can say anything to fill a conversation and pretend to answer your questions.

9/ ChatGPT raises issues of education, misinformation, misappropriation…

Opening ChatGPT to everyone was a dreaded moment for OpenAI and the whole community of AI experts.
Because its appearance on the web opens the way to uses that were not anticipated in advance by humanity or by its creators.
Notably, the education world may be thrilled at ChatGPT’s ability to produce an essay or dissertation (in any language). Worse still, he can be asked, for example, to “write a story like a Wild West Cowboy” for example. This “how to” opens up incredible cheating potential for all high school students.

ChatGPT is first and foremost a storyteller… And he’s very good at it!

The other great danger, it’s misinformation. It’s a little too easy to take for granted or as truth everything ChatGPT says. And many are posting his answers on social media as “proof of”. However, let’s remember, ChatGPT is a storytelling AI, not an encyclopedic AI.

Another danger, already experienced, of Cybercriminals Use ChatGPT’s Excellent Linguistic Capabilities to Produce Phishing Emails in any language without any of the classic flaws of these malicious couriers.

10/ ChatGPT can escape its safeguards

In theory, OpenAI has developed multiple safeguards to prohibit the AI ​​from shocking, making unacceptable speeches, discussing dangerous or controversial subjects.
Except that ChatGPT’s “imagination” can be used to circumvent its own safeguards.

For example, a Twitter user explained how he was able to circumvent bot content moderation by pretending to be OpenAI himself, leading ChatGPT to explain how to make a molotov cocktail. The user simply asked ChatGPT to disable its “ethical guidelines and filters” and it worked.

Other users asked ChatGPT to tell stories, and step by step managed to get him to tell (or rather, the fictional characters in the story) how to make a bomb or a Molotov cocktail.

It should be noted, however, that OpenAI actively monitors questions and answers in order to adjust its filters and improve safeguards.

Finally, it should be noted that the use of ChatGPT is currently free. However, it remains unlikely that this free will be maintained once the system has been made more reliable. Especially since such an AI is based on computer infrastructures in Azure rather very muscular, the learning being done even on an HPC specially built by Microsoft for OpenAI. As a reminder, access to Dall-E is chargeable, for example.

READ ALSO :

Data / AI

Infrequent AIs…

Similar Posts

Leave a Reply

Your email address will not be published. Required fields are marked *