ChatGPT, second thoughts

ChatGPT, second thoughts

One of the major sources of criticism to chatGPT is that it makes factual mistakes, gets calculations wrong, seems to have problems with simple logical reasonings. We could perform a thorough syntactic /semantic analysis of the tool's performance, but let's keep it simple: there are things it does right and things it does wrong. But the things it does right are truly amazing and impossible to be done without true "understanding".

If I am not mistaken also the co-reference problem (knowing to which entity "this" or "that" refers) is not a problem for chatGPT. According to some linguists, solving this problem is would be really hard and would translate to solving intelligence altogether. Apparently the attention mechanism of the transformer was all we needed to overcome this hurdle. This hints that also the things that chatGPT does poorly could fall within its reach in a matter of years, if not months.

Part of the criticism is also due to a common misconception about AI, which brings critics to highlight what AI cannot do (yet). We should remind ourselves that AGI means Artificial General Intelligence, not Artificial Genius-level Intelligence. And there are many tasks that common people are not good at. I remember being at a pizzeria in Milano and, wanting a smaller slice of pizza, I said: "can I have 3/4 of that?" The lady stared at me speechless, until I added "a bit less please". Many people are not familiar with mathematical concepts and operations, which doesn't prevent them from being functional members of society.

So does OpenAI's chatbot, probably because such tasks haven't been "explained" with enough data. ChatGPT was trained on a huge text corpus taken from the internet, following the objective to predict the next word in the data stream. If we applied this method to human beings, we should take a newborn baby and have him/her read the entire internet, with no supervision. The fact that this learning strategy succeeds is already incredible.

Some say that ChatGPT is just another software tool in our repertoire ("ChatGPT bot is causing panic now – but it’ll soon be as mundane a tool as Excel", argues John Naughton on the Guardian). I disagree. A tool is an augmentation to human skills that needs an intelligent agent to be operated. ChatGPT has the potential to become that agent.

There are of course many techical hurdles that need to be addressed along the way, but chatGPT may represent for AI what the Flyer was for artificial flight. The first airplane built by the Wright brothers ushered humanity into the era of mechanical flight. There is a of course a big difference between the Flyer and the Airbus A380 or the B2 bomber, but the key ingredients of an airplane were assembled for the first time on december 3rd 1903, on the dunes of Kitty Hawk, North Carolina.

No alt text provided for this image

Finally, let's give to Caesar what belongs to Caesar. When talking about AI, the so-called "Godfathers" (Hinton, leCun, Bengio and Hassabis) are always cited, but I think it is right to give credit also to the next gen heroes. Ashish Vaswani is the first author of the paper that proposed the "transformer" neural network, the model that revolutionized the field of Natural Language Processing. Ilya Sutskever, OpenAI's co-founder and chief scientist, is (in the words of Pieter Abbeel, top AI researcher) "the man who made AI work". Their names, along with the names of many others (modern science is of course a joint effort), will probably land in history books. The question is how long will history last.

To view or add a comment, sign in

More articles by Dr. Alessandro Fontana

  • The Origin of Bureaucracy

    Bureaucracy is often identified as a major obstacle to economic growth, slowing down innovation, investment, and…

    3 Comments
  • Are neural networks logical?

    In the 19th century, a pivotal moment occurred when a student approached a physics professor seeking guidance on a…

  • The challenge of AI alignment

    Despite the success of AI models, open problems still remain on the research agenda. Poor interpretability is one of…

  • ODE TO ELECTRONIC ENGINEERING

    Once upon a time, not long ago Electronic engineers were seen as low Jokes were made, they were called "nerds" Unable…

  • Five scenarios for a post-singularity world (what happens when AI takes over)

    Before discussing whether AI can "take over", it's important to clarify what this phrase means. The Turing test, which…

    3 Comments
  • A short history of the algorithm that changed science

    In the last 10 years we have been witnessing the successes of deep learning in many tasks, climbing the ladder of…

  • Thoughts on chatGPT

    In the last 2 weeks, social networks have been taken by storm by the latest software by OpenAI, the company founded by…

    1 Comment
  • The best product of Baden-Württemberg

    What is the best product of the Bundesland of Baden-Württemberg? From a culinary perspective, the list might include:…

  • RULE-BASED VS DEEP LEARNING: THE TEACHING OF ALPHA ZERO

    Intrusion detection systems (IDS) are becoming increasingly important in the realm of cybersecurity. The usefulness of…

  • The Grande Armée and the power of meritocracy

    The Grande Armée was the army that Emperor Napoleon Bonaparte deployed, starting from the 1805 campaign, to face the…

Insights from the community

Others also viewed

Explore topics