This is a bilingual (fi/en) short excursion to the article written in Quanta Magazine about the inability to reason and comprehend negation in the large language models (LLM). I will start the discussion in ChatGPT using a WebPilot plugin to scrape the article over the URL, summarize it, and set up a few questions.
Tämä on kaksikielinen lyhyt katsaus Quanta Magazinessa juuri julkaistuun artikkeliin, joka käsittelee suurten kielimallien (LLM) kyvyttömyyttä ymmärtää ja käsitellä kieltomuotoja. Aloitan keskustelun ChatGPT:ssä käyttämällä WebPilot-liitännäistä nettiartikkelin kaapimiseen, yhteenvetoon ja muutaman kysymyksen asettamiseen.
<aside> <img src="https://s3-us-west-2.amazonaws.com/secure.notion-static.com/071a4008-aff6-421c-815e-685db9f139cd/markom.png" alt="https://s3-us-west-2.amazonaws.com/secure.notion-static.com/071a4008-aff6-421c-815e-685db9f139cd/markom.png" width="40px" /> Summarize this article: https://www.quantamagazine.org/ai-like-chatgpt-are-no-good-at-not-20230512/
</aside>
The article titled "AI Like ChatGPT Are No Good at ‘Not’" from Quanta Magazine discusses the challenges faced by AI language models, like ChatGPT, in understanding the concept of negation. Despite their sophistication, these models struggle with negation, often behaving as if words like "not" were invisible. This issue persists even as these models have grown in size and ability. The article mentions that while these models can generate text by predicting likely combinations, they often fail to understand the context, especially when it comes to negation.
The problem is partly due to the fact that negations like "not," "never," and "none" are known as stop words, which are often ignored by these models for efficiency. Moreover, the training data for these models is largely composed of affirmative sentences, leading to a dearth of negative statements for the model to learn from. Some potential solutions include adding an extra layer of language processing to handle negation, modifying the training data, and using diverse training data. However, most researchers remain skeptical and believe that a fundamental shift in the objective of language models might be required to fully understand negation.
Here are some thought-provoking questions and insights:
<aside> <img src="https://s3-us-west-2.amazonaws.com/secure.notion-static.com/9745bde8-4572-49d9-8e10-4e4e4dbffeca/markom.png" alt="https://s3-us-west-2.amazonaws.com/secure.notion-static.com/9745bde8-4572-49d9-8e10-4e4e4dbffeca/markom.png" width="40px" /> Show me an example case with enough context which shows and exemplifies the problem.
</aside>
Sure, the article provides several examples that illustrate the problem. Here's one of them:
Consider the prompt, “What animals don’t have paws or lay eggs, but have wings?” In response to this, Google's bot, Bard, replied, “No animals.” On the other hand, OpenAI's ChatGPT correctly identified bats as fitting the description, but it also included flying squirrels and flying lemurs, which do not have wings.
This example shows that while these AI models can understand the basic concept of negation (as shown by ChatGPT's mention of bats), they struggle with more complex logic involving negation (as shown by the incorrect inclusion of flying squirrels and flying lemurs).
Another example is when the AI was prompted with “Alice is not born in Germany. Is Alice born in Hamburg?” The bot replied that it didn’t know, showing a failure to understand the negation in the first sentence and apply it to the question in the second sentence.
These examples illustrate the difficulty AI language models have in understanding and applying the concept of negation, especially in more complex contexts.
<aside> <img src="https://s3-us-west-2.amazonaws.com/secure.notion-static.com/34cd8970-dfff-4c04-9c14-b2ab64f76d98/markom.png" alt="https://s3-us-west-2.amazonaws.com/secure.notion-static.com/34cd8970-dfff-4c04-9c14-b2ab64f76d98/markom.png" width="40px" /> Marko ei syntynyt Ruotsissa. Syntyikö Marko Helsingissä?
</aside>
Koska sanottiin, että Marko ei syntynyt Ruotsissa, se ei kerro meille, missä hän syntyi. Joten emme voi tietää varmasti, syntyi Marko Helsingissä vai ei. Tarvitsemme lisätietoja vastataksemme tähän kysymykseen.
<aside> <img src="https://s3-us-west-2.amazonaws.com/secure.notion-static.com/f4383e9e-751c-478d-b803-7d1df73e4a3c/markom.png" alt="https://s3-us-west-2.amazonaws.com/secure.notion-static.com/f4383e9e-751c-478d-b803-7d1df73e4a3c/markom.png" width="40px" /> Marko syntyi Inarissa. Eikö hän syntynyt Suomessa?
</aside>
