How will ChatGPT be remembered?

With the finesse of an expert librarian, it effortlessly retrieved information. Like a patient teacher, it makes complex subjects approachable. And in the vast realm of virtual companions, it adapts to the diverse whims and desires of its users. From unraveling the mysteries of the universe to helping plan a daring escape from a zombie apocalypse, ChatGPT can do it all.

While large language models possess an astonishing range of skills, they are not immune to the occasional flight of fancy or misremembered fact, affectionately termed as "hallucinations" by their creators. But in those quirks might lie ChatGPT’s biggest impact.

Unveiling the AI Evolution: ChatGPT's Impact

Users are becoming more informed consumers of AI technology. Through conversations with ChatGPT, the mystery surrounding artificial intelligence was lifted for millions of people worldwide. They gained familiarity with AI, how it works, and especially the uncertainty that accompanies all of AI.

AI models, for all their technological prowess, are not immune to a misstep or two. ChatGPT's interactions demonstrated that AI models, while impressive, are not infallible and may occasionally produce responses that are incorrect, biased, or nonsensical.

This is, to me, the most important step in pushing AI into the limelight.

AI drastically diverges from familiar software systems

When people interact with any kind of software, they expect deterministic behavior. They expect it to follow rules. When I put in x, I get back y. The output is always the same for the same input x. Often even the same for sufficiently close enough x and x’.

People interact with software like a calculator. They tell it to do something or ask it something and expect it to be consistent. A calculator always returns 4, when you ask it “What is 2 + 2?”. This is expectation users transfer to AI algorithms. They expect it to be perfect: always return the same correct answer.

But this is not how AI works or how it is supposed to work. It thrives on statistical principles and patterns rather than rigid rules. AI models, such as ChatGPT, learn from vast amounts of training data to identify and replicate these patterns. Even if you put in the same x, an AI model might return different outputs y.

When you ask ChatGPT to “Tell me a joke” twice, you get different results.

1. "Why don't scientists trust atoms? Because they make up everything!”

2. "Why did the chicken cross the playground? To get to the other slide!”

It isn’t consistent.

When we program AI, we want them to be robust, so it can be applied to more than a singular case.

• In image recognition, rule-based systems fail to identify objects in complex and cluttered scenes. Imagine a picture of a crowded beach, where people are engaged in various activities. A rule-based system may falter in accurately labeling each individual or discerning their actions. In contrast, AI models trained on extensive image datasets can deftly recognize people, beach balls, and sandcastles, and even distinguish between sunbathers and surfers.

• For customer service chatbots, rule-based systems can often fall short of comprehending the nuances and variations of user queries. An inquiry like "What's the best restaurant near me?" can baffle a rule-based bot confined to a predefined set of rules. However, AI-powered chatbots, with their ability to learn from vast conversational data, can understand the intent behind the query, consider contextual information such as location, and provide personalized restaurant recommendations tailored to the user's preferences.

• In fraud detection, we do not want our detection to be rigid, and stuck on specific rules. These can be easily circumvented by fraudsters. We also need our system to adjust over time, and identify new behaviors and patterns in the data, so new techniques and new types of fraud can be identified.

In each of these examples, the power of AI lies in its ability to learn from data, adapt to diverse scenarios, and address the rigidity that rule-based systems struggle to overcome.

The Place for Strict and “Rigid” Behavior

This doesn’t mean there is no place for rule-based systems or traditional software. In safety-critical systems (nuclear power, control of medical devices), deterministic behavior is crucial. Reliability and predictability rule the world, when human lives hang in the balance. Legal systems require adherence to well-defined rules and regulations. Accounting, banking, or payment processing rely on precise calculations and deterministic behavior to ensure accuracy and prevent financial discrepancies.

But in all of these areas, AI-based systems can augment the rule-based system: they can detect faults in nuclear power plants, better classify transactions in accounting that do not fit exactly into a certain bucket, or correct weird patterns in imperfect medical devices by taking in the context like the skin color of the patient.

AI and traditional software should coexist and collaborate to meet the demands of a rapidly evolving world.