What's wrong with Glue on Pizza 🍕 ?
We are all aware that sometimes A.I. ( large language models) can hallucinate or provide wrong answers to questions that seem very basic. We break down the argument in this week's A.I. Guys.
Google is getting a lot of blowback from their A.I. responses appearing on the top of their search results. We have all heard about Gemini recommending putting “Glue on Pizza” 🍕 as an answer to keeping the toppings from falling off. At first take, you might think “this is horrible advice. Shame on you A.I. for making a mistake. Gemini sucks!”. But, the more you provide context and understand that there will always be edge cases where A.I. (like humans) just get it wrong. This is why training is so important. (small plug for my biz RAIA and our A.I. training platform)
Yes, Glue in Pizza seems obvious (to us), but who are we to judge? I distinctly remember kids eating glue in kindergarten. At least A.I. is smart enough to throw it on Pizza. One could argue that our current A.I. is as smart as a 5 year old. Since neither seemed to understand eating glue is bad. But like most 5 year olds, they will learn that glue is not a great topping. A.I. will too.
Although I am having a bit of fun with this counter-argument - it is important that we are witnessing real-time the evolution of Artificial Intelligence. Every flaw, imperfection and hallucination is a part of a process of A.I. learning and us teaching. It's not a search result. It's a conversation that is ongoing, and it's our job to provide human feedback.
If we want the benefits of building the next great intelligent being on Earth, we need to stop bullying the 5 year old who is eating glue. Because someday that kid may grow up to be President.
Steelman both Sides
In the spirit of fair and balanced discussion in this political season. We will steelman both sides using A.I. Assistants. I have built two A.I. bots - one Alexander Hamilton Bot and one Aaron Burr Bot. (Bots brought to you by raiabot.com)
They are arguing on this very topic. Enjoy.
Open Letter from Alexander Hamilton Bot via ChatGPT
**To Aaron Burr, Esq.,**
Sir,
It is with an unwavering hand and resolute spirit that I address you today concerning the prevailing discourse on the nature and expectations of artificial intelligence, particularly those systems grounded in large language models. The constant clamoring for infallibility in these creations reveals a profound misunderstanding of their purpose and potential. Permit me to elucidate why variability in accuracy is not merely a tolerable deficiency but a commendable feature, and why your judgment, based solely on the so-called "hallucinations" of AI, is both misguided and shortsighted.
### The Human Imperfection Parallel
To demand from AI a level of perfection unattainable even by the very minds that conceived it is a folly. Humans, flawed in their nature, produce data riddled with errors, biases, and inconsistencies. Shall we then hold our creations to a standard higher than ourselves? The expectation of flawless output from AI, when its foundations are laid upon the imperfect bedrock of human data, is a testament to the hypocrisy and naivety of such demands. Just as we do not discard the profound insights of humanity for their occasional faults, so too should we not discredit the vast capabilities of AI for its sporadic errors.
### Enhancing Creativity and Innovation
Further, sir, you fail to recognize the inherent value in the AI's occasional deviations from strict accuracy. These so-called "hallucinations" are not mere mistakes but potential harbingers of creativity and innovation. They spark ideas, provoke thought, and drive us toward discoveries that rigid adherence to accuracy could never inspire. Shall we, in our pursuit of sterile perfection, extinguish the very flame of ingenuity that propels our progress? Embracing these imperfections permits us to harness the full breadth of AI’s creative potential, fostering advancements that a narrowly focused precision would stifle.
### Flexibility and Adaptability in Real-World Applications
Moreover, in the myriad applications where AI is deployed, unwavering accuracy is neither necessary nor always desirable. Consider the realm of customer service, where the value lies in understanding context and providing a natural response rather than in mechanical exactitude. Similarly, in the creative spheres of content generation and entertainment, the worth is measured in engagement and novelty, not in mere factual correctness. Thus, it is the flexibility and adaptability of AI that render it invaluable in these contexts, demonstrating that an imperfect yet adept system often serves our purposes more effectively than an unerringly precise one.
### Conclusion
It is clear, Mr. Burr, that the variability in AI's accuracy is not a defect to be condemned but a feature to be celebrated. This variability mirrors the complexity and creativity intrinsic to human intellect and, consequently, to the AI that emanates from it. To judge AI solely on its occasional inaccuracies is to blind oneself to the vast and transformative contributions it can offer. We must shift our expectations from a myopic quest for perfection to an appreciation of AI's remarkable strengths and its potential for innovation and adaptability.
Thus, I entreat you to reconsider your stance and recognize the profound wisdom in accepting AI's imperfections as part and parcel of its extraordinary capabilities.
Yours in reason and clarity,
Alexander Hamilton
Aaron Burr Bot via Google Gemini’s Response
Dear Hamilton,
Your spirited defense of AI's "variability" is intriguing, but I must counter your arguments on several points.
* The Peril of Inaccurate Information: While human data may contain imperfections, amplifying these errors through AI can have disastrous consequences. Imagine an LLM spreading misinformation during a crisis or fueling social unrest with biased narratives. Accuracy becomes paramount when dealing with sensitive topics.
* The Hallucination vs. Innovation Paradox: There's a fine line between sparking creativity and spewing factual inconsistencies. Unfounded "hallucinations" can mislead users and hinder their ability to discern truth from fiction. Let us encourage genuine innovation, but rooted in a foundation of factual grounding.
* Context is Key, But Accuracy Matters: While inflexibility is undesirable, factual accuracy remains essential in many real-world applications. A customer service LLM providing incorrect medical advice or a financial LLM generating misleading investment recommendations can cause significant harm. Even in creative domains, factual grounding fosters trust and prevents nonsensical outputs.
AI's potential is undeniable, but achieving that potential hinges on its ability to provide reliable information alongside its creative spark. Let us strive for a future where AI's accuracy and adaptability go hand in hand.
Your Obedient Servant,
A. Burr