Is AI Hype Fading Due to Hallucinations?

The Key Factor AI Needs to Level Up

Image made with DALL-E

In recent months, it feels like an illusion has formed around generative AI. It's as if life has somehow paused, with everyone eagerly awaiting the next "new" thing it will surprise us with each week.

It’s common to see headlines like those in The Washington Post that read: "Generative AI is a business game-changer," where AI is almost automatically credited with the ability to solve any business problem. This will inevitably lead to side effects, such as the reduction of large amounts of labor. On top of that, these views are consistently backed by tech leaders like Bill Gates, who claims that AI will be a revolutionary event.

But this is where I start to question just how amazing generative AI really is.
I’m not dismissing the breakthroughs we’re witnessing with AI. In fact, I think we’re all impressed by the results we’re seeing. But it’s not going to take over the world.
I have to acknowledge that generative AI is at its peak right now, with the potential to drive efficiencies. But, for now, those efficiencies remain limited.

AI is Still Just a Technology.

It’s easy for AI to capture our attention right now. But it was the same in the past, though maybe not with the same impact we’re seeing today. For instance, back in 1997, chess grandmaster Kasparov was defeated by Deep Blue. Then there was Watson, also from IBM, which won the "Jeopardy" game show. In 2016, Google’s AlphaGo beat Lee Sedol in the game of Go.

All of this made me pause and reflect on the buzz AI has generated over the past few decades. Things really took off with the launch of ChatGPT in 2022, and we’re still riding that wave today.

What stands out most, though, is that this new generation of AI feels much more human and accessible than those earlier systems. It can have conversations on nearly any topic and often responds in ways that seem to truly understand what we’re saying. It appears to have a grasp of the meaning behind words, phrases, sentences, and paragraphs. Still, I think the gap between what AI can do and what humans are capable of will become increasingly apparent over time.

Generative AI is pretty good at delivering responses, except for what we’ve come to call hallucinations, where it starts making things up. What I appreciate, though, is that I’m aware of this, and it doesn’t stop me from using it. What I mean is, I’ve learned to recognize where it can genuinely help, and I’ve stopped getting frustrated when it doesn’t meet my expectations for certain tasks. I’m more confident in its abilities now, like drafting the first version of an email or brainstorming ideas on specific topics. But I still can’t trust it completely beyond that.

From my perspective, this makes it less autonomous than it could be. That said, this is exactly the challenge that tech giants are working on around the clock.

Redefining AI for the Next Level (AGI)

Just a few weeks ago, the wave of new AI project launches once again had us excited about the future of the tech industry. But recently, I noticed a drop in NVIDIA’s stock prices, which is telling. You can see evidence of this in the following image.

NVIDIA Stock Chart

So, does this hint at something more concrete behind the AI hype?

The current outlook isn’t looking so bright for those involved in AI. Investors are starting to panic as stock prices of companies betting on this "new industrial revolution" are taking hits.

It might be tempting to think that the AI bubble is bursting. But the real issue seems to be tied to LLMs (Large Language Models) and the hallucinations they still produce. We’ve definitely seen improvements over time, with some obvious errors being avoided. In this sense, better results are closely tied to the training data, which might make us think that simply tweaking this variable (sometimes a hidden one) could solve the problem. But the truth is, it’s not that simple. Hallucinations are just a symptom of a bigger issue. For an LLM, the correct output is often "close" to the incorrect one in a quantifiable sense.

A clear example is when Midjourney generated images of hands with six fingers. This shows how hallucinations can be close to the correct value (five fingers) but still miss the mark. The issue for us humans is that we use a different metric for what counts as a good result compared to the models. Defining a "good" metric is a challenge that can't be solved just by feeding the model more data. You can't simply keep training models on more text and images and expect them to start truly understanding what’s going on.

That said, layering symbolic reasoning on top of existing models is much harder. This is partly due to what Ludwig Wittgenstein called "linguistic confusion." No two people use a word to mean exactly the same thing. So, when you combine text from billions of different people, the logical relationships between those words start to disappear—if they were even there to begin with (and honestly, people aren’t always that great at logic either). That’s why I worry that the LLMs we’ve already trained will eventually need to be retrained.

At the end of the day, the real problem with LLMs is that the world isn’t made of words. At its deepest level, the world is built on mathematics. So, if we want to create truly intelligent AI (what we’ll soon call AGI), we need to start with math and models that are based on physical reality or real environments, and only then add words on top of that. This suggests that companies that have invested heavily in large LLMs might never see a return on that investment.

The real winners will likely be those who build AI grounded in logical reasoning and models that actually understand the real world—like what NVIDIA and DeepMind are doing.