110 reads

The Limitations of LLMs Like ChatGPT: A Straight-Talking Overview

by McMurchieOctober 23rd, 2023

Too Long; Didn't Read

Large language models (LLMs) like Chat GPT are impressive, but they're far from perfect. LLMs can make statements that sound right but are plain wrong. There's an insane amount of tech and money behind these models to get them to remember up to 8,000 tokens at once.

featured image - The Limitations of LLMs Like ChatGPT: A Straight-Talking Overview

Figure Above: A 4-minute bullet overview if you prefer this in video format.

Hey folks,

Let's dive right in and get to the meat of the matter without any fluff. I've heard a lot of hype about large language models (LLMs) like Chat GPT, and while they're impressive, they're far from perfect.

Here's a Quick Rundown on Why These Models Aren't the Future of AGI (Artificial General Intelligence).

Hallucination: LLMs, like Chat GPT, can make statements that sound right but are plain wrong. They can spit out info that fits a pattern without actually being connected to reality.
Memory Limitations: There's an insane amount of tech and money behind these models to get them to remember up to 8,000 tokens at once. But, they have a short memory span – after a certain point, they start forgetting what you said earlier. Not so great for larger databases or extended conversations.
No Golden Source: When ChatGPT shares a fact or tidbit, it can't cite its sources. It's been fed data from countless places, but it can't tell you exactly where each piece of info came from.
Heavy Manual Integration: All the cool visualizations, GUIs, games, and application integrations? That's manual work. Engineers putting in hours to make it look seamless. It's not all AI magic; it's just good old-fashioned human labor.
Ethics and Filtering: LLMs weren't designed with ethics at the forefront. They're optimized for performance. That means there's a ton of manual work done to filter out the unsavory bits to make them user-friendly.
Over-reliance on Data: Without data, these models are like empty shells. They don't actively learn in real-time. They get massive updates every so often, requiring billions of dollars and tons of hours. It's a colossal human effort, not the model spontaneously evolving.
They Get Outdated: Remember GPT-1 or GPT-2? No? That's because models age out and get replaced. It's a sure sign we're dealing with a product, not an evolving intellect.
Language ≠ Reality: These models have a strong grasp on language, but language isn't always a perfect reflection of the world. Just because Chat GPT can generate coherent sentences doesn't mean it understands the complexities of reality.

In essence, while the concept of a "super-intelligent parrot" might be overused, it's not entirely off the mark. LLMs are indeed powerful language models, but that doesn't equate to a deep understanding of the world.

So the next time someone gets all starry-eyed about LLMs being on the brink of AGI, arm yourself with these points. There's much more to say, but this should give you a solid foundation.

Stay curious, and thanks for dropping by!