Why Smaller Language Models Hallucinate More. I Turned Down $475,000.

0:00Here is why smaller language models are dumber and hallucinate more often than larger language models.

0:07If you're a software engineer and trying to demystify these AI models and understand how they work under the hood, then stop and follow my page, and if you learn something leave a comment below saying that you learned something.

0:23I'm a senior AI engineer and I've rejected job offers where I'd make over $475,000 per year.

0:29I also have a master's degree in software engineering from Carnegie Mellon, which is, according to U.S. News & World Report, literally the number one school in the entire world for software engineering.

0:41Not only that, but CMU is also the number one school in the entire world for artificial intelligence.

0:49I use AI so often that OpenAI literally gave me an award for it.

0:56It says for Austin Starks, honored for passing 10 billion tokens.

1:02To understand why smaller language models hallucinate more often than the larger language models, you have to understand how large language models are trained.

1:09And they are trained in two key phases.

1:12So number one is a process called generative pre-training.

1:15And this is the process you see all the time on TikTok.

1:18Essentially the language model tries to learn to predict the next word in a sentence.

1:23It's not 100% precise, but you can think of it trying to learn the next word in a sentence or the next letter in a word.

1:32You know, the next token.

1:34We're trying to learn to predict that, right?

1:36And we found that given a sufficient volume of data these models learn to speak like us very, very well.

1:46This is called generative pre-training.

1:48Now when you use a larger language model there are literally more weights and biases.

1:55And so because of this we can form more accurate mathematical representations from the inputs to the desired outputs.

2:05This will be more important later, but the second part of training a large language model is called fine-tuning.

2:12And there are different ways to fine-tune a language model.

2:15The most popular one is reinforcement learning with human feedback.

2:19You essentially get a human being and he says this response sounds better than this response and this response sounds better than this response, and then you train that into the weights and biases of your AI model.

2:31And so when you continue to train with a higher volume of data, when the model is much larger, it has more weights and biases to form the most accurate function from the inputs to the outputs.

2:47And when you have fewer weights and biases, then you have fewer accurate functions from the inputs to the outputs.

2:54All that to say is when you have a larger large language model, it is smarter.

3:01Same way when you have more data, when you fine-tune it on more data, it's also smarter, because again we're forming the most accurate functions from the inputs to the outputs.

3:13That's literally it.

3:14Now if you learned something, you made me a promise.

3:17Follow my page and leave a comment saying what you learned and what video I should make next.