What do language models really learn

Are language models truly creative, or just powerful mathematical optimizers? This talk reveals what LLMs actually learn beyond the hype.

#1about 7 minutes

The fundamental challenge of modeling natural language

Language models aim to create intuitive human-computer interfaces, but this is difficult because language syntax doesn't fully capture semantic meaning.

#2about 3 minutes

How deep learning models learn by transforming data

Deep learning works by performing a series of transformations on input data to warp its vector space until it becomes linearly separable.

#3about 3 minutes

Why the training objective is key to model behavior

The training objective, or incentive, dictates exactly what a model learns and can lead to unintended outcomes if not designed carefully.

#4about 8 minutes

From Word2Vec and LSTMs to modern transformers

The evolution from slow, non-contextual models like LSTMs to the parallel and deeply contextual transformer architecture solved major NLP challenges.

#5about 7 minutes

A practical demo of a character-level BERT model

A scaled-down, character-level transformer model demonstrates the 'fill in the blank' pre-training task by predicting masked characters in artist names.

#6about 2 minutes

What language models implicitly learn about language structure

By analyzing a model's internal weights, we can see it learns phonetic relationships and syntactic structures without ever being explicitly trained on them.

#7about 7 minutes

Why current generative models don't truly 'write'

Generative models like GPT are excellent at predicting the next word based on statistical patterns but lack the underlying thought process required for true, creative writing.

#8about 4 minutes

Exploring the future with Blank Language Models

Blank Language Models (BLM) offer a new training approach by filling in text in any order, forcing the model to consider both past and future context.

#9about 3 minutes

The need for better tooling to accelerate ML research

The complexity of implementing novel architectures like BLMs highlights the need for better infrastructure and compiled languages like Swift for TensorFlow to speed up innovation.