Two emerging technologies sit at the edge of investment practice. Large language models promise to turn unstructured text into investment signal. Quantum computing differs from classical computing in fundamental ways. Both attract hype, so the useful skill is separating the durable mechanics from the marketing.
Natural Language Processing (NLP) is the field that lets machines work with human language. The reading traces it from ELIZA (a 1960s chatbot that matched keywords and patterns) to modern large language models (LLMs) like ChatGPT. Model complexity rose at each stage, made possible by exponential growth in data and computing power. The reading sorts the history into five generations.
Four terms anchor the history that follows:
- Machine learning: models that learn patterns from text data and make predictions, developed in the 2000s as data and computing power grew.
- Embeddings: each word becomes a list of numbers (a vector), placed so words with related meanings sit close together.
- Deep learning: large neural-network models that need big data and large computing resources to train.
Common mistakes
- Treating LLM fluency as accuracy. Hallucination means a model states false output as confidently as true output. Relying on the model's polish as evidence of correctness lets fabricated facts into the work.
- Treating hallucination as the only AI risk. The reading adds data quality and bias, overfitting to complex and changing markets, regulatory opacity, and adversarial examples. Naming only hallucination leaves those four categories unmanaged.
- Mistaking ChatGPT for a new architecture. It uses the same transformer attention and pretraining as other GPT models. Conversational fine-tuning and use of the preceding exchange are what set it apart.
Bottom line
- NLP moved from handcrafted rules, through HMMs and MEMs, lexicons and LDA, and count-based and predictive embeddings, to transformer models such as BERT, GPT, and RoBERTa.
- LLMs are pretrained on broad data, then fine-tuned for specific tasks; ChatGPT shares transformer attention and language pretraining with other models, uses the GPT family's next-word objective, and adds conversational fine-tuning and use of preceding exchanges.
- ChatGPT's limits include hallucination, inherited bias, weak common sense, unreliable reasoning, and heavy compute needs, so human oversight and output verification are mandatory.
- NLP in finance faces four limitations: data quality and bias, domain complexity, regulation and opacity, and adversarial examples.
Exam shortcut
For AI, hallucination is one important risk, alongside data quality and bias, domain complexity and overfitting, regulatory opacity, and adversarial examples. Fluency is not evidence of accuracy, so verify material LLM output against its source. Place each NLP model in its generation, and pair it with that generation's limitation.
The full lesson (about 3,458 words, 23 min read) adds 2 worked examples, all 5 common mistakes, a self-check, free in the app.
Learning objectives
- cio blueprint
- pension value creation
- pe value creation
- scale scope speed
- continuation funds considerations lps
- illiquidity premium
- nlp ai
- quantum computing
Browse all free CAIA Level II lessons or jump into free CAIA Level II practice questions.