Language models are not knowledge bases
Why ChatGPT & co. are genuinely useful, but don't replace a search engine, and what we do at wonk.ai instead.
Models like ChatGPT keep both academia and industry talking. The idea is tempting: a machine colleague that drafts texts, answers emails and writes code. And ChatGPT, as a large language model, often is a real help. But to understand how these technologies work, and why they cannot replace traditional search engines, we first need to look at the underlying building blocks.
What are transformers?
Transformers are a deep-learning architecture introduced in 2017 by researchers at Google. They were built specifically for natural-language processing (NLP) and have changed the way models work with human language. The transformer is based on self-attention mechanisms that let it consider different parts of a text in parallel rather than strictly in sequence. As a result, it captures the context of each word in a sentence far better than its predecessors.
GPT builds on this. The name stands for "Generative Pre-trained Transformer". Developed by OpenAI, GPT is known for producing remarkably fluent text. The "pre-trained" part already hints at the recipe: the model is trained on an enormous corpus of text before it is fine-tuned for specific tasks. During that training, the model picks up both the nuances of our languages and a great deal of world knowledge, even though that wasn't strictly the goal.
What is ChatGPT?
ChatGPT is a specific application of GPT, tuned for conversation. It is trained to produce human-like replies in a chat context. Unlike traditional chatbots, ChatGPT is not built on hand-written rules or scripts but on the statistical analysis of billions of words and sentences from the web. Even so, it has clear limits:
- Training-data cutoff: ChatGPT is only as good as the data it was trained on. It cannot know about events or developments that happened after that cutoff.
- No real understanding: The model generates text and responds to prompts, but it does not truly understand the content. It emulates human communication based on statistical patterns, not genuine comprehension.
Plugins extend ChatGPT
ChatGPT plugins are extensions that give the model additional capabilities beyond pure text output. You can think of them as "eyes and ears" for a language model. They let it reach data that is not in its training set, current, personal or domain-specific. A browsing plugin, for example, lets ChatGPT look things up on the web and pull in fresh information well beyond its training corpus.
Why doesn't ChatGPT replace Google?
ChatGPT and Google Search are both impressive, but they serve different purposes and cannot simply be swapped for one another. Google is, first and foremost, a search engine. It is built to crawl the open web and return relevant results for a user's query, with links to actual pages (still mostly written by humans). That means you can land on excellent sources as well as on less reliable content.
ChatGPT, by contrast, generates answers based on the data it was originally trained on. It usually doesn't link to sources, so its answers carry a real risk of being inaccurate or misleading. And it's worth remembering that the training data itself contained plenty of wrong, biased or questionable material.
What do we do at wonk.ai?
We combine generic knowledge from public sources with the specific knowledge inside your company. Only then do answers become concrete, traceable and verifiable, instead of eloquently guessed. Once relevant sources have been collected and reviewed by your editors, we generate drafts that serve as a starting point for human work, not as the finished product.
If you are considering setting up an in-house language-model or retrieval solution, you are welcome to book a 30-minute intro call. Honest, no PowerPoint.