Article Transmogrification With Claude
LLMs are adept at summarization, categorization, and text generation. So, can Claude:
- Fact-extract a group of articles
- Mix and match the facts, organizing them by (new) topics
- Produce new articles based on the new fact combinations
Is it possible to use ChatGPT to replace Google?
Google is the overwhelmingly dominant leader in search. But "search" isn't necessarily the problem people are trying to solve, any more than elevator riders have the goal of "stand still in a metal box for a minute." Search is a tool for finding information. In this instance, I'm looking at the sub-use-case of "find the answer to a question," along with a somewhat "find information on the web" question
tl;dr -- Microsoft thinks that LLMs can disrupt search -- based on a few tests, I think they're onto something.
Read More
Using ChatGPT to generate code in an obscure language
I experimented with generating LiveCode using ChatGPT. LiveCode is fairly obscure, so ChatGPT has much less training data to work with than e.g. for Python.
tl;dr -- Despite the limited training data, ChatGPT does a good job of generating LiveCode and an excellent job taking direction. And some of its code was better than mine...
Read More
How implementing a BERT-based model improved precision and recall in recognizing global entities at LivePerson
Entities are the things, and their attributes, referred to in conversations. Some are recognized using a list: models of cars, for example. Machine-learning is useful for recognizing entities that aren't easily enumerated, like numbers, dates, times, places, currency, and people's names. The model that is trained on a large amount of data that has been labeled with the desired entities.
tl;dr -- implementing a BERT-based model enabled LivePerson to add new entity types while also improving F1 from 42% to 89%, doubling performance.
Read More
