OpenAI staff have told me that they didn’t specifically tune on these questions and I believe them. The improvements are seen even for newly made-up problems in the same style.
@goodside
-
Prompt Injection Discovery and Attribution: A History Correction
By
–
A history correction: I am not the first to discover prompt injection. I was merely the first to do so and discuss it publicly. PI was discovered independently by multiple teams. The first was Preamble, an LLM security company, whose find predates mine by several months.
-
AI Chatbot Generates Python Code for UMAP Visualization of Handwritten Digits
By
–
Using Poe Sage (GPT-3.5 chatbot from @Quora
) to ask a fast-talkin' wise guy from a 1940s gangster movie how to use UMAP to visualize the topological manifold of UCI ML handwritten digits. "Snazzy" Python code and Matplotlib plot shown in reply. https://
poe.quora.com/s/kjbVTU3hH7mL
b82H1L6S
… -
Discussion on GPT-3 Tokenization and Spacing
By
–
Just search “GPT-3 tokenization”. It hasn’t changed since GPT-2 days, I don’t think. I just used dashes because you need spaces to be spaces between words.
-
LLM Tokenization: Not Vectors or Letters, but Token Sequences
By
–
It doesn’t see words as embedded vectors or as letters, it sees them as token sequences — chunks of about 4 letters from a fixed mapping of strings to tokens. It does better on letter-based tasks if you ask it to first rewrite words l-i-k-e t-h-i-s.
-
Critique of AI Model Hallucinations: GPT-3 vs. ChatGPT
By
–
That isn't ChatGPT. GPT-3 hallucinates in many scenarios.
-
Prompt Engineering’s Future: From Tricks to Frontier Problems
By
–
Many tasks that require prompt engineering tricks today soon won’t. The scope of “just ask” will increase. But prompt engineering will move on to harder tasks. We won’t run out of problems on the frontier.
-
Prompting GPT-3 with the “Format Trick” for Complex JSON Synthesis
By
–
Prompting GPT-3 using the "format trick" (implemented as a Python function) to synthesize complex example JSON objects, e.g. for mock API responses. Output is shown in second screenshot.
-
AI Model Performance Comparison for Code Generation
By
–
For zero-shot prompts with instructions text-davinci-003 does better in my experience. The text models are actually trained on just as much code as code-davinci-002.
-
Instruction Template Method for Synthesizing Code Files in ChatGPT
By
–
A demo of my "instruction template" method for synthesizing complete code files in ChatGPT. Templates that mix literal code with inline placeholder instructions give better control over generations than using English prose.