I've hacked a prototype together that enables GPT to break down user tasks into manageable sub-tasks, and then schedule and oversee their execution on a Python VM while collaboratively addressing syntax and semantic errors through a back-and-forth message dialogue.

Example:

"Go to the https://ten13.vc/team website, extract the list of names, and then get me a summary of their LinkedIn profiles."

Will turn into GPT generated a-normal form Starlark code:

var1 = download("https://ten13.vc/team") # Step 1: Download the webpage var2 = llm_call([var1], "extract list of names") # Step 2: Extract the list of names from the webpage answers = [] # Initialize an empty list to store the summaries for list_item in llm_loop_bind(var2, "list of names"): # Step 3: Loop over the list of names var3 = llm_bind(list_item, "WebHelpers.search_linkedin_profile(first_name, last_name company_name)") # Step 4: Search and get the LinkedIn profile of each person var4 = llm_call([var3], "summarize career profile") # Step 5: Summarize the career profile of each person answers.append(var4) # Step 6: Add the summary to the list of answers answer(answers) # Step 7: Show the summaries of the LinkedIn profiles to the user

Which will then be executed statement-by-statement by the Python runtime. When syntax errors, exceptions or semantic issues occur, there's a error correction loop where GPT will get involved to identify the issue, regenerate code, and try again.

Lots of fun little programming language/compiler challenges in here. Happy to answer questions if you have them.