Conversation

Notices

1. Embed this notice
   Javi (javi@sfba.social)'s status on Tuesday, 09-May-2023 11:01:53 JST Javi

   How long until the programming languages we use today are regarded the same way we think of Assembly today: very low level tools that only very few people write directly, and instead we write programs using natural language, like plain English sentences describing the program requirements, and then something like Large Language Models "compile" that into actual code? (Or maybe straight into assembly)
   I feel like this may happen in my lifetime.

   • Embed this notice
     Paul Cantrell (inthehands@hachyderm.io)'s status on Tuesday, 09-May-2023 11:01:49 JST Paul Cantrell

     in reply to

     @Javi This of course is what programmers do. And the reason it’s hard is not because the syntax of code is hard, but because natural language — human thought! — is full of ambiguity, hidden assumptions, surprising implications, etc. which can only be resolved by communication and introspection.

     The question is whether machines will do •that• better than humans in our lifetimes.

   • Embed this notice
     Paul Cantrell (inthehands@hachyderm.io)'s status on Tuesday, 09-May-2023 12:12:47 JST Paul Cantrell

     in reply to

     • William Pietri

     @williampietri @Javi Yeah. To be clear, I do think LLMs will prove useful in development — for discovery, getting those first steps, rapid prototyping…. I also think they’ll make bigger messes faster, and the hard part (“What is this mess?!” “Is there a better way to…” “Are we even building the right thing?!”) will become even more essential.

     LLM ≈ low hourly rate outsourcing

   • Embed this notice
     William Pietri (williampietri@sfba.social)'s status on Tuesday, 09-May-2023 12:12:48 JST William Pietri

     in reply to

     • Paul Cantrell

     @inthehands @Javi As a kid I went to a computer conference. Call it 40 years ago roughly. And this fantasy of "just write English" was being used to sell tools even then.

     I think it will remain a fantasy until AGI. As you say, the hard part is understanding the human need and expressing it with precision. Document-driven development vigorously demonstrates how inadequate mere words are for explaining what software is supposed to do even when the readers are human.

   • Embed this notice
     jesse squires (jsq@mastodon.social)'s status on Tuesday, 09-May-2023 14:01:10 JST jesse squires

     in reply to

     • NeoNacho

     @NeoNacho @Javi why did they even need to make swift. should have just used AppleScript

   • Embed this notice
     NeoNacho (neonacho@mastodon.social)'s status on Tuesday, 09-May-2023 14:01:10 JST NeoNacho

     in reply to

     • jesse squires

     @jsq @Javi that’s right

   • Embed this notice
     Paul Cantrell (inthehands@hachyderm.io)'s status on Tuesday, 09-May-2023 14:01:10 JST Paul Cantrell

     in reply to

     • jesse squires
     • NeoNacho

     @NeoNacho @jsq @Javi
     if set firstElem to the o3ptional value of the first of array then
     tell (tell array to drop prefix 1) to reduce into firstElem the operator *
     end if

     It’s just that easy!

   • Embed this notice
     NeoNacho (neonacho@mastodon.social)'s status on Tuesday, 09-May-2023 14:01:11 JST NeoNacho

     in reply to

     @Javi AppleScript is right there

   • Embed this notice
     Javi (javi@sfba.social)'s status on Wednesday, 10-May-2023 08:13:40 JST Javi

     in reply to

     • Paul Cantrell

     @inthehands Yeah that is super interesting. But even programming languages aren’t perfect at that. Bugs and crashes are also disconnects from what we thought we told the computer to do and what we actually told it to do

   • Embed this notice
     Paul Cantrell (inthehands@hachyderm.io)'s status on Wednesday, 10-May-2023 08:13:40 JST Paul Cantrell

     in reply to

     @Javi Well…yeah. The defining characteristic of programming languages is that they have both machine and human readers, and most of the work of p-lang design (and of writing readable code) centers around helping those two readers understand the code in the same way.

     LLM-generated code adds risk — nondeterministic, surprising, and (for the foreseeable future of AI) bizarre — on •both• sides of that equation.