Ancestors

Toot

Written by Dan Luu on 2025-01-09 at 00:00

This exchange between some programmers and a non-programmer typifies what I was getting at in https://danluu.com/codenames/#appendix-writing-the-code-for-the-post

Programmer 1: GPT-4o and Claude Haiku are useless for programming

Programmer 2: Claude Sonnet is useless for programming

Non-programmer: What do you mean GPT-4o is useless? I don't know how to program and created an app that makes $10k/mo with GPT-4o

LLMs have allowed non-programmers to produce apps for years and programmers are calling these things useless for programming.

=> View attached media | View attached media

=> More informations about this toot | More toots from danluu@mastodon.social

Descendants

Written by Dan Luu on 2025-01-09 at 00:08

And, sure, the LLM generated code isn't great, but if I compare the LLM-generated game and "AI" in https://danluu.com/codenames/ to commercially successful online board game implementations that were created by human programmers, the LLM generated version is faster and less buggy.

Per the argument in https://danluu.com/customer-service/, AI doesn't have to be very good to replace humans in a lot of roles because, in practice, humans often aren't all that good (see also, https://danluu.com/p95-skill/).

=> More informations about this toot | More toots from danluu@mastodon.social

Written by Dan Turner on 2025-01-09 at 05:04

@danluu LLMs let people pass off others' work and ideas as their own, without credit, while convincing themselves they're not doing plagiarism. It almost doesn't matter how good they are if none of the benefits to the people who made them possible.

=> More informations about this toot | More toots from dan_turner@cosocial.ca

Written by Mikołaj Hołysz on 2025-01-09 at 00:17

@danluu I know some people who kind-of sort-of remember what an if statement is from their school days, are smart and very comfortable with tech stuff, the command line, server administration, following tutorials etc, but wouldn't call themselves programmers by any stretch of the imagination, and they've had great success with LLMs for basic / one-off apps.

=> More informations about this toot | More toots from miki@dragonscave.space

Written by Jeff Miller (orange hatband) on 2025-01-09 at 00:17

@danluu I have a programmer friend who is bullish on generative models and suggests that they provide opportunities for hand-written tests as specification, rapidly satisfied by having the generative model implement the application code.

Another paradigm shift moment was Jon Udell's report of synthesizing an application to convert a mobile phone photo of an event flyer into a calendar event, an app domain he had been working on for a long time (community calendaring). The generated app was a swift jump forward in capability without requiring highly technical specs.

I'm pretty negative personally because the signal for good generative model use cases is drowned out by the undiscriminating hype around the slogans.

=> More informations about this toot | More toots from jmeowmeow@hachyderm.io

Written by Jeff Miller (orange hatband) on 2025-01-09 at 00:37

@danluu "Not all of the money that's saved from phone trees is turned into profit — some goes into hiring support people who handle other tasks."

Your note on automatically generated support trees is part of the"everyone wins" scenario that a good support chatbot can deliver, making FAQs and standard requests fast for users and easy for support providers, allowing reps to concentrate on understanding rather than speed. Requests escalated to human reps come with user context ready at hand. Success metrics on the automated support interactions allow it to be tuned for faster or more effective responses, saving more human exception handling work.

But only a few orgs implement "everyone wins" levels of conversational agent support.

=> More informations about this toot | More toots from jmeowmeow@hachyderm.io

Written by Andrew Helwer on 2025-01-09 at 00:50

@danluu “my side project pulls in $X0,000 per month” posts on HN are probably the least believable type of post that regularly occurs there, doubly so if LLMs are involved.

=> More informations about this toot | More toots from ahelwer@discuss.systems

Written by th0ma5 on 2025-01-09 at 01:14

@ahelwer @danluu It is so common, like a new kind of logical fallacy, for people who are newly enabled by them to see anyone warning them about their show stopping faults to be trying to gatekeep, thus emboldening them to continue being a true believer

=> More informations about this toot | More toots from th0ma5@mastodon.social

Written by Donald Ball on 2025-01-09 at 01:16

@ahelwer @danluu This part. Hustlers gonna hustle.

=> More informations about this toot | More toots from donaldball@triangletoot.party

Written by soc on 2025-01-09 at 12:48

@danluu Now that claim only needs to be not made up!

=> More informations about this toot | More toots from soc@chaos.social