Some links to threads I pulled together, won't be complete of course: https://futureofcoding.slack.com/archives/CLYCGTCPL/p1679833469621219 https://futureofcoding.slack.com/archives/C5T9GPWFL/p1679642239661619 https://futureofcoding.slack.com/archives/C5T9GPWFL/p1679902227016569 https://futureofcoding.slack.com/archives/C5T9GPWFL/p1679892669316079

Oh! Did I put mine first? Completely unintentional obvs.

ChatGPT (4) and DALL-E seem to be very slow atm, almost unusable. What if the demand for these services grows so fast that the semiconductor industry can't keep up? Could it be that AI won't take over the world within the next couple years as either there won't be enough computational power to run all the queries, or the prices will be hiked by a lot to balance the demand for the available computation power?

just goofing around... pt: say a pithy comment about commenting about artificial intelligence ChatGPT: Commenting on artificial intelligence is like teaching a fish to swim - we’re all still figuring it out. pt: When will ChatGPT be trained only on code generated by ChatGPT? ChatGPT: “... pt: What is in-breeding? ChatGPT: ... it can also increase the expression of deleterious or harmful recessive alleles, leading to reduced fitness or genetic disorders. ... pt: Who said “you’re fired”? ChatGPT: “... pt: Who said “just goofing around”? ChatGPT: <crickets - generating for many minutes - answer did not include ‘22 Minutes’>

Found this tool to use LLMs on command-line with pre-define roles/prompts etc: https://github.com/sigoden/aichat Here are some sample roles/prompts: https://github.com/olalonde/chatgpt-roles/blob/master/roles.yaml

@Jimmy Miller what do you make of section 4 of https://www.noemamag.com/gpts-very-inhuman-mind/

Invokes Ryle and Dennett

Example of using LLM and classical verifier (a compiler for some DSL) to boost an LLM's code generation: https://arxiv.org/pdf/2303.14100.pdf

Bespoke UIs mixed in with Chat.

There are two technologies that I'd expect to be useful in the context of LLMs: 1. Controlled natural langues (see https://sites.google.com/site/controllednaturallanguage/ for a list), which are subsets of natural languages that can be assigned formal semantics. 2. Concatenative programming languages (https://concatenative.org/wiki/view/Concatenative%20language), which are formal language whose syntax shares the linear structure of natural languages that is the basis of LLMs. Has anyone seen such connections pointed out or explored yet?

I built this command-line bot that automates complex online tasks using both the browser and the APIs of cloud-services: https://github.com/aicombinator/bot It's built with GPT-3.5-Turbo, Puppeteer and ActivePieces.

So is it the case that

langchain

and

agents

are specifically built prompt templates that are backed with some fun interprocess coms, passing the results to various piped tools like a python / JS script to get the next ‘prompt piece’, and then looping until something ‘correct’ comes up? I feel https://github.com/mpaepper/llm_agents simplified the abstractions a bit to help me understand what was going on.

Thinking about multi-agent / multi-tool systems where humans and bots collaborate:

DoD calls this “human-machine teaming”

(Though it is assumed that there’s a human in the loop, so it’s not the same thing)

If we assume GPT is system 1, can we get it to "develop" a system 2 by asking it the right questions? 1. I and others have had success asking it to work step-by-step and talk out loud. 2. I've also had success asking it to provide a better answer after it answers. 3. I've tried doing a de Bono's Thinking Hats scenario to answer questions. I think there's some interesting work there to be done, especially regarding "frameworks". Like ask it to use a framework to answer the question. Ask it to use several different frameworks then synthesize a single answer. The cool thing is, it tends to already know the framework. You don't have to program it in, you just have to ask it to use the framework. I also don't know if there is any advantage to using multiple chats (that don't share history) to collaborate on an answer vs using a single chat with a single history.

Using GPT to generate an app's GUI on the fly. Basically, the design of the app is transient and created dynamically based on the exact need of the moment. https://mastodon.social/@stroughtonsmith/110193697691243762

LLM prompting with a visual programming language https://arxiv.org/abs/2304.08103

The end sample mentions injection of expert domain knowledge, and calls out psychology and mock interviews. Interviews? Sure. Psych I’m not sure.. I can hardly trust a human to understand my meanings and intent therapeutically, and the context build up is part of building session value up. That seems quite hard to stuff into a small sliding window.

I don't know if I want to laugh of cry 😂. This is done just by duct-taping a bunch of APIs and GPT-4 prompts together. Maybe the future of programming will be more English than code? https://chameleon-llm.github.io/. Paper (https://arxiv.org/abs/2304.09842) and Code (https://github.com/lupantech/chameleon-llm)

Been thinking about the infra needed for human-machine teaming, I think a lot of knowledge from the field of business process management can be reimagined, just with powerful natural language capabilities. • Natural language becomes the API interface between humans and machines. • No need to formally describe the workflow. Just talk to it and let the LLM create/modify/delete these. • These workflows are essentially Turing-complete, distributed program where the job "processors" are people and bots. For eg: BPMN already supported branching, sequential processing, parallel processing etc. • Open-source Zapier-equivalent systems can be a good starting point for these.

‘Godfather of AI’ Geoffrey Hinton quits Google and warns over dangers of misinformation” - maybe old news, but I haven’t seen it mentioned here... https://www.theguardian.com/technology/2023/may/02/geoffrey-hinton-godfather-of-ai-quits-google-warns-dangers-of-machine-learning

Perspective on AI from a music producer (He reminds us of Napster and concludes that networks, like Paramount+, Apple Music, Spotify, etc. will simply side-step artists and music labels altogether)

AI autocompletion like Copilot was just the beginning - we're going to be getting more agents trying to automate much of the process of software engineering. The future of coding is starting to look like management. https://github.com/smol-ai/developer/

• humans have cognitive limits, • e.g. the Magical Number Seven, Plus or Minus Two • https://en.wikipedia.org/wiki/The_Magical_Number_Seven,_Plus_or_Minus_Two • does GPT have similar cognitive limits? • is this a Golden Number? a universal truth? a universal constant? • is it necessary to arrange information in a hierarchy? • or, does GPT’s perfect memory change this pattern?