←back to thread

634 points david927 | 1 comments | | HN request time: 0.238s | source

What are you working on? Any new ideas that you're thinking about?
1. bsenftner ◴[] No.41346095[source]
I'm working in the final features for an office productivity suite with deep AI integration. I think the current slapping of a chat window on software is pretty tame, if not lame. So, in addition to that chat window that replies with conversational chat, I've got over a dozen integrated into software tools LLMs that when you ask them things, their output is programmatic and either modifies the data of what you're working on, or or performs some type of modification to the software tool in use. The tools include a complete word processor, where the AI knows the word processor's API and can directly manipulate the document in the word processor, likewise for a spreadsheet, and there is a "prompt engineer" interface where one can clone, modify, and create new LLM Agents that have deep access to the APIs of the tools, the data in the tools, and the wider application framework.

This has been in development for 3 years, and I've got an immigration law firm using it, with about 1/3rd of the LLM agents being immigration law specialist agents of some type.

I'm just wrapping up transforming the system from being immigration law specific to being generic, capable of operating in any industry. I previously made a "do it yourself: build your own home solar energy system" as a proof of concept that this framework could be modified in such a way, which I put online for a few months and then took down, having proved what I wanted.

I've got multiple creative writing 'bots: legal, technical documentation, creative writing, and code authoring. I've got multiple spreadsheet 'bots: create any standard spreadsheet form on demand, reverse engineer and explain complex spreadsheets, and co-author spreadsheets interactively with the human user, guiding them through the understanding of the spreadsheet being built. I've got foreign language translation agents that allow people without a common language to speak to one another through the voice transcription interfaces.

And the users are never copying and pasting LLM outputs from one place to another, that integration is built in to the business logic of the software: ask a chatbot to write a document, the output does directly to the word processor and the document is created, likewise for spreadsheets, likewise for talking to the "projectBot" and asking "what's the state of this project?" and a detailed report is generated.

I've also been making the app itself multi-lingual, and multi-skinned so it can be refaced for different cultures, demographics, and industries. I've been calling it "AI CMS" but that is meaningless to far too many. I'm considering calling it "Midom Office AI" because that sounds like "my dumb AI" and I'm generally sarcastic, considering an anti-gushing sarcastic marketing angle on the software. Rather than everyone's else's over praising, I'll have just some confident smuck referencing how he's got an entire team of AI experts helping him, enabling him to be calm and cool in the face of all the deadline pressures, he sips lemonade while his AI team works for him, and not him for it. We'll see what my "marketingBot" says...