Who's regularly using LLMs at work?

My org has a bot that generates AI summaries of all internal pull requests, and it's honestly been pretty helpful to compare that (a fairly accurate guess of what the PR will do) against the PR author's description (a statement of what the PR is meant to do). The main thing I look for in PRs is an alignment of intentions and actions, so the AI summary helps, especially for large PRs.

Of course, some engineers have stopped writing PR descriptions since the bot will do it for them. But that means that the only people who can effectively review that PR are the ones who already know what it's supposed to do, which is generally a small pool.

This has been a pattern I've seen repeated with workplace AI: they make something hard a little bit easier, but in a way that will the underlying problem worse over time.

I'm using it as a leverage for my side-project ideas. I was making excuses not to make open-source projects for years because of not having enough time.

Last week alone I launched two mini side projects, apparently using "code vibing" technique. Didn't know what I was doing had a name. - https://github.com/antonbelev/llm-fuse and https://github.com/antonbelev/hexo-bluesky-feed if you are interested.

I feel LLMs open the productivity door for me. Especially, outside my day-to-day work.

Yes, the tools I've built are not using the best practise, they may have little bugs here and there, but at least I can iterate quickly over my ideas, and get something out there.

So far I'm focusing on tools which I know at least I will be using for my personal blog.

I found it really useful for debugging errors. Missing an argument in a graphics call was always an easy fix. As I could rerun my code immediately and get feedback. I’d also use it for “give me a strawman document for …” and I’d then do the hard work by hand to fill it in. I never trusted text output but always worked over significantly. One thing I really enjoyed doing: I had to listen to a three hour presentation on IT governance. I captured the audio, converted to text, and then had copilot or gpt summarize and put together an outline. The audio was really not organized and there was no way I’d make it thru three hrs of note taking. I still spent some time going over the summary and filled in my content. I have a homeless friend who got an unfair parking ticket. I wrote up a short note and used a LLM to grind down the city’s parking enforcement to get the fee dismissed. So it’s good for responding to dumb organizational requests. I see a lot of people passing off auto generated work and it so often shows. To me it signals they are not to serious about what they are doing. I do find if I use it, I can quickly do things mindlessly and not fully understand what I did. I need to balance that as I like to do the work and stretch my brain and sometimes that does not happen.

I work in cloud consulting with a focus on app development. But I can do almost any vertical except Big Data/ML (learning that now). I use LLMs in every step of the process.

Anytime we talk to the client on a call we use Gong for text transcription (https://help.gong.io/docs/see-a-call-transcript). After the call is transcribed, it summarizes the call, gives you items to follow up and has chapters as the conversation topic changes.

Once I get all of the documentation, statements of work, any artifacts that we come up with, etc, I use Google’s NotebookLM. I put all of the artifacts in it including transcripts before I came in the project. You can then ask it questions and it will give you answers either with citations to your sources that you included.

I use ChatGPT along with NotebookLM to write the assessments and requirement docs along with a project plan. I’ve been doing this type of writing for awhile before LLMs were a thing so I do a lot of prompt iteration and editing so it sounds like me and not “AI Slop” (https://news.ycombinator.com/item?id=42909042).

After the project is signed, I then become a tech lead.

ChatGPT has been well trained on the AWS SDK for various languages, Terraform , the CDK etc. I use it to write scripts and Lambdas involving AWS. I don’t get a chance to do as much hands on coding as I use to between working with sales and as a tech lead.

I once used it to create a simple Hello World API. I was demonstrating to a Java shop how to deploy APIs to Lambda, ECS (AWS’s Docker orchestration service) and EC2 via Ci/CD. I was upfront with then about not knowing Java. I am a c#/Node/Python guy.

All of the tools I mentioned are specifically approved by my company and we use GSuite as our corporate standard and I think we added the pro version of NotebookLM to each seat a couple of weeks ago.

We have been using this system to combat scammers. Our team developed and trained an in-house LLM to initiate and manage conversations with scammers across various platforms, enabling us to continuously gather their information. In our company, our primary responsibility is to extract this data and share the scammer details with both the relevant platforms (for example, Meta) and law enforcement agencies.

We transitioned from a team of manual chatters, to hard-coded conversational scripts to an LLM-driven approach. This change has allowed us to handle interactions more accurately and scale to a much larger number of conversations per day.

Less grandmas are falling victim to crypto scams. Thanks AI.

Coding. Often the grunt work of iterating "what if this" and I get my code and try it, do some more "what if that" and go from there. Not rocket science but the sort of nit picky toil that I'd rather not expend brain on.

I use it for “introductory” programming, where it writes some script in an env I’m not familiar with and then I read the docs and clean all the bs it hallucinated into it. Often (but very not always) takes less time than learning-curveing it yourself.

I also get sysadm and networking ideas from LLMs. Although here their imagination has no limits and you have to fact-check everything. They imagine cli/config options, whole daemons and docsites, everything.

To sum up: for exploration and one-shot doing by example.

I'm actively building LLM tools and use Claude as both a sparring partner and UI mockup maker.

In the past, I've had good results combining the artifacts' React UI elements into Figma for mockups.

Some things are just faster with an LLM like making a dropdown and filling it with text, instead of doing it in Figma and manually typing out a bunch of dropdown items.

- I use them to create a vendor order each week. They take my voice and transcribe it into a list to text my sales reps.

- I use them to OCR PDFs from my bank

- I use them as a search/replace tool on CSVs

- I use them to replace Google searches

- I use them to code

- I'm trying to use them to view (still frames from) cameras periodically

I'll think of more uses, I'm sure. My work is very low tech and simple.

The main culprits are explicitly blocked on our VPN. When I WFH I use it on my personal machine for bash/regex/whatever stuff I don't keep in my brain for very long, never anything interesting.

I use it to generate some makefile or bas scripts and improve from my end.

I also use it to learn new stuffs e.g. Flink. I don't always trust it even if it works, though.

JMC wrapped with shallow JXR. Does the job perfectly & can't imagine working without them anymore.

We do. Can't imagine working without them anymore. JMC wrapped with JXR does the job perfectly.

I use them as a stack overflow replacement and it is also really good for generating tests

My tech team is using LLMs to build our product, leveraging them for tasks.

http://uss-express.com/