To experiment with multi-agent systems you either need a clear business case and big budget, or you need to be a jabroni with a high credit limit.
I'm just the jabroni for the job, so I taught Claude to play chess, and learned a ton along the way:
jfkirk.github.io/posts/claude...
Checking the timeline after being away for a bit
Randy, what is going on here?... nvm
😜
What your current best setup / best model / best practice when using aider?
Claude 3.7 Sonnet for everything? Is reasoning mode worth it?
The stress of millions of hours of whiteboard coding interviews has been boiled down into DeepSeek R1's coding monologues for you to experience on demand
Enjoy
This is what we were using LLMs for before we go the green light to use them in production
Roy's three levels of ML/AI enlightenment:
1. ML is amazing. Let's use it everywhere
2. We should probably avoid ML if we can
3. Ok. Fine. Sometimes there's also marketing and other non-technical value to a company in using "AI".
Somehow I didn't know there was a catchy name for this 🍲
I think there are some videos/pics of Jimi Hendrix tearing down the house while the audience is just seated staring at him
I'm using a random, open source signal processing app written 10+ years ago in C with zero documentation.
I was successfully able to ask GPT-4o mini (via duck.ai) to explain the default settings to me by copy-pasting the source code.
A happy LLM user over here 🖐️
save time with the shortcut `uvx sprint-planning`
A reusable McDonalds coffee cup from the Netherlands
These were more CPU heat sink style though
I kept a couple of those cups as souvenirs from the McDonald's in the Amsterdam airport instead of getting the Euro back
Spaß und Genuss
What are the best resources for learning how to interact with Bluesky via API?
Let me know if any of you have recommendations for other ways to send yourself notifications reliably.
This stack is just Python + bash + Keybase + cron
I decided to recycle my old script that uses Keybase to send me alerts on my phone and it still works out of the box 🎉
I'm glad/surprised that Keybase is still going. It seemed to have a lot of promise and buzz before it was purchased and put on life support.
gist.github.com/roycoding/87...
Alright, now that everyone is here, what are the new/cool data conferences that people are looking to attend in 2025?
If you were part of NormConf and I missed you, let me know so I can add you to this Starter Pack
The number of people that I couldn't find a couple days ago that are now here is not insignificant!
NormConf was such a vibe
Maybe there will be a 2.0 one day
youtube.com/@normconf
I just got here and I gotta say I’m getting choked up seeing everybody in one place again.
Very similar to my team's work
Definitely a lot of areas I'd stay away from right now.
We're using it for human-in-the-loop (user-in-the-loop) stuff, replacing task specific ML models we previously trained.
One nice thing is that we had "multi-language support" languishing on the future roadmap forever when we were still finetuning BERT-based models, but we basically just got it for free when we switched to LLMs.
We haven't even announced this feature to our customers yet, but it's there in prod.
What are some things people are using LLMs in production for these days (that are not chat)?
My team is using LLMs for discrete information extraction from PDFs. It lets us move really fast, but of course you still need to do all of the standard ML eval stuff.
What are you using LLMs for in prod?
I created a Starter Pack with people from NormConf 🎁
go.bsky.app/PhyvcCM
I've been playing around with uv and am definitely in the "this is going to be a game changer" camp.
Some one can run the entire Python project with a single uv command. Everything else is taken care of under the hood: Python version, virtual env, dependencies. Everything.
docs.astral.sh/uv/
Let's make Bluesky good!
I miss the old days of high signal Twitter. Maybe this will really be the new place...
My prompts usually end with "in the style of Chris Albon. Yes, that Chris Albon, smh"