Watermarking, harnesses and clever prompt compressions
What do watermarks, watches and answer boxes have in common?
Newsletter #4!
Welcome to the FOURTH newsletter blast for Answering Machines. Yes, I’m still doing this. Yes I wrote way too long of an article again and it took forever. I promise, it’s worth your time.
One new thing: hold music!
Before we get to this week’s (months?!! quarters?) article, I wanna try something new.
I love music, specifically rap music, and I love finding and listening to fun lyrics or verses that I wanna share with other people. So, I’m gonna reserve a tiny section each newsletter for this. I’m calling it hold music, and here’s this times selection, fresh off of J.Cole’s last album, 39 Intro from The Fall Off:
I'm pleased to report, after many attempts
I climbed out from the mouth of the gritty abyss
The antonym of idiot, who witty as this?
My mind frame has been trained with Nvidia chips
My rhymes are like shrines for the city I miss
Where, in time, I found beauty in the hideousness
39 Intro splits The Fall Off exactly in half. The whole album is J.Cole’s meditation on his career and why he spent so much time doing what he does. On this verse, he’s kinda just going off and having fun and I appreciate that so much.
Alright, back to what you came here for :)
Why are people so mad about watermarking?
There was an enormous uproar around Anthropic’s decision to implement watermarking in output to compy with EU regulation. I decided to write about why this has more in common than you think with watch springs and other mental models…
By the way, if you got this email from someone else, or found it on the web, please subscribe. I promise, it’ll be a good time.
Updates on Shikhu
No major updates recently, although I finally have time to revisit my project. I have an idea concerning evals that I think would be really fun to tool around with. I also wanna add OpenRouter support! More things coming soon!
Next, some interesting tidbits from the internet:
What is a harness? from the makers of Pi
Just a really good, clean writeup on what the heck an agent harness really is
Arjun’s Take:
Sometimes you just read something and you already know what it is, but it feels good to read it from another person. This is one of those things!
Professor Ng on what AI Engineers need to learn now
A breakdown from the professor on the skills AI Engineers need to know for the future
Arjun’s Take:
As always, really good stuff. Love the focus on evals. Hate the haters in the X comments, damn! Put some respect on the name…
Someone very familiar giving a talk on retrieval and birds!
An hourlong talk on the importance of retrieval evals in building RAG
Arjun’s Take:
I did a big talk with Gen AI Academy, and over 800 people registered to watch it! If you are curious about seeing me in action, take a look!
A new research paper demonstating that prompts don’t gotta be readable to be good
Arjun’s Take:
Fun paper, demonstrates that you can use pretty well-known linguistic algos to compress prompts and have them still work. BRB writing a hook for my Claude code sessions. BTW, what’s with Claude Code/Opus writing so terribly lately?!?!
See ya later, space cowboy
That’s it for this time, thanks for tuning in. And as always, if you enjoyed what you read, hit that reply and let me know, or share it somewhere on socials and tag me! Ya boy will appreciate it.
