Hello, happy Wednesday, and welcome to Issue #3 of Morgan.xyz, that actually rhymes a bit doesn’t it? I’m excited to dive into this week’s issue, but before I do, one quick update - I’m going to start sending this newsletter out over the weekend because I realized that Wednesday is a bit of a weird day since AI coding tool companies are shipping like crazy all week long.

So rather than give you what happens from last Weds to this Weds, it feels more logical to me to share what happened in the AI coding tool world last week, right? If you disagree please send me an email and let me know, I’m still learning as I build up this newsletter and reader feedback is everything to me at this stage.

Okay, now onto the good stuff - to call the last week a big week in the world of AI coding tools would be an almost insulting understatement. Last week was huge, with a new model from OpenAI, o3-mini going live on Friday, and instantly seeing support from top AI-powered coding tools.

And that’s where we’re going to start, and of course with the question that’s on everyone’s mind, is o3-mini the first model to finally give Claude 3.5 Sonnet a run for its money when it comes to coding?

So let’s talk about o3-mini

First, a little primer from OpenAI on what o3-mini is. Just for the fun of it I decided to plug their announcement into o1-pro and have it summarized in one paragraph, so fresh from an H100 cluster to you, here’s the TLDR; on o3-mini ⬇️

Of course it didn’t take long for people to start comparing o3-mini to Claude 3.5 Sonnet and for whatever reason the test that seemed to be the most popular was a falling letters test, here’s one of the most popular analysis.

I would suggest clicking the tweet above to see the animation so you can judge for yourself, but to cut to the chase - most people agreed that Claude 3.5 Sonnet was the winner here with @LinusEkenstam probably saying it best.

Without a doubt the most humorous tweet comparing o3-mini to Claude 3.5 Sonnet is this gem from @iruletheworldmo 🤣

As for how o3-mini performs on some coding-specific tests, it absolutely crushes Aiden’s Bench as you can see from the chart below. Note, you do need to be using o3-mini:medium or high to see this kind of performance.

And as you probably could guess, AI coding tools were quick to add support for o3-mini.

When Cursor made their announcement they also shared that Cursor devs still prefer Sonnet for most tasks 👀

All that being said, it’s still early and my guess is we’ll see a lot more performance reports coming out over the next week or two. For now I’m using o3-mini in Windsurf myself so I’m taking it for a spin.

Automate terminal commands in Windsurf

Last week Windsurf shared a pretty nifty new feature that definitely solves for something I’ve run into quite a bit with AI-powered IDEs, and that’s terminal commands.

The tweet above includes a short video walking through the feature but I’ll give you the dets here too. Essentially there’s a checkbox deep in Windsurf settings that you’ve probably never noticed before under Windsurf Editor labeled “Auto Execution Policy”

This comes in three flavors, off (default), auto, and shred, and lets be honest, shred is one of the most hilarious dropdown settings I’ve seen before. Auto means the model will decide whether to auto-execute based on the command. And shred, well, it means what you think - always auto-execute.

When it comes to getting and staying in flow state and just moving faster with AI-powered IDEs this seems like a pretty stellar feature to me.

Bolt has been busy ⚡

Okay, so let’s switch gears to Bolt who released so many new things last week I put together a thread just to capture some of them. You can give it a read to play catch up.

Get rocking with Bolt.diy

If you don’t know what Bolt.diy is, let’s change that. You all know Bolt, heck I just shared all the crazy awesome stuff they did last week above. Well Boly.diy is simply a way to run Bolt yourself, since Bolt is Open Source, so you can run it anywhere you want and swap out the model under-the-hood while you’re at it 👀

Well last week one of my favorite people on X, Melvin, put together an awesome video showing you how to get Bolt.diy going, and the video is literally just a minute and twenty seconds…so you have no excuses.

This tutorial is also a fun one because it shows you how to deploy using Railway which I’ll be honest, I’ve never used, and even though I already have Bolt.diy up and running, I’m going to try it again using Melvin’s tutorial above over the weekend. In case you don’t know what Railway is, here’s the skinny:

Okay, so let’s talk about Chunkr.ai and why it’s going to supercharge RAG

This week, thanks to this tweet from Sam Denty, I discovered Chunkr.ai and honestly, it kinda blew my mind. For anyone that has been doing stuff in the RAG space, you know that getting data out of PDF files is one of the biggest, if not the biggest pain points.

Now just to set expectations here, I am not a RAG expert, heck I wouldn’t even call myself a RAG beginner, I’ve played around with it a little but wouldn’t even give myself a one out of ten so don’t take this as any kind of expert opinion.

But, based on what I know about RAG, and challenges around getting data from PDFs, it does seem like Chunkr.ai might just be a game-changer. I wanted to include it in my newsletter this week for that reason, so if you do stuff with RAG, check it out and feel free to shoot me an email or @ me on X and let me know what you think.

Bonus round: what I think might be the slickest new AI-powered dev workflow

Okay, ending with another bonus round, this one is a tweet from Claire Vo, the CPTO of Launch Darkly and founder of ChatPRD. And shameless plug, ChatPRD is awesome and you’re honestly crazy if you don’t use it. And no I’m not being paid to say this, and it’s not an affiliate link, I use ChatPRD and love it, that’s all.

Okay now that you know who Claire is, here’s a tweet she shared this week that I think might just be the slickest new AI-powered dev workflow I’ve seen yet ⬇️

Claire is using Cursor as her AI-powered IDE, hence the <tab> <tab> <tab>ing her way to an 100 file PR. And then uses Devin.ai to go through and make comments on higher risk changes, and ends with human code review and testing.

I honestly think this is exactly the direction software engineering is going. You’ll start by writing code in a tool like Cursor or Windsurf, if you already know how to code. Bolt or Lovable if you don’t. Then you’ll use an AI Agent that’s essentially like a junior dev who will layer on from there, and you’ll end with a real human that does know how to code.

The good news for all of us is that nowhere in this process does the human go away, but what the human does, is a lot more specific, and wastes a lot less time doing things that AI can do faster, and better than us.

Thanks so much for reading! If you enjoyed my newsletter I would be eternally grateful if you shared it on X, LinkedIn, wherever you are active. Right now I have 55 subscribers, I’m hoping to have 550 by the end of the year, and you can help me hit that goal.

Until next time.

Morgan