Good morning and welcome to the second issue of the Morgan.xyz Newsletter. Since I have some new subscribers, I’ll start by saying - welcome 👋 I’m Morgan, the co-founder and CTO of Bold Metrics, a startup using AI to help companies unlock the power of body data. If you want to see us in action, you can see a bunch of our customers listed on our homepage, or you can pop over to one of my personal favs, Vuori and just click the “Find my Size” button on any product page.

This newsletter is about the intersection of AI + Coding, a major paradigm shift that I’m so fanatically excited about I can’t help but shout from the rooftops. And with that said, let’s start the shouting, I have a lot to share in this issue and you can bet DeepSeek is front-and-center. So let’s dive in 🐳

Let’s talk about the elephant whale in the room

DeepSeek has absolutely dominated the news this week, and not just in the AI world, everywhere, and for good reason, but how good of a reason we’ll likely learn over the next few weeks.

What has made DeepSeek so interesting is a novel new way of training LLMs that the company claims made it possible for them to train the model for less than $6M, and without using any fancy NVidia hardware.

This, not surprisingly, sent the stock market into a tailspin as investors worried that maybe demand for NVidia’s chips would wane, dramatically, and well, there’s no other way to put it - people went into full panic selling mode.

And yes, the stock did shed hundreds of billions of dollars but has since been creeping back up. AI experts have been quick to jump in and try to correct the narrative. The key confusion that investors seem to have is not understanding the difference between training and inference. Yann LeCun, the Chief AI Scientist at Meta, and a professor at NYU did a good job breaking down the confusion in this post on Threads ⬇️

At the same time, Matthew Carrigan shared this post on X yesterday with details specs for building out a machine that can run DeepSeek R1 locally, at a cost of around $6,000. And while that’s a pretty stellar price, I haven’t done a deep dive into the performance so take this with a grain of salt. Still, it makes you think.

So let’s get technical - why is DeepSeek so good?

DeepSeek is so good, and so interesting, because of a novel new approach to training, and interesting because the company made it Open Source so anyone can use it for free on their own hardware. Remember, OpenAI and Anthropic, the two leading foundational LLM companies in the US have closed-source models, you can’t run their models without paying them.

Chamath wrote an excellent post on X about why DeepSeek really is so special, and it has a lot to do with getting models to reason step-by-step without requiring a huge dataset. Here’s the post ⬇️

Now if you want to really geek out, Akshay did a pretty awesome comparison of DeepSeek R1 vs. OpenAI o1 using RAG that’s definitely worth taking a look at, the tweet has 1M views at this point so safe to say it’s getting a lot of attention 🤩

But DeepSeek has a deep privacy problem

As DeepSeek frenzy swept the world, people started to look into their privacy policy, and well, what they saw was pretty shocking. If you use DeepSeek through their website or app, you’re not only giving them all your data, you might also be giving them access to your email and can even collect keystrokes from your device.

Yes, collect keystrokes, that means anything you type, goes to them.

So this freaked people out a lot, and rightly so, but a lot of people (myself included) quickly looked at getting DeepSeek up and running locally, and when you run it locally, all of your data stays nice and safe.

So how can you use DeepSeek as an AI coding companion, safely?

There are a lot of ways you can (and probably should) be using DeepSeek as a coding companion safely, without giving away all of your data, code, emails, etc. If you want to run DeepSeek directly, on your own machine, I highly recommend using Ollama.

In fact, I actually taught my Dad how to run DeepSeek locally on his MacBook this week so if he can do it, you can do it. I put my walkthrough together on Medium so you can follow along and get it setup for yourself in a matter of minutes.

And of course, AI-powered coding tools have been quick to add DeepSeek, Bolt CEO Eric Simons confirmed two days ago that DeepSeek is coming to Bolt.new.

And people are also starting to share examples of DeepSeek in action in AI coding tools like Windsurf:

The key to using DeepSeek safely for coding is to not use the DeepSeek website or app, install DeepSeek locally, or use it through a third party that is running it locally, that you trust.

I haven’t seen any updates from Lovable or Cursor, and for whatever reason Cursor seems to have stopped tweeting which is a bit odd. Their last tweet was two weeks ago and that was their one and only tweet in January 👀

This has honestly been one of the biggest surprises to me this month as the AI coding community is absolutely on fire on X, and every AI coding tools tweets like crazy, and Cursor was tweeting a lot last year…but then they fell off the map. It will be interesting to see how this changes over the next few months, right now I have to imagine Windsurf is taking marketshare as a result.

Okay, so I’ve written almost 1,000 words so far and it’s all been focused on DeepSeek so I’ll end it there for now, but I’m guessing DeepSeek will be a core topic for a while now so more to come next week!

Cline.bot is now on my radar

I learned about a new coding tool this week and plan to do a deeper dive this weekend, but since it’s on my radar, I thought I’d also share it with all of you.

Cline is a VS Code Extension that I thought had some pretty sharp marketing, it’s not just an AI coder, it’s a “thoughtful” AI coder. It’s like having an AI coding partner that really takes the time to think before it codes, and yes - you can use any model you’d like with it so it’s not locked into one specific model.

Here’s my favorite tweet describing Cline:

The Cline team has been absolutely rocking it on X, sharing a lot of good nuggets, here’s one of my favorites from earlier this week:

This got my gears turning. Using DeepSeek (and yes, we’re talking about it again!) as a “code archaeologist” and creating sequence diagrams before actually writing any code is a super interesting idea. And yes, while you can use Claude 3.5 Sonnet to do this, it’s going to cost a lot more.

Still thinking about this more but I think this could be a really useful path to take when using AI coding tools. While we all want to just dive in and code, and it has been expensive to use tokens on planning, with DeepSeek, maybe we can rethink things a bit, and in the process, get the models to do more thinking 🧠

SupaBase gets Supa-Generous

I woke up this morning to see a pretty awesome tweet from Paul Copplestone, the CEO of Supabase. I became a huge fan of Supabase after Bolt.new added a really slick integration with them. You can think of a Supabase as a free firebase alternative, and today, Supabase is making free even more free.

This is a pretty big move and as someone who now uses Supabase daily, I’m pretty darn excited about this so thanks Paul, you rock!

Bonus round - AppWrite.io

I thought I’d add a little section at the end of my newsletter that I’m calling “bonus round” - essentially a section for new things I just discovered and thought other people might find interesting.

This morning, probably about an hour ago, I saw a tweet about a tool called AppWrite.io that is being called BaaS or backend-as-a-service, which is an acronym I haven’t heard before, but I dig.

So what is AppWrite? Here’s the skinny from Nick ⬇️

Okay and at over 1,400 words I’ll put a bow on this one. I’m doing my best to keep my newsletter at ~1,500 words. I know there are so many newsletters out there and none of us have enough time, so I’m hoping that by keeping mine in this range, it will become a regular weekly read for many of you.

And with that I’ll say goodbye, thanks for reading and I’ll see you next week!