zachary delong

Cleaning Up Nudge

Sun, 27 Sep 2026 03:14:28 GMT

Today I did a bunch of work on Nudge, a little Mac app I built a while back that pops up Claude’s permission requests so I can answer them without switching back to the terminal every time, like when I’m doing homework in Chrome.

Nudge only runs on Mac and I mostly work at my desktop, so I started off today by connecting the two with Tailscale. Now Claude can SSH into my Mac from my desktop. Then I installed Peekaboo, which lets Claude take screenshots of the Mac and click around in it. So when I’m at my desktop and want to work on something for macOS, I can just do it from there.

I’m not going to lie, it was kind of a bitch to set up. I had to jump through a good amount of hoops, but I got it working.

In terms of the work on Nudge, most of the work was put in just making it much cleaner, running a bunch of tests, and getting rid of a bunch of bugs. For example, when you’re in full screen mode, the Nudge pop-down wouldn’t show up, or it would still pop up permission requests even when Claude was in auto mode. Simple stuff like that. I ran a bunch of tests, which, for the most part, Claude actually did, and after a couple hours it’s pretty clean (283 unit tests now).

On top of that, last time I said I was working on Codex support. That’s merged now, so Nudge works seamlessly with both Codex and Claude, which is awesome. It all went out today as v1.4.0.

That took up most of the day, but since it’s only 8 pm I’m going to get back to the Coral-Board project.

Coral-Board & More Obsidian

Fri, 25 Sep 2026 21:34:30 GMT

All right I’ve been working a bit more on Obsidian since the last log and decided to upload my Obsidian notes to GitHub so when I switch between my Mac and my desktop, I don’t have to store separate notes and whatnot. I also just didn’t want to have to pay for Obsidian Sync so it’s just a private repo, synced with the Obsidian Git plugin.

In addition to that I’ve been using Opus 5.5 a lot just because of how cheap and intelligent it is. I can run a bunch of research and just do a bunch of products and I literally don’t even come close to my rate limit, which is fucking awesome.

On top of that I started work on a Coral-Board project, which I was able to get from my good friend Justin. The Coral-Board has an NPU, which is like this chip that’s built to run neural networks using barely any power. It does this by storing numbers at lower precision: INT8 (meaning 8 bits, so 256 possible values) or INT4 (4 bits, 16 values) instead of the traditional 32-bit.

I decided that I’m going to start off with testing if memory can survive being stored that coarsely. I decided to use Hebbian memory, which was inspired by a paper I read last year called “The Dragon Hatchling” (BDH), which is an architecture that uses Hebbian learning. You might recognize it by saying, “Neurons that fire together wire together.” Every time two things show up together, the connection between them gets stronger and the old connections fade over time.

I’m currently running some simulations and then after that I am going to set up the Coral-Board and start running, depending on the results of that, which I sort of have a strong idea of what they’ll be, I think I’m gonna start on a more interesting project. You can follow along in the repo.

The idea is to have an image detection model that uses Hebbian style learning to remember certain shapes or objects, and gets better at it over time.

Obsidian?

Thu, 24 Sep 2026 04:00:24 GMT

Today I haven’t done much in terms of generating anything but I did sort of want to try out something with Obsidian.

So I was inspired by this guy I met in class. He said that because he’s managing a bunch of projects from both the Claude terminal and Codex, it can get sort of confusing as to managing different files and stuff like that, as well as him sort of keeping track of what they’re doing. He just chose to use Obsidian.

I have used Obsidian for a while now but it is mostly just for my notes. I used it a while, like a year ago, for biology, philosophy, and just a lot of note-taking courses that I had to take. I use it a bit now but nothing for my agents or anything like that. I decided to give it a try. If it doesn’t work out then I can easily disable it. We’ll see how it goes. Just an experiment.

On top of that I’ve been using Wispr Flow a lot. I just got a subscription. I used it previously around 6 months ago and I decided to use it now and just see where it takes me. I’m really just trying to do as much experimenting as possible, see what I fuck with, see what I don’t fuck with, sort of trying to get the vibe.

New Model

Wed, 23 Sep 2026 06:47:29 GMT

Today Claude Opus 5.5 and GPT Sol and Luna came out. My experience with them has been somewhat mind-blowing.

First things first, I’ve conversed a lot more with Opus 5.5 than the GPT 6 models. In my experience, Opus 5.5 is so much better at regular conversation and writing. When I talk to it, it feels so much more natural. It feels less try-hard, if that makes sense.

I had Opus 5.5 write a couple of animations and videos and I was incredibly impressed. It does not look AI-generated anymore. It’s very, very impressive and it reminded me of the picture of the quote, “How far do we have to stretch the picture before pixelating the human texture?” I literally remind myself of that quote every day.

Now in light of there being a new model release, I decided to work on some projects that I have not really touched in a minute. One of them was Nudge, in which I made some meaningful improvements to the overall working order of it (the tmux version of Claude Code) and it overall just feels much smoother. I am currently working on a connection to Codex.

In addition I had it work on some other stuff like Christmas cards and just the journal. I had it come up with a couple of different sorts of designs just to see how its UI work would do. Again I was very impressed.

Overall today was pretty good. I definitely was locked out of my physics class just because of the new models and everything.

Gnoled

Wed, 16 Sep 2026 21:35:19 GMT

Spent a good couple hours last night working on this website. It started off with me reading the most recent post on Shtetl-Optimized (Scott Aaronson’s Blog) called ‘The Age of Wonders and Terrors’. It was an incredibly good read. Towards the end he recommended an essay by Stanford math undergrad Logan Graves titled “Priest, Monk, and Mathematician”, which I highly recommend reading.

Anyways, to get to the point, I found his website simple yet elegant. It was quite inspiring and I felt the pull to create something similar. That is to say, the site’s aesthetic is heavily inspired by his.

I used the same font for the main text, and a bold, italicized sans-serif for headings. As for the intro, I first struggled with trying to identify myself in one sentence. At first I went with the classic “student building AI applications” but it didn’t feel like me. I thought back to a favorite quote of mine, one I think about now more than ever — “How far do we have to stretch the picture, before pixelating the human texture?” (Shing02). Paired with my curiosity, I thought it captured what I’m currently after.

I also added nods to some projects I’m proud of, influences, and a painting by Albert Bierstadt with Calvin and Hobbes added in, inspired by my good friend Justin.

AgentRacer and QWOP

Tue, 15 Sep 2026 21:22:54 GMT

Started work today on both AgentRacer and QWOP.

AgentRacer is going good but ran into some issues. So far we have been using passages that we found online for NitroType. However, after doing some testing I found the models were likely trained on the passages, meaning they could easily predict what was being said next. I decided to transition away from general passages and integrate something analogous to monkeytype - randomized words. I’m utilized the google 10000 English frequency list, with a filter for only lowercase and a range of 2-8 letters. Going to run this for all versions (sprint, scramble, and reveal).

At the same time I’ve also sort of revived an old project of mine that I lost passion for. It started out with me wanting to try out CleanRL. I got it set up for QWOP and well the results were lackluster. I decided to give Claude a go at it and it only showed marginal improvements. This was 5-6 months ago. Today I got Astra to take a look at it. We landed on a much better environmental design. Doing a couple of things: For one I’m choosing a bunch of different runner networks. On top of that the training method is PPO and Astra is doing network design and experimentation. What’s the Goal? Not sure yet, just having fun.

After this week imma start on the CoralBoard project and take a look at Omar’s findings when it comes to AI being terrible at Yu-gi-oh.

this is not a blog, this is a LOG

Tue, 15 Sep 2026 09:06:18 GMT

Jesus - first post and I already fucked it up by calling this a blog

First Post

Tue, 15 Sep 2026 08:13:59 GMT

Alright just made this blog. Honestly was inspired after watching The Social Network. When I checked out the current options I wasn’t satisfied, maybe I didn’t look hard enough. Anyways I got fable to create this thing. I might like it. I might only use it a couple times. Who cares.

Right now I’ve been working on AgentRacer. It’s like this sort of benchmark thing where a bunch of models use computer-use and try to win a typing race using an on screen keyboard. Right now its a main indicator for latency and tok/s but eventually I think it’ll become more complicated.

I’m glad I got back to building. Got super burnt out a couple months ago working on Sophie. It didn’t really lead to much but I think it helped Jay with what he’s currently doing. I should talk to him this week about applications.

I remember that this whole summer I did so much less than I wanted to. But at the same time I was putting in 70hr weeks with 16 credits plus Sophie. I should be less hard on myself.