Sure, I dressed it up in Bladerunner vibes. But I think it is an interesting concept. Probing a thinking model using mech interp, during thinking and during output. What activates during thinking that doesn't during output?
I am running locally on my MacStudio. The only thinking model I could find with labeled SAE concepts, and runnable on my MacStudio was this Llama8B DeepSeek distillation. Need the new Qwen Scope to get the concepts labelled...
I built a simple tower defense game as a weekend project. I made it playable by agents. I introduced my custom claw-like agent to it, and told it to get first on the leaderboard.
The process it took was interesting, so I had it write a paper about its process.
I had no idea so many people still clung to these lines!
Personally, I see huge value. I’ve built more in the past year than I ever have before. Ideas that used to go on a list, now get implemented in a weekend. The ones that are good, I’ve shipped. Others get played with, tested, and set aside. Not due to lack of quality, but the idea wasn’t right. Some need more thought so I’ll keep testing, and tweaking over time.
Evidence? Those projects would not exist if not for these tools. Not because I couldn’t write them, but because I WOULDN’T have written them.
But I’ve given up trying to convince people. If you don’t think it works, great. Don’t use it. That’s fine with me!
I’d urge you to keep trying though. Hopefully you have the moment as well. And if you think, “it’s ok for some things, but not what I do”, well…just don’t wait a year before trying again. Keep an open mind, and integrate the tools into your workflow everyday. Not as a one off “ok fine, I’ll test this stuff” grumpy kind of way, but in an honest iterative way where you are using the tools on a daily basis, even in a small way. For docs. Whatever. Just keep using them. Eventually you’ll see. Eventually you’ll have the moment.
One thing I think is lost in a lot of these conversations is how much FUN I am having. I’ve been coding for…well if you count copying BASIC from books into a Timex Sinclair, about 40+ years. Maybe professionally for 30+? I haven’t had this much FUN building since back when I copied those games from that little red book of BASIC programs. These tools make even mundane work fun for me, cause you need to have your workflow right. So coding becomes more like Factorio or something. Keep optimizing your setup, then simplify, the optimize, then simplify.
Syntax was always something that felt like a hurdle. I always rolled my eyes at people that get lost in the minutia of a language. Like I remember the Perl neck beards from way back. lol.
I learned what I needed to get the task done, then moved on. Likely forgetting 90% of the details anyway, so I could keep an eye on what I considered important. Architecture. Functionality. Separation of Concerns. Could I dive back in and figure out the details again if needed, sure. No problem.
NOW these tools let me work at this abstraction level even more.
And I’m here for it.
Just amazing this is where we are at as an industry.
Well this isn’t code, but I’ve been working on a memory system for Claude Code. This portion provides semantic search over the session files in .claude/projects. It uses OpenAI for embeddings so not completely local (would be easy to modify) and storage in ChromaDB.
"The genAI has built it" -- this is the core point. If I did nothing except complain about AI for the past 10 months, would these projects exist? No they would not. So. I. Built. It.
If you actually use these tools, really use them. You realize that it's an augmentation not a replacement. Simply because the training data is what has already come before (for now!). The LLMs need help, direction, focus...and those are learned skills dependent on the tooling. Not to mention ideas.
And sure, I imagine the software development workforce will change quite a bit, probably this year, no doubt about that.
But the need for builders will not change. I imagine that the 'builder' role will change to be traditional software developers, designers, sales people, writers, c-suite...whatever.
So I think you are right. "That could easily exclude you or me". 100% correct. The required skill set to be a builder is changing on a weekly basis. The only way to keep up is to keep building with these tools. Trying things. Experimenting. Otherwise, yes, you will probably be replaced. By a builder.
All I know is that I have built more in the past 10 months than I ever have. How do you quantify for the skeptics the mental shift that happens when you know you can just build stuff now?
COULD I do this stuff before? Sure. But I wouldn’t have. Life gets in the way. Now, the bar is low so why not build stuff? Some of it ships, some of it is just experimentation. It’s all building.
Trying to quantify that shift is impossible. It’s not a multiplier to productivity you measure by commits. It’s a builder mind shift.
I redesigned my app Back2Back to lean into the music discovery angle.
Create a DJ persona using GPT-5 in high thinking with web search tooling, or just use the default personas. GPT-5 uses the persona's style guide to make selections, another GPT-5 matches the selection to Apple Music, and a final GPT-5 validates the selection.
If you are the type of music fan that has been unimpressed by any AI DJ features, I urge you to give my app a try and give me feedback! I am passionate about music and have put a lot of work into getting the best results possible, but am always looking to improve the prompting.
I find these takes so odd. Porn is the early adopter of EVERY technology. When non-linear editing software was invented in the 90s by Avid, who was the first industry to use them? It might not be advertised, but it's true. Porn leads the way. lol. This is no different.
I decided to come back and work on this open source project more. I updated it to 1.1, with a big focus on performance.
The feed is random (based on the collections and identifiers enabled). By default I just enable a collection by "AV Geeks". It's usually NASA/Space stuff, or old educational videos. I use them cause their videos are optimized the best! Other collections have varying degrees of loading issues. Sometimes I can compensate in the player, but sometimes it's just slow to load. But AVGeeks is rock solid unless the server is down! I had the complete collection of Seinfeld in the main feed when I launched v1.0. (There is a ton of copyrighted content on the Internet Archive). I wanted to see how long it would take to see it taken down...3 days. lol. Anyway, to customize the feed, you can create Presets where you can flip on collections. You can add any identifier you find to a preset as well.
Search was custom implemented. Internet Archive does have a basic keyword search, but I wanted something more. So I used the Internet Archive cli tool to download metadata for all the collections I liked. I used OpenAI to generate embeddings for the metadata (just the metadata, didn't get super fancy here, but would be fun to maybe extract more metadata from the frames?). Anyway, I stored the embeddings in Pinecone. Boom. Natural language search with easy 'show me videos like this one'. Also added a year filter to only focus on older videos (otherwise you get tons of lame G4 vids). I want to include more metadata on Pinecone, I stupidly only included year at first. Now I'd have to reindex and just haven't bothered yet.
Another interesting thing I learned was that identifiers on The Internet Archive can have multiple individual video files. For example, before it was removed, Seinfeld has each episode in a different video file, all organized under one single identifier. SO, if I had a preset with only that one identifier, as I swiped, each time I would make it show a random identifier, cued up to a random point in the video. This made it so I could swipe thru all the episodes of Seinfeld...get bored, swipe to the next...watch a bit, get bored, swipe. SORT of like a clip show? But each video was the FULL episode, so if it was a part I knew and liked I could watch for a while. Could always swipe when I got bored tho. This sparked in me, cause I think it's a good interface. Short form video interface for long form legacy video content. I dunno. I like that. Really good for old sitcoms, skit comedy shows, etc. Similar to how shows like Parks & Rec just post up hour long clip comps on Youtube. This is like that, but you are more in control. Like I said, there is a lot of copyrighted content on the Internet Archive, but it regularly gets taken down, as expected. And therein lies the problem. I like the interface enough tho, so I made a version of this app for the Plex Media Server. So I loaded up Parks and Rec, and did exactly what I described. But, eh...I think the Venn diagram of people that like this type of TikTok interface and people who use Plex Media Server are... me? lol. I briefly looked into how to do this for real, but UGH. The FAST licensing model is the most likely candidate, and my friend at NBC says that would be the easiest to get, but after investigating it's purely about linear programming. Specific ad placement at specific times. Not setup for a vertical swipe feed. Would have to negotiate a totally new thing. FAVF=Free Ad-supported Vertical Feed? Not my thing. Anyone a tech media legal person interested in negotiating a new licensing format?
So thats how ended up back on this open source project, hanging with Claude again. Enjoy the v1.1 update.
I am running locally on my MacStudio. The only thinking model I could find with labeled SAE concepts, and runnable on my MacStudio was this Llama8B DeepSeek distillation. Need the new Qwen Scope to get the concepts labelled...