Aaron Levin

@AbnerHout @SamuelBeek yea but this isn’t new. building in a silo was the saas 2.0 version of this for 15 years

@SamuelBeek tokenmaxxing can give you a false sense of accomplishment and feeling youre building something valuable

multiple friends of mine have quit their jobs and moved back in with their parents to vibe code full time. is this a thing?

@TheRohanVarma unreal x.com/awlevin/status…

Today I’m attempting to hit the 5-hour limit on the Codex $100 Pro plan. The method: build a MapleStory-like game from scratch. 30 minutes in, I already have a working game with sprites, maps, and assets generated with Imagegen. Unfortunately, I’ve only used 5% of the limit so far. At this pace, I may need to start building RuneScape in parallel just to make a dent 😬

@sama Eh

i keep thinking i want the models to be cheaper/faster more than i want them to be smarter but it seems that just being smarter is still the most important thing

Somebody has to see how long it takes AI to recreate RuneScape from scratch

Becky on GPT 5.5 ladies and gentlemen 😂

@ClementDelangue @OpenAI Didn’t know you could share session agent traces live on hugging face. That’s sweet!

I'm trying to build an office receptionist app for my reachy mini today with ml intern + @OpenAI GPT 5.5. Wish me luck! You can follow my session agent traces live here: huggingface.co/datasets/clem/…

@Skeegan123 @sama Bro lol

@sama Ouch

we will plan bigger parties for future releases. a lot more people wanted to come than we expected. thank you! gonna try to think of a really good idea for the next one.

@shrikardayalu Dude you could literally get 1M back by becoming the most prolific software contractor of all time. Or become over employed

Vibe coders, would you take 1. Claude Max in 2010 OR 2. $1,000,000

Everyone is arguing over who will survive: PMs, designers, or engineers. It finally hit me. We are all going to be more productive and, while these roles will evolve, none of them will cannibalize the others. Why? Well, product people have been able to make prototypes for years with low code apps. Then they get enough traction or internal buy-in to delegate to engineers and designers to make it more polished and more scalable. That feels like it will stay true— the floor has just been raised because vibe coding >> software 2.0 low coding (sorry @retool Claude code is eating our GUI builder lunch too) While PMs count spend time prompting to optimize DB performance / server load / optimal UX flows on every single page, I doubt that’s where they’ll want to focus their efforts. Thus, the engineering and design domains will continue to exist. And for those that prefer prompting to build fault tolerant architectures, that enable the PMs to one-shot prototypes closer to production quality and reduce cycle times, but rather not think about “what” to build — the engineering role will continue to exist as well. But what about when we get AGI/ASI? Surely then the PM won’t need the engineer to handle infra/devops? Well, when we reach that point we’ll have much bigger problems anyway.

@ATTlKA That’s insane

If I understand the studio behind it (Serene) correctly, the PCB & electronics are etched and fused into the interior, so that same surface also acts as the plate simultaneously. Switches are mounted and soldered directly into that surface to connect with the PCB. Also allows for full size switches in a low profile body, which feels ungodly.

Your email “finds me” replying on my brand new keyboard milled out of a solid aluminum block

@artcreativecode So creative! And really cute scene design

I’m excited to finally share my entry for the #threejsJourney Challenge! A tiny toy train adventure inspired by Snake, where you build the track on the fly and race to collect new wagons before time runs out. Built with #threejs + Blender. Play here: wagon-rush.artcreativecode.com

@andreeliasdev Nice work man! This post was a wholesome and informative read. Also appreciate the updates you were giving throughout the month. Got my follow. Best wishes to you and your growing family!

I've built a multiplayer survival game for the browser in 30 days and I didnt write a single line of code 🤯 So I'd like to write my notes about it once the deadline is done now. The last 20% of a project is indeed the hardest part of it all. In the last 2 days i've been working on the final version of Hollowlands for @levelsio's 2026 #vibejam I worked a lot! I fixed a lot of bugs, tested a lot playing with my wife. Fixed more bugs... But now, thank God, I have the final result! The game looks good, yes. I've spent half of the jam just tweaking every single detail of the procedural world generation. And It was worth it! The game still runs well on most devices and still has less than 20MB in size. Which is pretty impressive to me. The process of building was very straighforward: 1. I used @threejs + React to create the game. Before I started I created a huge document defining every rule i'd like the codebase to be implemented upon. It had clean code rules, archtecture decisions, ECS principles (good for games), and so on. This helped a lot constrain the AI to maintain the code maintainable and separated into clear domains (systems). This document was made partially by me given my previous experience building web games (I've made a lot of mistakes in the past and made sure they will not repeat again) and partially AI suggesting best practices. 2. For every feature I prompted the AI, I was very specific on details that I'd like implemented. If I knew what I was doing, I was more specific on the "how" I wanted to be done. If I didnt know what was doing I first asked the AI to brainstorm possibilities with me, based on industry standards, pros and cons, and so on. Then I would decide what path to go. 3. Every single feature I used Leva for tweaking the values. So for example, the size of the trees, colors, distance between objects, animation speed, etc. Everything I have a slider that I can adjust. 4. To make the game look good I used @tripoai to create the models. To be honest it's not 100% perfect. If you check the character in game you can notice a few issues here and there (the arms lol), but I believe that it will keep getting better over time. 1 Year ago I tried and it wasnt even close to what it's now. Then I used a few post-processing + shaders + particles techniques to bring the "wow" factor. 5. I did only 1 feature at a time. And every change it was a commit with a clear message of what was done. This helped the AI fetch previous commit to understand what changed and know exactly what happened in the past. If I opened multiple terminals at once and blasted prompts the code and commit history would be a mess in a few days. And for every change the AI did, I asked it to do a Code Review of everything that was changed and look for violations on the document I wrote in the beggining defining the codebase guidelines. I realized that I pretty much applied software engineer principles using AI: Code, review and merge, code, review and merge... 6. On the last week I implemented the multiplayer using @colyseus which was pretty nice. I used AI to fetch the whole documentation and create a skill.md file that helped a lot. The creator of the framework @endel also did a extensive research on how real multiplayer games are implemented. It's open source and it was a gold mine for the AI to come up with the implementation for my game. You can ask him the link :) And that was it I think. I'm extremely tired now. The last 2 days was insane. The game's final version is not what I had in mind in the beginning of the jam. I had to cut a lot of cool features. But given the 30 days deadline I think it was a huge success. Hands down my best game ever made. Now I'll take a few days to rest and come back with more updates on the game. Maybe it can become a big deal in the future with more updates. Who knows. And in the event I win first place in this game jam, I'll use the money to fund my own game studio + cover the expenses of my child that is going to be born in September :) Cheers guys! If you have any question, feel free to reach out. I'll answer every non-bot comment in here haha Note: The link to play is in my profile

@Fluffinu @sama gobby

@sama What's the name of this guy?

goblinblog dropped

@KurtBullock Now make one for marketplace

Meta just rolled out a CLI 👀 Launch ads, pull stats... agent compatible developers.facebook.com/blog/post/2026…

Woah this is the first time I’ve thought about this: how many lines of code exist in the world that are not being executed? We’ve built so much software, so many lovable projects and Claude code weekend projects, but how many are actually being used regularly, or at all? Does it matter?

Isn’t the contradiction between these two exactly the biggest problem with today’s task and coding-oriented AI? You’ve turned AI into a marketing gimmick about anyone becoming a builder, when in reality you’re just burning through their time to produce more cyber-garbage disconnected from real needs. What AI should actually be doing is augmenting human cognition, not replacing human work. Please, think hard about where the brilliance of the GPT-4 series really came from.

feels like a good time to seriously rethink how operating systems and user interfaces are designed (also the internet; there should be a protocol that is equally usable by people and agents)

Yeah, I’m skeptical of using a design resource from someone whose mobile site looks like this.

we run filesystem-based memory for personal AI assistants and see the exact same jump. older models append everything flat, newer ones build their own taxonomies and prune stale info unprompted. sonnet 4.5 staying flat at 43% regardless of budget but opus scaling to 84% without any changes feels like more evidence of that so... i guess memory format matters less than the model's judgment about what to keep, and we should just stop designing memory schemas and start letting the model decide? lol

across model generations, Claude gets better at using the filesystem to manage memory. i wrote abt one example here from Claude Plays Pokémon (c/o @DavidSHershey): Sonnet 3.5 v Opus 4.6 shows a striking difference in its memory files. claude.com/blog/harnessin…

Claude Managed Agents just added memory. memory files are managed by Claude, shared across sessions, and easily exported / edited. platform.claude.com/docs/en/manage…