Kurtis Welch (@hclewk_) profile picture

Founder of Recursive AI

Austin, TX
Joined February 2009
Photo by Kurtis Welch (@hclewk_)Photo by Kurtis Welch (@hclewk_)Photo by Kurtis Welch (@hclewk_)Photo by Kurtis Welch (@hclewk_)Photo by Kurtis Welch (@hclewk_)Photo by Kurtis Welch (@hclewk_)Photo by Kurtis Welch (@hclewk_)
My software factory spent the last month making 30 open source libraries. The article has a breakdown on how complex and innovative each one is. For example, OneLite is a SQLite-compatible database engine written in pure C#. 5/5 complexity, but 1/5 innovation - it's a lot of complex code, but it's literally nothing new. OneModel (Run and train ONNX models in pure C#, on CPU, GPU or in the browser) is 5/5 complexity, and 3/5 innovation. The concept isn't new, but there are some solid internal innovations. OneSpeech (Transcribe, time and diarize speech with Whisper and other open models, in .NET) is 3/5 complexity and 1/5 innovation - it just brings some developer friendly functionality to .NET. In general, there's more complexity than innovation in the set, with OneGame (a headless, agentic game engine) the only 5/5 for innovation in the lot, and OnePhysics (Simulate 2D and 3D rigid bodies, cloth, soft bodies, fluids and vehicles, bit-identical everywhere) and OneScope (Scrub, fork and compare runs of a deterministic simulation, recorded or live) the only 4/5's for innovation. recursiveai.org/blog/what-my…
1
28
I shipped more in the last week than I did in the first 20 years of my career combined. In fact, I'd wager that I shipped more last week than the full output of every engineering team I've every worked with, combined, over that same 20 year period. Crazy time to be alive. #dotnet #oss #agentic
2
93
Survived 102s and dropped 53 kills (Berserker) at 82% accuracy in the 3D Minecraft Aim Trainer. Real Minecraft sensitivity, full-360 horde — beat it. mcrpg.com/minecraft-aim-trai…
2
231
I love that I can be snarky with Claude and it gets me.
110
Your inefficient participation will slow the process.
58
Made a babysitter app for Alera, so one agent can babysit another and send me a message on telegram when it's done.
89
Agentic Operating System: Creating Minesweeper & FreeCell in Alera (Recursive AI) youtu.be/vKg34I_IkjY?si=cb2T… via @YouTube
105
I had an idea earlier this week for a self-improving AI architecture. I've been spending evenings this week designing it, and I finally had some time this evening to spin up 25 AI agents to build it. This is actually an iteration on a project I built last summer. The original project is called Flows. It took me about 6 weeks to build it - and I still use it all the time. This new iteration, called Routines, is significantly more complex. I spent probably 4 or 5 hours this week fleshing out the architecture for the project. I spent about 2 hours yesterday evening architecting the agent swarm to build it. Now I'm about 6 hours into managing the bots that are building the project. Essentially I'm checking in on them every 5 or 10 minutes, manually moving them through check points so I can keep tabs on the project. Issues I've run into: - I had to completely wipe the last module because the instructions the coding bot was given were pretty terrible. I had the master bot rewrite the instructions, and it performed vastly better the 2nd time - The AI *loves* to write things like `// TODO: implement this`. I have to prod it to make sure it doesn't get lazy - When writing the last module I realized a flaw in my architecture. This wasn't an AI error, but I had to go back and refactor one of the first modules to make everything work. Aside from that, it's been pretty smooth sailing. The agents are still chugging along as I type this, doing their coding, testing, documenting, and reviewing loop. The end is in sight, though! It's currently working through the final integration tests. Tomorrow I'll get to start building some routines and see if the architecture is actually effective. Tonight I'll sleep well knowing I didn't have to spend 2 months building this. If you're a software engineer and you want to learn how to use those same techniques I'm using to spin up a bunch of agents to write code for you, shoot me a DM. I'm hosting a series of classes this month to teach engineers how to build and use their own AI tools.
2
4
24,091
Example of why "ai coding is slower" is stupid. #vibecoding #metaengineering I have code writing a long list of products to a YAML file, which I will feed to an LLM. But, to help the LLM see those products as separate entities, I want a blank line between products in the YAML.
1
273
Me: 30 seconds to type a prompt Agent: 5 minutes to write, build, debug, test Me: 10 minutes to review, test, commit the code Start to finish, we're 15 minutes in, for a feature with a tiny impact on the end result.
1
191
The end result is that it took me 15 minutes longer than if I hadn't added the blank line. And it's something that I wouldn't have done if I wasn't using AI. Does that mean that AI slowed me down in this specific instance? Or did it give me leverage to be more thorough?
174
Made this: unclassified.news It's a strategic conflict analysis site that aggregates news and information on the Israel / Iran conflict and does scenario analysis and strategic assessment of the situation.
1
223
LLMs are capable of so much more than writing CRUD code. Just had a RecursiveAI Agent write enable headless agents on the RecursiveAI Agent platform. Took all of 5 minutes. Zero bugs.
1
210
The biggest mistake people make with AI is thinking it's plug and play and the results are what they are.
171
Kurtis Welch retweeted
Passwordless login (via email) *and* a place to store user data – with barely any lines of js. Looks… perfect? uthentic.net/documentation/q…
I liked a @YouTube video youtu.be/J5_bnZX5CHA?a "I'm Scared," Songs About Emotions by StoryBots
2
1
3