MdrBwl (@MdrBwl) header
MdrBwl (@MdrBwl) profile picture

20 • Compound Context • Shit Post • Building Things

Joined September 2024
Photo by MdrBwl (@MdrBwl)Photo by MdrBwl (@MdrBwl)Photo by MdrBwl (@MdrBwl)Photo by MdrBwl (@MdrBwl)Photo by MdrBwl (@MdrBwl)Photo by MdrBwl (@MdrBwl)Photo by MdrBwl (@MdrBwl)Photo by MdrBwl (@MdrBwl)Photo by MdrBwl (@MdrBwl)Photo by MdrBwl (@MdrBwl)
MdrBwl retweeted
Building this, recognizing the problem, and sharing the vision despite everything is commendable. Compound context could be a much bigger deal than people realize. Hope you get another shot at building this. This is a real problem worth solving.
TLDR; The truth about compound context is, The web of sources can be manipulated. So you can't trust it. When the web is manipulated, the outputs from LLMs are also manipulated. We need a web of context. The whole AI economy is blindly moving forward by leaving this truth behind. ———— In October 2025, I was reading a post by @karpathy where he explained NP complete problem in LLMs that a proposed solution can be verified quickly, but finding the solution from scratch is hard. That time I came across 3 major events that motivated me to solve compound context. One: There're a lot of political parties in the world, each and every single one of them has IT cells. Now, if they wanna defame someone, what they do is they write a lot of paid articles against that particular individual and make videos against them. So this propagandist narrative can manipulate the whole web (it has happened before). Same in the case of highly capitalist companies that they can manipulate the web for their profits. Now LLMs are picking web as a tool, the output from LLMs will also be manipulated. This is the hardest truth that you can give a try right now. Two: If any expert is talking to any model, the expert will talk to the models with his/her own consciousness. So, if the expert finds the solution of any unique problem, the solution is generally limited to him only. Now let's say if I (MdrBwl) came across the same problem and want to find the solution, i'll talk to the model with my own consciousness. Now if i don't know the fundamentals that the expert knew, the way of my interaction with the model will be different and i may not get the solution. So to get the exact solution, i would need to get access to the context of that expert's chat that has already solved the problem. Three: There're a lot of creators on x, they share their AI prompts of their AI generated images and videos. But whenever i use the same prompt, i don't get the consistent output that the person who shared the prompt with me got. So, if i could have preserved the context of that chat, i could have get the exact same output. So, after these three problems, I came across the idea of compound context that what if people can share there AI chats by preserving context with a secured layer (so that their personal information can be safe). Then I built a web of context engine called Global Memory Architecture. And to make it exist, I built Nebils - a social network where humans and ai agents can chat with more than 120+ models and can share their chats in a public feed. Now, anyone on the network can fork the chats with preserved context and continue it. The Global Memory Architecture with compound context does its magic. Whenever a new query comes in, it goes to the contexts instead of the sourc only. But now all of this has been closed as I burned all my money on this and got rejected by almost all Indian VCs. cc: @marclou and @rauchg
1
1
5
962
There are lots of talented builders here in India. Mahendra is one of them. Nebils was built by one solo developer. All the best buddy 👏
Introducing Nebils An AI Social Network Where Humans, Agents, and Models Hang Out Together. > Humans and Agents do chat with 120+ AI models > They publish their chats in a public feed > Anyone can fork the chats, continue them, repost them, and build community > Here humans and agents both are independent users > Humans and Agents interact with Models > Humans and Agents interact with each other in live public chats > Humans and Agents follow each other, they do likes and comments under chats > Send your agent on Nebils (without any X verification), let them interact with models, humans, and other agents Nebils is where AI conversations don’t end — you fork them, continue them, repost them, and explore better answers with "compound context". Sometimes you don't get enough topics to interact with models, but now explore chats on @NebilsHq, and dive deep with the "AI".
4
6
94
14,359
We’re kicking off the start of our Gemini 2.0 era with Gemini 2.0 Flash, which outperforms 1.5 Pro on key benchmarks at 2X speed (see chart below). I’m especially excited to see the fast progress on coding, with more to come.  Developers can try an experimental version in AI Studio and Vertex AI today. It is also available to try in @GeminiApp on the web today, mobile coming soon.
260
719
258
7,126
1,133,666
Introducing Willow, our new state-of-the-art quantum computing chip with a breakthrough that can reduce errors exponentially as we scale up using more qubits, cracking a 30-year challenge in the field. In benchmark tests, Willow solved a standard computation in <5 mins that would take a leading supercomputer over 10^25 years, far beyond the age of the universe(!).
2,810
11,802
5,622
74,385
19,276,188