Ikrrrrrrr context management is the only thing that can make an llm usable on a large project and its a only thing it can’t manage
Some studies were made to diffusion denoising to optimise the attention models but no big llm has actually delivered on it

To learn in public i first have to embarrass myself in public 😵‍💫😵‍💫🔥🙏

Oh no no… thank you 😭😭

Will share the repo once I’m sure claude didn’t accidentally made my envs public 😵‍💫

Looking for suggestions on documentation and tools for how to strategise my cold start training ai without pre-existing UGC and user interaction data

I also wanted to build an ai that trains and builds on realtime user data which I have no experience working with so that would be a crazy learning experience

Basically i wanted to learn a bunch of tech like
1. Streaming using dash-mpeg or hsl protocol
2. Serving videos through a kafka service
3. Stats for nerds like youtube
4. Content Recommendation Engine
5. A lightweight server with high concurrency

Tech Stack

1. Client - react router + vite + shadcn + htmx
2. Server - Go + Kafka + RabbitMQ
3. DB - Postgres + postgres_vector + redis + typesense

Ohh yeah it happens i next js sometimes but mainly in go it gives me solutions for older code

And not just claude even chatgpt does it
Even today it started using log.printf instead of slog which is a newer feature for server logs in go

It’s just the second week of the year

For instance, DASH-MPEG or HSL protocols are heavily used throughout big tech but usually developers just outsource their streaming so building a production grade streaming server is a challenge for claude code (at least in my experience)

It can build a next js application easily due to the heavy documentation and community but sucks at building native solutions due to lack of in depth documentation or community discussions

AI (for now) is terrible at coding, it sucks at building logic but excels at giving you a base or boilerplate over which you can make your optimisations
I think devs who understand this and are able to make learned decisions about when to use ai and when to code are the only one’s who’ll survive