Wrapped up Stanford CS336 (Language Models from Scratch), taught with an amazing team @tatsu_hashimoto@marcelroed@neilbband@rckpudi. Researchers are becoming detached from the technical details of how LMs work. In CS336, we try to fix that by having students build everything:
Late night vibe coding Stripe :D
Ok.. not entirely vibe coding, I would say "steering AI"...
I find claude code to be more interpretable, as I have to approve each call and read through the files very attentively. On the contrary, Cusror hides stuff when in Agentic mode. So I understand my codebase less. But with claude code i understnad more. Interesting :)
Worked extensively on the infra stuff for the agentic webapp generator. Here is the current setup and the approach to separate main project VPS from the machine(s) which will contain users' apps. Turns out nice :) Keep building π
Been struggling with integrating langgraph to the backend + streaming on the frontend. My setup is next.js + fastapi. I need to manage task running somehow efficiently. So trigger a task for langgraph from frontend , with ainvoke on the backend. And Then also stream updates to frontend using astream_events on backend. And i need to control these tasks, e.g know when to stream, how to stop etc. currently i am using in-memory TaskManager on backend, with some state tracking in postgres. Celery might be a good alternative, buuut i donβt want to add another service for a broker (like Redis)β¦ anyone did a backend for langgraph themselves? What do you think?
More progress on BuildBare app UI + Backend
β Handle individual dockers for each app
β Message history retrieval
β Switched UI framework to NiceGUI
β Auth and App List
ππππ
@thepatwalls Yes. Upside that it writes way less code and donβt clutter your codebase. But in terms of problem solving - questionable. I switched back to sonnet a couple of hours in trying to fix the algo logic, and it was successful from 2nd try .. soβ¦ π₯΄ letβs see
Made some progress recently for my mini app generator side project. Did the UI, backend. Refining. Picked a name, I don't know but should be fine... Also filed for sole proprietorship in Germany. Bureaucracy is not that bad π#AgenticAI#SaaS
I discovered βComputer Useβ by Claude today. And this is wild. LLM basically controls computer screen and can surf the web, make notes, execute commands in bash etcβ¦
It will be an essential part of testing ai-generated web apps and working with websites as part of development (eg understand how to parse them).
Here is a link to the demo:
(You will need anthropic API key)
https://t.co/XJQCpfHGGS
Small update to the this little app called xliff parser. I did some context engineering, but most importantly - switched gpt4o to Claude Sonnet 4. And it is SO much better. It created an app that works instantly. The cost is higher for the initial creation, but the app works!!
Implemented memory/context management for my miniapp generator agent... Works good! Created a custom miniapp to convert localization files from xliff to excel. Overall cost of this solution 0.26 USD π and it can be greatly optimized. Niiiice!! #localization#xliff#interpreters #AI #AgenticApps #AgenticAI
@jackfriks@postbridge_ There is a book called "The Mom Test" by Rob Fitzpatrick; it discusses how to talk to customers to get honest feedback. It might be helpful. You will solve this! πͺ
Did an overhaul of my MacBook Pro late 2016. this one is still on intel. I replaced battery, usb-c ports (connectors were loose) and thermal paste + deep clean everything. Gonna serve me well for another couple years I guess.ππ₯³ Repair in public π