True reason why we need something other than current architecture for LLMs is they lack curiosity! With the vast amount of knowledge it has, when you ask it to build a simple app that's on GitHub, a human would as why do you want to build it? Just get it from Git.

Why do you think @CerebrasSystems withdrew their IPO?

I need feedback about both the website and the product. Any data analyst wanting to help me?🙏 I will share the link for the app. Quickly whipped up the landing page so it is professional. https://t.co/kNA6fB7ZPq

True reason why we need something other than current architecture for LLMs is they lack curiosity! With the vast amount of knowledge it has, when you ask it to build a simple app that's on GitHub, a human would as why do you want to build it? Just get it from Git.

I am convinced that small models and well engineered tools are the way to go for enterprise automation. Qwen3 4b instruct was able to document a @MATLAB simulink model with over 2 million blocks! Any ideas on how to generate task specific LLMs? Teacher SFT vs custom RL envs?

Last month I started VibeJams with my team. All automotive engineers, never built anything other than simulation models using Matlab. Tomorrow I am going to be introducing how to build a simple chatbot with @ollama and streamlit all using @code GitHub copilot.

For the first time ever, I want to buy an @Apple iPhone air. Never saw the value when compared to androids. This feels like a very well engineered phone and I can see value! The hardware part at least. I hope I don't regret the lack of subtle features that android offers.

#code-supernova does not seem to follow instructions properly. And it never reads the files even when you add them to context. It assumes and edits wrong files. Anyone facing these issues? It's freaking me out!

I don't see anything other than hate on X. @elonmusk I am all for free speech but if 8/10 posts are people hating each other, maybe you should take a deeper look at your algorithm and if this platform really is helping humanity!

Had to rethink the entire app because I missed an important test case.! Even though it is much easier now to rewrite with the help of AI, standard principles such as V-models still help you a lot to reduce later stage bottlenecks. Is there any AI coding app that follows these?

It is incredibly difficult to find people who love working on AI projects outside of twitter/X! Anyone in Michigan looking to collaborate?

Are coder models inherently better at agentic RAG compared to a regular instruct model? What models are you using for RAG applications in production?

The main model on @GoogleAIStudio has become significantly faster and noticably dumber than before. @OfficialLoganK are you testing gemini 3 flash as a replacement to gemini 2.5 pro?

If a million AI powered apps are being created a DAY by 2025 only using AI Studio, where is the edge for people creating apps? There must be other ways to capture this storm. Any ideas?

A goal for @GoogleAIStudio: Enable builders to create 1 million AI powered apps a day, by the end of 2025 🤯

One of the easiest ways to manage manage context length is to hide the output from a frequent tool call in history. Example: You have a to-do which gets updated every time a change is made. You don't need the entire history, Just the last status! @AnthropicAI @alexalbert__

Is there scope for a product that can produce slides like this by just uploading your raw dataset? What are the must have features? Long way to go, but feels good!

Dark mode or Light mode? Or both? Any suggestions to improve the aesthetics?

Double shocker🤯 @Alibaba_Qwen Qwen3-4b-instruct-2507 is the most impressive model for engineering automation🔥🚀 It completes tasks over an hour autonomously! Unexpected twist-> Q4S is much better than the darling Q4M at tool calling reliability!🤯 Give it a try thank me later!