@reach_vb Thank you so much. It opened up so many things in my mind. Maybe this post will matter in the future. The last line hit my nerves hardest: "Relentlessly prune bullshit, don't wait to do things that matter, and savor the time you have. That's what you do when life is short.".

The most challenging part in RAG pipelines is - replicating what people present online as small POC to Production. Trust me why? Because I have implemented a production of RAG pipelines in a multi-tenant environment. Isolation, multiple models, smaller infrastructure, list goes.

Someone on LinkedIn posted a perfect RAG pipeline. Please don't call for those. There are no perfect RAG pipelines. Everything is just a recommendation on productions unless you sit with one and implement yourself.

Building Generative AI product for Indian Client with se eral languages? Try fine tuning this for your usecase. You will not be disappointed.

RT @Alibaba_Qwen: We’re officially releasing the quantized models of Qwen3 today!. Now you can deploy Qwen3 via Ollama, LM Studio, SGLang,….

Qwen 3 is cooking. Soon to be released this week.

mistralai/Mistral-Small-3.1-24B-Instruct-2503 · Hugging Face. mistralai/Mistral-Small-3.1-24B-Base-2503 · Hugging Face.

Released under Apache 2.0, Mistral Small 3.1 is now available on Hugging Face and via API on La Plateforme.

🖼️ Excels in multimodal tasks (MMMU, ChartQA, DocVQA). 🌍 Strong multilingual support across European, East Asian, and Middle Eastern languages. 🚄 Lightweight—runs smoothly on a single RTX 4090 or 32GB Mac.

Boom! @MistralAI Meet Mistral Small 3.1, the top multimodal, multilingual, open-source model in its class!. ✅ Key Highlights:.📈 Beats Gemma 3, GPT-4o Mini, and Cohere Aya-Vision. ⚡️ 150 tokens/sec inference speed. 📚 128k context window.

Same thing is also applicable to most of the optimisation in inference time. While they are good tweaks, it takes 10x efforts to get that working in production.

Observing a lot of twitter space is flooded with a lot of AI tools, models, and research releases, while hardly 1% is actually the approach to take everything to production. For example - it's very easy to find ollama and vllm servings, but they are not fully suitable.

It would have taken easily 3 days for me to figure out some things about Nvidia & CUDA kernels, which #GROK3AI did for me in 3 minutes.

Humorously fine. Good luck though!🤣😅.

I have canceled my #Claude & #OpenAI Subscriptions, saved $40 from them, and got #Grok3 Super for $30. Saving $10 while being super at coding, deep research, and impressive thinking. Smartest AI so far until the next one.

If you are paying low-code no-code CHATBOT websites for your organization, stop it right away. I will show you a full-flow method to build a CHATBOT from scratch. You will thank me later for how easy it is to build one nowadays with so many open-source Small-LLMs.

I have been a developer for 5+ years now. Starting a Practical Python Publication series. Just sharing knowledge and learning from this amazing community. #pythonprogramming #programming #python #developer.

Why does Manchester City's John Stones look like @CillianMurphyOG on stage? 🤯.@ManCity are you sure it's not him? 🤣

5. Most importantly, It provides more accurate and to-the-point responses. Earlier without explicit instructions, I never was on point. Now it has a better understanding of how the response should be. A lot of human feedback has gone into this.

3. Vision Capabilities have been enhanced. Testing on wild images has more accurate results now. Wondering what's changed here. Will read more on this. 4. Speech Synthesis is better, yes, BUT, Open-Source has more customizations. It shouldn't be hard to build something similar.

Within just a couple of hours of testing GPT-4o, here are my observations - .1. It's actually fast. I don't have to switch over tabs and come back now to see if it's completed. Big Win. 2. It is just a minor upgrade, slightly smarter, as @reach_vb rightfully pointed out earlier.