Sai Yashwanth

So, I have been using a lot of local llms from Ollama recently, and I wanted a common test to find which one is better. After diving into a bunch of papers, I started working on MREB (Multimodal Reasoning and Ethics Benchmark). Checkout the blog here: https://t.co/eyPXAeh0OP Github repo: https://t.co/Ht8hUm5HoD

I used to write consistently and it was such a good period of my life. Need to bring that dawg in me. https://t.co/DNBsA0CYdC

It was indeed fun. Blown away by the scale. I am happy something like this is happening in India.

Found an old raspberry pi. Lets try some fun experimental projects on this. AI + Hardware. Any ideas? https://t.co/YBsyRXvxwi

Today is going to be fun https://t.co/2khG4nO7YP

Saving up my salary so I can buy a good 3d printer

Literally what I am doing from quite some time and I can vouch this is the best thing one can do if they want to "break" into tech. Everyone has a completely different path. There's no competition or any version of it. Enjoy life, be curious, read books, learn stuff, build, break, write about it, talk about it. This is the "system" which you should build, instead of having goals which die out after u reach at some point. For starters you can identify and follow people who are doing the same thing. Learn from them, mix and make your own recipe. You can check out mine: https://t.co/eKoYZaDpUd

gpt4o vs grok https://t.co/8TQrJhgPKB

Really likes the docs of @AgnoAgi Very well documented. Trying to implement a toy project using agno framework. Lets see how it goes

Seems very interesting. What if we also have a - Personal journal tool to document its findings. - Common place book - Second brain? So many experimental ideas

RT @yashwanthsai29: MCP is a standard way of giving arms, legs, and context to an LLM. With all of these, LLMs can do much more than they c…

RT @yashwanthsai29: The reason I’ve been in a creative drought and not posting is because I’ve been pouring every last scrap of my mental e…

From the past 2 weeks, I have been using this model QwQ. Havent touched claude at all. This was a restriction on myself to not rely on the coding abilities of claude. QwQ is fantastic as a daily driver. But it has a lot of shortcomings compared to claude. I think the restriction made me realise how much I was relying on LLMs for code and its a shame. Well, I am shameless. I will be using claude heavily in my work, but anything apart from work like passion projects, its QwQ. 2 Reasons for QwQ: - I dont want to do boring and redundant tasks myself. - QwQ is very good at overthinking. And i really like that. I read the thought and it gives me a good perspective.

Openai agent sdk + Anthropic's MCP setup = AI Agents AI Agents was/is the best space to bet on. I am going all out lol https://t.co/ekx1P2PGjr

This was the reason why my agents were failing to execute. Assistants api was literally unusable last month

Great Stuff

The reason I’ve been in a creative drought and not posting is because I’ve been pouring every last scrap of my mental energy into writing this book. Thanks @wolefizzy for giving me this opportunity. Writing a book as an undergrad that too with a big publication is not a small feat. One of the best parts of my day is sitting down and diving into it—those moments when the ideas just flow. It’s coming along nicely, and honestly, I’m genuinely excited. The book’s going to unpack a ton of experience which I acquired while working at vuhosi over the past 7 months. Note: @crewAIInc is mentioned in chapter 2 but theres going to be many more latest technologies which we will cover.

I wake up and another powerful model drops If claude hits limit, go to grok, if grok hits limit, go to r1 or openai go to QwQ ...... No upper limit to using intelligence https://t.co/jwPQmaKTcs

MCP is blowing up my timeline! Seems like many still have questions about it. So, I’m starting a mini-series here on X to break it down and explore it together. Tutorials, tips, experiences, and much more Join me—retweet if you’re in! #ModelContextProtocol #AIDevelopment https://t.co/R5ApARyzKk

MCP is a standard way of giving arms, legs, and context to an LLM. With all of these, LLMs can do much more than they could without them. We could provide all of these (limbs and context) to an LLM in the traditional way—by giving access to functions as tools, and for context, we could use a RAG tool. However, what makes MCP useful is the ease of integrating these limbs, especially for someone who is a non-coder. You don’t have to code. Just add MCP servers provided by other cool developers into your client and enjoy. Another reason why everyone likes MCP is that it serves as a standard for creating these arms and legs for LLMs. Previously, we had to write a tool in Python and provide that to an LLM (via code, of course). Now, we have a standard process for this. Everyone in this space seems to loce this open protocol. Think about it: MCP makes it easy to build agents now. Agents = LLMs + memory + limbs. This is what I have been discussing in all my previous posts on AI agents (do check them out). Maybe I should start a similar series on MCP, like I did with AI agents. (Correct me if I am wrong about the diagram)