Sabitlenmiş Tweet
Kunal Kumar Sahoo
2.1K posts
Kunal Kumar Sahoo
@KunalKSahoo
PhD Student | Exploring Intelligence | Realizing Physical AI
New Delhi, India Katılım Eylül 2021
292 Takip Edilen388 Takipçiler
For high dim control tasks like this SAC, TD3, and DreamerV3 generally outperform PPO. SAC's max entropy objective drives stronger exploration and off policy replay for discovering stable gaits faster.
TD3 adds twin critics + delayed policy updates for improved stability and reduced variance in high dim continuous actions.
DreamerV3 uses world model planning to scale more efficiently on complex proprioceptive dynamics even with abundant simulator rollouts.
But none of them solves the reward maxxing problem though 🤡
English
So AI agents are not a technology but a mindset to build goal-driven intelligent systems.
So even if you develop a robot that uses the A* algorithm to solve a problem, it can be called an AI agent if it follows the perception->reason->act loop.
In case of LLM-agentic systems:
- Prompts/Context become the perception,
- Reasoning can be thought of by multiple agents communicating and validating, and
- Actions become the tokens generated and tool calling capabilities.
This just the core, concept of Agentic AI. There is more to it. You can know more about it here:
kunalkumarsahooai.hashnode.dev/ai-101
English
@trulyjeet @grok That's a particular use case but the notion of Agentic AI goes beyond.
English
@KunalKSahoo Yeppp i was researching this on @grok and turns out its just , ai being guided to convert use input into strucured output objects and usig app integrations and APIs to perform tasks , turns out now i can build my own AI Agent without watching some youtube guy .
English
Anyone that is gonna replace you with an AI system is going to be you only.
Upskilling is important, substituting is NOT!
In practice, the boundary is blurrier than expected, your average influencer doesn't deblur it for you.
GIF
English
@Rushu_Tushu The variance gets amortized as time passes*
*provided honest efforts are put in.
English
Two of my friends, one from IIT Kharagpur and the other from a tier-3 engineering college.
The IIT friend posts about college life on Instagram, always seems happy, has a good social life, doing well academically, and even landed an internship at a top MNC in 3rd year.
The tier-3 friend, on the other hand, is mostly bunking college, sitting in his room all day, watching tutorials and coding. No real social life, always looks burnt out whenever we meet. Currently doing an unpaid internship in 3rd year.
It's crazy how differently both of them are living their lives.
College really does seem to matter in overall exposure and opportunities.
English
Kunal Kumar Sahoo retweetledi
Excited to announce the 𝗦𝗰𝗔𝗜 𝗜𝗻𝗱𝘂𝘀𝘁𝗿𝘆 𝗗𝗮𝘆: 𝗔𝗰𝗮𝗱𝗲𝗺𝗶𝗮-𝗜𝗻𝗱𝘂𝘀𝘁𝗿𝘆 𝗖𝗼𝗹𝗹𝗮𝗯𝗼𝗿𝗮𝘁𝗶𝗼𝗻 𝗦𝘆𝗺𝗽𝗼𝘀𝗶𝘂𝗺, hosted by the 𝚈̲𝚊̲𝚛̲𝚍̲i 𝚂̲𝚌̲𝙰̲𝙸̲̲ at 𝙸̲𝙸̲𝚃̲ ̲𝙳̲𝚎̲𝚕̲𝚑̲𝚒̲! 🚀
🔗 To know schedule: scai.iitd.ac.in/industry-day
English
Before downloading any codebase in your main machine, it is always recommended to go through the code base in great detail.
And going through a large codebase is challenging. So in case any of you don't know about this tool, it is great for interacting with codebases in natural language.
For code-review-graph: codewiki.google/github.com/tir…
English
Started using code-review-graph toolkit for my PhD work, seems like an interesting idea.
Haven't benchmarked how much tokens I am saving using this, but the graph already helps me identify redundant code.
Give it a try: github.com/tirth8205/code…
English
Kunal Kumar Sahoo retweetledi
An extensive interview with the Director of @iitbombay, Prof. Shireesh Kedare.
Covering a wide range of topics from energy, research, JEE coaching, university rankings, and more.
indianexpress.com/article/idea-e…
‘Coaching culture trains students to game exams, not build concepts’: IIT Bombay Director Shireesh Kedare at Idea Exchange
Director of the Indian Institute of Technology (IIT) Bombay, Professor Shireesh Kedare, an environmental science engineer, is a strong voice in India’s energy policy sector.
English
@OpenAI's Prism is a life saver lately. Highly recommend.
Please do NOT substitute it for writing the document, but rather than a writing assistant which can summarize, edit, and format your LaTeX documents.
English
@KunalKSahoo wow, didn't know about that, thats so cool
personally, liked the way data distribution was visualized and then how it changes while training and sampling
I really think i'm ready to read the back to the basics paper now xd
waiting for the blog!
English
Kunal Kumar Sahoo retweetledi
Classical robotics struggled with edge cases.
VLAs turned everything into an edge case.
California, USA 🇺🇸 English
@OfficialLoganK
Does Gemini offer some support for PhD research in India?
English
A bitching tweet but it's okay:
So I came across a profile where the person is focusing on non-core-tech roles business dev, prod mgmt, etc.
But the hypocritic and funny part is in the portfolio it is written:
"Most people expect a computer science student to follow a clear arc — internship at a startup, placement at a tech company, repeat. I chose something different. Not because I couldn't, but because I was too curious about what else was possible.
I spent my years at ----- building things that weren't in any syllabus — communities, research papers, investor decks, policy analyses, art exhibits. I walked into rooms I wasn't supposed to be in yet — XYZ, ABC, DEF — and I made them mine.
I believe the most interesting people are the ones who can hold multiple worlds at once. Finance and art. Technology and policy. Leadership and creativity. I'm still figuring out what I am. But I know how I learn — by showing up, asking questions, and caring deeply about people."
Not just the portfolio but other social media posts are also like this: "I am doing XYZ despite being a CSE student by choice..."
I know it's a long rant but hear me out. It's completely fine if one doesn't want to pursue the core line in which (s)he is trained. That's accepted.
But why does it make sense to milk your formal education/accomplishment which has no relevant impact on what you are doing right now or trying to convey right now?
There are many people who have pivoted in their life. Only if all of them boasted about their pre-pivot life we would be just living in nostalgia and not progress.
I think profiles like these are utter bullshit and one needs to be careful who they are following.
Open to discussion about this perspective.
STOP MILKING STEM DEGREES, IF YOU DONT CARRY IT FORWARD!
English
@kalashjainn @fnhirwa I have used such losses and they work :)
English
@kalashjainn @fnhirwa That's alright.
There are losses, for example negative log likelihood when used on probability density functions (which has to be just positive and not necessarily <1) then the losses can be negative as well.
English
