Prof, EECS, IISc, IITD, Co-founder LatentforceAI | https://xitter.twitterapi.xyz/t.co/6xoL0JR37m Disclaimer - All my opinions are personal.
Bengaluru, India
Joined October 2009
- Tweets1.4K
- Following189
- Followers28.1K
- Likes540
Pinned Tweet
We (@VanditPatel002) just completed a "Hello-world" task for Weight-Space Programming and open-sourced it.
Can an LLM figure out what a neural network was trained to do, simply by inspecting its weights without running a single forward pass?
We tried this on a small MNIST classifier with random labelling. The results were fascinating.
A thread 🧵
prathosh A P retweeted
We (@VanditPatel002) just completed a "Hello-world" task for Weight-Space Programming and open-sourced it.
Can an LLM figure out what a neural network was trained to do, simply by inspecting its weights without running a single forward pass?
We tried this on a small MNIST classifier with random labelling. The results were fascinating.
A thread 🧵
WHAT DID WE LEARN?
The agents uncovered fascinating fingerprints in the weights:
• The original dataset could be identified.
• Pixels that were never trained could be recovered.
• Dead neurons and digit-specific detectors could be identified.
Yet, recovering the exact task remained surprisingly difficult.
A coherent explanation of weights need not be a correct explanation of the network's behaviour!
"There is an enormous treasury in Indian thought. Only now are we mining it and discovering its endless depth. It requires a certain amount of confidence; we don't have to wait for an external stamp."
Swami Sarvapriyananda ji on the relevance of Indian logical/philosophical traditions in the age of AI.
In the same discussion, he also mentions an observation of mine connecting Anvitābhidhānavāda with the Attention Mechanism in modern neural networks, calling it "an interesting insight"
A wonderful example of the conversations that emerge when classical Indian thought engages with contemporary ideas.
आ नो भद्राः क्रतवो यन्तु विश्वतः।
You can watch his insightful talk here: youtube.com/live/MXP04031qYM….
Video courtesy of Sri Ramakrishna Math Chennai
He is a Hindu monk whom Anthropic invited to discuss the repercussions of AI on humanity.
Many students asked me for the PyTorch notebooks that are used in the tutorials of my course on GenAI.
Sharing them here: github.com/Chandan-IISc/IITM…
Thanks to my student, Chandan, for putting it together.
A lot of people ask me how to learn Sanskrit properly.
Here is a source -vidyasvam.in/samskrita-karya…
Dr. Sridhar Subbana's method is one of the most methodological that I have come across.
They have both an online + short term residential program.
I have personally seen people from ages benefitting immensely from it.
A reply to my paper on the Gita was attempted. Here is my rejoinder: xitter.twitterapi.xyz/prathoshap/status/2119…
Hermeneutics should be based on epistemology, not opinions. If there is further interest, I shall answer them only via a live, well-moderated, pramana-based debate.
My reply to @prathoshap's paper on the Gītā’s dehī, written after his conversation with Acharya Prashant.
It proves the dehī isn’t the body, which AP never denied, but not that one individual passes from birth to birth.
Full reply: tiny.cc/9scd101
Discussion welcome.
A reply was attempted to my article on Gita. Welcome, academically.
It first concedes - dehī differs from the body, then relocates the Vipratipatti to: no individual survives; only an ego-tendency re-forms.
Rejoinder गीतोक्ताहङ्कारातिरिक्तजीवसिद्धिः zenodo.org/records/24081746 🧵
5/
Chapter 15 makes it explicit.
The word अहङ्कार does not occur in that chapter at all. The one who departs and arrives is called अंशः, ईश्वरः, पुरुषः.
15.7: ममैवांशो जीवलोके जीवभूतः सनातनः.
सनातन is said of the jīva itself.
6/
15.8: गृहीत्वैतानि संयाति. Having taken these, it goes. An absolutive, so one agent, and what takes is not what is taken.
15.10 puts departing, abiding and experiencing in three accusatives on one object.
15.16 to 15.18 counts three puruṣas: क्षर, अक्षर, and the उत्तम who is अन्यः, other than both.