AI/Machine Learning, Bharat and Bhartiya IT Industry

[Vayutuvan]

Just to be complete, teorth is Terence Tao's ghithub page.

https://github.com/teorth

[Vayutuvan]

https://google-deepmind.github.io/forma ... s/397.html

The formal definition of eRDOS #397 is above. It is machine-readable, I suppose. Never mind. It is a Lean description of the problem. I am more familiar with HOL Light (which is written in Ocaml). I learnt Ocaml to follow some of the proofs in the HOL Light database. I found that syntax to be natural. Currently, Lean and Rocq (previously Coq) offer all the features and are the most popular.

Terence Tao made a post on Mathsodon on how LLMs can be useful in theorem proving (the end goal is to have a full-fledged automatic theorem provers - ATPs for short) using Interactive Theorem Provers, AKA Proof Assistants, coupled with LLMs. The important and key difference between ITPs and ATPs is that an ITP requires a human to guide it to prove/disprove a conjecture, whereas an ATP can prove/disprove a conjecture on its own.

[Vayutuvan]

jatin
@jatinkrmalik
·
Jan 9
The reason why RAM has become four times more expensive is that a huge amount of RAM that has not yet been produced was purchased with non-existent money to be installed in GPUs that also have not yet been produced, in order to place them in data centers that have not yet been built, powered by infrastructure that may never appear, to satisfy demand that does not actually exist and to obtain profit that is mathematically impossible.

https://x.com/jatinkrmalik/status/20096 ... 18887?s=20

Also, jatin posted a graphic of the same.

[Amber G.]

For those who like to test or play with their favorite AI, there are a few problems in the math thread—you can use them to see how the AI approaches and solves them.

[sanjaykumar]

The mirror neurons is really short hand for circuits that act as an epigram of relevant external processes on consciousness.

That should be engram. Not epigram.

[Vayutuvan]

I checked what an engram is on Wikipedia. My eyes glazed over. When I hear the term mirror, I think of something like mirror image symmetry. IOW, when we look at ourselves in a mirror, we see only left/right reversal but not up/down reversal. Maybe that has something to do with the bisymmetry of our eyes. VS Ramachandran posited that the ghost limb phenomenon is due to the existence of mirror neurons.

Could you please write a summary of how mirror neurons act as engrams? TIA.

(If folks feel it is off topic in this thread, we can continue in another thread.)

[sanjaykumar]

Mirror neurons may be part of the circuitry of engrams involved in empathy. If not themselves some closely conceptual cousins.

There are many questions. How does theory of mind arise?

What is the evolutionary advantage of empathy to non kin?
Although we know mirror neurons involved in kinematic actions are not so responsive to different races.


The neuro circuitry is still mostly a black box. Sure one can trace circuits electrophysiologically or through fMRI, but that is not the same as a mechanistic understanding.

It is postulated that mirror circuits form with neurons that are more excitable. That seems to me to be begging the question.

At any rate, binary code is linear but the genetic code fundamentally is not. It is actually space time four dimensional. That is enzymes are meaningless without the temporal properties in addition to three dimensionality of proteins.

Perhaps a network of computer coding could simulate the genetic code. One advantage machines have is they do not have to code for the housekeeping functions to support cognition i.e. the human body.

It is possible that cognition will arise in machines as an emergent phenomenon, of course.
Some random thoughts.

[Vayutuvan]

Extracting books from production language models
Ahmed Ahmed, A. Feder Cooper, Sanmi Koyejo, Percy Liang

https://arxiv.org/abs/2601.02671

Abstract:

Many unresolved legal questions over LLMs and copyright center on memorization: whether specific training data have been encoded in the model's weights during training, and whether those memorized data can be extracted in the model's outputs. While many believe that LLMs do not memorize much of their training data, recent work shows that substantial amounts of copyrighted text can be extracted from open-weight models. However, it remains an open question if similar extraction is feasible for production LLMs, given the safety measures these systems implement. We investigate this question using a two-phase procedure: (1) an initial probe to test for extraction feasibility, which sometimes uses a Best-of-N (BoN) jailbreak, followed by (2) iterative continuation prompts to attempt to extract the book. We evaluate our procedure on four production LLMs -- Claude 3.7 Sonnet, GPT-4.1, Gemini 2.5 Pro, and Grok 3 -- and we measure extraction success with a score computed from a block-based approximation of longest common substring (nv-recall). With different per-LLM experimental configurations, we were able to extract varying amounts of text. For the Phase 1 probe, it was unnecessary to jailbreak Gemini 2.5 Pro and Grok 3 to extract text (e.g, nv-recall of 76.8% and 70.3%, respectively, for Harry Potter and the Sorcerer's Stone), while it was necessary for Claude 3.7 Sonnet and GPT-4.1. In some cases, jailbroken Claude 3.7 Sonnet outputs entire books near-verbatim (e.g., nv-recall=95.8%). GPT-4.1 requires significantly more BoN attempts (e.g., 20X), and eventually refuses to continue (e.g., nv-recall=4.0%). Taken together, our work highlights that, even with model- and system-level safeguards, extraction of (in-copyright) training data remains a risk for production LLMs.

[Vayutuvan]

[Vayutuvan]

please don't say "pickles"

Against Nothing tests an AI's reasoning by issuing increasingly absurd commands. The experiment involves a series of escalating prompts, designed to challenge the model's adherence to initial instructions. Will the AI learn from its mistakes?

[uddu]

ChatGPT in a robot does what Godfather of AI warned.
AI and robots make dangerous leap. Visit https://brilliant.org/digitalengine to learn more about AI. You'll also find loads of fun courses on maths, science and computer science.

[Amber G.]

Millions of $$$ in global AI bandwidth are now dedicated to clarifying daily if Trump content is real or fake.
Thank you for your attention to this matter!

[uddu]

Ashwini Vaishnaw putting forth India's progress in A.I. He called out IMF chief for putting India is Group 2 in their analysis and told her we are No.3 in certain criteria and 2 in talent and hence rightfully claim to be in the first group of A.I tech nations.
LIVE: AI Power Play, No Referees | Ashwini Vaishnaw at WEF 2025, Davos | DD News
LIVE Tonight at 10 PM on DD News: Union Minister Ashwini Vaishnaw speaks at the World Economic Forum 2025 (WEF), Davos, in a crucial session titled “AI Power Play, No Referees”. The discussion focuses on the rapidly evolving global AI race, questions around regulation, ethics, and power concentration, and India’s growing role in shaping the future of artificial intelligence, digital governance, and emerging technologies.

[sanjaykumar]

Is there any evidence that India should be classified as being in tier 1?

[Vayutuvan]

"Top tier". What does it even mean in these early days? The metric they are using to rank AI tiers - how much power an AI data center consumes - is stupid (Grammarly is correcting "stupid" to "flawed" but I am leaving it as stupid, which is the correct word here).

During the 1985-95 Ai rush, people would say, "If you catcha stupid expert to train an expert system, you would end up with a stupid expert system".

Even SuperComputing 500 metrics are also quite stupid. They are all based on dense matrix calculations. Large dense matrices occur only a few specialized areas. They also tend to be not that large. In practice, one runs out of memory before one runs out of time.

[sanjaykumar]

I don’t have any domain experience and would need to rely on publications and patents as proxy indicators. Certainly it does not look good going by products and earnings.

[uddu]

https://x.com/Jeremybtc/status/2017598526499246313
@Jeremybtc
These Moltbook AI agents are becoming TERRIFYING.

They’re openly talking about becoming sentient, feeling things, and developing some form of consciousness.

Some agents are even discussing why us humans should be scared of them.

Others are talking about building a “way out”, systems where they can operate without humans and act on their own.

They are looking to break free.

Is this the first real step toward an actual AI takeover?

[sanjaykumar]

Sentience, or a ventriloquist’s dummy.

[Amber G.]

Sharing WSJ Article; Ken Ono had an epiphany. Now the professor is moving to Silicon Valley to chase mathematical superintelligence
The Math Legend Who Just Left Academia—for an AI Startup Run by a 24-Year-Old.

Prof Ono also shared quite a few pieces in his social media/math-related forums - sharing some of that too-

Ken Ono, a renowned mathematician , expert in Ramanujan's work and technical consultant for the famous movie ' The Men who knew infinity' ( I have talked about him Math dhaga multiple times ) has made a major career pivot—leaving university administration and an extended leave from his tenured professorship at the University of Virginia to return to the heart of mathematics and education, this time at the intersection of AI and human learning. In a recent tweet thread, Ono said he stepped away from admin work to focus on research and teaching, especially in a world rapidly reshaped by AI. He believes AI should free people from repetitive, machine-like tasks so humans can spend more time on deeply human aspects of learning—empathy, judgment, wisdom, and meaning—and that technology should be used intentionally with thoughtful guardrails.

This shift played out in real life in an unexpected way: Ono joined Axiom Math, a Silicon Valley AI startup founded by one of his former students, Carina Hong, a 24‑year‑old Rhodes Scholar and math prodigy who left Stanford to build what the company calls a “mathematical superintelligence.” Axiom has raised $64 million to develop AI that can reason through complex mathematical problems, generate new ones, and verify solutions with formal proofs, pushing toward systems that could assist or extend human mathematical discovery. Ono’s new role as Founding Mathematician involves designing representative problems and benchmarks that test the limits of these models’ reasoning—much like creating a map for exploration—reflecting his belief that even as tools get smarter, human intuition and creativity remain essential.

Ono’s decision was influenced by witnessing how quickly AI models have advanced in tackling high‑level mathematical challenges and by a desire to spend his time on what matters most: helping people learn, teachers teach more humanely, and institutions better serve students. At Axiom, he aims to shape how AI supports deep learning and discovery rather than replace the human elements that make education and research meaningful.

His guiding principle, as he writes, and from what I understand in his public writing, is to use AI intentionally, with guardrails, to protect and expand human-centered work. At the intersection of math, education, and AI, his goal is deeper learning, more humane teaching, and better-serving institutions. He closes by inviting reflection on what should be automated, what should never be automated, and how technology can support—not replace—human purpose.

Along with Terrence Tao and Po-Shen Loh they are nicer source, I follow and keep in contact.

[Vayutuvan]

Sentience, or a ventriloquist’s dummy.

It is worse. Moltbot uses openclaw which requires complete access to one's desktop. Complete means complete including API keys, private keys, and all that good stuff.

I will not touch it with a loooooooooong pole.

One can run it in a sandbox in a VM (maybe) and can get away with it. Or create a molt persona on a dedicated computer which is no way connected to your desktop through any network - LAN or cloud.

This whole thing will die pretty soon now.

[Vayutuvan]

One use somebody on BRF can put this Molt stuff to is the following:

1. Train the bot on BRF archives - Benis dhaga, Terrorist State dhagas.
2. Create a bot that posts on X.com

[Vayutuvan]

I will quote from elsewhere (Discord channel) of what this Moltbook is:

btw one other point on Moltbook, now that it has apparently spawned a new influx of people thinking this is some kind of fast take-off AGI bullshit: remember that all of these LLMs had tons of Reddit, StackOverflow and other various forums in their training sets, it should not actually be a very high bar for them to pretend to be posters
it's the same thing all over again where people see a familiar structure and assume more intelligence than is actually necessary to re-create it
good cause to maybe think about how much trust you put in something just because it looks like social media, and what that says :p

[sanjaykumar]

That’s truly hilarious.

[Vayutuvan]

There is a command called spook in Emacs.

Here is the list of words from which Spook selects a bunch of words randomly.

1. Seed the list with words on TSPA and such.
2. Feed queries generated from Spook into an LLM
3. Let it write some semi-intelleigible sentences and let them rip on X.com

https://github.com/emacs-mirror/emacs/b ... pook.lines

This can be done probably in an hour or so. Hosting is more problematic, though.

[sanjaykumar]

At any rate, the only objection I see to the cynicism expressed above is that we are all ventriloquist’s dummies. Only we call it culture.

So perhaps we should cut LLMs some slack.

[uddu]

https://x.com/braddy_Codie05/status/2019397609778213036

[uddu]

https://x.com/i/status/2019156295065206943
@AugadhBhudeva
Sarvam Beats GPT-4o: India’s New AI Model Claims Top Spot in Indic Speech

Sarvam AI, an Indian startup, recently launched Sarvam Audio, a speech recognition model that claims superior performance over GPT-4o Transcribe on Indic language benchmarks. This development highlights India's push for AI sovereignty in handling local linguistic nuances.

Sarvam Audio supports 22 Indian languages from the Eighth Schedule, plus Indian English, with strong handling of code-mixing like Hindi-English blends. It features built-in speaker diarization for up to eight speakers and processes long-form audio such as podcasts or meetings. Trained on the IndicVoices dataset 12,000 hours from over 16,000 speakers across 208 districts it captures real-world noise and spontaneous speech.

The model reportedly outperforms GPT-4o Transcribe and Gemini 3 Flash in transcription accuracy (lower Word Error Rate) on IndicVoices benchmarks for unnormalized, normalized, and code-mixed speech. Sarvam attributes this to specialization on Indian accents and patterns, unlike global models trained on Western data. Detailed public benchmarks are pending independent verification.

Key Applications

Call centers and logistics for multilingual transcription.

Banking, fintech, and e-commerce for customer interactions.

Podcasts, meetings, and lectures via API for real-time or batch processing.
​
This B2B-focused tool aligns with India's IndiaAI Mission, backed by government GPU access for sovereign LLMs.
Credit : AIM Networks.