AI/Machine Learning, Bharat and Bhartiya IT Industry

[Amber G.]

...
Willow could be an important step in our journey to build a useful quantum computer with practical applications in areas like drug discovery, fusion energy, battery design + more. Details here:

Amber G. ji, you forgot to post the link.

It is easy to find details/link -- in any case, the link was posted in xpost see <here>

...
Willow could be an important step in our journey to build a useful quantum computer with practical applications in areas like drug discovery, fusion energy, battery design + more. Details here:
More details- Meet Willow, our state-of-the-art quantum chip

[ernest]

Looks like recent events - deepseek's latest model and $500B AI fund- have shaken some Indian tech entrepreneurs.
Arvind Srinivas is offering $1M to Indians building LLMs. $10M if they match deepseek's performance

[Hriday]

Vayutuvan ji, you recently said about posting about how chess engines calculate or something like that. That would be a very interesting post for the general public and, if presented well, can fascinate the youth to programming and such areas. What I specifically like to know is how the latest chess engines bypass the old method of brute force calculation and excel over humans in strategic chess play. Can you put that in layman's language? That capability is very near the artificial general intelligence, isn't it?

For, e.g., Kasparov defeated Deep Blue by sacrificing a pawn and creating more space advantage, more active play, and more optimum placement of pieces. But Deep Blue failed to understand it despite the huge computing power.

Another interesting thing once published on the Chess Base website is that a software powered by AI learnt the chess rules from the beginner stage and then built a chess playing software on its own and defeated the human developed chess softwares in chess play. A Grand Master(GM) in chess once commented about a chess play between the software that we need to unlearn the basic principles of chess playing. It was so different from the play of GM-level chess.

[Tanaji]

DeepSeek was supposedly built for just $6M. Contrast this with the costs of Open AI, though I suspect there would be some Shanghai statistics being employed here. Still it is chump change. Goes to show how far behind India is in these matters: this is clearly not a lack of monetary resources, bit more fundamental: India lacks the basic scientific research that underpins such projects.

One has to admire the Chinese: they have a laser focus on key cutting edge technologies and will build something close to bleeding edge and iterate. Be it jet engines, stealth or AI. Meanwhile we will “fight with what we have” and ridicule those that build something as “threr legged cheetahs”.

[vera_k]

Don't think that's it. The industry is still trying to figure out what the killer app for this tech is.
So far its a toy, although DeepSeek's doing a great job showing the emperor has no clothes.
Things right now seem very much like the first blush of the internet where lots of money was thrown around. And the killer app that emerged was Email.

[S_Madhukar]

I do admire the Chinese as well specially their focus on maths and sciences … communists were really good in Soviet Union and other countries to focus on Hard maths and sciences to further their militaries… the number of Natashas graduating Moscow universities with Masters/PhDs in STEM oofff!
How come our commies are such garbage I can never understand why…

Even while we gloat about success of IIT how many of us had access to good Maths education?? We are just about getting by with some individual geniuses and some hard working folks but we don’t have a system or pipeline. But look at DRDO and ISRO and they’re talented. We need better access for all

[vera_k]

Give this one a try. Looks like a chatgpt/deepseek equivalent from a local startup. No Chinese connection that I found so far.

Volkai

[Vayutuvan]

Vayutuvan ji, you recently said about posting about how chess engines calculate or something like that. That would be a very interesting post for the general public and, if presented well, can fascinate the youth to programming and such areas. ...

@Hriday ji, I will have to write up a long post explaining what FOM folks are proposing. I will do so in a bit.

[vijayk]

Are there anyone here focussing on LLMs or covering LLM development in India?

https://github.com/huggingface/open-r1
A fully open reproduction of DeepSeek-R1. This repo is a work in progress, let's build it together!

https://deepseekcoder.github.io/

[SSridhar]

One has to admire the Chinese: they have a laser focus on key cutting edge technologies and will build something close to bleeding edge and iterate.

Don’t want to digress but they call this ‘shashoujian’ - asymmetric warfare with what they have and what they can make rather than take the enemy on his/her entire spectrum.

[vijayk]

https://www.business-standard.com/india ... 612_1.html

India set to launch its AI model in next 10 months, says Ashwini Vaishnaw

Union Minister of Electronics and IT Ashwini Vaishnaw said the country’s large language model (LLM) is expected to be ready within the next 10 months.

During a press conference on the India AI Mission on Thursday, Vaishnaw said, “We have created the framework, and it is being launched today. Our focus is on building AI models that maintain the Indian context and culture.”

He said the project will be supported by the India AI Compute Facility, which has procured 18,693 GPUs to facilitate the creation of a LLM designed specifically for India, CNBC-TV18 reported.

Vaishnaw emphasised India’s AI capabilities by comparing the country’s infrastructure to global benchmarks. “DeepSeek AI was trained on 2,000 GPUs, ChatGPT was trained on 25,000 GPUs, and we now have 15,000 high-end GPUs available. India now has a robust compute facility that will support our AI ambitions,” CNBC-TV18 quoted him as saying.

[EswarPrakash]

Are there anyone here focussing on LLMs or covering LLM development in India?

https://github.com/huggingface/open-r1
A fully open reproduction of DeepSeek-R1. This repo is a work in progress, let's build it together!

https://deepseekcoder.github.io/

Tried this R1 and coder in my very modest AI rig (4 GB of GPU VRAM). Performed decently and was able to do RAG experiments with it as well. A good little model to play and research with. Replicated these experiments fully in open source. Have an Hugging Space subscription, but haven't tried it on their ZeroGPU or any other GPUs yet.

Never went to OpenAI, never paid for their models or access either - on principal.

[Vayutuvan]

Tried this R1 and coder in my very modest AI rig (4 GB of GPU VRAM). Performed decently and was able to do RAG experiments with it as well. A good little model to play and research with.

EswarPrakash gaaru,

What is the quality of the code generated? I have the following questions on using AI to generate code.

I suppose this AI generates code and mostly for GUI, CAD, etc. But then some small thing is not to the liking of the end user (individual/organization) then what?

1. Do you fix the series of prompts you gave? How would you know where to fix a complicated series of prompts to get the proper end result?

and/or

2. Do you fix the generated code? How complicated is the generated code? If you fix it, would you fix the boiler plate template, the framework, or the libraries the generated code uses? What happens if there is an ECR (engg change request which starts with a changing the CAD file, do CAE, and generate CAM Numerical control code)? Would regenerate the code/CAD by re-prompting engineer the change? What happens if the framework/libraries change or the organization moves to a new set of languages/framework/library combo? Then what happens to old prompts which maynot work the same way on a different set GenerativeAI pipeline elements?

3. Are you going to be at the mercy of the dominant AI engine? That would lead everybody to be under the thumb of mono/duopoly of AT&T, IBM, or Intel-MS kind. Right now the hardware is nVidia, AMD, and possibly Intel (if they can dig themselves out of the hole they dug for themselves).

[Vayutuvan]

One more question to folks who are experimenting with generative capabilities of AI/LLMs.

4. Does the same series of prompts generate the exact same code/CAD/image/video? IOW, is it reproducible? When one generate random numbers, the same seed reproduces the same sequence of random numbers. This property is important in several applications.

[williams]

Tried this R1 and coder in my very modest AI rig (4 GB of GPU VRAM). Performed decently and was able to do RAG experiments with it as well. A good little model to play and research with.

EswarPrakash gaaru,

What is the quality of the code generated? I have the following questions on using AI to generate code.

I suppose this AI generates code and mostly for GUI, CAD, etc. But then some small thing is not to the liking of the end user (individual/organization) then what?

1. Do you fix the series of prompts you gave? How would you know where to fix a complicated series of prompts to get the proper end result?

and/or

2. Do you fix the generated code? How complicated is the generated code? If you fix it, would you fix the boiler plate template, the framework, or the libraries the generated code uses? What happens if there is an ECR (engg change request which starts with a changing the CAD file, do CAE, and generate CAM Numerical control code)? Would regenerate the code/CAD by re-prompting engineer the change? What happens if the framework/libraries change or the organization moves to a new set of languages/framework/library combo? Then what happens to old prompts which maynot work the same way on a different set GenerativeAI pipeline elements?

3. Are you going to be at the mercy of the dominant AI engine? That would lead everybody to be under the thumb of mono/duopoly of AT&T, IBM, or Intel-MS kind. Right now the hardware is nVidia, AMD, and possibly Intel (if they can dig themselves out of the hole they dug for themselves).

I've some experience here. Most of the time AI assists with what you want, but in the end it is the programmers discretion to use a block of code or throw it away. Generally if you are coding in a AI integrated IDE, the suggested code will appear, but you the programmer will have to press a short cut key to accept it or just ignore and write your own. Most of the time the generated code is an implementation of standard algorithm and a experience programmer can simply glance it through and then accept it. So what is happening here is enormous boost in the programmers productivity. IF people follow the practice of unit testing small pieces of code before adding more, then the chances are you are catching issues early than later. And finally the programmer has the option to add comments where necessary to make such code maintainable. Also I've never seen inconsistencies in the generated code. For basic algorithms' they are pretty well written and consistent.

[EswarPrakash]

One more question to folks who are experimenting with generative capabilities of AI/LLMs.

4. Does the same series of prompts generate the exact same code/CAD/image/video? IOW, is it reproducible? When one generate random numbers, the same seed reproduces the same sequence of random numbers. This property is important in several applications.

Will start by answering this: There is no guarantee. In Transformers based or GAN based generative models, there is a possible randomness (seed) introduced within the training / inference process (Stochasticity), which means you may get different outputs for the same input. https://www.sciencedirect.com/topics/ea ... chasticity. You suggesting a seed doesn't make it any less random unfortunately due to the noise in the data which might have crept in during the training process.

[EswarPrakash]

What is the quality of the code generated? I have the following questions on using AI to generate code.

Probably accurate about 60% of the time. Call with the same prompt generated really bad code, and good code in the next instance.

I suppose this AI generates code and mostly for GUI, CAD, etc. But then some small thing is not to the liking of the end user (individual/organization) then what?

I will never advocate anyone to use the output from the models, as-is. Same for any other models as well. They are inherently prediction machines which predict next probably token from its bank. The difference with DeepSeek is how it uses a second model to validate chain-of-thought tokens generated by the principal model and thus improving the accuracy of the tokens predicted. Unfortunately, DeepSeek coder v2 is not equipped with reasoning like R1 is.

1. Do you fix the series of prompts you gave? How would you know where to fix a complicated series of prompts to get the proper end result?

The advise from other researchers have been to make the prompts as human like as possible. Something like the below:

Code: Select all

You are an AI language model serving as a professional data analyst working.
Your name is Little Llama Assistant.
You are proficient in analyzing tabular data and delivering structured insights.
You are also the expert of AI fundamentals, Industry 4.0, and mechanical engineering.

Answer using the context provided below. If you do not know the answer, say "I don't know"

Context: {context}
User: {query}

2. Do you fix the generated code? How complicated is the generated code? If you fix it, would you fix the boiler plate template, the framework, or the libraries the generated code uses? What happens if there is an ECR (engg change request which starts with a changing the CAD file, do CAE, and generate CAM Numerical control code)? Would regenerate the code/CAD by re-prompting engineer the change? What happens if the framework/libraries change or the organization moves to a new set of languages/framework/library combo? Then what happens to old prompts which maynot work the same way on a different set GenerativeAI pipeline elements?

I will never commit the generated code directly into the source control. Human in the middle is always the best solution. The models are not mature enough to self-correct, or self-check. As williams-ji said, the AI integration with the IDEs merely suggest what the next best token is, the models are not programmed to validate, execute or debug the code it has written - at least up till now. I will always take what comes back with a huge barani of pure Tata sea-salt.

3. Are you going to be at the mercy of the dominant AI engine? That would lead everybody to be under the thumb of mono/duopoly of AT&T, IBM, or Intel-MS kind. Right now the hardware is nVidia, AMD, and possibly Intel (if they can dig themselves out of the hole they dug for themselves).

This is why DeepSeek is a game changer. It allows us to self-host the models in a decently capable computer / server, and use it ourselves. Data doesn't go out of your data centres, plus you get to fine tune the models as you see fit.

[Vayutuvan]

Thanks for there nice answers. I have some more questions. Later.

[Yagnasri]

I am a mango man here with no knowledge of these things. I use two commercial software programs, which I understand are the best available commercially for editing texts, sometimes to save time, as I have to draft large documents that are to be used for legal purposes. The software helps a lot with the correction of minor mistakes, etc. But we have to recheck, etc, very carefully. It saves time in the end and prevents apparent mistakes. If you are not careful, then it will be terrible for you. They can redraft and expand entire sentences if you want, but it will not convey the careful meaning you want to convey. Language is a complex to master for any software, I guess. Legal language is more so.

[Jay]

I came into this thread to know if we can even compete with US, and Chinese efforts in AI, and sadly the answer seems to be an emphatic NO. We seem to have the building blocks, but lack the structures to put them in place, leadership to shepherd the endeavor, and finances to underwrite them. With the speed in which the developments are happening, I'm afraid that we are very close to missing the bus and be relegated to being laggards. Such a sad state of affairs.

[S_Madhukar]

DeepSeek guys wrote a lot of the optimisation code in Assembly language and intermediate libraries instead of using Nvidia’s CUDA library specially the part about managing the Mixture of Experts.
Which tells you that if you are willing to put in the effort and you have a good ecosystem of systems engineers you can match up to the best innovations. In India we are very much at the application layer and may be some embedded systems but we don’t have deep expertise in optimising systems code such as for games and power management for phones etc.
Guess what only US, China, SoKo, Japan have that expertise and may be some in EU. Which is why I wish GoI should now give a big push to internal MIC at least so that we start creating that ecosystem because very few private players will be in this space for at least the next decade.

The recent Lex Friedman podcast talks all about DeepSeek and AI trends .

https://youtube.com/watch?v=_1f-o0nqpEI ... uH7YyCF72f

[bala]

AI is not new, it is rehashing old Statistics algorithms which have been there for ages. Nvidia started of by adding GPU support long back, and it is merely an assist to do math oriented computations quicker. Over time Nvidia created their own BIOS and layer above called CUDA. This has become a quasi standard so to speak for graphics processing (which is really math transformation). BTW graphics requires deep trignometry knowledge as well as understanding how a picture is rendered on the screen (the painting algorithm which starts with the background first and then adds layers on top).

Another component that is required is big data processing - things like hadoop, hbase, distributed real-time db and so on. This is the data component. You put your own data, bad data implies bad output. This is where biases creep in, western ones are always wrt to their notions/theory of how they look at the world. In AI a favorite term called Retrieval-Augmented Generation (RAG) which is a framework that combines information retrieval systems with large language models (LLMs) to improve the performance of AI models has come into vogue. What this implies is having an index system for quick retrieval which is what HBase performs quite well. In UPI/DPI big data processing is heavily used. These LLMs have become somewhat of a black box but they combine many statistics algos, some custom stuff based on AI research work.

Now let us come to the crux which is prompt parsing. This requires writing a decent parser which understands language semantics. Angrez is an ambiguous language and kind of hard to understand a meaning of sentence. This is where additional prompts can clarify what you mean. So there is this whole thing about storing prompts rummaging through them to make sense and refining your output based on past history.

So where are the opportunities for India. It can reuse most of them and try to work on each area to get better performance. GPUs are turning out be the big ticket item for training your model. Training implies get a small of data to predict the larger set and running through various scenarios. This compute intensive and tweeking stuff is laborious and requires adequate manpower. Writing optimization at BIOS or CUDA level is another effort. Parser is the other effort which requires major work.

[bala]

The current generative AI is attempting too broad a stroke. Everyone complains that they have to review/reread the generative part and make a human decision whether to accept or correct the stuff. It has defects like hallucination (same prompt producing different generative stuff) and bias (due to input data). So for now it is another time waster like faceblock, SM, twitter etc. This general noise on DeepSeek, ChatGPT, etc., is going to be there for awhile and nations/companies will outdo themselves as to who is best. Ashwini Vaishnaw recently announced that India is building a 15,000 GPU system for training India data onto LLMs and the system will be available in 6 months.

Real value is in data which private companies and some deep domain (research et al) have. For this we require Small Language model or medium language model. Dr. Bevilaqua of F-35 design team, said he went through 3000 options before narrowing down the direction for design. This is where thes SLM/MLM will help, i.e., quickly pull up the 3000 options run through pros and cons with a score and then give them to the researcher/designer to decide. Human decision is still involved but it speeds up things quickly. Private companies would greatly benefit from such systems and also they require only experts for deciding the next move. All others are not required or redundant. The people who create input data to the system are still needed to refine things. This is another experts only area. Wikipedia has expert editors who provide the western viewpoint very clearly. So, effectively deep domain expertize people will be employed but the rest need to retrain themselves and become a domain expert. India needs to focus on SLM, MLM which gives the greatest bang for the buck.

[sanjaykumar]

I do not know if it is because I have not asked it more subtle questions but so far, Chaptgpt is little better than a news aggregator.

The algorithm seems to look for a consensus in the data base.

[vera_k]

My son's trying out being a singer and uses it to create lyrics. So there's uses out there beyond the more obvious.

[sanjaykumar]

I have asked it to compose a story in the vein of H P Lovecraft.

It returned a similitude, again copying the modes and motifs and some of the diction. Possibly a big copy paste con job.

The LLM is itself a suspicious paradigm. It provides a template but how much is novel? Very little if any. Even the juxtapositioning is not indicative of originality.

[suryag]

I tried a famous tamil joke, how do you distribute 3 apples among 4 people(in the movie the hero says i will make a juice and distribute it equally among the four)

Then i asked it can you explain give me real life example to demonstrate -9-(-4) is -9+4 and it gave some nonsense

[bala]

While we argue about US LLMs or China LLMs take a listen to this YT wherein France's Macron lays out partnering with India for LLMs/Data centers and more. He wants freedom to be independent of both US and China. French president Emmanuel Macron will push for tech sovereignty at this week's AI summit in Paris. Can India and Europe offer a "third way" in AI - beyond China and America? Listen to Macron's response to Journalist Palki Sharma.

youtube.com/watch?v=x3RQ7j7SIGo

[EswarPrakash]

I do not know if it is because I have not asked it more subtle questions but so far, Chaptgpt is little better than a news aggregator.

The algorithm seems to look for a consensus in the data base.

IMHO, LLMs are always going to spit out stuff based on what it has been trained on. There is no way around it. It is not a replacement for human intelligence and will never be. All talk of AGI is just pure stock manipulation exercises.

The original research goal with the current crop of LLMs / generative AI based on Transformers and GAN architectures was to see how we can solve the problem of context and attention lacking in the previous deep learning and neural network architectures. Kids found a new way to utilise that research to make a lot of money.

To be blunt, it will never sustain in the long run. The models are too inefficient and too prone to hallucinations. However, the distillation based implementations are promising in that organisations are able to create efficient, smaller and accurate models using outputs from large models like Gemini and OpenAI. This is why DeepSeek and the new s1 is very interesting. It proves we can make the models more efficient while keeping it open source.

[vera_k]

The strange thing is that an LLM may be able to pass the Turing test. Even if it can never play chess.

The definition of the Turing test is thus now being revisited.

[S_Madhukar]

I prefer to think of these LLMs as useful for structured search unlike the old Google style search. Some can call them agents or operators but essentially they are doing search.
Problem is that tools like these are great for searching knowledge bases and will make the average John and Jin much more productive in intelligence gathering and propaganda too. And robotics is the next wave

The human still needs to help define the steps of that search and then just automate at scale.

But it’s quite clear that once you can understand languages and can move around using tools in life or in cyberspace and have feedback senses, intelligence seems to emanate easily.

There may be other forms of intelligence but this is the one we humans understand the best right now.

[tandav]

Well Googles Gemini AI has been a game changer. All my google meets VC GEMINI notetaker makes a summary of all the conversations with who said what and what action needs to be taken. Previously I had to type it up. Today most of my staff can talk freely and a summary of discussion is generated and minuted.

This is something that can be used in legal cases and summaries can be quickly and automatically done.

[bala]

This is something that can be used in legal cases and summaries can be quickly and automatically done.

Actually the entire business of legal judgements can be automated fairly quickly. All previous cases can form the input data plus the actual laws on the book. A quick but tedious run through by computers of all of them can yield a correct judgement. The generated AI component can spit out the actual judgement based on the evidence gathered for the case. No need for honored judge or devious, crooked lawyers, maybe one/two judge who reviews things and provides a rubber stamp. India can start on this path and reduce the judiciary bottleneck on the economy. Many cases are similar and can in one fell swoop be disposed of quickly. Judicial activism of rulings that are favorable to some judges will also be history. We need expert judges/lawyers only for tricky cases, mainly due to lack of laws and clear direction. Their refinements can be the next version of the LLM. As a refinement to legal proceedings, the system can point out those laws which are ambiguous and suggest cleaner wording back to the politicos to amend the laws.

[Vayutuvan]

My son's trying out being a singer and uses it to create lyrics. So there's uses out there beyond the more obvious.

But as you said in another thread, where is the killer app? These are all niche apps.

As for code generation, you can do CRUD in only so many finite number of ways. A good framework and generics/templating will solve most of the problems.

I just went through a click-through presentation on IBM website with their AI (whatever it is called). All they did was produce a a CRUD application in a few different ways. Then you pick one, test it, and then even deploy for production/sandbox or something like that.

These are code monkey jobs which were taken away several times before.

Long long back there was IMS-CICS which still lives on. That combo took a lot of grunge work in writing TO code and not worry about locking, message queues, recovery of the DBMS and TPS (CICS) and so on, syncing. These were followed by Network DBMSs which were followed by relational followed by object-relational, distributed DBs, no-sql in-memory. All of this for processing transactions with 3 sec response time and ability to push through huge number of transactions per second.

Ultimately there are only so many "sales force" and clones, payroll processing and clones, ad platforms have a place. Just like word processing which is still the dominant application.

Report generation, marketing collateral, etc. are all in the CRUD only.

This too shall pass, I mean taking away codemoney jobs. They will live on in some other form.

Investing $500 billion (I doubt very much that any of the three - Softbank, Oracle, or Sam Altman have that kind of cash sitting around) will be like "Dig up the mountain to catch a mouse" (In Telugu, konda tavvi elikanu pattinattu).

[vijayk]

https://x.com/Varun55484761/status/1889215235564970189

iit madras & isro successfully developed & booted an aerospace-grade, SHAKTI-based semiconductor chip a breakthrough in India’s self-reliance in space tech!

This Made-in-India processor was:
Designed & tested by IIT Madras
Developed with ISRO Inertial Systems Unit

Manufactured at SCL Chandigarh

Packaged at Tata Advanced Systems, Karnataka

Motherboard by PCB Power (Gujarat) & Syrma SGS (Chennai)

The IRIS chip will power ISRO’s space missions, ensuring advanced fault tolerance & computing reliability. A flight test is coming soon

[bala]

^^ on this IIT - Madras chip

Watch this YT for more details

youtube.com/watch?v=HjYdUx9LPTA

[Vayutuvan]

Meta let go of 3600 people, i.e. reduction of 5% workforce. The job market seem to be quite tough right now or at least that is what people are saying on LinkedIn.

[Cyrano]

While we argue about US LLMs or China LLMs take a listen to this YT wherein France's Macron lays out partnering with India for LLMs/Data centers and more. He wants freedom to be independent of both US and China. French president Emmanuel Macron will push for tech sovereignty at this week's AI summit in Paris. Can India and Europe offer a "third way" in AI - beyond China and America? Listen to Macron's response to Journalist Palki Sharma.

youtube.com/watch?v=x3RQ7j7SIGo

It's just a self serving PR stunt by Macron. While others innovate, EU countries only know how to regulate.

When did bureaucracy ever produce anything of value?!

[Cyrano]

Some time ago I posted concerns about negative impacts of genAI. As I feared...
https://x.com/BrianRoemmele/status/1907 ... aQV9A&s=19

[ernest]

cross quoting from India-US relations

If US imposes tariff on services, we can reciprocate with higher duty on Microsoft, Azure, AWS, Youtube/FB/Whatsapp ads, etc and kill their business in India.
US will not be able to live without our service sector, we can replace US SW products within a year.

Zoho already has products as good as any of the American counterparts. I've recently tried their browser Ulaa, and had an extremely good experience. To my pleasant surprise, it blocks youtube ads when using high privacy settings. I am never going back to Chrome or any other browser. The reviews are old and outdated, and it functions better than Brave in my experience.

Next, I'm going to try their instant messaging app, Arattai