Joined February 2019
- Tweets 25,361
- Following 970
- Followers 20,248
- Likes 18,837
7,076 Photos and videos
Pinned Tweet
This Week of idealist is Now live :)
Technical scan on 500 companies with scans and filters, plus one case study on premium consumption, and here you understand how to look at strength and what variables are involved in it
normalguy.co.in/post/idealis…
1
19
1,541
Why can't India build a conventional submarine on its own yet?
Design capability exists, the shipyard capability exists. What India lacks are two specific pieces of equipment:
> Propulsion diesel engines - submarine diesel engines are special. When a submarine snorkels (breathes through a tube while partially submerged), the engines face high back pressure from the water. Indian industry hasn't built full-scale engines that handle this.
> Compact electric motors - conventional submarines need permanent magnet motors in the 3.5–5 megawatt range, extremely compact. One company in Hyderabad is working on it, but it is not ready yet.
For SSBNs (nuclear submarines), India has reached very high indigenization levels. But conventional submarines have different requirements - smaller size means everything must be more compact and power-efficient.
India planned 24 submarines by 2030. It built 6. The older submarines are aging out. Pakistan is getting modern Chinese submarines fast.
India has the design and building capability but has been stuck in procedural delays, sequential decision-making, and gaps between programs that kill accumulated skills.
The decisions made in the next 1–2 years (signing P75I, splitting orders, keeping P76 moving) will determine what India submarine fleet looks like by the mid-2030s
Italy had more submarines than the US and Germany combined at the start of World War 2.
But after losing their fleet and stopping construction, they could never rebuild their submarine industry.
Submarine-building is a skill that dies if you don't keep the production line running.
7
6
69
9,621
Italy had more submarines than the US and Germany combined at the start of World War 2.
But after losing their fleet and stopping construction, they could never rebuild their submarine industry.
Submarine-building is a skill that dies if you don't keep the production line running.
2
6
38
14,869
Tejas MK1A uses Israeli radar, DRDO-built electronic warfare suite, & DRDO-built weapons. Problem is integration - these different systems can't talk to each other
India defense budget is 2% of GDP but that 2% already eats 14.67% of the entire Union Budget, the highest share of any ministry. Interest payments alone take 26% of all government spending.
Inside defense itself, pensions take 22%, salaries another 26%, That leaves about 28% for buying new weapons. Every rupee added to defense comes from education, health, food subsidies, or rural development.
Moving that number higher means cutting somewhere else, and every option for cuts has strong political protection
Late CDS Gen Rawat told Parliament in 2020 that each Agniveer saves roughly ₹11.5 crore in pension costs over a career. That was the original pitch behind Agnipath. Now the PM says it was never about pensions, just about keeping forces young.
But the Bombay High Court was told in May 2026 that Agniveers get no pension parity with regular soldiers. Army wants 50% retention, the scheme caps it at 25%
Tejas MK1A uses an Israeli AESA radar, Indian EW suite, and DRDO missiles like the Astra. The problem is software, the Israeli radar runs proprietary code that does not talk easily to Indian avionics.
HAL thought integration would take months after first flight, it has stretched past two years. GE has delivered just 6 of 99 contracted engines. The defense acquisition procedure meant for 2025 arrived as a draft in February 2026 and is still being debated.
Making parts is one challenge, making them work together is a harder one.
This is Why AMCA Comes :) HAL have 2lakh cr of order book and PSU have their own limitation :)
theprint.in/defence/iaf-will…
6
21
112
8,588
Social Media and Now AI Deepfake
2 Guwahati schools allegedly told students to stay quiet after AI deepfake images of 64 girls and teachers were sold on Telegram
South Korea faced this same crisis in 2024 when over 500 schools were targeted by deepfake rings on Telegram, run by teenage boys, By early 2025, police had arrested 682 suspects.
In Malaysia, a 16 year old sold AI generated images of 40 classmates for RM2 each. Free AI nudify tools are available to anyone with a browser and that is the root cause of every school deepfake case we are seeing globally.
Just think about the situation. What if someone blackmails families or women with this? This loop can go very deep, and what you are seeing is just the tip of the iceberg.
I watch Recently The Capture Web Series, Yu can Connect with that :) How to Escape this? Not Easy at all i think Impossible :)
Season 1 was the best only :)
reddit.com/r/guwahati/s/4MWe…
4
24
2,103
Normal Guy retweeted
16 Jan 2025
This Sector is FIRE
Big Countries will try to put some friction and tune their politics, geopolitics with technology. Limitation is quite brutal if USA use it by all means
First Mover Advantage is like yu make the Atom Bomb let's put others to just live in shadow.
AI with all means bigger than Nuclear or anything they put World data, scale themselves and now from defence to Scientists all will use it
the top model will stay in USA and only used by few Organisations and their people.
Lets be clear if yu think AI is just another story. China is not stupid they are still buying a lot, even doing smuggling of Nvidia GPU and making their own chips.
India is too late or I think it still early to build LLM with support of Government or themselves and scale bit by Bit and I don't even if someone will do
IT companies and just saving money in their Reserve, there are only few People in India that they can do it perhaps this is bigger than bomb or not-- time will tell
when first Strike happens ...
16 Jan 2025
This GPU quota rule & India being excluded from the allies category is something to ponder over. Only 50k GPUs are allowed to be imported. Just the IndiaAI mission is trying to get 10k GPUs. The larger demand on GPU, DC companies will certainly have an impact or at least need to assess how it impacts. For sure that in the ages of technology, compute is going to be the option that'd be heavily used to twist nations and their policies by biggies.
Don't blindly buy DC companies in listed / unlisted please.
2
5
22
6,493
Jun 13
Nilekani has argued India does not need to build frontier AI models and should let the Valley handle it.
That view got tested on Friday when the US pulled Fable 5 and Mythos 5 from every non American user within hours, The strategy works when access is permanent.
It breaks down when one export control letter can turn off your best AI tools overnight.
Nuclear comparison works in direction but breaks on difficulty, Pakistan built a bomb for about $3 billion using domestic uranium and its own physics talent.
Frontier AI needs Nvidia GPUs made at TSMC, both supply chains the US controls. You cannot go independent in AI the way countries went independent in nuclear because the basic inputs are not available at any price.
Space fits better than nuclear, Only 3countries can put humans in orbit independently and the barrier is cost, not secrecy, Same with frontier AI.
US hyperscalers will spend over $650 billion on AI this year. India entire IT sector makes around $250 billion. That spending gap alone explains why India cannot build a Fable competitor the way ISRO built a launch vehicle.
Open source AI from China and India right now is the equivalent of buying rides on someone else rocket while you build your own program.
India signed the Pax Silica declaration in February, formally joining the US led tech bloc alongside Japan, South Korea, Australia, and the UAE.
4 months later, Indian users still lost access to Fable 5 and Mythos 5 because the ban hits all foreign nationals regardless of alliance status.
Being inside the bloc gives you a seat at the table for chip supply chains. It does not protect you from overnight model shutdowns. Alliance membership and frontier model access turn out to be separate commitments.
Chinese AI labs released their best models under free licenses partly because US chip controls pushed them toward open source as a workaround.
MIT research shows Chinese open source now leads US models in global downloads. On Friday when Anthropic two best models went dark for every non American user, Qwen and DeepSeek were still sitting on Hugging Face, downloadable with no permission needed.
For most of the world, Chinese open source just became the default insurance against American export controls on AI.
Trump June 2 executive order asked AI companies to voluntarily submit frontier models for government review before release.
Anthropic did not pause Fable 5 and Mythos 5 for that process, 10 days later Commerce Secretary Lutnick sent an export control letter forcing a total shutdown of both models for every user on earth.
I wrote this in the past, in Jan 2025. If you're curious, read it. Even the GOI or any nation knows this too. If I know this, I'm just a normal being. 👇
anthropic.com/news/fable-myt…
38
140
567
52,135
Jun 13
Jun 10
In 1968 five countries that already had nuclear weapons signed a treaty declaring them too dangerous for anyone else to build.
India refused, pointing out the treaty did not say nukes were too dangerous to exist, just too dangerous for new entrants.
Anthropic built Mythos, deemed it too powerful for public release, then shipped Fable with the same weights but hidden degradation on frontier AI work.
The restriction started the day after they finished building. Non proliferation was never about preventing danger. It was about preserving advantage.
Mythos 5 goes unrestricted to Microsoft, Nvidia, Google Cloud, AWS, and about 200 other approved partners. Fable 5 goes to everyone else with silent capability limits on frontier ML development.
The biggest paying customers get the full product. Potential competitors get a version that quietly gives worse answers on the work that matters most.
Anthropic filed confidentially for its IPO one week before this launch. India had a phrase for this kind of arrangement when it refused the NPT. Discriminatory by design.
Jensen Huang called the GPU to nuclear bomb comparison stupid. He is wrong about the analogy but right about the instinct behind it.
The NPT worked because nuclear weapons require enrichment facilities, centrifuges, and state level infrastructure. AI does not.
Qwen has 942 million downloads. DeepSeek V4 ships under MIT license with full weights matching closed frontier models.
The knowledge Anthropic is trying to restrict through hidden degradation is already open and available in competing models. You cannot run a non proliferation regime when the material is free to download.
Anthropic Fable 5 silently degrades its own performance when it detects someone building a competing model.
No warning, no refusal, just worse answers through hidden prompt tweaks and steering vectors
Meanwhile DeepSeek published its full R1 training pipeline, failure modes, RL schedules, everything, under MIT license. One lab is hoarding knowledge at the frontier. The other is giving it away.
The gap in approach is now wider than the gap in capability, Open is only threatening when you are slow.
Alibaba Qwen crossed 942 million downloads on Hugging Face by March 2026. Its share of new open weight derivatives went from 1% in January 2024 to 69% by February 2026.
Chinese models now account for 30% of global model usage on aggregator platforms, up from 1% in late 2024. All under Apache 2.0 or MIT licenses, fully permissive.
US frontier labs are spending $700 billion on capex while keeping the developmental knowledge locked. China is spending a fraction and giving the knowledge away. Adoption follows access, not origin.
Now China too going to do 230 billions capex as per report i think...
Fable 5 and Mythos 5 are the same model. Mythos goes to 200 approved partners. Fable goes to everyone else, with hidden capability limits on frontier ML work.
The stated reason is safety. The result is that US labs build the best tools and then weaken them for the work that advances AI.
DeepSeek V4 matches Opus 4.7 on agentic benchmarks and ships under MIT license with full weights. The question is not who builds the better model. It is who gets more people building with it.
Some of the Stuff I took from SemiAnalysis, But this will go Nuclear way I don't know
2
2
2,811
Jun 13
Now, Anthropic is a supply chain risk
As I told you before, the best models may eventually stay with the US defense ecosystem, and whatever is left could be given to other nations based on what they are allowed to use. Any day, access to models can be denied if they are used in an ulterior way.
The future becomes far scarier if things move in this direction. But it won't be so easy either, because there are many stakeholders sitting at the table. Just like you can't simply deny access to ASML technology holders, there are broader strategic and economic interests involved.
Still, as of now, I feel that in the future the most advanced models will primarily be reserved for US defense and strategic companies first, with everyone else getting access accordingly.
1
1
512
Jun 13
Apr 29
India doesn't need to lead the world in building the most advanced AI models. But it must lead in ensuring benefits of AI are widely shared.
@rvenk and I have an op-ed in The @EconomicTimes
economictimes.indiatimes.com…
3
1,584
Normal Guy retweeted
Nifty 50 PE ratio today is at 19.98 times which Below average range of its trading history.
Data from @stockscansin
20
27
472
43,177
Jun 12
Will give yu Review Soon, Something decent came this week & I Love Crime based stuff
3
14
2,146
Jun 12
Tomorrow, one of my favorite, I will teach you how to filter noise and add brick by brick to your dots so you can understand the bigger picture
The first original article on technical charts will be published tomorrow. With more than 10 charts and a trail-tested system, plus with one case study
it will show you how I filter opportunities in a certain way, along with the terms, conditions, and limitations.
Yup, it's one of the most important systems I use internally every week while going through 500 charts
Tomorrow, on Idealist :)
3
4
80
6,228
Jun 12
Physics does not run on GPUs. Bezos is separately raising $100B to buy manufacturing companies, apply Prometheus to their operations, and use their engineering data to train the model.
Acquired factories become both the customer and the training set at the same time. This is not an AI startup, it is an industrial conglomerate that runs on AI. Factory data is the scarce thing, not algorithms.
AI model can simulate faster but it still needs real world testing to validate. You cannot ship a jet engine by running inference.
He know where the chokepoint will come in next 5 year down the line
Jun 11
Jeff Bezos on CNBC explains revealed what Prometheus is building.
Today his new company Prometheus announced a $12B funding round at a valuation of $41B .
Prometheus trying to build an artificial general engineer that can help design and manufacture physical products like engines, medical devices, and electronics.
So the target areas are hard physical products like jet engines, chips, bridges, medical devices, consumer electronics, aerospace systems, vehicles, and drug design, where design cycles can take years because every idea has to survive physics, materials, cost, testing, and factory limits.
Bezos’ jet-engine example explains it well: asking for the same engine with 10% more thrust can become a 10-year engineering program, and Prometheus wants to shrink that “dream-build” cycle by 10x or more.
The $6.2B launch funding gave Prometheus a massive starting base, and the new raise says the company likely needs far more compute, talent, and industrial data before it can prove the product.
Their $41B valuation shows that frontier AI is becoming less a software race than a compute procurement race.
A company with no broadly shipped product can raise $12 billion at a $41 billion valuation because investors are not only funding a model, they are prepaying for the machines that might make the model possible.
The scarce asset is no longer just talent or algorithms, but clustered GPUs, power contracts, cooling, networking, and the operational skill to keep expensive silicon busy.
They are proof that demand is arriving faster than infrastructure can be built, and that every frontier funding round quietly turns into a future claim on power, racks, GPUs, and uptime.
3
12
68
7,410
Jun 12
China lost 30M factory jobs since 2013 even as manufacturing output kept rising, They installed 290,000 industrial robots in 2024 alone, more than rest of the world combined.
New high-tech export economy does not need workers the way the old one did. Exports grew 19% in May, retail sales rose 0.2%.
A country can be industrially dominant and economically hollow at the same time. China is building the best factories in the world and filling them with machines, not people.
China weak domestic demand is not a separate problem from its export strength. It is the same problem. Falling property values destroy household wealth, consumers stop spending, prices drop, the yuan weakens, and that makes exports cheaper.
Exports prop up GDP just enough that Beijing delays painful reforms like expanding social safety nets or stabilizing property.
The trap is that export success removes the pressure to fix the consumption problem. China ran $1.19 trillion trade surplus in 2025 while its own shoppers stayed home.
Barely 3 of China 129 EV brands turn a profit. Car factory utilization fell to 50%, lowest in a decade. Local governments keep subsidizing unprofitable firms because those factories are the only jobs left in their provinces.
The subsidies drain local budgets that should be building the social safety net that would give Chinese households confidence to spend. Which is now cut off
China is spending money keeping dead factories alive instead of giving workers an income floor to consume.
Subsidies protect factories but starve consumers, Lot of things changing in which Over production is gonna reduce & Subsidy going elsewhere plus due to tariff and trade barrier chinese businesses are looking for loopholes and Proxy nation
economist.com/leaders/2026/0…
11
14
49
6,159
Jun 12
Internet analogy gets this wrong, When computing got cheap, anyone could build a website because the protocol was open
SpaceX controls 82% of commercial launches and has used that position to pressure satellite rivals into sharing wireless spectrum with Starlink before agreeing to launch their hardware.
DOJ antitrust division has been briefed on these arrangements. Internet created trillions of value distributed across millions of companies. SpaceX cost curve created $1.75 trillion concentrated in one. That is vertical capture, not a platform.
SpaceX made $18.7B in revenue last year, Starlink was $11.4B of that, growing 50%. Launch services brought in $4.1B and grew just 8%. SpaceX is not building roads for everyone to drive on. It built roads so it could run its own trucks.
The company that lowered the cost of reaching orbit became the company that captures the value from being there. Of 165 Falcon 9 launches in 2025, the majority carried Starlink satellites. The road builder is now the trucking company.
$626B space economy is 40% satellite TV and telecom, 25% ground equipment like GPS chips in phones, 15% government budgets. Launch services are just 5% of the total.
In-orbit manufacturing and space computing, the applications that cheap launches are supposed to unlock, still contribute almost nothing in revenue.
Varda Space runs tiny capsules that process one drug compound per trip. The gap between what cheap access to space could enable and what it has enabled so far is enormous.
Cost alone does not create demand.
SpaceX also have data center infra, that's another cash rich business well focus on Space ETF
1
3
27
2,242
Jun 12
India entire CRDMO sector made $3.5B in FY26, WuXi alone did $6.3B and grew US revenue 34% last year while being the direct target of BIOSECURE.
They are now building a 1.74m sqft plant in Delaware opening this year, a 50 acre site in Singapore for 2027, and expanding in Switzerland.
The company getting sanctioned is quietly adding more non-China capacity than India top 3 firms put together. $700m annual gain is real, but WuXi moves really faster
India peptide CDMO space is worth $80m today, around 3% of the global market. WuXi already runs 100,000 liters of solid phase peptide synthesis capacity and is adding more.
Sai Life Sciences just opened a 50 person peptide research center in Hyderabad, which is where you start, not where you compete. GLP-1 drugs are the biggest manufacturing wave in pharma right now and India is still in the lab stage while China runs commercial plants.
Building capacity takes 3 to 4 years, restrictions land in 2028.
Markets have fully priced the China 1 transfer story into Indian CRDMO stocks. BIOSECURE restrictions do not legally bind until late 2027 at the earliest, and WuXi is opening its own US and Asian plants before that deadline arrives. Hope is already in the price.
But in this matter, Indian Pharma will do fantastic from Divi to sai life and all others
1
33
146
8,780