deepak_kannan

PS: short answer is, big LLM modern architecture and hardware memory limitations. Microsoft etc are building millions tokens/sec at different scales ie. multiple high-speed chained datacenters to chip. Microsoft CEO Satya Nadella (@satyanadella) has posted about a 1.1M tokens/sec “datacenter rack”. This same tokens/sec question, funnily enough, can be asked of @bunjavascript. Why use Bun, when @ThePrimeagen style “blazing fast” options exist?, why even use Javascript? It is not as if, AI labs care primarily about efficiency, they use slower Python infra, which to be fair, runs on faster C++ libraries. Do software shops & labs care about “faster Javascript”, slower compilation cycles, or slow CI builds? But the real reason is of culture, I think. People are betting the horse race on AI getting faster and better, like Microsoft bet on CPUs exponentially rising under Moore’s Law. One AI lab had a quip that they do not care about “software debt” because they are betting on AGI fixing their codebases, in the future! Historical LLMs and transformers are one of the approaches winning out. And on top of that, AI labs have decided that it is a good idea to have, since approx ChatGPT-3, even larger-and-larger AI models. The shift to larger LLMs since ChatGPT-3 reflects a belief in emergent capabilities outweighing efficiency, much like early software ignored optimization for raw Moore’s law backed compute gains. The very large LLM architecture itself seems to be an historical accident. Same with focusing on language for AI. And another cultural fact is that everyone wants the problem to be shifted downstream or upstream, eg. is the current LLM architecture the best? Do we have the electricity to run these huge AI datacenters? etc. It’s ignorance-abstraction turtles all the way down, where each tier offloads complexity upstream or downstream. Something similar to how UI design does not need ergonomics expertise, or databases can be coded without being an expert in optimisation maths. But this modern model works very well. Even though vintage programmers were betters on tiny chips, without the internet, and the modern learning ecosystem. Or I can write this tweet without being an AI researcher. I am putting words in other peoples mouth, but George Hotz (@realGeorgeHotz) and Jim Keller (@jimkxa) believe it be to a software/compiler problem, along with hardware obviously (not patronizing). Lastly high tokens/sec is a memory chip issue, of handling very large LLMs. - rising costs - memory capacity & bandwidth - top speed and future growth - bus and memory interconnects - legacy sockets’ large trace lengths - ECC errors rate Memory (and HDD) has not kept pace with Moore’s Law. Manufacturing memory chips also suffers from production issues and competition from GPUs. Which people are trying to solve in-part and in-combination by: - finding better algorithms (eg FlashAttention) for present architectures. - using and finding, new better derivative architectures (eg. Mixture of Experts MoE). - using symbolic maths, tool calling etc rather than pure LLMs for AI. - using small LLMs (SLMs) - experimentally lobotomise LLMs by trying to remove superfluous weights, other techniques like quant. distillation, etc to reduce a model (sidestepping the memory problem), using routers and software layers to eg. serve simpler questions using smaller models. Although the “ultimate argument of kings” answer was by George Hotz @realGeorgeHotz (but useless in the short term) on the neural capacity and electrical efficiency of the human brain. The human brain, with ~86 billion neurons at ~20W, achieves exponentially higher efficiency than today’s 100kW+ datacenters, hinting at a biological upper bound we’re epochs away from matching. George Hotz has the actual physics and speed of light answer ie. light-speed signal propagation and thermodynamic efficiency. Physics is the law, everything else is a recommendation. — Elon Musk (@elonmusk)

@jukan05 Memory makers should just create an OPEC (oil) like body at this point!

Samsung and SK Hynix to Cut, Rather Than Increase, NAND Flash Production This Year… “Maximizing Profits in Super Boom Period” Samsung Electronics and SK Hynix, which together hold more than 60% of the global NAND flash market, are expected to reduce their NAND flash production this year. As competition in inference-type artificial intelligence (AI), led by Nvidia, intensifies, tight supply of NAND flash—a key component—is likely to drive price increases across servers, PCs, mobiles, and other sectors. Analysts say this will be a major positive factor for improving the operating margins of Samsung Electronics and SK Hynix, on par with DRAM. According to data from market research firm Omdia obtained by Chosun Biz on the 20th, Samsung Electronics has slightly lowered its NAND wafer production from 4.9 million wafers last year to 4.68 million this year. This is a further reduction from the production cuts implemented last year due to a sharp drop in NAND profitability in 2024. SK Hynix is also projected to follow a similar path, reducing NAND production from around 1.9 million wafers last year to 1.7 million this year. Given that the NAND market this year is seeing strong AI-driven demand, supply adjustments by the leading suppliers—Samsung Electronics and SK Hynix—are highly likely to exacerbate shortages not only in AI servers but also in mobiles and PCs. According to Citigroup, Nvidia’s next-generation AI accelerator “Vera Rubin,” set for mass production in the second half of this year, will feature solid-state drives (SSDs) with a capacity of 1,152 terabytes (TB)—more than 10 times that of the current “Blackwell” product. With Vera Rubin shipments projected at 30,000 units this year and 100,000 units next year, this translates to new demand of 34.6 million TB in 2026 and 115.2 million TB in 2027. The reduction in NAND production by Samsung Electronics and SK Hynix is largely attributed to lower investment priority for NAND facilities compared to DRAM, which currently shows the highest profitability. Additionally, as demand grows for high-capacity SSDs in AI data centers, the transition to quad-level cell (QLC) technology inevitably leads to natural production losses. Switching from existing triple-level cell (TLC) technology to QLC, which is better suited for AI data centers, involves various factors such as facility setup, stabilization periods, and initial yield rates. Management at Samsung Electronics and SK Hynix reportedly see no reason to aggressively increase NAND production. For years, NAND has suffered from poor profitability, requiring both companies to focus on price defense. Now, in this memory super-boom cycle, they aim to maximize profits as much as possible. An industry insider commented, “Whether the NAND production cuts by Samsung Electronics and SK Hynix are intentional or natural, the benefits from reduced supply will be greatest this year.” Some analysts suggest the moves are also a response to growing supply of commodity NAND from China. Unlike Samsung and SK Hynix, China’s YMTC has been increasing its presence and output in the NAND market since last year. This is interpreted as a strategy to counter low-price competition from China by reducing supply in mobile and PC segments to protect profitability while increasing higher-margin server and enterprise volumes to manage the product mix. Major market research firms are forecasting a strong NAND price uptrend starting in the first quarter and closely monitoring production discipline by key suppliers. TrendForce predicts NAND flash contract prices will rise 33–38% quarter-on-quarter in Q1, noting that Samsung Electronics and SK Hynix are maintaining a conservative production stance. IDC also expects NAND supply growth this year to be around 17%—lower than the average of recent years. $SNDK

Today I learnt that in 2009, neuroscientists placed a dead Atlantic salmon into an fMRI scanner, scanned it, and that this has apparently implications for AI interpretability. 🐟 They showed the dead fish pictures of humans in social situations and "asked" the fish to determine the emotions of the people. When they ran their standard statistical software, the results showed "brain activity" in the fish that correlated with the emotions. Obviously, the fish was not thinking; the "activity" was just random noise. The point of the study was to show that if you don't correct for statistical noise and use rigorous controls, your tools will find patterns where none exist. This paper claims that the same lesson should be applied in interpretability work: many researchers use various tools to explain what is happening inside a neural network (e.g. probes, SAEs etc). But some of these convincing-looking explanations can also be extracted when applied to randomly initialized and untrained AI models (the dead salmon equivalent): saliency maps remain plausible after weight randomization, sparse autoencoders find interpretable components in random transformers etc. The authors propose that we stop treating interpretability as "storytelling" and start treating it as statistical inference: doing null hypothesis testing, quantifying uncertainty more systematically, interpreting explanations as a simplified surrogate model etc. Although they also acknowledge that finding some signal in random networks doesn't automatically invalidate finding stronger signals in trained ones. I'm not interpretability researcher myself but would be curious to hear takes! arxiv.org/abs/2512.18792

Nvidia’s biggest fear is hyperscalers’ own chips – I would speculate that Nvidia didn’t want to risk one of its big hyperscale customers acquiring Groq (10/x)

Nvidia is certainly selling something. And analysts want to spice up their report with poetry. 5/5

It’s more like a few oil barons funding their own drilling ops while the toolmaker (Nvidia) gets paid upfront and hedges the downside. Makes for great headlines, less so for a clean historical parallel. 4/5

Seems like a dumb way to put it! Who was buying the GPU shovels before? And where are the individual prospectors in this GPU gold rush?, none! Who writes these Bank of America (BofA) reports, AI! (sarcasm obviously). 1/5

#Live at 8 PM IST Venezuela’s Heavy Investment in Arms Couldn’t Stop the US — Who Is to Blame? • #TheGunnersSight Lt Gen Ravi Shankar (R), Lt Gen Raj Shukla (R), and Lt Gen Harimohan Iyer (R) @palepurshankar @Gen_RajShukla - youtu.be/s7cu4tm8TIw

Yin-Yang of global economy: China exports deflation (overcapacity + dumping + weak Chinese consumption); USA exports inflation (reserve-currency dollar hegemony & oversupply + deficits). Push-pull.

Whatever happened to Chinese chip designer Chen Jin of Hanxin?

Wonderful to see that INSV Kaundinya is embarking on her maiden voyage from Porbandar to Muscat, Oman. Built using the ancient Indian stitched-ship technique, this ship highlights India's rich maritime traditions. I congratulate the designers, artisans, shipbuilders and the Indian Navy for their dedicated efforts in bringing this unique vessel to life. My best wishes to the crew for a safe and memorable journey, as they retrace our historic links with the Gulf region and beyond. @INSVKaundinya

Delighted to receive this picture from the team of INSV Kaundinya! Heartening to see their enthusiasm. As we are all set to usher in 2026, my special greetings to the INSV Kaundinya team, which is on the high seas. May the rest of their journey also be full of joy and success. @INSVKaundinya

World's first transnational corporation was formed in South India in 9th century CE! Before the East India Company, before the Hanseatic League, even before Roman trade guilds matured into semi-formal chambers, the world’s first true corporations emerged—not in the West, but in the heart of South India. Their names echo through copperplate inscriptions and temple walls: Ayyavole 500, Manigramam, Nanadesi. These weren’t mere caravans or merchant clans. They were institutional bodies with rules, councils, seals, and sovereignty—what we would today call transnational corporations. The Ayyavole 500, based in the city of Aihole in Karnataka, declared themselves protectors of dharma and trade alike. They operated across borders—from Tamilakam to Sri Lanka to Southeast Asia—governing their own members, settling disputes, and building temples with their surplus wealth. Like modern corporations, they had identifiable seals and legal personhood. They could sign contracts, give donations, and be held accountable. In fact, their agreements—etched in stone or inscribed on copper—read like the memoranda of incorporation. The language was different; the structure, strikingly familiar. As stated in one 10th-century inscription from Tamil Nadu: "We, the lords of Ayyavole, protect the path of trade, remove obstacles, maintain tanks, and construct roads." A fragmentary Chera inscription datable to 1000 CE in the reign of Bhaskara Ravi, found on three broken stones in a mosque in Pandalayini-Kollam (near Kozhikode), refers to Valanjiyar and other merchants found in the assembly of Ayyavole-500 trade guild. The Ayyavole-500 were simply called Ainuruvar in Tamil. In Visakhapatnam, three inscriptions were found, two in Telugu and one in Tamil. They were on the Ainuttava-perumballi (500 perumballi) in Visakhapatnam and dated to 1090 CE under the reign of the Ganga king Anantavarmadeva. Another Telugu inscription records a similar grant to the same Ainuttava-Perumballi by the Chief Mahamandaleshvara Kulotungga Prithvisvara. What made the Ayyavole 500 formidable was not just their internal cohesion but their capacity to construct a seamless, transregional web of trust. From their heartland in Aihole, they spread through key Chola ports like Nagapattinam, Tamil sacred cities like Kanchipuram and Thanjavur, and political capitals like Gangaikonda Cholapuram. These cities were more than waypoints—they were secure nodes in a network built on temple infrastructure and royal patronage. Temples such as Brihadeeswarar in Thanjavur and the Shiva temple in Kanchipuram served as treasury vaults, legal arbitration halls, and social registrars for merchant records. Donations were recorded not only as acts of piety but as instruments of financial trust. Temples were equipped to issue receipts, hold reserves, and maintain accounts with meticulous precision. To the north, their caravans reached Vengi and in the south even across the Bay of Bengal to Lanka. Beyond lay the outer sphere of their influence, facilitated by Chola naval expeditions—Kalinga, Odra, and Pala territories in the Indian subcontinent; and overseas to Pagan, Pegu, and Burmese kingdoms. Their ships, often protected under Chola naval wings, landed at Chenla (Khmer), Champa, Cahaya, and Kadararm (modern Kedah). Further south, Pana and the imperial court of Srivijaya became trading partners, as evidenced by copperplates found in these zones bearing Ayyavole merchant seals. What sets these merchant guilds apart is the seamless fusion of spirituality, commerce, and civic life. Far from the profit-driven logic of today's boardrooms, these early corporations upheld a moral economy. Ethical codes were enforced: no cheating, no false weights, no exploitation. Trust, not just capital, was their currency. The temple acted as both moral arbiter and financial trustee, holding deposits, endowments, and even dispute resolutions. Merchants stored their wealth in temple treasuries not just for safety but to align with divine order. Their donations weren’t charity—they were strategic investments in social infrastructure, building roads, tanks, inns, and granaries. The temple was the first community bank, the sanctum a notary office, the hundi an IOU that circulated across kingdoms. Their reach was breathtaking. From the port of Nagapattinam, they shipped textiles and spices to Java and Sumatra. They had representation in Angkor and even had trade protections under foreign kings. Copper plates found in Southeast Asia bear their names and marks, proving their multinational presence centuries before Dutch or Portuguese galleons circled the Cape. As noted in "The Ancient South Indian Merchant Guilds and the Modern Corporation" (Arthik Charche, 2021), these guilds exhibited all the hallmarks of modern enterprise: asset management, long-distance contracting, and intergenerational capital transfer. And they lasted. Not for decades, but for centuries. Through dynastic changes and invasions, these guilds endured—because they weren’t built on coercion, but on consensus, credibility, and the dharmic obligations of trade. Their decline came not through internal failure but through the slow grinding gears of colonial disruption. The tragedy is not that the world forgot them. It is that India herself did. In our textbooks, we study limited liability from 17th-century London, ignoring 9th-century Tamil Nadu where merchants had far greater legal protections and obligations. We dissect Wall Street’s governance reforms, while copperplates from Chola-era Thanjavur outline ethical audit procedures and trade arbitration. For instance, a 12th-century Manigramam charter specifies: "Any guild member found falsifying accounts shall be fined and his name removed from the records." These ancient Indian corporations did not chase monopoly or profit. They embodied a different imagination: one where trade was sacred, accountability was communal, and prosperity was shared. In remembering them, we recover not just a lost history, but a lost possibility—for an economy anchored not just in balance sheets, but in balance. Reference: Champakalakshmi, R. (1996). Trade, Ideology and Urbanization: South India 300 BC to AD 1300. Oxford University Press. Yazdani, G. (1960). Early History of the Deccan. Oxford University Press. Subbarayalu, Y. (2008). South India under the Cholas. Oxford University Press. "The Ancient South Indian Merchant Guilds and the Modern Corporation," Arthik Charche, ISEC Bangalore, 2021. Narayanan, M. G. S. (1996). Perumals of Kerala. CosmoBooks. Two medieval merchant guilds of south India (South Asian studies / Heidelberg University. South Asia Institute. Meera Abaraham, 1988 Map as usual is open source at kaggle.com/code/shreep/va… I will be putting all the links for the publicly available texts in the reply here next week.

So what is life like aboard @INSVKaundinya ? Modern amenities are minimal except for safety and communications. This is how is looks under the deck. As you can see, there are no cabins and just a dark hold for storing supplies. Mostly we sleep on the open deck with sleeping bags, but in rough weather we will have to take turns to sleep here between the boxes. 1/n

I wonder how the night sky looked when the Vedas were composed, no smog, no lights. Just Akasaganga, the Heavenly Ganges, a radiant river of stars bright enough to cast shadows. We’ve lost that wonder, we never look up anymore at the sky, and wonder.

A chicken in every pot and a robot to cook it for you?

this is very interesting to me my parents (& many Indian parents) used to make me rinse out my nose with salt water when I was sick. I hated it so I asked our family doctor, who said it was a bad idea and would just irritate the sinuses Turns out my parents were right

They are talking about the ancient Indian yogic practice of Neti which has already been scientifically proven to be effective in preventing colds, flus, reducing allergies & asthma by Indian scientists FIRST. It is now researched by western teams to again prove its effectiveness. This is as usual - Yogic and Ayurvedic remedies and techniques are always first researched by Indian scientists. But Indians themselves will ignorantly think and claim we do not do such research, so western scientists deserve the credit. Be it Turmeric or Ashawaganda, Yogic breathing techniques or poses, the research is first done in India and whatever effectiveness proved. Only then, seeing those papers, the world takes over and reexamines evidence. It would be wise to stop assuming and claiming ignorantly that Indians aren't scientifically researching our own traditional knowledge. So somehow make it deserving of others to appropriate it. And, as usual, most of the articles and studies on this Neti process now doesn't attribute it to ancient Indian practice in Yoga and Ayurveda. Slowly the name is also being changed from Neti to Nasal irrigation or Nasal cleanse. This is how cultural appropriation works. Neti is one of the six purification practices mentioned in Hatha Yoga Pradipika to cleanse the body of impurities, stay healthy, and prepare the body for Hatha yoga. But soon it will be known to everyone as just Nasal cleanse and Indians themselves will claim "but this is not Neti, it was researched and invented in a lab in Stanford," and say "only a special Greek kosher salt will be good to be used mixed in Evian water in a special Chinese ceramic pot for Nathan's Nasal cleanse."

🚨 #BreakingNews: Islamists are running a special survey to detect "Indian citizen" within Bangladeshi press community. In their perception, anyone criticizing Islamists, jihadists and Pakistan are Indian nationals.