Deepth Dinesan

Maths !

@NandanNilekani @rvenk @EconomicTimes India must not build nuclear reactors, instead must consume electricity. India must not make electric engines for cars, instead should consume it. India must not invest in R&D, just like Infosys, and instead should consume other's research. India must not be like China, not lead

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/opinion/et-com…

@policytensor You got a small dick, don't you?

The rest of our DM.

It is highly probable that @pati_marins64 is not who she says she is. That does not mean that she is an intelligence operator working for an adversary. But what seems highly probable is that she is lying about her identity. This requires serious investigation. I was alerted to the potential fraud by a reader (first screenshot). They said that her Portuguese is suspiciously poor. I had already been alerted by her refusal to join any audio or video discussion, claiming that a health condition made it impossible for her to join spaces. I shared the reader's concerns about her identity. She refused to provide any proof at all. I asked her for a quick video call. She refused. I asked her for a link to any appearance of her anywhere. She refused. Happy to be proven wrong. But this deserves investigation for all the obvious reasons. @dwallacewells, @jeremyscahill, @MazMHussain, @MJMBlackburn81.

@alexandr_wang Where is the Technical Report?

8/ read more about our research and muse spark here: ai.meta.com/blog/introduci…

1/ today we're releasing muse spark, the first model from MSL. nine months ago we rebuilt our ai stack from scratch. new infrastructure, new architecture, new data pipelines. muse spark is the result of that work, and now it powers meta ai. 🧵

@casper_hansen_ Doesn't it need multi-objective hyper parameter optimization? ai21.com/blog/scaling-v…

every inference engine should have a section in their docs with exact commands to achieve best possible tokens/s on the most popular models i'm told kimi k2.5 can run at 300 tokens/s on B200s if you run nvfp4 with speculative decoding in open-source

@shanaka86 Spielberg is also arriving to shoot his next movie "Saving Private Marines"

BREAKING: Nearly five thousand United States Marines are converging on the Persian Gulf. USS Tripoli, an America-class amphibious assault ship carrying 2,200 Marines from the 31st MEU with F-35Bs and Ospreys, transited the Malacca Strait on March 17 and is now crossing the Indian Ocean toward the Gulf of Oman. USS Boxer, a Wasp-class carrying 2,500 Marines from the 11th MEU, departed San Diego on March 18 under accelerated orders with USS Portland and USS Comstock. Their destination has not been confirmed. Their capability is not ambiguous. Kharg Island handles 90 percent of Iran’s oil exports. The military sites have been struck. The oil terminals have been deliberately spared. The Marines are not coming to destroy Kharg. They are coming to hold it. The problem is the water between them and the island. Hormuz is mined. IRGC vessels laid ordnance before they were sunk. Those mines remain on the seabed. CENTCOM destroyed the mine-layers and cleared coastal batteries, but clearing mines requires dedicated sweepers in waters where suicide drones and submersible torpedo boats still operate. Two littoral combat ships, USS Tulsa and USS Santa Barbara, have docked in Singapore and are likely heading to Hormuz for mine countermeasure duty. The insurers who fled on March 5 could not price the threat density. The Marines are heading into the density the actuaries could not model. The strategic path is layered. Carrier aircraft and Tomahawks suppress remaining coastal positions. Mine countermeasure vessels clear a corridor. This is the slowest and most dangerous phase. A single mine strike on a transport could produce casualties that reshape the domestic political calculus overnight. Then Ospreys launch from Tripoli carrying Marines in vertical assault directly onto Kharg, bypassing the mined waters entirely. F-35Bs from both ships provide close air support. Boxer’s well deck deploys landing craft with heavy equipment once the perimeter is secured. Vertical insertion is the key. Tripoli was built for this. The America-class sacrificed its well deck to carry more aircraft. It is an aviation-first assault ship. The Ospreys fly over the mines. The F-35Bs fly over the drones. The initial force arrives by air while the surface echelon follows through cleared corridors. You do not drive through the minefield. You fly over it and secure the objective from above while engineers clear the water below. The risk is the Iranian response. Kharg sits 25 kilometres off the Iranian coast. Every remaining IRGC asset can be directed at the island once Marines are on it. Drones. Ballistic missiles from the Yazd tunnels that are still launching. Submersible torpedo boats. Speedboats loaded with explosives. The Marines who land will be holding an island within the engagement envelope of a military that hit Arad and Dimona from 2,000 kilometres away. Kharg is 25 kilometres from the launchers. But the leverage is worth the mathematics. Holding Kharg does not require destroying it. It requires standing on it. The Marines become a bargaining chip no diplomatic cable can replicate: reopen Hormuz or lose 90 percent of your oil revenue permanently. The island is the card. The Marines are the hand that holds it. The strait is 21 miles wide. The mines are on the seabed. The Ospreys are warming up. And 4,700 Marines are heading toward an island that Iran cannot afford to lose and America cannot afford to fail on. Full analysis: open.substack.com/pub/shanakaans…

@jordan_ross_8F OPENCLAW

The founders who figure out OpenClaw in the next 90 days are going to look like geniuses in 2027. The problem is most agency owners don't have time to figure out the install, the security risks, where to start, or what to actually hand it first. So my team built a 48-page beginner's guide that does it for you. Inside: — The exact prompts to hand it on day one — Plain English setup for Mac and Windows — How to secure it so it doesn't burn your business down — 42 copy-paste workflows across sales, marketing, ops, and finance Your competitors are sleeping on this. Comment OPENCLAW and I'll send it.

Kerala is one of the smallest states in India in terms of area. Malayalam-speakers make only a small fraction of India’s overall population. Despite this, we punch above our weight in several fields, one of which is Modern Medicine. 1/26

@dwarkesh_sp @dylan522p 1GW datacenter build = $50 Billion? Anythropic's revenue = $10 Billion? I still don't get where the pay off from $5T AI datacenter buildout will come from - any ideas?

.@dylan522p thinks Anthropic will get to 5–6 GW by year end. But at a far higher cost than it would have had to pay if it had gone crazy on compute commitments early like OpenAI did. Ant will get the compute either through cloud partners like Bedrock and Vertex who take ~50% gross margin on revenue, or shorter-term deals where rates have climbed well above what long-term contracts locked in a year or two ago cost. They’ll probably also have to destroy demand by pushing up model prices.

@dwarkesh_sp Really? ASML has 10,000 suppliers for the EUV Lithography equipment?

EUV machines are the most complicated tools humans make. Their supply chain has over 10,000 individual suppliers, and any one of them not scaling fast enough can bottleneck the entire AI industry. An EUV tool fires lasers at a tiny tin droplet three times in precise sequence, blasting it hard enough to emit EUV light. That light bounces off 18 multilayer mirrors onto the wafer. Meanwhile, the two platforms inside the machine - one holding the stencil, one holding the chip - are flying back and forth at 9Gs in opposite directions. The successive passes have to land on top of each other to within 3 nanometers. If any part of this is off, yield goes to zero. Take just one component. The mirrors are mostly supplied by Carl Zeiss, who have probably fewer than a thousand people working on them. In turn, Carl Zeiss rely on machines from Switzerland to deposit each of the layers, and use a coating process co-developed with a different German company. None of these companies have woken up. They’re gradually increasing production, but nowhere near the levels necessary for what the labs want by the end of the decade. @dylan522p predicts production can't scale beyond about 100 EUV machines per year by 2030, no matter how much money gets thrown at the problem. In the medium term this is the key bottleneck on scaling.

@imetatronink @TerryMichaelTX 😁👍

@TerryMichaelTX So your assessment is that the Iranians are unquestionably inferior in comparison to their mighty American adversaries? The objective evidence disputes your conclusion.

🔥 Iranian "359" Loitering Missile The smaller "358" has shot down many US drones in the past 10 days, but we have yet to see the "359". In the hands of a competent platoon in the Iraqi desert, the "359" could pose a serious threat to AWACS and tankers. deepnewz.com/middle-east/ir…

@caitoz Why does Australia support Epstein Coalition against Iran?

Give me some questions and comments you’d like me to address in a podcast today.

@gdb Would be great if you guys publish details of h/w s/w co-design. How does the gb200 architecture assist model architecture and vice versa. Would COP/optical backplane better for MoE,RL. Where does rail opt-topology fail?

gb200 has really been enabling us to do some amazing things

@lmsysorg @modal @Alibaba_Qwen hard to join in person from Australia ;)

@deepth_d @modal @Alibaba_Qwen Unfortunately this is not recorded 🥲 but you’re very welcome to join in person to catch the vibe and meet great people next time 🙂

Great discussion last Thursday on High-Performance Inference for Open LLMs with @Modal, @Alibaba_Qwen, and SGLang🔥 More than 300 founders, AI engineers, and ML systems researchers had registered for the event. Key takeaways from the talks: 💻Cold start really matters: infrastructure choices directly shape user experience ⚡️Open models scale fast when infra keeps up — tight collaboration between model builders and inference engines is what unlocks SOTA performance at much lower cost 🧠 Our core dev speaker, Qiaolin Yu, walked through how SGLang scales open models to enterprise levels, covering techniques such as prefill–decode disaggregation, expert parallelism, speculative decoding, and rack-scale optimization. Big thanks as well to @charles_irl for the excellent live demo showing how to serve SGLang with low latency on Modal 🔥 More conversations like this coming soon — the future of inference is open 🚀

@INCIndia aijac.org.au/update/modi-an… Modi is a puppet of Israel, explains why he is so anti-muslim.

Extremely shameful!! Narendra Modi’s name has surfaced in the Epstein files. America’s serial rapist, child sex offender, and human trafficker Jeffrey Epstein wrote in an email on July 9, 2017: 👇 “The Indian Prime Minister Modi took advice and danced and sang in Israel for the benefit of the US President. They had met a few weeks ago. IT WORKED!” Epstein clearly states that Modi took advice from him and went to Israel, where he danced and sang for the benefit of the US President. He also said that it worked. 👉PM Modi was on a visit to Israel from July 4 to 6, 2017. Epstein wrote this email just three days later. 👉Just before the Israel visit, on June 25–26, 2017, Modi had met US President Donald Trump in America. 👉 Connecting the dots in Jeffrey Epstein’s emails reveals that Modi went to America in June 2017 and took advice from Epstein there. 👉 One week later (July 4 to 6, 2017), Modi arrived in Israel and, as per that advice, danced and sang there, and it WORKED! Now it is clear that Prime Minister Modi has a very deep and long-standing connection with Jeffrey Epstein, which is shameful for India. This is a matter of national dignity and international reputation, for which Prime Minister Modi must answer. ❓What kind of advice was Narendra Modi taking from Jeffrey Epstein? ❓Why was Modi dancing and singing in Israel? How did Trump benefit from it? ❓ Epstein wrote, “IT WORKED!” So what does that mean? PM Modi, the nation wants to know: What is your relationship with Jeffrey Epstein? 👉Link to the document on US Department of Justice Website: justice.gov/epstein/files/…

wrote some thoughts about scaling models and hardware! apologies if it’s a bit dense but I talk about test time scaling, llm vs diffusion inference, and relate it to current and future hw.

@chrisbarber in Enterprise Labs examples in South Korea - KT ( korea telecom), SKT T-Brain team, LG CNS, Samsung AI research. In Japan PFN( Preferred Networks), PFN develops models, builds chips, datacenters, and also products and services

@chrisbarber NeoLabs could be Agent Labs( harness,scaffolding around models, but not developing models) , Model Labs( developing Models, specialized intelligence instead of general), in Enterprise AI Labs( AI Labs within large enterprises who build models, build products and services too)

what are the main sub-groups at the big labs seems like meta = tbd, msl, fair gdm = deepmind, gemini openai = pre-training, everyone else anthropic & xai don't have major sub groups?

@chrisbarber What is the difference between NeoLabs and Frontier Labs? Maincode in Australia

i asked people for suggestions of neolabs. here's the list, which am i missing? - ssi - prometheus - thinking machines - manifest ai - vmax - goodfire (interp focused) - reflection (enterprise focused, US open models) - humans& - periodic - ai2 - lila sciences - richard socher new lab - another stealth new lab - mistral - morph labs - adaption labs - ami labs - world labs - induction labs - moonlake ai - poolside (enterprise focused) let's say neolab = the success/failure of their business primarily depends on how good they are at training models, and they're not a frontier lab yet