
Non-Linear Thinker,Futurist,#solopreneur,Crafting #AgenticAI & #GenAICoach,Speaker, Founder - https://xitter.gq/t.co/OiU7DUhEWy Chief AI & Tech Officer at https://xitter.gq/t.co/Pxt6Vb5Tsx
Mumbai & Kolkata
Joined May 2007
- Tweets80.6K
- Following5.4K
- Followers3.2K
- Likes46.3K
Pinned Tweet
Many might not understand but this one single tweet is a incredible demonstration of technology -
#AI in #Space 😱🤯
A prompt sent to @google TPU in space satellite and got a response 😎 – at Kolkata, India
0x_arjunghosh_ai retweeted
Google put an AI computer in outer space, and it has to shut itself off about every 15 minutes because it's overheating.
Outer space is around minus 455 degrees Fahrenheit, so that sounds backwards, but the emptiness that makes space so cold is the same thing cooking the chips.
Think about how your laptop stays cool. A little fan blows air over the chip and the air carries the heat away. Big data centers do the same thing on a massive scale, with water pipes running out to cooling towers. All of it works because there's air and water around to soak up the heat. Up in orbit there's basically nothing. A vacuum is one of the best insulators there is, which is the whole reason your thermos keeps coffee hot all day.
So the only way a satellite can get rid of heat is to glow it off.
Anything warm gives off infrared light, you just can't see it, and in space that faint glow is the one way out. A physicist named Josef Stefan worked out the math back in 1879. The heat a surface throws off goes up with the fourth power of its temperature, so if you double the absolute temperature it sheds 16 times as much. And while all this is happening, the sun is beating down on the other side of the satellite.
Google's test satellite is about the size of a fridge. It has four of Google's AI chips on board and solar panels that make around one kilowatt, roughly what your microwave pulls. Even on that little power, the chips can't run nonstop. They work for about 15 minutes, then shut off and wait for the radiators to catch up. The haiku Sundar posted came out of one of those runs.
Now scale that up to what a real AI data center runs on. One rack of Nvidia's GB200 servers pulls about 120 kilowatts, and pretty much all of it turns into heat. The International Space Station cools itself with six radiators, each about 76 feet long, and the whole system can dump 70 kilowatts.
So one rack of AI servers puts out almost twice the heat the entire space station was built to get rid of.
Google's long-range plan has around 80 satellites flying in a cluster about a kilometer wide, some only 100 to 200 meters apart, each one loaded with dozens of chips. A solar panel in the right orbit makes up to 8 times the power it would on the ground, and every one of those extra watts eventually becomes heat the satellite has to glow back out into space.
SpaceX has filed for a million satellites of its own, and Starship keeps pushing the price of a launch down. Cheaper launches mean you can afford to send up bigger radiators, and whoever builds the lightest radiator that dumps the most heat is going to own AI in orbit. For now, the satellite in Sundar's video still has to stop and cool down every 15 minutes or so.
First AI computing in space. Made by Google launched by SpaceX. What an incredible time to be alive.
0x_arjunghosh_ai retweeted
A ‘hello world’ milestone for Project Suncatcher 🚀
After launching our prototype satellite a few weeks ago, we’ve now successfully sent a command up to orbit, booted up the TPU, ran the prompt, and received our first complete response back on earth - a haiku!
Our TPUs just phoned home... from space.
In early October, we sent four TPUs into orbit to test whether our AI hardware could operate in space. The initial goal was to see if our chips could turn on, process a prompt, and transmit a response back to Earth.
This week, we got our answer.
0x_arjunghosh_ai retweeted
Fighting abuse & spam on a consumer app is a perpetual cat-and-mouse game:
Every solution you create is a bandaid for 3-12 months -- until the adversaries find a workaround. When money is involved, people are tenacious to circumvent your patch and keep the money flowing.
While large language models have made attacks more intractable, there are tools you can build to protect your product:
(1) An in-house Risk Scoring stack that assigns a 0-100 score of likelihood someone is a spammer or bot. This should be informed by machine learning, IP & email reputation data, device telemetry, among many other signals.
(2) A near-bulletproof method to "checkpoint" your highest risk users, such as liveness verification or phone number auth. This will have to be reinvented every year (e.g., CAPTCHA tests are obsolete now with today's models).
As new attacks emerge, you can move the risk threshold lower to checkpoint more people. However, you want to be cautious so you don't add unnecessary friction for too many people and impact usage.
0x_arjunghosh_ai retweeted
Andrej Karpathy’s 1-hour Stanford lecture on AI engineering is one of the best explanations I’ve seen of how AI systems actually work.
The progression is simple:
10% - LLM
30% - Prompt
50% - Agent
70% - Loop
100% - Graph
The key takeaway: AI engineering isn’t just about writing better prompts.
It’s about building systems around models, giving them context, memory, tools, feedback loops, and data flows.
“Delete everything, keep Graph.”
Definitely worth watching if you’re building with AI agents.
Watch and bookmark it.
0x_arjunghosh_ai retweeted
Chinese open-source models have peaked and will decline from here.
“I think the Chinese open-weight models are pretty much over in the US.
I think it is the beginning of the end.
DeepSeek is beginning to be banned across all federal agencies.” @jasonlk
Love to hear your thoughts @mattsheehan88 @deanwball @jordanschnyc @jjding99
Very few podcasts will actually make you smarter AND make you laugh.
No politics, no egos, just tech.
The biggest news this week:
- Anthropic Files S1 Ahead of $2TRN IPO
- Instinct's Rumoured $30BN Fundraise
- Microsoft Lays off 27,000 While Meta Taps Equity Markets for $95BN Capex Spending Spree
My notes with @jasonlk and @rodriscoll
1. Analysis of Anthropic S1
Anthropic's S1 beat expectations on margins and operating income, even after stock-based compensation, leaving the burden of proof on the naysayers. It also looked strong on customer concentration, compute efficiency and cash. Some questions remain, including the undisclosed Q3 revenue number.
2. Is the Value Extraction Requirement From the Latest Models a Benefit or a Curse of Anthropic?
Eighty percent of Anthropic's revenue comes from models less than six months old, and 70 to 80% of compute costs go to training. Models are expensive to build and highly profitable to serve, but only for six to twelve months. Few companies can write a $10 to 12BN compute check, so this short shelf life can favour Anthropic if it stays aggressive.
3. Brutal Reality Is Microsoft Have Just Admitted They Don't Want 27% of Their Workforce
Whatever gets said on LinkedIn, Microsoft's move says it does not want 27% of its people. That is a worse scarlet letter than a pure layoff, because the company has told these employees they do not meet its minimum bar.
4. Chinese Open-Source Models Have Peaked and Will Decline From Here
Chinese open-weight models are close to finished in the US, and this is the beginning of the end for them. Despite DeepSeek's latest raise, it is starting to be banned across all federal agencies.
5. Why No CIO Will Take a Chance on a Chinese Model
Even if confidential data reaches Alibaba servers in Asia only one time in 100,000, without consent, no CIO or CISO will approve the model. If it keeps happening and Alibaba stays silent, anyone using these models will lose their job.
6. Why OpenAI Is an Underappreciated Asset to Invest in Today
OpenAI is mispriced at the $1.25 TN round. Brief market panic has created a dislocation like Anthropic's round at around $183BN, even though OpenAI has closed all its competitive gaps.
7. What Is the Upside of OpenAI at $1.25 Trillion?
OpenAI and Anthropic are both expected to become $10 to 20TN companies. Anthropic takes only 0.1% of Fortune 500 budgets today, and even conservative estimates put that at 2 to 3%. Because OpenAI is not priced to perfection, the worst case is doubling the money.
(links in comments)
0x_arjunghosh_ai retweeted
there are way too many voice ai startups
how do i know which one is the best?
0x_arjunghosh_ai retweeted
Reminder... no one ever got rich by playing it safe.
Elon Musk had to borrow money from friends to keep Tesla alive in 2008.
Walt Disney borrowed against his own life insurance policy to help fund Disneyland.
Daymond John mortgaged his mom's house to fund FUBU.
Sara Blakely put her entire savings into Spanx.
0x_arjunghosh_ai retweeted
if you have an MCP, this is a HUGE day
Google just launched its Connected Apps Store
this means your MCP can now reach their 4.3 BILLION active users across both Gemini and AI Mode
this is yet another app store moment
not only can your product can be discovered and used by those people, without them ever having to leave AI Mode or Gemini
but Google is now ORGANICALLY surfacing these MCPs at the moment of intent after prompts
what a time to make your product agent-first with an MCP in every major store
if you want to track the ecosystem daily and see whose going live in your niche, go to: agentdiscoverability .com/track
Gemini will be live imminently...
Ofcourse ex government employees do not want this as they want safety of not doing job while other suffered, customer of bank!
IDBI Bank employees' forum opposes Fairfax stake sale, seeks parliamentary review
economictimes.indiatimes.com… – at Kolkata, India
0x_arjunghosh_ai retweeted
A founder friend, an early Blinkit exec, has freshly raised $3.7M from top VCs and founders to change how India eats: fresh ghar ka khana for every daily meal, from a nearby kitchen you trust like your own.
He’s hiring ex-founders with solid chops for the Founder's Office in Gurugram. If this sounds like an exciting next big bet for you - DM me.
He's also my go-to person for all things growth! Having played a key role in scaling Blinkit (then Grofers) from day 0, he knows what it takes to build at a 10x pace. I can vouch that you'll learn a tremendous amount working with him.
0x_arjunghosh_ai retweeted
Did you know Twitter links can reveal which device was used to share them?
Check the URL ending:
s=46 → iPhone
s=20 → Web Browser
s=19 → Android
Try it yourself by copying a tweet link.
Technology is interesting, isn't it?
0x_arjunghosh_ai retweeted
Replying to @arjunghosh @piyushbhargav22
@arjunghosh Saluti, here is your unroll: threadreaderapp.com/thread/2… Share this if you think it's interesting. 🤖
Sarvam is going to be an inference player? There's competition coming soon:
E.g. TogetherAI has committed to grow India DC to 250MW. Launching within 6-9 months. B200/300 from what I hear.
Within India options, do buyers care it's from S or just the best? We'll find out!
0x_arjunghosh_ai retweeted
An unintended consequence of consumer chatbots: they may slow the growth of knowledge on the Internet.
The old loop: people read forums, answered a few questions along the way, and authors earned social status from likes and replies.
Now readers consume it through a chatbot. Fewer chances to contribute, zero validation for authors, so fewer reasons to write.
The models are trained on that commons. If the commons dries up, the next generation of models has less to learn from.
Someone needs to build attribution and reward loops back in: credit, traffic or payment to the people whose answers get used.



























