ashu garg retweeted
IIT is MIT on steroids.
No other school matches the rate IIT graduates become founders, CEOs, and CTOs.
The entrance exam is one of the hardest math, physics, and chemistry tests in the world.
Over 1.4 million people take it every year, and most score zero. Roughly 1% of the t get in. The average MIT student wouldn’t get into IIT.
The test measures 2 things: extreme intellect and extreme grit. Performing at the top level requires years of disciplined preparation. Raw academic talent alone isn’t enough.
That’s why IIT alumni punch so far above their weight, especially in startups. A material number of them end up in Silicon Valley as founders, CEOs, CTOs, and VPs of engineering.
- Sundar Pichai (IIT Kharagpur) → CEO of Alphabet
- Vinod Khosla (IIT Delhi) → Co-founder of Sun Microsystems
- Aravind Srinivas (IIT Madras) → Founder of Perplexity
And in the founders I’ve backed too - @ashutoshg and @mohitaron studied at IIT Delhi and @krishnanvijay studied at IIT Bombay.
I try to meet as many IIT founders as possible. When a founder has passed IIT’s filter, I already know two things: top 1% intellect, top 1% grit.
No other school matches the rate IIT graduates become founders, CEOs, and CTOs.
The entrance exam is one of the hardest math, physics, and chemistry tests in the world.
Over 1.4 million people take it every year, and most score zero. Roughly 1% of the t get in. The average MIT student wouldn’t get into IIT.
The test measures 2 things: extreme intellect and extreme grit. Performing at the top level requires years of disciplined preparation. Raw academic talent alone isn’t enough.
That’s why IIT alumni punch so far above their weight, especially in startups. A material number of them end up in Silicon Valley as founders, CEOs, CTOs, and VPs of engineering.
- Sundar Pichai (IIT Kharagpur) → CEO of Alphabet
- Vinod Khosla (IIT Delhi) → Co-founder of Sun Microsystems
- Aravind Srinivas (IIT Madras) → Founder of Perplexity
And in the founders I’ve backed too - @ashutoshg and @mohitaron studied at IIT Delhi and @krishnanvijay studied at IIT Bombay.
I try to meet as many IIT founders as possible. When a founder has passed IIT’s filter, I already know two things: top 1% intellect, top 1% grit.
14
5
105
11,226
ashu garg retweeted
@JayaGup10 and @ashugarg make an important argument here: agents create an expanding execution workload around inference.
My view: this changes the architecture of the cloud itself.
An agent’s task can outlive its process, its model connection, and the person who started it. The infrastructure has to preserve progress, authority, and context across all three.
Durable execution goes deeper than keeping a sandbox alive. An agent needs to pause, checkpoint, fork, and resume. But restoring a microVM doesn’t establish whether an external transaction committed. If a tool times out after submitting a payment, blindly replaying the step can duplicate the payment.
You need recorded step outcomes, idempotency where supported, and reconciliation when the result is uncertain. Recovery semantics become part of the agent architecture.
Model routing and execution placement are coupled.
The cheapest model endpoint can become expensive if switching to it requires resending a long history, rebuilding a warm KV cache, or moving context across regions.
A router should consider model capability, queue time, cache state, data location, and the probability of completing the task. A locally cheaper decision can make the whole trajectory more expensive.
That’s why inference, runtime, and data placement need to be designed together.
A sandbox isolates code. Every tool invocation also needs to establish whose authority the agent is exercising, which action is permitted, and whether approval is required.
Broker credentials at invocation, enforce policy outside the model, and connect the action to the session’s audit trail. Forking an agent should never silently expand its permissions.
Production traces should feed a tested improvement loop.
Connect model calls, tool responses, state changes, human corrections, and outcomes in one trace. Turn recurring failures into evaluation cases. Replay candidate fixes in simulations before changing production behavior.
That gives execution history a second purpose: evidence for improving the next run.
The economic objective follows directly: minimize total cost per successful task, including inference, active compute, persistent state, failed attempts, and recovery.
More capable models can support longer tasks and more parallel work. Falling token prices can therefore expand demand for the infrastructure surrounding each call.
At @DigitalOcean, we’re building around that entire loop: think, act, remember, improve. Open harnesses and models, governed execution, durable state, and customers retaining control of their accumulated intelligence.
The agent’s full lifecycle is becoming the unit of AI Native cloud's architecture.
My view: this changes the architecture of the cloud itself.
An agent’s task can outlive its process, its model connection, and the person who started it. The infrastructure has to preserve progress, authority, and context across all three.
Durable execution goes deeper than keeping a sandbox alive. An agent needs to pause, checkpoint, fork, and resume. But restoring a microVM doesn’t establish whether an external transaction committed. If a tool times out after submitting a payment, blindly replaying the step can duplicate the payment.
You need recorded step outcomes, idempotency where supported, and reconciliation when the result is uncertain. Recovery semantics become part of the agent architecture.
Model routing and execution placement are coupled.
The cheapest model endpoint can become expensive if switching to it requires resending a long history, rebuilding a warm KV cache, or moving context across regions.
A router should consider model capability, queue time, cache state, data location, and the probability of completing the task. A locally cheaper decision can make the whole trajectory more expensive.
That’s why inference, runtime, and data placement need to be designed together.
A sandbox isolates code. Every tool invocation also needs to establish whose authority the agent is exercising, which action is permitted, and whether approval is required.
Broker credentials at invocation, enforce policy outside the model, and connect the action to the session’s audit trail. Forking an agent should never silently expand its permissions.
Production traces should feed a tested improvement loop.
Connect model calls, tool responses, state changes, human corrections, and outcomes in one trace. Turn recurring failures into evaluation cases. Replay candidate fixes in simulations before changing production behavior.
That gives execution history a second purpose: evidence for improving the next run.
The economic objective follows directly: minimize total cost per successful task, including inference, active compute, persistent state, failed attempts, and recovery.
More capable models can support longer tasks and more parallel work. Falling token prices can therefore expand demand for the infrastructure surrounding each call.
At @DigitalOcean, we’re building around that entire loop: think, act, remember, improve. Open harnesses and models, governed execution, durable state, and customers retaining control of their accumulated intelligence.
The agent’s full lifecycle is becoming the unit of AI Native cloud's architecture.
10
8
23
2,874
ashu garg retweeted
💥 Introducing Warp Sciences, a new company applying AI and robotics to help the world develop new medicines faster.
We’ve raised a $215M Series A led by @lightspeedvp to bring frontier AI out of the browser and into the lab, and build scientific capacity here in America.
I co-founded Warp with my long-time friend @dpatil, FAIR co-founder @manohar_paluri, and drug hunter Sol Reisberg. The early team is full of incredible people like @bkspears9, who helped achieve Q>1 fusion at LLNL.
We’ve raised a $215M Series A led by @lightspeedvp to bring frontier AI out of the browser and into the lab, and build scientific capacity here in America.
I co-founded Warp with my long-time friend @dpatil, FAIR co-founder @manohar_paluri, and drug hunter Sol Reisberg. The early team is full of incredible people like @bkspears9, who helped achieve Q>1 fusion at LLNL.
240
185
2,404
554,273
IIT is MIT on steroids.
No other school matches the rate IIT graduates become founders, CEOs, and CTOs.
The entrance exam is one of the hardest math, physics, and chemistry tests in the world.
Over 1.4 million people take it every year, and most score zero. Roughly 1% of the t get in. The average MIT student wouldn’t get into IIT.
The test measures 2 things: extreme intellect and extreme grit. Performing at the top level requires years of disciplined preparation. Raw academic talent alone isn’t enough.
That’s why IIT alumni punch so far above their weight, especially in startups. A material number of them end up in Silicon Valley as founders, CEOs, CTOs, and VPs of engineering.
- Sundar Pichai (IIT Kharagpur) → CEO of Alphabet
- Vinod Khosla (IIT Delhi) → Co-founder of Sun Microsystems
- Aravind Srinivas (IIT Madras) → Founder of Perplexity
And in the founders I’ve backed too - @ashutoshg and @mohitaron studied at IIT Delhi and @krishnanvijay studied at IIT Bombay.
I try to meet as many IIT founders as possible. When a founder has passed IIT’s filter, I already know two things: top 1% intellect, top 1% grit.
No other school matches the rate IIT graduates become founders, CEOs, and CTOs.
The entrance exam is one of the hardest math, physics, and chemistry tests in the world.
Over 1.4 million people take it every year, and most score zero. Roughly 1% of the t get in. The average MIT student wouldn’t get into IIT.
The test measures 2 things: extreme intellect and extreme grit. Performing at the top level requires years of disciplined preparation. Raw academic talent alone isn’t enough.
That’s why IIT alumni punch so far above their weight, especially in startups. A material number of them end up in Silicon Valley as founders, CEOs, CTOs, and VPs of engineering.
- Sundar Pichai (IIT Kharagpur) → CEO of Alphabet
- Vinod Khosla (IIT Delhi) → Co-founder of Sun Microsystems
- Aravind Srinivas (IIT Madras) → Founder of Perplexity
And in the founders I’ve backed too - @ashutoshg and @mohitaron studied at IIT Delhi and @krishnanvijay studied at IIT Bombay.
I try to meet as many IIT founders as possible. When a founder has passed IIT’s filter, I already know two things: top 1% intellect, top 1% grit.
14
5
105
11,226
Huge congrats to @venkyganesan @mmurph @shawnvc on @MenloVentures 50th anniversary. Key to a great firm is a great partnership!
PS: I was the black sheep at the Black Tie event:-)
PS: I was the black sheep at the Black Tie event:-)
4
0
62
5,693
When cars first arrived, there were no stop signs or speed limits, and few paved roads.
We built all of that over the following decades, and now most people drive without a second thought.
Electricity was the same.
The outlets in your kitchen and bathroom have a small reset ar power makes a short circuit more likely.
We added extra protection where the risk is higher.
AI needs the same kind of safety work. The companies building it and the people using it each have to decide how much risk they’re willing to take, what they let it do, and where the safeguards go.
If you’re nervous about an AI agent, don’t give it your credit card.
If you run a startup, decide which systems it can touch and put the controls in place before you need them.
We built all of that over the following decades, and now most people drive without a second thought.
Electricity was the same.
The outlets in your kitchen and bathroom have a small reset ar power makes a short circuit more likely.
We added extra protection where the risk is higher.
AI needs the same kind of safety work. The companies building it and the people using it each have to decide how much risk they’re willing to take, what they let it do, and where the safeguards go.
If you’re nervous about an AI agent, don’t give it your credit card.
If you run a startup, decide which systems it can touch and put the controls in place before you need them.
0
1
3
466
Putting a nuclear reactor in your house sounds ridiculous today.
In 20 years, it might not.
Right now, reactors need hundreds of staff and safety systems to run safely. So I have solar panels.
But if someone builds a reactor small enough to power my neighborhood, and safe enough to sit down the street, I might switch.
We make that kind of decision about every powerful technology.
We weigh what it does for us against what could go wrong, and the answer changes as the technology gets safer.
AI is the same.
Some use cases may feel too risky today. As the technology and the safeguards around it get better, many of them won’t.
That’s why I’m optimistic about the future. We have the agency to build the systems that make it possible.
In 20 years, it might not.
Right now, reactors need hundreds of staff and safety systems to run safely. So I have solar panels.
But if someone builds a reactor small enough to power my neighborhood, and safe enough to sit down the street, I might switch.
We make that kind of decision about every powerful technology.
We weigh what it does for us against what could go wrong, and the answer changes as the technology gets safer.
AI is the same.
Some use cases may feel too risky today. As the technology and the safeguards around it get better, many of them won’t.
That’s why I’m optimistic about the future. We have the agency to build the systems that make it possible.
3
0
12
1,189
ashu garg retweeted
Worth reading, this article makes a lot of valid points, especially around where the economics of agentic AI may end up.
• The agent/assembler could end up controlling a lot more of the economics because it decides which model gets used, where the work runs and which supplier gets paid
• More AI usage does not automatically mean more value flows to the same GPU or infrastructure providers
• Open models, routing, caching and execution can change the workload mix pretty quickly
• AI can keep winning overall while some of the infrastructure investment theses still turn out to be wrong
The part I’m still thinking through is how much value really moves away from infrastructure. Frontier training, hard reasoning and high-end inference will still need a lot of expensive compute, so my instinct is that the value gets spread across more layers rather than simply moving away from infra.
The question I’m left with is: as agents get better at routing work across models and infrastructure, who ends up with the strongest economics: model labs, hyperscalers, infrastructure providers, or the application layer?
@JayaGup10 @ashugarg
• The agent/assembler could end up controlling a lot more of the economics because it decides which model gets used, where the work runs and which supplier gets paid
• More AI usage does not automatically mean more value flows to the same GPU or infrastructure providers
• Open models, routing, caching and execution can change the workload mix pretty quickly
• AI can keep winning overall while some of the infrastructure investment theses still turn out to be wrong
The part I’m still thinking through is how much value really moves away from infrastructure. Frontier training, hard reasoning and high-end inference will still need a lot of expensive compute, so my instinct is that the value gets spread across more layers rather than simply moving away from infra.
The question I’m left with is: as agents get better at routing work across models and infrastructure, who ends up with the strongest economics: model labs, hyperscalers, infrastructure providers, or the application layer?
@JayaGup10 @ashugarg
0
4
10
2,025
ashu garg retweeted
@JayaGup10 is batting a thousand when it comes to spotting emerging AI trends. Ignore at your own peril.
1
2
4
1,167
ashu garg retweeted
Intriguing piece by @JayaGup10 and @ashugarg on where costs are headed.
Getting the judgments/decisions right at each stage is paramount.
Getting the judgments/decisions right at each stage is paramount.
0
2
1
468
Cyber risk is going parabolic right now.
None of our existing security systems were built for agents.
The security industry has historically worked by isolating risk and writing rules in advance. Can employee B access this document, yes or no?
That works when you know the path in advance. A firewall can police network traffic because traffic follows known routes.
An agent doesn’t always follow a known route. Ask one to book a family trip and it might pull your passport number, your credit card and your kids’ birthdates, send them to a model, write them into logs, then pass them to Stripe.
Often the route isn’t tightly mapped so your next request may take a different one.
Yes-or-no access rules don’t scale when employee B has thousands of agents invoking tools on their behalf.
So, as @anshublog puts it, you invert the problem.
Instead of trying to secure every possible path an agent might take, how do you secure the data itself - whatever route it takes?
That’s the premise of @SkyflowAPI.
Learn more here: youtube.com/watch?v=_OwWEl…
None of our existing security systems were built for agents.
The security industry has historically worked by isolating risk and writing rules in advance. Can employee B access this document, yes or no?
That works when you know the path in advance. A firewall can police network traffic because traffic follows known routes.
An agent doesn’t always follow a known route. Ask one to book a family trip and it might pull your passport number, your credit card and your kids’ birthdates, send them to a model, write them into logs, then pass them to Stripe.
Often the route isn’t tightly mapped so your next request may take a different one.
Yes-or-no access rules don’t scale when employee B has thousands of agents invoking tools on their behalf.
So, as @anshublog puts it, you invert the problem.
Instead of trying to secure every possible path an agent might take, how do you secure the data itself - whatever route it takes?
That’s the premise of @SkyflowAPI.
Learn more here: youtube.com/watch?v=_OwWEl…
Cyber is having a moment
Across 21 major software companies, including Apple, AWS, Microsoft, and Google:
- Reported critical vulnerabilities never cleared 100 per month in four years
- Since spring they've jumped to over 600 per month
2
0
7
1,407
ashu garg retweeted
@chamathinvestor update, earlier in the year:
"As we noted last month, we've tightened our definition of signed ARR to be as truth-seeking as possible: monthly and quarterly recurring deals no longer count as ARR (we now call that "run rate" instead), and ARR is now strictly defined as enterprise revenue that's committed for a full year and renews automatically, with a x-day opt-out period toward the end of that contract year."
"As we noted last month, we've tightened our definition of signed ARR to be as truth-seeking as possible: monthly and quarterly recurring deals no longer count as ARR (we now call that "run rate" instead), and ARR is now strictly defined as enterprise revenue that's committed for a full year and renews automatically, with a x-day opt-out period toward the end of that contract year."
0
1
1
367
ashu garg retweeted
@SenSanderss pay the same price as everyone else for practically everything they buy. Government is the glaring exception, where they already pay considerably more.
The audacity of pretending it's an injustice for someone to receive benefits proportional to what they paid into a system is remarkable.
Bernie apparently believes that the wealthier you become, the less entitled you are to the fruits of your own contributions, and the more entitled everyone else becomes to them.
It's parasitism incarnate.
The audacity of pretending it's an injustice for someone to receive benefits proportional to what they paid into a system is remarkable.
Bernie apparently believes that the wealthier you become, the less entitled you are to the fruits of your own contributions, and the more entitled everyone else becomes to them.
It's parasitism incarnate.
6
5
112
5,935