ashu garg
@ashugarg
Enterprise VC @FoundationCap | Early investor in @databricks @tubi & 6 other unicorns- @cohesity @eightfoldai @turingcom @amperity @alation @anyscalecompute
2.6K Following 12.7K Followers
Every security company: “how do I stop the bad guy?”
@anshublog: “how do I make my data useless to them?”
That inversion is the whole premise of @SkyflowAPI , and Anshu thinks it's a $100 billion company.
In the podcast we get into:
- Common misconceptions about security in AI
- How Skyflow rewrote its problem statement when LLMs arrived
- What he learned working with Marc Benioff at Salesforce
Listen to the full conversation here:
Show more
My new pod:
How to solve AI’s security problem
with @ashugarg
None of our security systems today were built with agents in mind.
In our latest B2BaCEO podcast, @SkyflowAPI co-founder and CEO @anshublog joins @ashugarg to talk about what he’s learned securing sensitive customer data across data stores, models, and agents.
It aligns with a problem he’s been chasing for over two decades. “At some point you realize, you know what? There's probably a $100B company to be built just protecting the most sensitive data for customers at all these companies,” he shares.
Tune in to the full episode:
Spotify:
Apple Podcasts:
YouTube:
Show more
New pod coming.
Hoping to see more such founders from/in India. Congrats Devakumar and @neerajKh_ and wish you the best👏🚀
He spent 13 years at ISRO working on the cryogenic engines behind Chandrayaan and Gaganyaan. Then he left to build the one engine ISRO had never attempted, and that only one company on Earth has ever flown.
His name is Devakumar Thammisetty.
After more than a decade inside India's space programme, he went to EPFL in Switzerland to study a single question: how do you make a rocket engine that can fly again and again?
The answer he came back with is called full-flow staged combustion. It burns methane with liquid oxygen and routes every drop of both propellants through the turbines before they reach the chamber, which makes the engine more efficient, gentler on its own components, and far better suited to being reused.
It is also brutally difficult. Only a handful of these engines have ever been tested anywhere in the world. Only SpaceX's Raptor has flown one. Astrobase calls it the hardest rocket engine cycle ever built.
The man funding it made his money somewhere unexpected.
Neeraj Khandelwal is an IIT Bombay graduate who spent the last decade co-founding CoinDCX and scaling it past 15 million users and a $2.1 billion valuation. He is Astrobase's primary investor. "Building India's first FFSC rocket engine means building the factory that can build the engine," he has said. "No shortcuts."
They founded Astrobase in Bengaluru in 2024, along with two more former ISRO scientists, Pawan Kumar and Prashant M.
A sub-scale engine hot fire in September 2025. High-speed turbopump tests in January 2026. The full engine design and manufacturing milestone in August 2026. A 21.5-acre propulsion test facility in Anantapur. In June 2026, IN-SPACe chose them from 43 applicants for its Technology Adoption Fund.
SpaceX and Blue Origin each spent billions and most of a decade getting their staged combustion engines to work. Astrobase has raised around $10 million.
The full engine hot fire is next. Four countries have ever tested an engine like this and India is not one of them yet.
Show more
If you’re the CEO of any software company and you’re not offering open-weight models as a SKU right now, you’re asleep.
“Pace the frontier” may be the greatest invitation software incumbents have ever gotten. While the labs debate how quickly intelligence should advance, software companies should be racing to commoditize the intelligence we already have.
Pharma and banks are already picking up open weight models partly for margins, partly because a revocable lab API is a dependency they increasingly don’t want!!!
AI natives and tech companies that care about COGS are doing the same.
Most software companies that tried had failed attempts because the open weight models sucked. But now open weight models are good enough.
I believe that every major software company should become a model factory for its vertical: own the evals, post-train open weights on the workloads it uniquely sees, serve those models to its customers, and use production feedback to continuously improve them!
Show more
To get to ASI we likely need auto-meta-research, not just auto-research.
Auto-research hill climbs within the current recipe. Minimize pretraining loss, maximize post-training evals.
Auto-meta-research defines new objectives. An outer loop that searches across paradigms. Outside deep learning, maybe even outside gradient descent. Not just scaling transformers + RL.
The inner loop optimizes the recipe. The outer loop questions the recipe.
Show more
All the AI labs are peddling something similar: ”we’re building a god-like system that will eat all knowledge work.”
One piece of advice I’d give to app-layer founders:
Labs have an incentive to tell a maximalist story where raw intelligence is all that matters.
Today, the entire U.S. economy produces roughly $33T a year. If Anthropic predicts a $30T TAM for AI, there's only one way to read a number that size. AI models so advanced, they eat nearly all the knowledge work in the economy. Only a god-like model gets to that scale.
It’s a convenient fiction.
The labs have to tell the god-model story to justify their IPO valuations.
In the future they’re painting, every job, every workflow, every company is a problem that Anthropic’s model will solve. So the entire application layer becomes obsolete.
I believe this future is wrong.
Cursor, Claude Code, and Replit all run on essentially the same models. But they're three completely different products because each was built around a different opinion about how developers want to work.
Cursor bets that developers want AI inside the editor they already use, suggesting help as they type.
Claude Code bets developers want to hand off the whole task and come back to finished work.
Replit bets people shouldn’t need a professional workshop at all. A beginner can instead build in the browser in minutes.
Each company is an example of customers choosing products rather than models.
While a lab needs to preserve capability across a wide range of tasks, an app company only needs to optimize intelligence around the particular job it cares most about. This is what @Jonsid, founder of @turingcom, calls “artificial narrow intelligence.”
Knowing which job to focus on - and what a great product for it looks like - is the founder's real edge. The model supplies the intelligence, but the p.o.v. on how to apply it needs to come from the founder.
In this month’s B2BaCEO, I share my full advice for founders building at the app layer:
Show more
9/11 was a turning point for many of us. Despite growing up in India, Sudan and Nigeria, all of which are hotbeds of terrorism, 9/11 was different for me. It felt personal, it made me more “American” and it made me really angry with respect to terrorism/terrorism. When I look back, it was a turning point for me, for many of us and for America.
Show more
Turing's focus right now:
1.Automate AI research
2.Automate engineering
3.Automate knowledge work
4.Automate scientific discovery
The models come from frontier labs. The training signal comes from us. Data, evals, and RL environments built from real workflows. Hard enough that today's best models still fail.
AI research sits at the top for a reason. Automate that, and everything below it compounds.
If you're training models toward any of these four and need environments that don't saturate, DM me.
Show more
There is something unbearably cruel about this.
Chinmoy Krishna Das spent months behind bars while his mother grew critically ill.
His family pleaded for him to be allowed to see her one last time.
He couldn’t.
She died.
Only then was he given a few hours of parole to attend her funeral.The video of him breaking down beside his mother’s body is devastating.
Whatever your politics are, there has to be room for basic humanity. A final goodbye should never become a privilege granted too late.
Show more
Kudos to @gargi_kand & @darrenmarble for organizing a great event. @JoshConstine : Why does one need to hire painters, cellists, poets or sculptors to make an event tasteful? Why generalize your opinions/preferences ?
Show more
Hire painters. Hire cellists. Hire poets. Hire sculptors.
Your tech event doesn’t have to be tasteless. Props to @gargi_kand & @darrenmarble for this one.
Valid concerns, but the solution is to democratize AI and not regulate it. A $50B commitment by the US Govt (which is less than 3 days of burn) to fund open weights models will go a long way. And as for @OpenAI and @AnthropicAI, perhaps they are the new @IBM and @BellLabs, which pioneered the PC, internet and most of what we take for granted and yet became irrelevant
Show more
I resigned from Anthropic today. I spent the last three years doing pretraining research at both OpenAI and Anthropic. Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives. More thoughts below.
Show more
Valid concerns, but the solution is to democratize AI and not regulate it. A $50B commitment by the US Govt (which is less than 3 days of burn) to fund open weights models will go a long way. And as for @OpenAI and @AnthropicAI, perhaps they are the new @IBM and @BellLabs, which pioneered the PC, internet and most of what we take for granted and yet became irrelevant
Show more
I resigned from Anthropic today. I spent the last three years doing pretraining research at both OpenAI and Anthropic. Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives. More thoughts below.
Show more
Meta's new AI Agent Muse is here.
Foundation Capital Partner Jaya Gupta shares how the assistant handled complex real-world tasks like apartment negotiation and finding parking in San Francisco:
“The amount of things it's been able to do has been pretty crazy. I also really like…how much they thought about security and privacy, all those things on the Muse side. I'm super impressed already.”
Show more
Everything you need to know about model distillation:
Open models like Kimi K3 and DeepSeek V4 now match frontier models on some benchmarks. The explanation everyone reaches for is distillation.
Before you buy that explanation, there are 3 questions to answer:
What is distillation, really?
What kind is even possible on a frontier model?
And can it explain the recent success of open models?
1. Traditional distillation (the real thing)
Training a student model on the full outputs of a teacher model. Not just it’s output - its reasoning traces and its confidence on every next word: what it chose, what it almost chose, and by how much.
Done right, the student inherits the teacher's performance at a fraction of the cost.
But this requires full access to the teacher's internals. Only a lab distilling its own models can do it. Nobody outside the building gets the probability distributions.
2. Cross-lab "distillation" (what people actually mean)
Frontier models stopped returning reasoning logic last year. They hide or heavily summarize the chain of thought before outputting.
What's left is behavior parroting: training on finished outputs rather than the full decision-making process.
Useful synthetic data? Yes. True distillation? Not close. You cannot do end-to-end distillation of a locked-down frontier model through an API.
3. So does distillation explain the success of open models?
No.
K3 launched weeks after Fable, and 7 days after GPT-5.6. You can't extract enough high-quality signal through a restricted API to match frontier performance - especially in that short a window.
Distillation is inevitable and it isn't a sufficient explanation.
As @GavinSBaker put it, the "Sputnik moment" for Chinese open source hasn't arrived yet. When it does, open models will be more performant AND cheaper than closed ones.
The gains in performance are undeniable. And they can’t be attributed to distillation.
Show more
Phenomenal takes in here
“But here's an interesting question - why has nobody truly been paying close attention to matching capabilities to the unlock? Where is this discourse happening? The root cause of all the shallowness in thinking is that the wrong category was crowned far too early: the AI-native ERP. That is fundamentally the wrong abstraction for what AI changes in finance.”
Show more
Model capabilities have unlocked every category
Heres why finance is having its inflection moment
Finance is entering its Legora moment
Finance is entering its Legora moment
Welcome to my space, Zuck.
