Karthik Subramanian

Tech Exec, AI & Startups. Board Member. Operator across four continents. @zalora cofounder, CTO+CMO, @gfg ex-CTO.

[Karthik Subramanian @KkarthikS]

The comparison to open source software has a big problem. Open source gives you the actual recipe. You get, read and change the code. You can build it again.

Open weights give you the finished model. The recipe, training data and training process stay closed. You get the result, not the method.

You cannot check a model the way you check code. A person can read a function and understand it. A weight file holds billions of numbers. No person can read those numbers and understand what the model will do.

Open source succeeded for another reason. Many people could help build it. A laptop and free time were enough. This let thousands of people improve Linux together.

Frontier models do not work this way. A single training run can cost hundreds of millions of dollars. No outside person can add a contribution to that process. Open weights let more people run the model. They do not let more people build the model.

The security argument also breaks down. In open source, a person finds a bug, the team fixes the bug, users update their software. This process does not exist for open weights. Once a company releases the weights, nobody can patch a copy that is already in use. Safety training is also easy to remove. A small amount of compute can strip it out.

So this does not rebuild the open source community. It looks like a different strategy. A company releases a base model at no cost. This builds an ecosystem around the model. The company then sells services on top of it. This is a valid business strategy. But it is not the same story as open source software.

[Jensen Huang @JensenHuang]

For my first post, I’m sharing a letter @NVIDIA signed on why open models matter.

AI will transform every industry, power every company, and be built by every country.

Open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty.

The world needs both frontier closed models and frontier open models.

https://t.co/AUKzoQ5Ikb

[Karthik Subramanian @KkarthikS]

@elonmusk @AravSrinivas Et tu Elon?

[Karthik Subramanian @KkarthikS]

Open weights give more companies access to AI. They increase competition. Paying twice: once with money, once with proprietary data, is fair critique of the closed models.

BUT look at the list of signatories. The letter comes from chip makers, cloud providers, security firms, and deployment platforms. All of them earn more money when many open models run on their infrastructure.

Who is missing? OpenAI and Anthropic. Both filed for IPOs in recent weeks. Both have valuations near one trillion dollars. Both need closed weights to protect their pricing power.

So the timing is not an accident. The infrastructure companies want to reduce the value of the model companies. And they want to do it just before the market prices those companies.

Microsoft plays both sides. It holds equity in closed OpenAI. It also runs a public campaign that weakens closed-model pricing. If that campaign succeeds, more inference moves to Azure.

One more thing. The security argument changed direction. It once supported closed weights. Now it supports open weights. But it always lands on the side that helps the speaker’s margins. This is far from a principle and clearly a strategy.

The ideas deserve support while this coalition deserves scrutiny.

[Karthik Subramanian @KkarthikS]

Open weights give more companies access to AI. They increase competition. Paying twice: once with money, once with proprietary data, is fair critique of the closed models.

BUT look at the list of signatories. The letter comes from chip makers, cloud providers, security firms, and deployment platforms. All of them earn more money when many open models run on their infrastructure.

Who is missing? OpenAI and Anthropic. Both filed for IPOs in recent weeks. Both have valuations near one trillion dollars. Both need closed weights to protect their pricing power.

So the timing is not an accident. The infrastructure companies want to reduce the value of the model companies. And they want to do it just before the market prices those companies.

Microsoft plays both sides. It holds equity in closed OpenAI. It also runs a public campaign that weakens closed-model pricing. If that campaign succeeds, more inference moves to Azure.

One more thing. The security argument changed direction. It once supported closed weights. Now it supports open weights. But it always lands on the side that helps the speaker’s margins. This is far from a principle and clearly a strategy.

The ideas deserve support while this coalition deserves scrutiny.

[Karthik Subramanian @KkarthikS]

@svpino Totally inconsistent with Dario’s constant doom-mongering about the end of days of SWEs 🤦🏾‍♂️

[Karthik Subramanian @KkarthikS]

The timeline and scale analogies are a bit optimistic, and turning those evaluations into something reliable (especially for subtle things like deception or loss of control) is going to be much harder in practice than it sounds on paper. A US-led framework could still shape norms globally through market and compute leverage, but getting meaningful coordination across labs and countries in the middle of this race won’t be easy. Laying out a specific, adaptable proposal is useful. We need more of this kind of detailed thinking from people who actually understand the tech.

[Karthik Subramanian @KkarthikS]

@ssmumbai beautiful and thought-provoking

[Karthik Subramanian @KkarthikS]

Really enjoyed this piece. Reads different from the doom-mongering of the frontier labs.

The Hayek/Polanyi framing on dispersed, tacit knowledge is a strong reminder that productive expertise lives in people doing the actual work, not in some central repository. The critique of benchmarks that only measure what AI can do alone, instead of measuring what humans and AI can do together, is a push worth pursuing.

That said, the world isn’t really choosing between fully centralized frozen models and completely decentralized ones. We’re already in the messy middle: big base models provide the raw capability, and then customization, fine-tuning, and better interfaces let organizations shape them to their own context.

The piece makes a compelling case for why keeping human judgment in the loop (especially where knowledge is local and changing) is valuable, but I suspect the most effective systems will be hybrids that combine frontier-scale intelligence with deep, ongoing human adaptation rather than trying to distribute everything from the ground up. The emphasis on building AI that actually strengthens the people using it is genuinely worth pursuing.

[Mira Murati @miramurati]

Today we share the worldview behind our mission.

Human values don't average out. Local knowledge can't be centralized. The good future has many AIs, raised in different places, shaped by the people they serve, disagreeing with each other the way we do.

https://t.co/A14SurOM2K

[Karthik Subramanian @KkarthikS]

@dkfromdk Solid piece. The core insight on margin leverage is sharp and the focus on coordination overhead (dispatching, exceptions, admin layers) as the real target feels right. Most AI tools today are built for tech-savvy teams that will adopt new interfaces, so calling that out as a non-starter for these industries is fair.

That said, I think the durability of those margin gains is overstated. In competitive, low-differentiation markets, efficiency improvements often just get competed away through lower prices rather than turning into lasting higher margins. And while embedding AI as invisible infrastructure sounds ideal, the reality of messy operations is that exceptions, judgment calls, and oversight don’t go away. Making agents reliable enough to truly reduce coordination burden without creating new risks is still very hard.

The direction makes sense though. These businesses have the strongest incentive to make it work.

[Karthik Subramanian @KkarthikS]

@MarkPommrehn Thank you!

[Karthik Subramanian @KkarthikS]

In the wave of agent loop and workflow hype, two real costs keep getting ignored:

- Iterations lack provenance, so governance and drift become impossible at scale.
- Components burn tokens and latency on constant state reconciliation (the hidden “orchestration tax”).

The missing primitive is simple but powerful: pair a versioned decision ledger (every loop step as a commit with full provenance) with shared query-able memory that uses explicit contracts between agents/steps.

This turns loops from opaque improvement machines into auditable evolution systems while slashing coordination overhead. This could be the difference between flashy demos and compounding production systems.

The next leap in agentic work will come from this kind of lightweight systems architecture.

[Karthik Subramanian @KkarthikS]

@chamath Really liked this. The core warning about AI making code cheap while accelerating drift and incoherence is spot on, and something a lot of the current tooling hype is conveniently ignoring. The historical framing around actual software factories is useful, and the five tests do a good job of separating real production systems from just another set of powerful developer tools. Who actually takes the call when systems break is the question we need to ask any vendor now.

That said, the framing feels a little binary. Treating it as “true factory or it’s just a tool” sets an extremely high bar that even many successful traditional systems never fully cleared. Most organizations are going to live in the messy middle for a while: strong governance layers, review processes, and orchestration on top of AI acceleration. Can anyone afford to wait for something that hits every test perfectly? Hybrids are probably going to win more often than pure factories.

Also, the legacy modernization example is impressive, but the harder thing (and the one that will actually determine whether this model scales) is maintaining coherence forward under continuous change with mixed human + agent teams over years, not just extracting rules from old code. Curious how you’re thinking about proving that closed loop at enterprise scale.

Overall though, this is a useful corrective. As code gets cheaper, the real advantage shifts to whoever can keep systems understandable and accountable. That part feels right.

[Karthik Subramanian @KkarthikS]

@JayaGup10 Really enjoyed this. The core point about judgment and institutional know-how being the actual scarce asset (not just the data) lands cleanly, and it ties together nicely with the context graphs work you’ve done before. The insurance and pharma examples make the risk feel concrete, and the question at the end is the right one for any CEO to ask before plugging high-stakes workflows into a shared model.

That said, I think it leans a bit hard into the inevitability of leakage and value erosion. In practice, commercial/enterprise agreements today default to no training on your usage, with short retention windows and strong contractual protections. A lot of the “learning” risk can be contained through architecture: heavy RAG, private memory layers, distillation into your own models, or running sensitive agents on sovereign setups. Not everything funnels back into the frontier weights the way the piece suggests.

The TikTok analogy is vivid but doesn’t map perfectly. Companies are paying customers here, not free content creators, and most aren’t handing over clean, high-signal judgment without friction. Labs trying to turn aggregated patterns into direct competition will still run into massive regulatory, capital, and distribution walls in fields like insurance or drug development. They’re better positioned as tool providers than as full replacements.

The real takeaway from your piece for me is to treat generic productivity work as fair game to pool, but ruthlessly own the decision traces and learning loops where your edge lives. Adopting hybrid approaches using frontier models for the heavy lifting with proprietary orchestration and context graphs for the judgment, I think, will let you capture more of the upside without sleepwalking into commoditization.

[Karthik Subramanian @KkarthikS]

@FT Europe took 10 years to build EES. India does airport biometrics in 30 seconds. Guess which one The Economist calls backwards: https://t.co/fivGbMuZ65

[Karthik Subramanian @KkarthikS]

Europe spent a decade building the EES. India registers your biometrics at the airport in 30 seconds.

If you’ve ever applied for a Schengen visa, you know the drill. An Indian company with a near monopoly on processing (the irony writes itself). Rejection rates that punish legitimate travelers. “Premium” fees to escape a booking system designed by sadists. Embassies treating every applicant as a future overstayer. Spain excepted, credit where due.

This should be blindingly obvious but merits repeating:

Visa holders surrendered their biometrics years ago. EES still dumps them into the same queues as visa-free travelers. The people who crawled through the frying pan get zero benefit for it.

This mess was invisible for years because the passports it hurt had no voice. Now that privileged passports can’t buy their way past the queue, suddenly it’s a crisis worthy of op-eds and appeals to “relax measures.”

Meanwhile India, Sri Lanka and China issue e-visas in days. The average Asian airport clears more passengers before lunch than some European hubs manage all day, with no meltdown coverage. And yes, the same India that The Economist loves to lecture on tech converts a 5 minute online form into e-gate access in 5 days flat.

Singapore deservedly gets a pass for being the platinum global standard. Everyone else lecturing Asia about the stone ages should sit this one out.

A decade of implementation. Overstuffed agencies that don’t talk to each other. Airports that were creaking long before EES went live.

So before you nod along at the next “Asia needs to modernize” take, ask yourself who actually shipped.

Today’s FT editorial on the topic is in the comments.

[Karthik Subramanian @KkarthikS]

The FT puts it more eloquently: https://t.co/e5FIeoFh1o ‘It’s whack-a-mole’: how Europe’s smart border melted down

[Karthik Subramanian @KkarthikS]

Europe spent a decade building the EES. India registers your biometrics at the airport in 30 seconds.

If you’ve ever applied for a Schengen visa, you know the drill. An Indian company with a near monopoly on processing (the irony writes itself). Rejection rates that punish legitimate travelers. “Premium” fees to escape a booking system designed by sadists. Embassies treating every applicant as a future overstayer. Spain excepted, credit where due.

This should be blindingly obvious but merits repeating:

Visa holders surrendered their biometrics years ago. EES still dumps them into the same queues as visa-free travelers. The people who crawled through the frying pan get zero benefit for it.

This mess was invisible for years because the passports it hurt had no voice. Now that privileged passports can’t buy their way past the queue, suddenly it’s a crisis worthy of op-eds and appeals to “relax measures.”

Meanwhile India, Sri Lanka and China issue e-visas in days. The average Asian airport clears more passengers before lunch than some European hubs manage all day, with no meltdown coverage. And yes, the same India that The Economist loves to lecture on tech converts a 5 minute online form into e-gate access in 5 days flat.

Singapore deservedly gets a pass for being the platinum global standard. Everyone else lecturing Asia about the stone ages should sit this one out.

A decade of implementation. Overstuffed agencies that don’t talk to each other. Airports that were creaking long before EES went live.

So before you nod along at the next “Asia needs to modernize” take, ask yourself who actually shipped.

Today’s FT editorial on the topic is in the comments.

[Karthik Subramanian @KkarthikS]

Did the Economist's editorial team fail to get an e-visa (I have applied for e-visas for foreign family members >10 times in 10 mins flat)?

Perhaps the editorial team needs to review their Internet/browser literacy skills before taking a potshot at the e-visa site?

[The Economist @TheEconomist]

The visa website of a country renowned for its people’s IT skills resembles a Geocities page from 1999. Register for free to discover why it is unlikely to improve any time soon https://t.co/EDAOT33dgG

[Karthik Subramanian @KkarthikS]

@TheEconomist Is the official editorial policy of The Economist to be openly derogatory and racist towards India?

[Karthik Subramanian @KkarthikS]

@Sudhir_Voleti @erbmjha True, but they an ever-thinning have a list of grouses and end up digging so deep that even by their biased standards, this looks ridiculous.

[Karthik Subramanian @KkarthikS]

@ishaantharoor Painfully detailed in Jonathan Kwitny's Endless Enemies.

[Karthik Subramanian @KkarthikS]

Nice! Skills that persist and can be versioned/audited feels like the next layer after memory. The real production challenge is making these reusable workflows traceable and low-coordination. Shared this take on versioned decision ledgers + contract-based shared memory to reduce orchestration overhead: https://t.co/QgnK7mr4cm

[Karthik Subramanian @KkarthikS]

In the wave of agent loop and workflow hype, two real costs keep getting ignored:

- Iterations lack provenance, so governance and drift become impossible at scale.
- Components burn tokens and latency on constant state reconciliation (the hidden “orchestration tax”).

The missing primitive is simple but powerful: pair a versioned decision ledger (every loop step as a commit with full provenance) with shared query-able memory that uses explicit contracts between agents/steps.

This turns loops from opaque improvement machines into auditable evolution systems while slashing coordination overhead. This could be the difference between flashy demos and compounding production systems.

The next leap in agentic work will come from this kind of lightweight systems architecture.