Arjun Malhotra

If you need an investor that can put on their big boy pants and lead your round, call me.

Microsoft won the browser wars - completely dominated. Netscape was dead by 2000. But then it turned out not to matter at all. The value of the internet was created by Google in search, by Facebook in social and by Amazon in commerce. People used the browser to get to those places - but the browser itself was just the door. ChatGPT might be in the same position. 900M users, the defining brand of the AI moment, but potentially just the thing people walk through to get somewhere else. I don't think the real value of AI will be captured by whoever owns the chat interface, even though that's what it seemed like at first. It'll be created in the vertical tools, embedded workflows, products people actually build habits around. Most of that hasn't been built yet. But when it is, I'd bet most of it won't come from OpenAI.

DAUs & MAUs made sense when every session had a human behind it. Now, half your product usage might be agents acting on someone's behalf. The human is still there, somewhere upstream - but they're not the one opening the app. What does "active user" mean in that world?

A year ago, Claude projects would hallucinate numbers quite a bit. You couldn't trust outputs without double-checking everything. Now I can give Claude Code vanilla instructions and walk away. I'll often come back to find it had already tried the standard approach, flagged that it was using too many tokens, listed three alternatives, picked the most efficient one, and kept going - all without me prompting again. Humans were the error-correction layer, but you now don't have to play that role as much.

Story time: We made an offer to a founder a couple of weeks ago. A competing fund soon came in at a better price - same dilution, more capital. The founder still chose us to lead the round. Part of it was that we'd already reached an agreement, though I don't think that alone stops most founders from walking away. He later mentioned he'd done his fair share of ref checks. The feedback was mostly positive, with a note of some minor inefficiencies we're working to solve. What struck us wasn't that he "chose" us but how much effort he'd put into figuring out what we'd be like as investors and not just as a source of capital - which made us think harder about every interaction we have with founders who don't end up in our portfolio. Most of us don't think about those interactions nearly enough.

India did what it was asked. We aimed to reduce russian imports & reorient toward the middle east to buy some goodwill on the absurdly high US tariffs. But then US went to war in the region India had just pivoted toward. Diversification works when one is spreading risk across genuinely independent sources. This wasn't that - because a single power (US) is implicated in both sides of the supply equation. Makes you think it's so hard to negotiate protection against the decisions of the country you're trying to please.

AI fluency isn't something you can read your way into. Newsletters and demo videos can show you what's possible, but they can't transfer the actual understanding, which only comes from sitting with a problem long enough to break it a few times. The automation that saves you 6 hrs every week will probably take an entire weekend to build. That process feels inefficient in the moment, but it's not really separable from the outcome. The people who've genuinely gotten a lot out of AI aren't the ones who stayed most up to date - they're the ones who were willing to spend unstructured time with it before they had a reason to. The hours you save later are basically a return on the hours you spent confused earlier. There's no shortcut to that part.

@RohitMalekar very well articulated!

Corollary: The uncomfortable truth is that most individuals aren't blocked by technology. They're blocked by themselves. By the inability to articulate what they actually do, why they do it, and what good looks like. AI doesn't solve that problem. It exposes it.

@neerajnathany Yes, and I think that context gap actually grows the longer you've worked on something

Absolutely true. I particularly struggle in communicating to AI the universe of context around why a particular variable/feature/view/idiosyncracy is how it is and how the AI should account for/ignore it while figuring the solution. Besides, a coder infuses several facets of his/her personality while building, by virtue of all the logical/infrastructural/architectural/lexical choices he/she continuously makes. Incorporating these via a .md file is so much trickier than coding oneself or even communicating to a fellow human honestly. We'll all eventually have to overcome this.

AI can give you five clean solutions to any engineering problem confidently and quickly. All of them will also be technically correct. But someone still has to decide: which of the five solutions to actually ship, which one fits how the team currently works, and which one doesn't become someone else's problem in a year? That decision requires judgement, experience, and knowing your internal systems in ways that don't fit into a prompt. That's the part that was always contextual, and hence, always human.

Didn’t know that activity required a specific kind of tissue

@TheOneDit hey Vandit, the role is now closed

@BadCapitalVC Hey Arjun, Did you manage to find your CoS or is that role still open? Asking for a friend 😄

about time an AI x social app makes it to the top of this list

Maybe the greatest content irony of this generation - reels killed our attention spans and cinema isn't making business as it used to, yet newsletters and podcasts have grown immensely over the last few years.

Here's a less discussed graph from @AnthropicAI's research paper. After ChatGPT launched, young workers (22-25) in AI-exposed roles saw a 14% drop in new job starts. AI might not be taking jobs, but it is taking job offers.

Expertise traditionally meant having superior domain knowledge + the ability to derive conclusions independently. An expert’s judgment was trusted because they had seen more, remembered more, and could connect patterns faster than everyone else. But in the last few years, expertise has become the ability to work effectively with systems that generate probabilistic outputs. A 'skilled operator' today is one who understands how to rightly interrogate models, stress test assumptions and contextualise AI generated outputs in the real world. This is because memory & raw processing (given they're now abundant) matter less than judgment, framing and system literacy. The inversion is that we used to command the machine to compute. Now we evaluate what it computes and decide whether to follow.

If you use Arc Browser, you're probably not a fun person in real life.

As AI keeps taking over more of your work, the only competitive differentiation left in the workplace will be not being fucking boring.

Enough has been said about how boring businesses are where the money is. The less discussed part is why this works so well in India - unsexy utility problems stay unsolved here & incumbents are too comfy to fix them. So founders who are willing to go deep on a boring use case with a tech-first mindset often build the most enduring companies.