Marco

@prayanks > "That's not a product feature. That's a compounding asset." > I've written a full framework on this — covering the four depths of personalisation, The emdashes, the "that's not X, that's Y".. prayank this is some A grade slop, I think you're cool, don't outsource your thinking

You just raised $5M to build someone else's moat. Here's what I mean. Most AI startups I meet are competing on the wrong axis. They're obsessing over model choice — Claude vs. GPT vs. Gemini. They're fine-tuning on domain data. They're building slick interfaces on top of state-of-the-art APIs. None of that is a moat. All of it can be replicated in weeks. The founders I'm most excited about are competing on a completely different dimension: time. Every session a user spends inside a well-architected AI system is a deposit. The system learns their editing patterns, their risk tolerance, their preferences — implicitly, without being told. After six months of daily use, that system knows how you work in ways you couldn't fully articulate yourself. That's not a product feature. That's a compounding asset. The architectural decision that separates these two worlds is simpler than most founders think: stateful vs. stateless agents. A stateless agent resets after every session — all that signal, discarded. A long-running agent retains it, learns from it, gets harder to replace every single week. The switching cost of a great stateless AI product is zero. The switching cost of a great stateful one, after two years, is enormous — not because of contracts, but because leaving means starting over. I've written a full framework on this — covering the four depths of personalisation, the three RL signals that drive compounding, and where the research frontier is heading. Link in the comments. One question for founders building in this space: are you designing for state accumulation from day one — or is that an afterthought?

@umanghome wow that's the saddest childhood dream I've come across :(

Ever since I was a little child I dreamt about building software while out for a walk in the sun

@peebsvibin @sidsm01 @akm1410 peebs, siddhant’s comment was satire. He hasn’t actually called his friends. He is mocking the original post. Hope you have a nice day

@sidsm01 @velocitycake @akm1410 I have been a long time subscriber (literally from day 1) and swear by their reportage. I’ll sponsor your subscription and that of 3 friends of yours. There you go- forget splitting. Game?

I have been calling VCs asking them if they will fund Sarvam after the AI summit considering sovereignty and desh are such big themes. Four out of four declines so far. I plan to reach out to at least the top ten who haven't yet invested in the company to cover more ground.

@chirag @akm1410 That would be an actual sensible question, but it’ll make the headline look bad. So we won’t answer that.

@akm1410 Why?

@sidsm01 @akm1410 Siddhant can you please not add nuance to this whole thing? The subscriptions need to sell. Thanks.

@akm1410 Would have been better if you also put at least a 1 liner with "why" (which I am sure you would've discussed in your calls) rather than just this sensational part of the conversation

@akm1410 what is this shoddy performative journalism ugh There could be a 1000 reasons why they wouldn’t invest in a company despite having a thesis in the sector😭 Are you aware this is net negative to society or you actually think this is somehow constructive?

@isenditbacc @shantanugoel Ah I’ll take the other side of the bet - as models become smaller+smarter, everyone unknowingly will do some local inference. Pretty sure apple/microsoft already bake in some tiny model in the os Btw is arm vs x86 a contention for your dev machine?

Recent procs from both intel and amd have dynamic allocation of ram for vram use. On the 285h is 50% max alloc and amd 395+ it's 70% (iirc). So you can do inferencing on some mid size models like 30B params on both. But unified memory gives you way bigger bandwidth so faster inferencing. But but but... I'm planning to use only for dev work and not running models, so for me specifically it doesn't matter 😁 (I'm willing to bet most people won't do much local inferencing either once the initial novelty value wears off)

Mac mini with M4 max, 64gb ram and 1 tb storage is 2.5 lakh Nuc with intel 285h, 64gb ram and 1tb storage is almost as much "Apple tax" has reduced?

@shantanugoel @isenditbacc Also aren’t these 2 very different in terms of AI inference (if that’s the use case), Apple has unified memory and those 64GB can be used to run models, whereas that’s not true for regular RAM?

@isenditbacc They haven't changed their prices after ram shoot up. So apple tax is covered by ram tax

@pratykumar this is the @simonw 's pelican on bike test result for this model

Drop 14a/14: Over the past few days, we’ve been putting our models, products, and research out there. The positive feedback has helped more people agree with our long held belief that #IndiaCan be a builder in this space. Today, we are rolling out Indus - a chat interface to experience Sarvam 105B. Here is what all you can do:

our last line of defense really

@NirantK nirant man, you're f'king around and you know the nuance, right? It's not that everyone rolls out their own CRM, it's that it is cheap enough for *someone else* to roll out a CRM and that erodes margin and it's next to free to have a really good system of record

I wish every single one of my competitors makes their own internal: ATS, CRM, Issue Tracker — and why even stop there, payroll, payments, email and messaging app inhouse! For email — the protocols are even free! Bring back SMTP! /s

@shantanugoel This reminds me of all the people who knew google-fu, and searching the web was an actual skill .... till the tech caught up and the gap vanished. Your prompt might give you some temp edge, but - your prompt alpha is their RL opportunity

"Show me the code" might soon change to "Show me the prompt" to judge how good you are

@rvivek "That's the bar" hahhahaha brother you changed an API endpoint and a few strings, bar seems quite low. You don't have to do QA, it literally works out of the box

Claude Sonnet 4.6 just dropped and HackerRank already supports it in AI-assisted interviews. Less than a day from release to production. That's the bar. 🚀

@regularanomaly @NirantK @adityaag maybe it does? it feels disingenous, like who know how much of it actually happened, versus him just prompting a model to come up with some nonsense viral post also immediately feels low quality, under edited slop, too many words, hardly any point made

@NirantK @adityaag Does it matter anymore?

It is a strange time to hire. I'm seeing this play out across the SPC community in ways that surprise me. We're in the middle of the biggest shift in how knowledge work gets done — maybe ever. And it's made me realize that the old playbook of filtering for pedigree and years of experience is breaking down in real time. I keep on coming back to: the thing that matters most is someone's relationship with change. Not whether they've seen change before — but whether they run toward it. One of our members just did ~20 work trials. Their finding: zero correlation between tenure and adaptability to AI. Younger skewed slightly better, but they watched 15-year industry vets pick it up and absolutely crush it. The divergence in performance wasn't marginal — it was extreme. Another member shared what actually predicted success in their hiring. Cool personal website. Side projects. Builder energy. What didn't predict it? FAANG on the resume. Name-brand university. None of it. A third told me their coding interviews have become a clean filter: you can immediately tell who's actually using Claude Code or similar tools in their daily work vs. who's just been reading about AI. The performance gap isn't 10%. It's 10x. Their interview task is too long to do by hand — so it separates the practitioners from the tourists instantly. These are three different founders, three different contexts, all arriving at the same conclusion: what someone knows today matters far less than how they learn tomorrow. I think about this a lot as an investor too. The founders I'm most excited about aren't the ones with the perfect resume. They're the ones who seem constitutionally unable to stop tinkering. Who get restless when things stay the same. Who treat every new tool like a puzzle they have to solve. That energy is what I'd hire for right now. Not the resume. The restlessness.

Hehe every two month openai gets a vibe shift, I felt like it was so over but now they’re in the so back days

@NirantK @dwarkesh_sp What exactly do you think he said in the podcast that he wouldn’t have said to a private investor? If anything, he had to pad a bunch of numbers and hide true revenue figures

pretty sure if Amodei had gone out to raise this round after the @dwarkesh_sp podcast came out, he'd get more bang for buck

From now on, hype-centric splashy launches will likely be strongly uncorrelated with success. If by the time you launch you don’t have escape velocity, you will likely get Sybil attacked¹. Agents will spin up 10 competing products with your same interface. Start with an audience of 1 and get confirmation that it works. Then expand to a circle of friends or design partners. By the time you go public, your moat needs to be deeper than it’s ever been. ¹ en.wikipedia.org/wiki/Sybil_att…

@cboyack > human effort redirects toward higher-order problems that the tool can't solve alone. Do you have any example of problems that AI won’t be able to solve?

@kr0der i always use 9000 subagents rather than 2 because LLMs aren't deterministic so i like having 9000 reviewers to catch more bugs

get Codex to review its own work with this prompt: "when done, review your changes with 2 subagents, fix any issues, then repeat until no issues found" i always use 2 subagents rather than 1 because LLMs aren't deterministic so i like having 2 reviewers to catch more bugs

@shantanugoel far less...what? tokens? compute resources? secret third thing?

picobot works! Single rust binary, multiple providers, whatsapp, TUI, Websocket integrations, lots of tools, good permissions system. Will add browser use tools soon and i think this will do most of the things that openclaw does but for far less.