Recording intelligenceAI-generated brief · check the source for context

Small Language Models, Big Bets | Prithvi Raj - 360Labs.ai

35:44 recording · AUTO · 2 speakers

Watch the original

Executive Summary AI
  • Prithvi defines AI native as rethinking every single process of what would have happened before an LLM existed, and says his shop qualifies even on projects with no AI features because every engineer comes from an AI research or AI engineering background and every coding workflow runs on something like cloud code or a codex.
  • He illustrates the 24-hour prototype claim with a call the day before yesterday with a public steel company that gave a one-sentence brief for an AI-native CRM and ERP: the CTO parsed their website and competitors into a single PRD, a senior engineer sprinted for two hours, and the demo call left the client asking which competitor the product had been sold to before.
  • Asked how a team of 10 scales against TCS with 600,000 employees and about $28 billion in revenue and Infosys with about 300 plus 1000 employees, he answers that scale is solved through team, citing Steve Jobs calling the Macintosh team his greatest product and a public Danish logistics company where his team sat with the control tower manager and shipped a SaaS platform within a week.
  • On India's AI bet he argues language has always been an abstraction of reality rather than reality itself, so future foundational models will be world models or physics-based models, and India's better use of resources is fine-tuned, on-device, multilingual small models — his own two senior engineers work from 12 midnight to 4am on an AI-native Linux operating system with SLM 360 embedded, with de
  • He places India's play at the application layer, notes that people don't like paying for software in India (GPT dropped its price to 300 rupees) and that after Dr. Saab, a WhatsApp-first AI doctor, they stepped back from consumer because a company putting 50 people on one product will always outscale them, returning instead through Omni AI, Crawl360 and hardware plans.

Brief overview

AI-native means rethinking every pre-LLM process; small fast teams beat IT giants, and applications beat foundational models.

  1. AI native is a process rewrite, not a featureEven an app with no AI in it counts if every workflow and hiring decision was rebuilt around what exists post-LLM.
  2. Speed is the only defensible advantageHe rejects moats outright: nothing stops Infosys spinning up a boutique arm, so the answer is always team and always being faster.
  3. India should build applications, not foundation modelsSarvam raised 53 million with government backing, but he'd rather see fine-tuned, on-device, Indian-language small models and consumer products.

Questions this recording answers

10 questions, each answered where it is said
What does being AI native actually mean operationally?

It means rethinking every process as it looked before LLMs and making it fundamentally different after. Even for an app with no AI features, every engineer has an AI research or engineering background, the tech stack is chosen for what will last, and all coding workflows use tools like Claude Code or Codex to ship faster.

How can a team deliver a working prototype within 24 hours?

For a public steel company that gave a one-sentence brief for an AI-native CRM and ERP, the CTO parsed their website and competitors, compressed it into one PRD, and a senior engineer sprinted for two hours. They got on a demo call immediately; the client asked who it had been built for before. It was custom.

How can a ten-person team compete with IT giants like TCS and Infosys?

Through speed. For a public Danish logistics company wanting something in two to three weeks, the team sat with the control tower manager, documented his day, turned it into a PRD and had a SaaS platform within a week. Large IT services firms have too many layers of decision making to move that fast.

Is there a moat against big IT firms creating small boutique teams?

He rejects the idea of a moat, especially for software companies. The only moat is staying ahead: outworking, being faster, and hiring the best people. Exceptional engineers, seen across 3,000+ applications, want ownership and end-to-end products, so talent will condense into small teams. But nothing stops Infosys from starting a boutique consultancy.

Should India invest in building foundational LLMs?

He doubts large foundational language models are the best bet. Language is an abstraction of reality, so future foundational models will be architecturally different: world models, physics-based or robotics models. More Indian LLMs like Sarvam would be great, but the better use of resources is fine-tuned, on-device, multilingual models for India-native use cases.

Why would anyone need on-device AI when connectivity is everywhere?

Two reasons. First, low or no internet: pilots with the Indian Army, Indian Navy and a ministry need real-time feedback off grid, as do aviation, shipping and logistics. Second, data privacy: executives holding critical data want intelligence without their chats leaving the device or relying on third-party API calls.

What AI-native operating system is the team building?

Two senior engineers work from midnight to 4am on a Linux-based operating system with SLM 360 embedded, meant to run on any blank Android phone. Owning both the OS and AI layers lets it handle complex instructions like finding the third photo from last month or a specific follow-up email in a thread.

Why build products like Mem360 and Med360 instead of niching down?

Seeing patterns across many sector builds shows which foundational products help all clients. A Mem0 subscription cost upwards of 1,500 a month, so they built Mem360, a memory layer for AI agents used across builds. Med360 compresses a 1.6–1.7 trillion parameter model's medical intelligence into a model a quarter or fifth the size.

What is Omni AI and why was Crawl360 built?

Omni is an assistant CEOs asked for: aware of Gmail, WhatsApp, Google Chat and Slack, able to send emails, make calls and reservations, with custom skill files. Building it required a world-class scraper, so they built Crawl360, which returns any URL as JSON, PDF, YAML or audits, via a five-tier captcha-resilient agent.

Where should India focus if not on foundational models?

The application layer, since no one understands Indian consumers better than Indians. Consumer is hard because many Indians do not pay for software, as their WhatsApp AI doctor attempt showed. Other areas: foundational AI tools like memory layers, developer tools, small models and Indian-language native models.

Key Quote

“When we say AI native, so that effectively for us means rethinking every single process of what would have happened before an LLM existed.”

— Prithvi, Small Language Models, Big Bets | Prithvi Raj - 360Labs.ai, at 0:13▶ watch at 0:13
Key Quote

“So I actually don't think the future belongs to LLM.”

— Prithvi, Small Language Models, Big Bets | Prithvi Raj - 360Labs.ai, at 1:15▶ watch at 1:15
Key Quote

“the main job of the founder is effectively just like recruitment and team building”

— Prithvi, Small Language Models, Big Bets | Prithvi Raj - 360Labs.ai, at 8:56▶ watch at 8:56
Key Quote

“Like the more exceptional the engineer, the less he actually wants to contribute to a large code base.”

— Prithvi, Small Language Models, Big Bets | Prithvi Raj - 360Labs.ai, at 13:53▶ watch at 13:53
Key Quote

“So language is always been like an abstraction of reality, not reality itself.”

— Prithvi, Small Language Models, Big Bets | Prithvi Raj - 360Labs.ai, at 16:46▶ watch at 16:46
Key Quote

“the main interest that we got for SLM has been from defense”

— Prithvi, Small Language Models, Big Bets | Prithvi Raj - 360Labs.ai, at 19:49▶ watch at 19:49

Cite this

“Small Language Models, Big Bets | Prithvi Raj - 360Labs.ai.” https://www.youtube.com/watch?v=kU7PRu0mFbM. Transcript summary by WhipScribe, https://whipscribe.com/transcript-pages/small-language-models-big-bets-prithvi-raj-360labs-ai.html.

Report this page
What is wrong with this page?