The Cognitive Revolution

Snowflake: The Great Data Unlock

"80 to 90% of enterprise information was trapped in PDFs and documents. We are making it queryable for the very first time."

Nathan Labenz

I’d love to go into how AGI-pilled Snowflake is today. Are we talking about a simple database upgrade, or is the core philosophy shifting?

Barish Gultekin

For us, we’re quite practical. We serve large enterprises. The goal is high-quality AI agents that create positive ROI—and that is happening today. Our philosophy? Bring the AI to the data, not the other way around.

90%

of enterprise data is unstructured.

Until now, this was "dark data"—trapped in contracts, research papers, and emails.

The Growth of Searchable Information

From Pipelines to Natural Language

In the "Before" era, you'd build complex pipelines just to classify data. Now, you can extract structure from 100,000 contracts with a simple prompt. You aren't just looking at a dashboard; you're talking to your equities research.

// The Wealth Management Agent Example

"Look up what the stocks are doing in a structured table, but cross-reference it with the sentiment in those 50-page PDFs."

"Data has gravity. Our customers don't want to move it. They want AI to live where the data breathes."

Next: Text-to-SQL and the nuances of semantics

Beyond the Hype: Text-to-SQL

We’ve moved past the "Snowflake 101." Now, we tackle the holy grail of data: asking a question in plain English and getting an answer that doesn't hallucinate your revenue.

"It’s been traditionally very difficult... the margin of error is zero."

I asked where we actually are with Text-to-SQL. Historically, it was a minefield of "tribal knowledge." You have columns nobody uses, schemas that look like a crime scene, and nuances that only live in one analyst's head.

Speaker 1

"If you ask 'What's my revenue?', there is only one answer. The expectations of quality are incredibly high. In the real world, we're talking about thousands of tables and hundreds of thousands of columns."

The Breakthrough

"In the last six months, reasoning models have hit a turning point. We're seeing tremendous demand for Snowflake Intelligence because the quality is finally at a place where you can deploy to business users, not just data scientists."

The "Reasoning Model" Leap (Past 12 Months)

Visualizing the "substantial gains in quality" described by Speaker 1.

How do we solve the "Tribal Knowledge" problem?

Metadata & Schemas

Scanning table names, column descriptions, and the raw data underneath to find the "definition of truth."

BI Connectors

Plugging into Tableau or Omni to scrape existing semantics. If it’s defined in a dashboard, the AI should know it.

Query Logs

Analyzing historical SQL queries. "These are hints for agents to build semantic models."

The Open Semantic Interchange

Snowflake is pushing for an open standard. If you build a semantic model in one platform, you should be able to reuse it in another. It’s early days, but they’re working with Tableau and Omni to ensure your "definition of revenue" isn't locked in a single vendor's basement.

"The walls are coming down."

I brought up the "AGI-pilled" view—are we entering an era where everyone becomes "frenemies"? If Snowflake can do what Tableau does because the pace of software development has exploded, do the old partnerships survive?

Speaker 1's Take: The Pie is Growing, Not Splitting.

There’s a shift in the air. The "lock-in" strategy is dying. When I pressed on the competitive dynamic—the idea that Snowflake might colonize the niches of its partners—the response was surprisingly optimistic.

In an AI-driven world, differentiation isn't about hiding your data; it's about the speed of execution and value creation. With open standards, the customer wins because they can move their semantic models freely.

"There is no lock-in anymore. That makes product development really important. Speed of execution is everything."

It’s a bold claim. Whether the "growing pie" is enough to keep everyone friendly remains to be seen, but for now, the silos are being sledgehammered by reasoning models.

Previously: Snowflake 101 & AI Basics Up Next: RAG, Embeddings, and the Model Layer →

Beyond the Query

We’ve mastered the art of Text-to-SQL, but the real enterprise beast isn't just rows and columns. It’s the ocean of unstructured PDFs, tables, and images. Welcome to the RAG era.

"We’ve been through the eras of chunking strategy and graph databases. But what is actually driving results in the RAG paradigm today?"

The Neva Legacy

At Snowflake, we brought in the Neva team—web-scale search experts. The secret sauce? It’s not just one thing. It’s the quality of the embedding model, the hybrid search layers, and the re-ranker sitting on top.

The PDF Problem

PDFs are messy. Multiple columns, embedded tables, weird image placements. We are moving toward agentic document analytics—where the AI doesn't just find a page, but extracts ten years of revenue data across fifty documents and runs the math for you.

"We’re getting to a point where the 'manual' work of the practitioner is being automated away. The complexity is dissolving into the platform."

The Pareto Frontier: Cost vs. Capability

Eighteen months ago, you needed a specialist model for everything—a table extractor here, a specific embedder there. Today? Claude 3.5 Opus or Gemini 1.5 Pro just solve the problem off-the-shelf.

"If you’re processing a few documents, use the giants. But if you’re processing hundreds of millions of documents? You can’t send that through Claude. You’d go broke, and it would take forever."

This is where the frontier shifts. Snowflake’s document extraction models are orders of magnitude smaller and faster. It’s about Amortized Intelligence—having a model that does one thing perfectly for a fraction of the cost.

The Hot Take

"Fine-tuning is on the decline."

We’re seeing a massive shift. Starting with a frontier model makes the most sense for 90% of use cases. You only go "custom" when the scale is so massive or the data is so unique that the general models fail.

Look at Cursor. At their scale, it makes sense to have a custom model for autocomplete. But for the average enterprise? A well-tuned RAG solution beats a fine-tuned model almost every time.

THE DIALOGUE

HOST

"The Chinese ecosystem—models like Qwen and DeepSeek—is open-sourcing faster than the US. Are customers actually comfortable using them?"

SNOWFLAKE

"It’s idiosyncratic. Some tech companies won't touch them; others see them as incredibly powerful starting points for fine-tuning. It’s a hyper-competitive space, and no one model dominates the proprietary or open-source world right now."

As we solve the "how" of RAG and model selection, the next challenge is structural: how do we bring the models to where the data actually lives?

MongoDB

Think outside rows and columns. Build AI apps fast on a flexible, unified platform trusted by the Fortune 500.

mongodb.com/build

Serval

Cut help desk tickets by 50%. Describe what you need in plain English and let AI automate the repetitive tasks.

serval.com/cognitive

Bringing Models to Data

We've spent the first half of this deep dive deconstructing the RAG stack and embeddings. But the real friction isn't just in the tech—it's in the trust. How do you give enterprises the "latest and greatest" from Anthropic, Google, and OpenAI without the data ever leaving the building?

The Host

"Is there an xAI relationship? If not, why not? And does it have anything to do with them putting women in bikinis all over the place?"

The Guest

Let's start with the narrative violation. Our customers don't want to move their data out of the security boundary. AI needs to come next to the data.

The Architecture of 'In-Situ' Inference

When we started this journey two and a half years ago, the message was loud and clear: enterprise data is stationary. By bringing model inference inside the Snowflake security boundary, we essentially reverse the traditional SaaS gravity. The models become "subprocessors."

There is no state saved. No data is leaked back to the model providers for training. The model providers (OpenAI, Anthropic, Google) own the IP, but the execution happens within a series of guarantees that ensure data residency.

This is where the cloud providers become the linchpin. Since companies like Amazon or Google provide the underlying physical infrastructure Snowflake sits on, we can draw a "dotted line" security boundary that encompasses both the data storage and the GPU-driven inference.

"My working heuristic is: everybody’s hacked, everybody’s pwned... and yet, we haven't seen the weights of a frontier model leak. How?"
— The Host's Skepticism

The guest’s take: It's not about human trust; it's about the "airtight" design of the execution environment. These systems are set up so you simply cannot do anything with the weights other than run inference through them. Access is limited by design, not just by HR policy.

The Myth of Model Commoditization

📍

Data Residency First

Choice is often dictated by corporate geography. If you're an Amazon shop, you might stick to what's inside that cloud boundary, even if a "better" model exists elsewhere.

The "Prompt Lock-in"

Switching isn't just a toggle. Once you have thousands of lines of system instructions and prompts optimized for Claude’s nuance or GPT’s logic, the "leapfrog" effect has to be massive to justify the re-tooling cost.

"The deltas between models aren't huge right now, so the familiarity of the interface and the reliability of the output becomes the real moat."

Moving Up the Stack

The guest argues that while model quality is central, differentiation is shifting to the application layer. Just like ChatGPT or Claude-for-coding, the stickiness comes from the workflow, not just the raw intelligence of the engine.

The Goal?

"How do we bring the most value to customers quickly? We start with the data at the core."

Previously: RAG & Embeddings

Next: Designing enterprise AI agents →

The Agent Spectrum

Moving from "Bringing Models to Data" to "Letting Models Act."

We’ve figured out how to bring the models to the data. But once they’re there, how do they behave? I tend to think about agents on a spectrum. On one end, you have the "Choose Your Own Adventure"—the Claude Code style where you give a goal and the agent just... figures it out. On the other, you have a Linear Structured Workflow.

Claude Code is undeniably awesome, but for most businesses? Don't try this at home. Structured workflows are how you actually get what you want, faster, in a way that doesn't keep the CTO up at night.

The Reliability vs. Autonomy Trade-off

The Persona Play

If you're a product manager asking for last week's usage, you don't want a "blank slate" agent. You want an optimized tool that knows your data schema cold. Structure is a feature, not a bug.

The Scale

Snowflake Intelligence is already hitting 5,000 sellers with this. It's ChatGPT for company data, but with the guardrails of a sales assistant who knows what a "renewal" actually is.

Speaker 0

"Are we doing one big 'long-running' agent, or are we doing the handoff? OpenAI loves the handoff; Tasklet says 'just let Claude cook.' What's actually working?"

Speaker 1

"Even when you 'let Claude cook,' you're developing skills. You're modularizing. In the enterprise, you have a Salesforce agent and a Snowflake agent. They have to talk. MCP and A2A protocols aren't just buzzwords—they're the plumbing for the future."

Status: Early

A2A (Agent-to-Agent)

Customers aren't asking for it by name yet, but they're asking for the result: interoperability.

"One of the biggest considerations is they do not want to be locked in."

Open Standards > Proprietary Silos

The Bridge

MCP

Currently being used as the 'hack' to bridge different continents of data.

"EVERYTHING IS ISOMORPHIC TO EVERYTHING ELSE."

Intelligence is flexible, fungible, and subdividable. Whether you call it an MCP tool or a sub-agent is just a choice of architecture. We're playing 'hide the intelligence' until we find the layout that sticks.

Next up

Trust, governance and guardrails →

The Trust Stack: Governance and Guardrails

We've talked about designing the soul of an enterprise agent, but none of it matters if the C-suite doesn't trust the output. In the enterprise world, 'pretty good' is a liability.

Respecting the Invisible Lines

The most elegant way to solve AI governance isn't by layering a hundred new filters on top. It’s by running AI directly next to the data. At Snifflick, we’ve found that if you respect the granular access controls already in the data platform, the agent follows suit by design.

"The agent should only ever be as smart—and as privileged—as the person using it."

If an HR manager asks for salary data, they get it. If a salesperson asks the same agent about the same colleague, the agent should effectively play dumb. It's the same model, the same orchestrator, but the data governance layer acts as the ultimate guardrail.

Dimension 01

Ease of Use

Trust begins with transparency. If a solution is too complex to build, it's too complex to trust.

Dimension 02

Quality Layers

Utilizing high-quality retrieval (RAG) ensures the context passed to the agent isn't hallucinated noise.

Dimension 03

Owner Verification

Bringing the "Human-in-the-loop" to the UI. Answers tagged as "Verified by Owner" build immediate user confidence.

The Enterprise Trust Hierarchy

Enterprise priorities move from core security outward to monitoring.

S0

Are we holding these agents to a "human-plus" standard? Like self-driving cars, do they have to be significantly better than humans before we adopt them?

S1

I don't think it requires "human-like" intelligence for mass adoption. If talking to an agent feels natural, our expectations skyrocket. But even at current levels, the value being created is so high that adoption is already exploding. It doesn't have to be perfect; it just has to be better at the specific task than the manual process.

"LLMs as Judges."

The frontier of guardrails isn't just static rules; it's using models to score other models across every dimension of a conversation in real-time.

With the safety net secured and the trust stack built, the conversation turns to the horizon.

Up Next: Agents and the Future of Work →

Current Block

Beyond "Vibe Coding": The Rise of the Agentic Workforce

We’ve moved past the era of governance guardrails and into the raw acceleration of 2025. The discourse has shifted: it's no longer about hitting a wall with LLMs—it’s about the threshold where AI starts doing the work for us.

The Observation:

"People have said, oh my god. The coding experience now is like it's not just vibe coding and eventually hitting a wall. You can actually really make this work. Are we all gonna be agent managers, or are we all gonna be talking verbally to agents while getting lots more exercise?"

The Reality Check

"I’d love to see the world where I don't need to do anything and I can just go get more exercise. I actually see the opposite. We're able to do a lot more, and we end up doing a lot more. Everyone is sprinting."

The New Build Philosophy

The traditional UI-first roadmap is crumbling. In a world of coding assistants, the "Working Prototype" has become the only currency that matters.

●Kill the Mockup: Don't design a UI; build a "skill" in a day.
●Rapid Feedback: Put the functional skill in front of customers immediately.
●Solidify Later: Only build the consumer experience once the logic is proven.

"The way we build products has to change. If my users are living in coding assistants, I just need to build a skill for them to call."

The Anatomy of an Autonomous Agent

Capabilities are crossing the threshold from "Copilot" to "Intelligence Entity."

"The first AI Colleague drops in Q2."

The prediction is specific: A remote worker with a name, a Slack account, and email access. While skeptical of the "anything and everything" autonomous agent, the industry is moving toward highly-scoped entities that exist right alongside us in our digital workspaces.

Democratization & The S&P Alpha

Ultimately, for any company, the differentiation isn’t the model—it’s the data. We are seeing two massive shifts in how enterprises handle their internal goldmines. First, the democratization of access: natural language has finally broken the barrier between the C-suite and the database.

Second, we're seeing the creation of "Alpha" from unstructured noise. Take S&P, for example. They are now using AI to analyze earnings calls—not just for what was said, but how it was said.

"They analyzed when CEOs respond directly vs. indirectly to analysts... using that as alpha to determine which stock to buy. This stuff becomes easy to build."

As we look toward the horizon, the focus shifts from the agents themselves to the platforms that house them. If data is the fuel and agents are the engine, the next battleground is the infrastructure that determines who captures the value.

The Supernova and the Black Hole

We’ve moved past the novelty of agents into the cold reality of market dynamics. Are we witnessing a flourishing of new ideas, or are we just doing unpaid R&D for the next generation of mega-platforms?

"I’m no astronomer, but my experience with platforms is that after the supernova comes the black hole... and all that value created on the edges gets sucked back into the center."

The Skeptic

"I told a web agent to scrape 600 pages of Intercom docs, fed them to Gemini, and then used a coding agent to export all my data with one prompt. It's never been easier to unplug. Where are the motes? What prevents platforms from running into trouble when switching costs drop to zero?"

The Strategist

"We love it. It’s great for product teams. If customers aren't happy for a quarter, they leave. That pressure is the ultimate incentive. We're embracing open formats like Iceberg so your data isn't locked in. The engine that processes it—the performance and cost—that's the real competition."

The Death of Lock-in?

There’s a shift happening in the "moat" philosophy. Historically, software companies thrived on inertia—the pain of moving was greater than the pain of staying. But AI agents have essentially automated the "migration" process.

Old World

Proprietary formats, 100-page API docs, and high switching costs.

New World

Open standards (Iceberg), agentic data extraction, and performance-based loyalty.

The "Middle" Erosion

We’re seeing a hollowed-out middle. Traditional businesses that simply "encode business logic" are in the danger zone. If you can "vibe-code" a custom extractor over a weekend, why pay a SaaS company to do it?

Where Value Accrues in the AI Stack

Infrastructure and specialized applications are expanding, while generic business logic "facilitators" feel the squeeze.

The ultimate takeaway? We are entering an era of horizontal platform dominance balanced by hyper-specific vertical applications. The generic SaaS applications—the ones that exist purely to best-practice you to death—are being replaced by agents that can assemble that logic on the fly.

As we look toward the enterprise, the question isn't just about competition; it's about how these massive organizations actually deploy these models without losing their competitive edge. If the "middle" is eroding, how does a Fortune 500 company maintain its unique logic? That's where we're headed next.

Previously: Platforms & Value

The Myth of "Company GPT"

"Why don't we see a PfizerGPT or a GE GPT yet? If these companies have a century of data, why aren't they baking it into the weights?"

The Observation

Databricks bought Mosaic ML for a billion dollars to do "continued pre-training." Then Ali Ghodsi says they basically killed that product. Why? If I’m a 100-year-old company with millions of employees of accrued data, wouldn't a model that has "world knowledge" of *only* my company be insanely valuable?

The "Search vs. Weights" Pivot

Barish explains the shift. It’s exactly like ChatGPT moving from a "knowledge cutoff" to using Web Search. We aren't training the information into the brain anymore; we're teaching the brain how to use a library.

RAG Retrieval Augmented Generation: Cheaper, tunable, and keeps the model "fresh."

Training Codifying info in weights: Expensive, rigid, and goes stale the moment training stops.

Enterprise Adoption Preference (Conceptual)

The Exception to the Rule

Barish notes that training *is* coming back, but for Small Models. High throughput, low cost, task-specific. If you have a specific domain no foundation model has ever seen, you train. For everything else? You retrieve.

"I know my Gmail better than a search engine because I know when I've found it."

The speaker argues that pre-training creates a "familiarity" that RAG can't mimic. It's the difference between looking something up in a book and *knowing* it in your bones.

Internal Benchmarks

Barish ignores public leaderboards. He watches internal quality and latency for specific Snowflake tasks. The signal is in the application, not the ARC AGI score.

Contrarian Take

"Product building needs to radically change."

Coding agents are so capable now that we shouldn't even 'build' the product until it's validated through automated iteration. The old ways are too slow.

Safety Paradigm

"Safety through narrowness."

A fleet of narrow, specialist AIs is easier to control than one god-like general intelligence.

The Executive Mandate

For enterprise leaders still "being careful" about adoption: Stop. The race is on. Barish’s advice is clinical and urgent:

01 Break the Silos: Your data isn't ready for AI if it's trapped in a department-specific cage. Investing in core foundations is the only way to enable the model.
02 Scale requires Ease: If the AI is hard to deploy, it won't be used. Snowflake's entire design principle right now is making high-quality AI "invisible" at scale.
03 Build Intuition Natively: You can't lead an AI transformation if you don't use it yourself. Native usage changes your trajectory.

That wraps up our deep dive into the enterprise engine. Next: Closing thoughts and final takeaways.

Closing Thoughts

"As we move from the complex architectures of enterprise AI and its future outlook, we find ourselves back at the most important element of this journey: the community navigating this revolution alongside us."

1

Thank you, Nathan. Thanks for having me.

0

And thank you to everyone who listens for being part of the cognitive revolution.

Support the Show

If you found value in today's deep dive, consider sharing with friends or leaving a review on Apple Podcasts, Spotify, or YouTube.

Connect with us:

cognitiverevolution.ai

The Turpentine Network

Part of a sixteen z, where experts tackle technology, business, geopolitics, and culture.

Production Credits

"Produced by AI Podcasting. From the moment you stop recording to the moment your audience listens."

aipodcast.ing

Keep Evolving.

Transitioning to final conclusions...