Everyone is watching the model leaderboards, but that’s the wrong battle. Today’s newsletter breaks down the signal in the noise of AI productivity.

New Cohort of AI Engineer HQ Starting on 24th September 2025, 8:30 PM IST. Register Here(starting soon).

In today’s edition:

• Signal vs Noise— Model vs. Model [Not the Real AI War]

• Build Together— Here’s How I Can Help You

[Sponsor Spotlight]

It’s go-time for holiday campaigns

Roku Ads Manager makes it easy to extend your Q4 campaign to performance CTV.

You can:

• Easily launch self-serve CTV ads

• Repurpose your social content for TV

• Drive purchases directly on-screen with shoppable ads

• A/B test to discover your most effective offers

The holidays only come once a year. Get started now with a $500 ad credit when you spend your first $500 today with code: ROKUADS500. Terms apply.

Get started today.

[Signal vs Noise]

Model vs. Model [Not the Real AI War]

Every week, my inbox is full of benchmarks.

GPT-4o vs. Claude 3.5 vs. Llama 3

Everyone wants to know who is winning the AI model race.

To be honest, this is the biggest distraction in AI right now.

The companies that will dominate the next decade of AI are not the ones building slightly better models.

They are the ones building the plumbing, the infrastructure that makes AI actually useful in the real world.

Let’s signal in.

[The Noise] Foundation Model Wars Are Ending

The idea that one company will own a single “best” model is already outdated.

Foundation models are quickly becoming a commodity, like electricity or cloud computing.

You don’t win by having slightly better electricity, you win by what you build on top of it.

The signs are everywhere:

• training a frontier model is getting 10x cheaper every year

• deepseek’s R1 model showed GPT-4-level reasoning at a fraction of the cost

• just a year ago, GPT-4-level inference cost ~$60 per million tokens

• today, you can get similar performance for a fraction of that

• aws bedrock now offers more than 100 models, swapping between them is as simple as changing a line of code

• for most practical use cases like reasoning, coding, analysis, the difference between top models is small enough that businesses barely notice

The Model vs. Model debate is quickly becoming irrelevant.

That’s the noise.

[The Signal] Infrastructure Is the Real Battlefield

If the model is the engine, then infrastructure is the rest of the car.

The transmission, the chassis, the steering system.

Without those, the engine is useless.

In AI, infrastructure is what allows agents to actually do things:

• connect to tools

• reliably run workflows

• integrate with company data

This is where the real war is being fought.

I want you to understand these 2 pointers:

1) Frameworks

Frameworks are the developer toolkits for building AI agents.

Each comes with a different philosophy:

LangChain

• connects to over 70 tools and integrates with everything

• perfect for complex enterprise workflows

• it can get complicated and introduce security risks if not handled carefully

CrewAI

• specializes in multi-agent setups where different AI agents play specific roles

• great for structured, collaborative tasks

LlamaIndex

• shines when agents need to connect with large collections of documents or knowledge bases

• boosted retrieval accuracy by 35%, making it the top choice for RAG-heavy systems

The choice of framework has a bigger impact than the choice of foundation model.

2) The 3 Pillars of Agent Infrastructure

To move beyond demos into production, AI systems need three layers of plumbing:

1 - Tools Layer

How agents interact with the world.

APIs, secure authentication, browsers, and new standards like the model context protocol (MCP).

Without this, an agent is just a chatbot stuck in a box.

2 - Data Layer

The agent’s memory and knowledge.

Vector databases, real-time data access, context routing, and private company data pipelines.

This is where intelligence compounds.

3 - Orchestration Layer

The system brain.

It manages tasks across agents, recovers from errors, balances workloads, and ensures the whole thing works at scale.

Infrastructure Is Harder Than It Looks

Switching from GPT-4 to Claude 3.5 is easy.

Changing one API call and you’re done.

But building secure, scalable infrastructure that integrates across multiple systems

That’s the real engineering challenge.

Enterprise surveys back this up:

• 53% of leaders and 62% of engineers cite security as their top concern with agents

• 42% of enterprises require connecting to 8+ different data sources before AI can be useful

• 86% say they need upgrades in their tech stack before deploying AI agents at scale.

They are infrastructure problems.

And solving them is where the winners will emerge.

Actionables

So how should you adapt to this shift?

Here’s a practical guide based on your role.

For Leaders, PMs, and VPs

• move investment from which model to infrastructure and integration talent

• instead of Are we using GPT-4o?, ask

  • how reliable is our orchestration layer?

  • can our agents securely access our databases?

• run small experiments with langchain, crewAI, and llamaIndex to find the best fit for your workflows

For Engineers

• master one.

• expertise in LangChain or LlamaIndex is more valuable now than just prompt engineering

• create secure reusable tools (a connector to shopify, salesforce, or your company’s internal wiki)

• learn logging, monitoring, error handling

• production AI lives or dies on orchestration reliability

Final Thoughts

The foundation model wars are winding down.

They’ll keep making headlines, but they are noise.

The signal is that the real value is moving up the stack, to infrastructure, frameworks, and orchestration.

The companies that dominate the next decade of AI won’t be the ones with the best model.

They’ll be the ones that solve the hard infrastructure problems and build reliable systems that connect AI to the real world.

Stop obsessing over the engine.

Start mastering the plumbing.

The model war is already over.

Winners won’t be model makers.

Happy AI

Signal vs Noise

This is the series in which I take the single AI topic of the week and deliver my sharp and original analysis.

If you want my analysis on any topic, reply with your topic to this email.

Want to work together? Here’s How I Can Help You

• AI Engineering & Consulting (B2B) at Dextar—[Request a Brainstorm]

• You are a leader?—Join [The Elite]

• Become an AI Engineer in 2025—[AI Engineer HQ]

• AI Training for Enterprise Team—[MasterDexter]

• Get in front of 5000+ AI leaders & professionals—[Sponsor this Newsletter]

I use BeeHiiv to send this newsletter.

Starting a new cohort of AI Engineer HQ on September 24th, 2025 [8:30 PM IST]