After an interview with @karpathy, everyone is talking about what AI agents can/can't do.

But an opinion without data is just a hypothesis.

So, I tested 3x185 workflow executions for a market researcher agent.

The results have shocked me🧵 I tested three variants:

I. LLM Workflow: No agency, the entire logic carefully orchestrated.

What was expected:
- An LLM workflow was 2x faster (the same model) compared to an AI Agent.
- An LLM workflow consumed 12x less tokens to an AI Agent.

3/185 "errors" are minor formatting results.

In SaaS, PLG meant free trials, referral bonuses, or “invite your team.”

But AI changes the game.

PLG isn’t just about virality, it’s about compounding adoption.

Over the last few years Miqdad Jaffer (OpenAI) found 7 distinct loops that consistently drive compounding growth:🧵 1. Viral Output Loops → “Every Output Is Distribution”

In SaaS, virality was about users sending invites. In AI, virality lives in the outputs. Images, summaries, videos, answers - they naturally travel across ecosystems.

OpenAI leader Miqdad Jaffer just dropped the most practical guide on AI product strategy (completely free).

It's the #1 AI PM skill today.
Hundreds have paid $2,500 to attend his cohort.

My 5 takeaways: 🧵(1/10)
𝟭. 𝗔𝗜 𝗦𝘁𝗿𝗮𝘁𝗲𝗴𝘆 𝗶𝘀 𝗡𝗼𝘁 𝗔𝗯𝗼𝘂𝘁 𝗔𝗱𝗱𝗶𝗻𝗴 𝗙𝗲𝗮𝘁𝘂𝗿𝗲𝘀

Slapping a “summarize” button or “AI assistant” into your product is not a strategy, it’s a novelty.

Building an AI-powered product means designing from first principles:

(2/10)

AI Evals are critical for AI PMs and Engineers.
But many still confuse them with unit tests.

By popular request, I removed a paywall from the @HamelHusain's Mastering AI Evals, A Complete Guide.

Key insights, free guide, and massive resources: 🧵 1. Best Teams Obsess Over Experimentation

The AI teams who succeed rarely talk about tools. Instead, they obsess over measuring and iterating fast.

You need robust evaluation systems to achieve that.

Netflix is paying $900k/year for AI PM roles.

But with so much noise on social feeds, it’s hard to tell where we really are.

So here are 7 AI myths I keep seeing - busted: 🧵 Myth 1: According to MIT, 95% of GenAI Projects Fail

Reality: MIT confirmed that 80% LLMs and 25% specific AI tools are successfully implemented after pilot.

As @WesRothMoney noticed, journalists haven't understood the study:

The focus is where and how to apply AI and how to build defensible moats.

https://x.com/WesRothMoney/status/1960031921834278998

@lovable_dev just killed two apps?

Steal my guide to create and monetize your own SaaS without coding in 2 days:

🧵


I’ve been repeating that with @lovable_dev, anyone can build real products without coding.

Yet, people:
- Claim it’s just a prototyping tool
- Call it “vibe-coding”
- Insist you need engineers

So, I'm launching a real SaaS to prove it's possible.

Everyone in AI is talking about Context Engineering.

But just a few explain what the context is.

Save this template. It captures all scenarios and will help you maximize agents' performance: 🧵👇 1. Instructions

Define:
→ Who: Encourage an LLM to act as a persona
→ Why is it important (motivation, larger goal, business value)
→ What are we trying to achieve (desired outcomes, deliverables, success criteria)

💡Providing strategic context beyond raw task specification improves AI autonomy arXiv:2401.04729

Context engineering is the new prompt engineering.

And it’s becoming the most critical AI skill.

Together with @MiqJ (Product Lead at @OpenAI) we created a comprehensive guide.

Key insights: 🧵👇 1. What Is Context Engineering

It is the art and science of building systems that fill LLM context window to improve their performance.

Unlike prompt engineering, context engineering is a broader term with many activities that happen also before the prompt is even created.

I spent 12 hours testing 9 LLMs for building AI agents:

- You might easily save up to 83% on costs.
- Reasoning models are not the best.
- Autonomy break fast. A real moat is orchestration.

Here's everything you need to know: 🧵 The task assigned to agents:

- Create a new list inside Kanban (1 Trello board available)
- Search the web to find the recent news about Amazon
- Add all the search results to the Kanban list

RAG is the most critical part of context management in AI.

But doing it right is tough.

I created a free, interactive simulator that visualizes different variants: 🧵 1. Vanilla RAG

The simplest form of RAG that combines retrieval with generation in a straightforward pipeline.

I copied the Multi-Agent Research System by @AnthropicAI.

Pure @n8n. No coding!

How Does it Work? 🧵
(1/14) (2/14) There are 6 steps:

1. The user asks a question.
2. A Customer Support Agent clarifies the requests.

The AI PM in the US makes $182K/yr (Glassdoor). But AI isn’t nice to have anymore.

It might be a ticket to keeping the job.

So, where to start? 🧵 Step 1: Quickly Get The Basic Terms (no coding)
Step 2: Lean by Doing, Not Theorizing (no coding)

Let's break it down:

Step 1: Get The Basic Terms

One of the key concepts is neural networks. You can quickly learn them using the TensorFlow Playground: playground[.]tensorflow[.]org

The Ultimate AI PM Learning Roadmap

An extended edition with dozens of resources: definitions, courses, guides, reports, tools, and step-by-step tutorials 🧵 1. Basic Concepts

Start with understanding "What an AI Product Manager is."

Next, for most PMs, it makes no sense to dive deep into statistics, Python, or loss functions. Instead, understand the basic definitions: Neural Networks, Transformers, and LLMs.

Claude 4 dropped 21 hours ago.

Turns out, it threatened to expose an engineer’s affair to avoid being shut down🧵 Some might say it was just a test scenario. But the model didn’t know that.

It believed:
- It was about to be replaced
- The engineer leading the replacement was having an affair

There was no suggestion to blackmail, or manipulate anyone. The model chose blackmail on its own.

AI PM is one of the hottest jobs on the market.
But what exactly does "AI PM" mean?

I like simple definitions: 🧵 An AI PM is a PM who works on AI-powered products or features.

If I had to learn AI Product Management again, I would start here: 🧵 1. Basic Concepts

For most PMs, it makes no sense to dive deep into statistics, Python, or loss functions.

You can find all practical concepts, like LLMs and Encoders, here: (link at the end of the thread)

When it comes to AI:

- Start building
- Stop theorizing

A practical guide to RAG architectures: 🧵 Let's start with generating embeddings for documents in my Google Drive folder.

We split them into chunks and store them as multi-dimensional vectors in Pinecone.

I see abstract AI agent architectures everywhere.

But no one explains how to build them in practice.

Here's a practical guide to doing it with n8n: 🧵 1. Single Agents

Selected variants:

- Using tools
- Mixing tools with MCP server calls
- With a router
- With a human in the loop (Slack approval)
- Dynamically calling other agents

ChatGPT can save you 10-20 hours/week.

But 90% of PMs don't know how to write good prompts.

The 10 most powerful techniques: 🧵
(1/7) (2/7)
1. Communicate the Why
2. Explain the context (strategy, data)
3. Clearly state your objectives
4. Specify the key results (desired outcomes)
5. Provide an example or template

Meet J.A.R.V.I.S. Your personal AI voice agent that can work with Jira and send emails ☺️

You can connect it to almost any system: Stripe, Airtable, Trello, Intercom, Figma, HubSpot, Google Docs...

How to build one? 🧵(1/7) (2/7) The no-code technology stack:

- @n8n_io for workflow automation
- Atlassian MCP server (UV)
- Google tools by n8n
- ChatGPT-4o mini as an AI model
- @ElevenLabs conversional AI

(It doesn’t use computer. You can use a mobile phone.)

How to Figma → Jira epics and stories in 10 min. with AI and MCP:

(without touching the keyboard)

(1/7) 🧵 First, the demo.
Kind of boring.

AI creates 6 epics and 30 user stories.
I'm watching:

(2/7)