Is the era of Python-only AI over? Mastra CTO Abhi Aiyer breaks down why 1.2 million developers are shifting to TypeScript to build production-ready AI agents, the brutal realities of Y Combinator, and why the "let AI code while you go to the bar" myth is complete BS.

[Main Description]
We’ve always been taught: If you want to build AI, you learn Python. But as the ecosystem shifts from training models to building functional, production-ready AI Agents, the requirements are changing rapidly.

In this episode of The Merge, we sit down with Abhi Aiyer, Co-founder and CTO of Mastra (YC W25), to unpack the wild journey of building one of the fastest-growing open-source AI frameworks. We cover their pivotal rewrite at the Crafty Fox Ale House, the struggle of having zero users at the start of YC, and their brilliant "pocket-sized book" marketing tactic that took over San Francisco.

If you are a web developer, an open-source maintainer, or just trying to figure out how to actually deploy AI agents in production—this is a masterclass you don't want to miss.

🎙️ In this episode, we cover:

Why "Python trains, but TypeScript ships."

The reality of YC: What happens when you get in, but nobody uses your product.

How Mastra scaled to over 1.2 MILLION monthly downloads.

The truth about multi-agent workflows and the "CloudBot" hype.

The commercial open-source playbook: How to monetize and manage 100+ maintainers using CodeRabbit.

⏱️ Timestamps:
0:00 - The "Go To The Bar" AI Coding Myth
1:25 - Welcome Abhi Aiyer: The Origins of Mastra
4:40 - LangChain Frustrations & The Need for TypeScript
7:15 - The NextConf Pivot & The Crafty Fox Ale House Rewrite
10:30 - The Y Combinator (YC W25) Experience & Early Struggles
14:50 - The Viral Pocket-Sized AI Agent Book Strategy
18:15 - Python vs. TypeScript: Why TS is Winning the Agent War
24:30 - Moving AI Docs into the Modules (MCP Innovation)
28:40 - How to Make an Open-Source Company Profitable
33:20 - Managing a Massive OSS Community (Shoutout CodeRabbit!)
40:15 - Real-World Multi-Agent Workflows & Future Predictions
45:30 - Rapid Fire Questions

🔗 Links & Resources:

Check out Mastra: https://mastra.ai

Follow Abhi Aiyer on X: https://x.com/abhiaiyer

Automate your code reviews with CodeRabbit: www.coderabbit.ai

👇 Join the Conversation:
Which side are you on? Are you building your AI agents in Python or TypeScript? Let us know in the comments!

#AIAgents #TypeScript #Python #SoftwareEngineering #YCombinator #OpenSource #WebDevelopment #Mastra #TechPodcast #CodeRabbit

Is the "Benchmark Chasing" era over? With the release of Gemini 3.1 Pro and the specialized Deep Think mode, Google isn't just releasing a faster model—they are introducing a fundamental shift in machine reasoning for real-world developer workflows.

In this episode of The Merge AI Newsroom, live from CodeRabbit’s San Francisco studio, applied AI expert Erfan Al-Hossami (ex-Stability AI, LLM researcher) breaks down why this is Google’s most significant release of 2026.

What we cover in this episode:

The ARC-AGI-2 Breakthrough: Why a 77.1% verified score (and Deep Think hitting ~85%) is the first credible proof of fluid intelligence.

Developer Workflow Shifts: Why task definition and problem framing now matter more than raw syntax coding.

Benchmark Deep Dive: Massive leaps on Humanity’s Last Exam, SWE-Bench Verified, Terminal-Bench, and Codeforces.

Model Strategy: Deep Think vs. Gemini 3.1 Pro—when to use which, plus a breakdown of cost vs. performance trade-offs.

The Future of Agents: Real-world implications for autonomous code review, debugging, and agentic task execution.

Timestamps:
00:00 - Intro: Why Gemini 3.1 Pro feels different
01:41 - ARC-AGI-2 Explained: The most credible AGI benchmark
03:42 - Deep Think vs. Gemini 3.1 Pro: Architecture & UI differences
05:00 - The 2026 Benchmark Gauntlet (SWE-Bench, HLE, & more)
08:40 - Impact on Developers: How your daily workflow changes
15:16 - Context Window Tips & Custom Thinking Controls
19:34 - Token Economics: Model selection & cost strategy
21:19 - What’s next for Google DeepMind + Final Thoughts

Watch the full conversation with Erfan Al-Hossami now 👇

🔗 Join the CodeRabbit Community:
→ Website: https://coderabbit.ai

About The Merge: The Merge AI Newsroom provides expert AI analysis with zero hype. We go beyond the headlines to show you how frontier models actually perform in production environments.

#Gemini31Pro #DeepThink #GoogleAI #ARCAGI #TheMerge #CodeRabbit #AICoding #ArtificialIntelligence #AIBenchmarks #SoftwareEngineering2026

🎙️ The Merge Episode #2: From Psychology to 12,000 Stars with Herrington Darkhome

In this episode of The Merge, Hendrik sits down with Herrington Darkhome, the creator of ast-grep, a lightning-fast structural search and rewriting tool written in Rust.

Discover how a self-taught programmer with a background in cognitive psychology went from discovering Vim on a Chromebook to becoming a core maintainer for Vue.js and building a tool used by tech giants like Microsoft and Amazon.

We dive deep into why Regular Expressions (Regex) fail for large-scale codebases, how Abstract Syntax Trees (AST) are the secret to "ground truth" for AI agents, and why Harrington believes the "open source for love" myth needs to die.

🔍 Inside This Episode:

• Structural Search vs. Regex: Why treating code as a tree is more precise than treating it as a sequence of characters.
• The Rust Advantage: How ast-grep achieves blazing-fast performance and stable concurrency.
• AI & Open Source in 2026: Why human communication and intent are more important than just writing code in the AI era.
• Scaling Knowledge: Using linting as a way to dynamically inject team knowledge into AI agent contexts.
• Monetizing Open Source: The reality of building sustainable, "serious" projects in today's ecosystem.
  

🚀 Level Up Your Code Review

This podcast is brought to you by Code Rabbit, the AI-first code review platform that uses tools like ast-grep to ensure high-fidelity, context-aware reviews.

• Try Code Rabbit for Free: https://coderabbit.ai/
• Star ast-grep on GitHub: https://github.com/ast-grep/ast-grep

🛠️ Resources & Links:

• ast-grep Official Website: https://ast-grep.github.io/
• Follow Code Rabbit on Twitter/X: @CodeRabbitAI
• Join the Discord: (Link found in ast-grep's official docs)

Enjoyed the episode? Support the show by Subscribing and hitting the Bell Icon 🔔 to stay updated on the latest in open source and AI.

#OpenSource #RustLang #ASTGrep #CodeReview #AIAgents #SoftwareEngineering #TheMergePodcast

THE MERGE - AI NEWSROOM
GPT-5.3-Codex vs. Claude Opus 4.6: Benchmarks and Best Agentic Workflows

OpenAI and Anthropic just changed the game for February 2026. But as these models get more "agentic," the stakes for code quality have never been higher. Today on the AI Newsroom, we’re pitting GPT-5.3-Codex against Claude Opus 4.6 to see which model actually earns its keep in a production monorepo.

We’re moving beyond simple autocomplete into the era of "Code Review as the New Coding." We break down the latest benchmarks (SWE-Bench Pro & Terminal-Bench 2.0) and reveal how CodeRabbit’s own internal metrics show a 1.7x increase in defects when AI-generated code isn't properly validated.

WHAT WE COVERED:

GPT-5.3-Codex: Why it’s the "Founding Engineer" of models (speed, iteration, and CLI mastery).

Claude Opus 4.6: The "Senior Architect" approach—handling 1M token refactors without losing the thread.

The CodeRabbit Eval: How we benchmarked these models on signal-to-noise ratio and bug detection.

Agentic Workflows: Parallel "Agent Teams" vs. Hierarchical Orchestration.

🕒 TIMESTAMPS:
0:00 - The Feb 2026 AI Collision 1:45 - GPT-5.3-Codex: 77.3% on Terminal-Bench 2.0 4:10 - Opus 4.6: Why a 1M Token Context window changes refactoring 6:30 - The "AI Code Crisis": 1.7x more defects in AI PRs?
9:15 - CodeRabbit Metrics: Precision vs. Noise in GPT-5.3
12:00 - Pricing Breakdown: $5 vs $25 - The "Intelligence Tax"
14:40 - Pro-Tips: High-context prompting for Senior Devs
17:05 - The Future of Code Review in 2026

💡 KEY TAKEAWAY: GPT-5.3 is built to DO, while Opus 4.6 is built to THINK. At CodeRabbit, we use both, but we always treat their output as a "draft" that requires agentic validation.

🔗 LINKS & RESOURCES:

Our Latest Report: State of AI vs. Human Code Generation 2026 [ https://www.coderabbit.ai/blog/state-of-ai-vs-human-code-generation-report ]

Sign up for free! https://www.coderabbit.ai/

Join our Discord: https://discord.gg/coderabbit

#CodeRabbit #AINewsroom #GPT5 #ClaudeOpus #AgenticCoding #SoftwareEngineering #CodeReview #AI2026

2025 was chaos in the best way: DeepSeek cracked open the model monopoly and proved world-class open weights don't need infinite budgets. Vibe coding went mainstream - prompt your way to an app without staring at code - unlocking ideas for non-engineers but flooding repos with bugs (our data shows AI code spawns ~1.7× more issues than human-written). Agents evolved from demos to long-running beasts, CLI tools like Claude Code let AI run wild in terminals, Cursor/Windsurf supercharged IDEs for pros, Gemini 3 stormed back with killer reasoning, Anthropic scooped Bun, and MCP + Agent Skills started standardizing the agent wars.

Hosted by Hendrik (CodeRabbit Dev Advocate) with David Loker (VP of AI), we dissect the timeline, the hype vs. reality, and why blind vibe coding is creating a maintenance nightmare. David drops hard truths and real predictions on 2026...

Try CodeRabbit: https://www.coderabbit.ai
Blog: https://www.coderabbit.ai/blog
Join our Discord: https://discord.gg/coderabbit

Subscribe and drop topics you want us to test next.

Click here to view the episode transcript.