Most software engineering mistakes are easy to miss. Performance engineering isn't one of them.
In this edition of No Vibes Allowed, Dex and Vaibhav show why performance optimization remains one of the hardest tasks for AI coding agents. They walk through benchmark suites, profiling strategies, and runtime optimizations while explaining why small mistakes can completely derail performance improvements.

Key Takeaways
• Why performance engineering demands tighter feedback loops than most software work
• How benchmark suites help AI coding agents make better decisions
• Why AI models struggle with highly sensitive optimizations
• Why copying proven ideas beats reinventing everything
• Why serialization and cached benchmark results matter
• The hidden cost of profiling and measuring performance
• Why one bad assumption can invalidate an entire optimization
• How to use AI without trusting it blindly

Summary
Performance engineering is one of the few areas of software development where intuition isn't enough. Every optimization must be measured, every tradeoff matters, and even a single incorrect assumption can erase hours of work.
In this edition of No Vibes Allowed, Dex and Vaibhav walk through their approach to optimizing the MAML virtual machine and building new profiling infrastructure. Along the way, they show how benchmark suites create the feedback loops needed to make better decisions and why mature runtimes like V8, CPython, and Lua are valuable sources of inspiration.
The conversation covers benchmark design, cache efficiency, memory allocations, string optimizations, profiling architectures, and the challenges of using AI on tasks where tiny mistakes have outsized consequences. The lesson is simple: AI can accelerate research and implementation, but humans still need to provide the judgment that makes performance engineering possible.

TIMESTAMPS
00:00 Why Performance Engineering Can't Be Faked
00:26 Introducing the Episode
01:45 Introducing No Vibes Allowed
02:10 Why AI Struggles With Performance Engineering
04:29 Understanding the VM Execution Pipeline
07:02 The Three Rules Behind Faster Software
08:23 Why Measurement Comes Before Optimization
10:54 Building Benchmark Suites That Actually Matter
16:10 Learning From the Engineers Who Already Solved It
21:30 Why Performance Work Forces You to Read Every Line
23:34 Designing Better Feedback Loops for AI Coding Agents
28:58 Inside the Tricks That Make Modern Languages Fast
33:05 Where AI Actually Excels
33:47 Building a Profiler Instead of Waiting on Benchmarks
36:31 The Tradeoffs Behind Modern Profilers
40:18 The Decisions Humans Still Need to Make
43:08 How One Bad Assumption Can Ruin Everything
45:56 Why Tiny Mistakes Have Massive Consequences
48:09 Turning Research Into Production-Ready Tickets
49:37 Turning an Entire Design Process Into One Diagram
51:50 Why Performance Engineering Makes You a Better Engineer
55:07 How to Start Learning Performance Engineering
58:04 The Fast Inverse Square Root and Doing Different Work
1:00:15 Final Thoughts

TOPICS COVERED
• Performance engineering
• AI coding agents
• Claude Code
• Benchmarking
• Profiling
• Virtual machines
• String optimization
• CPython
• V8
• Lua
• Memory allocation
• Caching
• Runtime design
• Software engineering

HASHTAGS
#AIThatWorks #NoVibesAllowed #PerformanceEngineering #AIEngineering #SoftwareEngineering #ClaudeCode #DeveloperTools #CodingAgents #ArtificialIntelligence