Insight | Speed is a feature: why top engineers obsess over latency
FFmpeg's handwritten assembly, Cursor's KV cache and speculative edits, and Perplexity's work on indexing and latency all show teams treating milliseconds as product quality. The pattern explains why performance work, often dismissed as premature optimization, is central to software people love.
What do the FFmpeg, Cursor and Perplexity teams have in common about speed? Show answer
They treat speed as part of the product and design the architecture around it from the start instead of tuning at the end.
How much faster is FFmpeg's handwritten assembly than C, according to Kieran Kunhya? Show answer
Typically 10 to 50 times faster, with one function shown running 62 times faster.
What is unusual about the makeup of VideoLAN's AV1 decoder? Show answer
It contains roughly 30,000 lines of C and 240,000 lines of handwritten assembly, built under a motto that every cycle matters.
Why does software decoding speed matter so much for video players? Show answer
Kunhya estimates up to 45% of files cannot be decoded on the graphics chip and need a software fallback on the processor.
What keeps FFmpeg's aggressive optimization safe? Show answer
Bit-exactness, the rule that every decoder must produce identical output, enforced by an automated test system across many operating systems, compilers and chip architectures.
What design rule does the Cursor team follow about fun? Show answer
A code editor should be fun, and much of fun comes from speed. They throw out experiments that work but are not fun.
Why does Cursor reuse a cache across keystrokes? Show answer
So the model only has to process the newest input. Without it, the team says, they would both destroy latency and overload their graphics processors.
What are speculative edits in Cursor? Show answer
A technique that feeds the existing code back as a draft, which the model confirms in parallel until it reaches the part to change, so edits stream faster.
What latency habit did Aravind Srinivas take from Larry Page? Show answer
Testing under poor conditions. Page tested Chrome on old laptops, and Srinivas tests Perplexity on in-flight Wi-Fi and benchmarks it against rivals.
What does Perplexity measure instead of average latency? Show answer
The slowest 10% and 1% of requests at every component, along with time to first token and throughput for the language model.
Where do the FFmpeg engineers and the Cursor team differ? Show answer
The FFmpeg engineers see low-level optimization as a human craft that current AI models handle poorly. The Cursor team bets that AI will make such engineering much easier.
What evidence gap does the article point out? Show answer
None of the sources gives data on how much latency changes user retention. The claim that speed drives love rests on the teams' own experience.





