machine-learning.
29 writings found
Latest Archives
How Meta Rewrote Recommendation Systems From Scratch With Index as Model
Inside SilverTorch: Meta's radical shift from microservices to a single neural network for retrieval
SilverTorch and the Death of the Recommendation Microservice
Meta's new Index-as-Model paradigm replaces microservice retrieval with a single PyTorch neural network, and the implications for AI developers are massive.
SilverTorch: How Meta Rewrote the Rules of Recommendation Systems
Meta's radical shift from microservices to a unified neural network transforms retrieval at scale.
The Hidden Complexity Behind Meta's Friend Bubbles
A deeper look at why seemingly simple features demand the deepest engineering work
When Simple Features Hide Complex Engineering: Lessons from Meta's Friend Bubbles
Exploring the hidden complexity behind Meta's Friend Bubbles feature and what it reveals about modern social platform engineering.
Why Simple Features Break Engineering: Lessons from Meta's Friend Bubbles
Friend Bubbles seemed straightforward, but required deep ML work. What this teaches us about 'simple' features in production systems.
Facebook's Search Problem: Why Keywords Can't Find Cupcakes
Meta rebuilt Facebook Groups search from scratch, blending dense embeddings with inverted indices. Here's why that matters for community search.
Project Maven: How AI Turned War Into a Database Query
The military's AI targeting system can now hit 5,000 targets daily. We should talk about what happens when warfare moves at database speed.
Facebook's Hybrid Search: When Keywords Meet Neural Embeddings
Meta rebuilt Facebook Groups search by merging traditional keyword matching with dense vector embeddings, then used Llama 3 to validate the results at scale.
Facebook's Hybrid Search: When Keyword Matching Isn't Enough Anymore
Meta's Groups search now blends lexical precision with semantic understanding. Here's why traditional keyword matching is dying and what comes next.
Facebook's Hybrid Search: When Keyword Matching Meets Neural Understanding
Meta rebuilt Facebook Groups search by blending traditional inverted indices with dense embeddings, then used Llama 3 to grade results at scale.
Sony's Ace Robot Is Beating Professional Table Tennis Players, and That's Actually Harder Than Beating Humans at Chess
Sony's AI-powered Ace robot defeats pro table tennis players using 12 cameras and 8 joints. Why physical games are harder for AI than Chess or Go.
Meta's KernelEvolve: When AI Writes Its Own Performance Code
Meta's KernelEvolve system uses AI agents to automatically optimize low-level hardware kernels, achieving 60% performance gains in hours instead of weeks.
Meta's Adaptive Ranking Model: The Real Cost of Serving Trillion-Parameter Ads
Meta scaled ads recommendations to LLM complexity while keeping latency under a second. Here's why their inference trilemma solution matters beyond advertising.
Why Your AI Benchmark Is Probably Wrong: The N,K Trade-off
Google Research reveals why using 3-5 human raters per item isn't enough for reproducible AI evaluation. The depth vs breadth problem explained.