Marko
@marko-tech.bsky.social
Writing about large-scale AI architecture, recommender systems, and getting ML from notebooks into production.
80% of enterprise AI projects never reach production.
And selling AI will never be as easy as it is today.
Here's why the window is closing fast 🧵
And selling AI will never be as easy as it is today.
Here's why the window is closing fast 🧵
February 6, 2026 at 9:41 AM
80% of enterprise AI projects never reach production.
And selling AI will never be as easy as it is today.
Here's why the window is closing fast 🧵
And selling AI will never be as easy as it is today.
Here's why the window is closing fast 🧵
1/ Snapchat's "People You May Know" is eerily accurate. Behind it is a map of 900 million people and 16.8 billion connections. Just storing the IDs eats 800GB of RAM, before the AI even starts. Most research breaks at this scale. Snapchat cracked it. Here's how 🧵
February 6, 2026 at 8:36 AM
1/ Snapchat's "People You May Know" is eerily accurate. Behind it is a map of 900 million people and 16.8 billion connections. Just storing the IDs eats 800GB of RAM, before the AI even starts. Most research breaks at this scale. Snapchat cracked it. Here's how 🧵
50% chance of AGI till 2035, 50% chance of Chapter 11.
February 6, 2026 at 6:40 AM
50% chance of AGI till 2035, 50% chance of Chapter 11.
What if the best way to teach AI to reason is to never tell it when it's right—only when it's wrong? 🧵
New Princeton research flips everything we thought about reinforcement learning on its head.
Here's why punishment might beat praise:
New Princeton research flips everything we thought about reinforcement learning on its head.
Here's why punishment might beat praise:
February 5, 2026 at 6:35 PM
What if the best way to teach AI to reason is to never tell it when it's right—only when it's wrong? 🧵
New Princeton research flips everything we thought about reinforcement learning on its head.
Here's why punishment might beat praise:
New Princeton research flips everything we thought about reinforcement learning on its head.
Here's why punishment might beat praise:
X open-sourced Phoenix, the latest version of its Recommendation System. You can see how it "thinks", but not what it "knows". Phoenix replaced 100+ handcrafted features with a transformer that learns from your last 522 actions. Your first 30 mins now decide everything.
shorturl.at/ojG83
shorturl.at/ojG83
X Open-Sources Its Recommendation Algorithm: Architecture, Gaps, and What They're Not Telling You
As Elon Musk previously announced, X has just published the latest version of its recommendation algorithm, Phoenix (source: https://github.com/xai-org/x-algorithm).
www.linkedin.com
February 5, 2026 at 9:47 AM
X open-sourced Phoenix, the latest version of its Recommendation System. You can see how it "thinks", but not what it "knows". Phoenix replaced 100+ handcrafted features with a transformer that learns from your last 522 actions. Your first 30 mins now decide everything.
shorturl.at/ojG83
shorturl.at/ojG83
Where did all the “this is huge” breakthroughs from last week go?
February 4, 2026 at 4:16 PM
Where did all the “this is huge” breakthroughs from last week go?