RESEARCH

→ View original post on X — @tunguz

AI Adds Y Axes to Misleading Graphs

By

@tunguz

–

22 April 2026 18h32

AI can now add the Y axes to every misleading graph out there.

→ View original post on X — @swyx

128k Context Window Ablation Study on Billion-Parameter MoE Models

By

@swyx

–

22 April 2026 18h32

any references come to mind for 128k context on a 1Bish MoE? not sure i understand this area very well. i guess the primary ablation i would want is on context length quality vs width dimension…?

→ View original post on X — @mustafasuleyman

Trillion-fold compute growth: AI training scales toward 2028

By

@mustafasuleyman

–

22 April 2026 18h26

Since I began work on AI in 2010, training compute for frontier models has grown by one trillion times. Now we're looking at something like another thousand-fold growth in effective compute by the end of 2028. 1000x the existing 1,000,000,000,000x. Extraordinary stuff.

MIT CSAIL explores efficient models for complex reasoning at ICLR

By

–

22 April 2026 18h18

This sampling of MIT CSAIL papers at ICLR shows a common need for efficient models that can reason about complex, real-world problems. More compute helps, but the ways these machines "think" also need refinement. You can find more info about these projects on our website:

MathNet: World’s Largest Olympiad Math Problem Dataset

By

–

22 April 2026 18h18

“MathNet” MIT, KAUST & HUMAIN have built the world's largest collection of Olympiad-level math problems. The dataset can help students prepare for competitions & revealed how AI models can improve at math problems, esp. ones w/visual reasoning: https://
bit.ly/4vBr2kl

MIT Improves Reasoning Model Calibration Through Reinforcement Learning

By

–

22 April 2026 18h18

“Reinforcement Learning w/Calibration Rewards” What makes top reasoning models overconfident? MIT found that in these models, RL rewards correct answers, not certainty. Training models to estimate confidence improved calibration while maintaining accuracy:

MIT CSAIL Advances Reliable AI Systems at ICLR Conference

By

–

22 April 2026 18h18

This week, MIT CSAIL will join other top ML researchers at ICLR to tackle a shift in focus from more powerful AI to more reliable systems Our papers at the conference show how to potentially make AI models stronger critical thinkers, more honest, & better at math

MIT Harvard Study AI Agents Critical Thinking Battleship

By

–

22 April 2026 18h18

“Collaborative Battleship” MIT & Harvard developed a collaborative version of Battleship to see if AI agents are as good at asking questions as answering them. They found that many LMs struggle w/critical thinking, but Monte Carlo inference strategies can help even tiny

→ View original post on X — @learnopencv

Position Encoding: How Transformers Understand Data Order

By

@learnopencv

–

22 April 2026 18h06

Position Encoding: How Transformers Understand Order in Data

In this episode of Artificial Intelligence: Papers and Concepts, we explore Position Encoding, a fundamental concept that enables transformer models to understand the order of information. Since transformers process… pic.twitter.com/FS8RcTE7Dt
— Satya Mallick (@LearnOpenCV) 22 avril 2026

Position Encoding: How Transformers Understand Order in Data In this episode of Artificial Intelligence: Papers and Concepts, we explore Position Encoding, a fundamental concept that enables transformer models to understand the order of information. Since transformers process