Research

10 articles

briefingMar 22, 2026

Can Google's free AI assistant actually save me time?

Can Google's free AI assistant actually save me time?.

3 min read
briefingMar 21, 2026

What happens when you cut a model's thinking short?

What do new nuclear reactors mean for waste?. Plus: 14 more stories and 4 research papers and 5 X posts from AI accounts

4 min read
briefingMar 20, 2026

What does Cursor's in-house model mean for everyone who builds with AI?

What do new nuclear reactors mean for waste?. Plus: 14 more stories and 4 research papers and 8 X posts from AI accounts

4 min read
briefingMar 19, 2026

What does difficulty-aware compute allocation change for you?

What do new nuclear reactors mean for waste?. Plus: 14 more stories and 4 research papers and 6 X posts from AI accounts

5 min read
researchMar 13, 2026

Scaling Test-Time Compute: How Thinking Tokens Change Everything

A new generation of 'thinking models' allocates compute dynamically at inference — the research behind this shift and what it means for system design.

4 min read
researchMar 10, 2026

Multimodal Reasoning: Vision-Language Models Move Beyond Description

The latest vision-language models don't just describe images — they reason about them, opening new categories of practical application.

4 min read
researchMar 06, 2026

Reasoning Without Chain-of-Thought: What Implicit Reasoning Means for Efficiency

New research demonstrates that large language models can maintain most of their reasoning accuracy without generating explicit chain-of-thought tokens — and the implications for inference cost and latency are significant.

4 min read
researchFeb 22, 2026

Synthetic Data Hits the Wall: What the Scaling Limits Mean for Model Training

Multiple research groups have independently documented diminishing returns from synthetic data at scale, revealing a ceiling that will reshape how the next generation of models is trained.

4 min read
researchFeb 18, 2026

The Alignment Tax, Measured: What Safety Training Actually Costs in Capability

New research from Anthropic and DeepMind independently quantifies the capability cost of alignment training, finding it lower than feared but unevenly distributed — with important implications for how we think about the safety-capability tradeoff.

4 min read
analysisFeb 03, 2026

The 2026 International AI Safety Report: Frameworks Doubled, But Defences Remain Bypassable

The second annual International AI Safety Report finds frontier safety frameworks have doubled since 2025 but remain vulnerable to sophisticated attackers — here's what practitioners need to know.

4 min read