Why Is AI Writing So Hard to Read? A New Paper Measured It
Why is AI writing so hard to read? A new paper measured 14 LLMs on six features linked to reading difficulty and traced the cause to length pressure in RL.
Research
AI Carb Counting Accuracy: I Tested Soba on 400 Meal Photos
AI carb counting accuracy, measured: Soba got 68.5% of meals within 10 g on 400 research photos. Where it fails, what fixes it, and how 8 LLMs compared.
Faster Than Dijkstra: The New Shortest-Path Algorithm, Explained
A new shortest path algorithm finally beats Dijkstra on sparse graphs. How it breaks the 65-year sorting barrier, and why Dijkstra still wins in practice.
Context Engineering: Why Less Context Makes Better Agents
Context engineering explained: why long-running AI agents suffer context rot, what 2026 research shows about pruning and compaction, plus a Python demo.
Stanford AI Index 2026: 12 Findings, Dissected
Stanford's AI Index 2026: SWE-bench hit 100%, entry-level dev jobs down 20%, $581B invested, US-China gap at 2.7%. 12 findings, broken down.
Project Glasswing: Claude Found 10,000 Critical Bugs in 8 Weeks
Project Glasswing gave Claude Mythos to 200+ orgs; it flagged 10,000 critical bugs in 8 weeks at a 90.6% true-positive rate. Why Anthropic keeps it locked.
Efficient LLM Reasoning: 7 Papers That Cut Token Costs by Up to 84%
Seven papers fix LLM overthinking: Sketch-of-Thought cuts tokens 84%, shorter chains boost accuracy 34.5%, and budget-aware prompting halves costs.
THINC: How a 4B Model Beat 235B Qwen3 by Reasoning in Code
THINC trains a 4B parameter model to reason entirely in code. It scored 78.1% on competition math, beating Qwen3-235B at 75.2%. Here's how the method works.
4.7x Faster Long-Context LLM Inference: AsyncTLS Explained
Long-context LLM inference collapses past 96k tokens. AsyncTLS pairs two-level sparse attention with async KV offloading for 1.3-4.7x more throughput.
Recursive Language Models (RLMs): How MIT's REPL Trick Handles 11M-Token Prompts
Recursive language models keep a huge prompt in a Python REPL and recurse over the pieces. RLM(GPT-5) scores 91.3 on 6M-11M-token inputs plain GPT-5 can't fit.