Artificial Intelligence
OpenAI's Math Retractions Expose the Verification Bottleneck
OpenAI had to pull three AI math papers within 24 hours, proving that producing reasoning is cheap while rigorous verification remains the real bottleneck.
Topic
Machine Learning
A collection of essays to explore, one idea at a time.
Essays
Newest first
AI Agents
Why AI Agents Must Edit Their Own Context Files
Context Language Models prove that letting AI agents edit their context window like a live file beats brute-force append-only transcripts.
AI Agents
Why Computer Use Needs System 1 Models, Not Massive LLMs
Massive frontier models are overkill for GUI clicks. CUA-S1 proves tiny 706k-parameter models offer faster, cheaper desktop automation.
Artificial Intelligence
Ternary LLMs Shrink 27B Reasoning to Laptop Scale
PrismML's Bonsai 2 27B achieves 98.2% parity with FP16 models in just 5.9 GB, proving ternary weights can run high-grade AI locally.
Artificial Intelligence
Why AI Is Mining Mathematics as a Non-Renewable Resource
Terence Tao warns that automated AI solvers risk exhausting fruitful open math problems, threatening open science and genuine mathematical insight.
AI Architecture
Continuous Diffusion LLMs: Breaking the Autoregressive Mold
Continuous Diffusion Language Models (CDLMs) are staging a major comeback, challenging the left-to-right next-token generation paradigm.
AI Security
Stealing CoT: The Hidden Attack Target in AI APIs
How attackers extract hidden reasoning traces from proprietary LLM APIs, and why chain-of-thought security is becoming AI's biggest battleground.
AI
AI Benchmark Saturation: Why 99% Scores Mean Less
As AI models hit 90%+ on popular benchmarks, static tests are losing meaning. Here is why benchmark saturation is forcing a reality check for AI evaluation.
AI
Is AI Reasoning Right for the Wrong Reasons?
New studies reveal reasoning LLMs often reach correct answers through faulty intermediate logic. Here is why it matters for engineers.