Large Language Models
AI Agents Can Invent Their Own Biases
AI agents may do more than reproduce the biases buried in their training data. As they explore, act, and learn from …
Large Language Models
AI Agents
AI Safety
Qwen
4 min read
The Last Useful Bit: Why 27B Models Hit a Wall Below 4-Bit
Qwen3.8 27B shows why 4-bit quantization can make local inference practical, while pushing a conventionally trained model to 1-bit risks a sudden collapse in quality.
Recent Dispatches
Curated insights from the frontier.
Artificial Intelligence
When AI Solves the Problems Mathematicians Were Saving for Later
Terence Tao has compared unsolved mathematical problems to a nonrenewable resource. As AI becomes capable of discovering …
Artificial Intelligence
Mathematics
Terence Tao
AlphaGenome
The Human Genome Is Becoming Searchable
Google DeepMind’s AlphaGenome Atlas aims to turn the genome’s complex regulatory machinery into a high-resolution map. …
AlphaGenome
Google DeepMind
Precision Medicine
OpenAI
A Million-Dollar Proof Needs More Than a Press Release
OpenAI’s reported Navier–Stokes breakthrough has triggered questions about verification, originality, and credit. Before …
OpenAI
Artificial Intelligence
Mathematics
LibreOffice
The Hottest AI Feature Is an Off Switch
LibreOffice is attracting attention by leaving generative AI out. The real story is not an anti-AI backlash, but growing …
LibreOffice
Generative AI
Open Source
Mistral AI
Europe’s €3 Billion Bet on Owning Its AI Stack
Mistral AI has secured €3 billion to pursue a European alternative to US- and Chinese-controlled AI. The harder question …
Mistral AI
Open Weights
European AI
RSA
The Weakest Link in PKI Is the Trust Nobody Retired
A decades-old CA key may be easy to crack with modern computing. The bigger danger is the legacy software that still …
RSA
PKI
Cybersecurity
VMware
VMware’s Lock-In Problem Now Extends to the Exit
Broadcom’s restrictions on VMware VDDK downloads could make migrations harder long after existing tools stop working. …
VMware
Broadcom
Cloud Migration
Linux
When the Compiler Lies, Clean Source Code Means Nothing
A malicious compiler can quietly inject backdoors while leaving source code and signatures looking legitimate. The …
Linux
Supply Chain Security
Compilers