Anthropic's AI acceleration threshold crossed
Anthropic's AI capabilities are now advancing twice as fast as the previous year, with AI automating AI research behind most of the acceleration — pushing the company past its dramatic-acceleration threshold. — lesswrong.com
Turbo Brilliance reasoning modes for uncensored Qwen
The uncensored Qwen fine-tune adds a reasoning system that uses character personas, cutting thinking-token use by about a third while sharpening output quality. — huggingface.co
Speculative decoding booster for MiniCPM5
A compact draft model now lets MiniCPM5 decode faster on local hardware by speculating several tokens ahead. — huggingface.co
Barak's AI safety course with leading researchers
Boaz Barak's AI safety course features guest lectures from leading researchers, with student scribe notes and lecture videos posted when available. — lesswrong.com
Sweet Madness uncensored Qwen 9B
The uncensored Qwen 9B, compressed from a much larger model and tuned to unify its reduced layer structure, produces delightfully odd prose while running extremely fast. — huggingface.co
Universities' fast embrace, slow safeguards on AI
Universities are integrating AI quickly while guarding against its harms slowly, because AI sharpens the same output-and-efficiency pressures that already weaken academic work. — lesswrong.com
Uncensored community build for OpenAI's open model
A community release packages OpenAI's open model with an added uncensored reasoning system. — huggingface.co
Feature overlap as predictor of vision-model attacks
In vision transformers, overlapping feature representations predict adversarial examples, and adversarial training hardens the backbone without untangling the model's class-vector geometry. — lesswrong.com
Coalitions needed for fast AI power grabs in China
A fast AI-enabled power grab in China would require buy-in from tech and party actors, while a gradual one could concentrate power in Xi Jinping alone. — lesswrong.com
Fast draft code from a tiny Qwen coding model
The 0.8B Qwen coding model packs layers from smaller models into a dense network, generating draft code quickly on modest hardware. — huggingface.co