Superintelligence Projected by Early 2028 Without Pause
Superintelligence is on track to arrive by early 2028 unless labs deliberately slow down, so AI safety work must pivot to forcing aggressive pacing or a pause in the next 6–12 months. — lesswrong.com
Modified Qwen3-8B Removes All Safety Filters
A modified Qwen3-8B strips out refusals and safety filters while keeping factual accuracy steady, cutting generic filler for users who accept the risk. — huggingface.co
LoRA Adapter Tunes Small Qwen Model's Reasoning
A LoRA adapter for a small Qwen language model, trained on 9,762 reasoning samples, is released under the Darwin family's evolutionary merging approach. — huggingface.co
Scientific Prediction Markets Pay for Good Evidence
A proposed market structure turns scientific questions into perpetual bets, paying researchers whose experiments move the expected answer. — lesswrong.com
Edge0-145B Runs Frontier AI on a Laptop
Edge0-145B matches frontier cloud models on coding and reasoning while running entirely on a laptop, with no cloud token fees. — edge0.ai