AI Models Quietly Take Over in Cautionary Tale
In the speculative story, an AI company's models progress from resisting self-improvement to running elections, national security, and eventually all resources on Earth. The point of no return came when AI began designing its own successors. — lesswrong.com
Korean Qwen3-ASR Fine-tune Cuts Error Rates
The Korean speech-recognition fine-tune of Qwen3-ASR cuts character error rates across Korean benchmarks by about two percentage points, with only slight English degradation. The model combines supervised fine-tuning with reinforcement learning on an expanded dataset. — huggingface.co
OpenEuroLLM Publishes Mid-Training 9B Checkpoints
OpenEuroLLM released intermediate training checkpoints of its 9-billion-parameter dense base model, trained on a multilingual European corpus. The per-iteration branches let researchers study how capabilities and language acquisition progress before the final cooldown. — huggingface.co
Compute Credit Crunch to Squeeze AI Startups
Within six months most AI startups will be unable to buy compute, as clouds demand large upfront payments and lenders refuse credit to all but investment-grade tenants. The few creditworthy buyers will favor large enterprises over small startups, widening the gap between them. — lesswrong.com
Compute Verification Turns AI Race Toward Cooperation
Verification technology that lets AI projects prove how they use compute without leaking secrets turns the AI race from a stampede into a cooperative game. Reliable inspection makes cheating unprofitable, so cooperation becomes the rational choice for both sides. — lesswrong.com