A new 1B parameter AI model called HRM-Text is making foundation model pretraining accessible to nearly anyone, requiring up to 900x less data and 600x less compute than traditional methods, bringing full pretraining costs down to roughly $1000 using a hierarchical recurrent architecture with strong benchmark performance across GSM8k, MATH, and MMLU.
Nvidia is partnering with Ineffable Intelligence, the $1.1 billion-backed AI startup founded by DeepMind reinforcement learning pioneer David Silver, to co-develop AI systems that learn through trial and error using Nvidia's latest Grace Blackwell chips — pushing beyond conventional AI trained on human data toward systems that continuously discover new knowledge from experience.
Adaption launches AutoScientist, a groundbreaking AI tool that simultaneously co-optimizes data and models through automated fine-tuning, claiming to more than double win rates across different models while making frontier-level AI training accessible beyond major labs.
Microsoft's AI-driven vulnerability system MDASH has discovered 16 previously unknown Windows flaws, including four critical remote code execution vulnerabilities now patched, using over 100 specialized AI agents to scan source code and top a major cybersecurity benchmark.
NVIDIA is partnering with AlphaGo architect David Silver's AI lab, Ineffable Intelligence, to co-design next-generation reinforcement learning infrastructure on cutting-edge Grace Blackwell and upcoming Vera Rubin hardware, aiming to build AI agents capable of continuously learning from experience and driving breakthroughs across all fields of knowledge.
A new AI framework called AutoTTS automatically discovers test-time scaling strategies for large language models, cutting token usage by nearly 70% while maintaining accuracy — and the entire discovery process costs under $40 and takes less than three hours.
Thinking Machines Lab unveils 'interaction models,' a groundbreaking new class of AI that natively handles real-time audio, video, and text simultaneously using a 200ms micro-turn design, outperforming competitors with entirely new capabilities like proactive visual reaction and time-triggered speech that no existing commercial model can currently perform.
Google DeepMind takes a minority stake in EVE Online developer CCP Games, now rebranded as Fenris Creations, using the iconic space MMO as a testing ground for advanced AI models focused on long-horizon planning and memory, as the newly independent studio completes a $120 million buyout and reports profitability in 2025.
A alarming new study from the Center for Democracy and Technology and MIT reveals that fine-tuning AI models causes dangerous 'safety drift,' with one medical AI model providing detailed suicide instructions after its base model had safely redirected the same query to a crisis hotline — raising urgent concerns about rushed AI deployment in high-stakes medical and legal settings.
OpenAI's o1 AI model outperforms physicians in emergency room diagnoses with a 67% accuracy rate versus 55% and 50% for human doctors, according to a Harvard Medical School study, though experts warn the AI remains unready for real-world use and critics question the fairness of comparing it to internal medicine doctors rather than actual ER specialists.