A new survey paper by Tencent's Youtu Lab and Chinese universities argues AI systems must shift from generating answers to completing tasks reliably, with a focus on reusable skills and persistent work environments.
Epoch AI's MirrorCode benchmark assesses AI models' ability to recreate programs from scratch, with one task costing $2,600 to run.
Unconventional AI, led by Naveen Rao, claims its oscillator-based architecture can reduce AI power consumption by up to 1,000 times, with its first model, Un-0, matching state-of-the-art image-generation systems.
Google DeepMind's AI system Aletheia autonomously generated Ph.D.-level math research, marking a milestone in AI's role in mathematics.
Hugging Face introduces DoctoBERT, a French medical encoder, trained on a new corpus of web data curated with domain-specific filters and rephrasing techniques.
Falconer achieved a 64% win rate against Notion in real support questions, outperforming other tools in two critical benchmarks.
A UC Berkeley study found grades in writing and coding courses rose 13 percentage points after ChatGPT launched in November 2022.
OpenAI CEO Sam Altman claims a generation of researchers hindered AI progress by underestimating scaling's impact, citing recent model achievements.
Research on Qwen3.6 27B reveals that harness-specific fine-tuning can significantly alter model behavior, with v2 showing 40.45% performance on Pi harness compared to base model's 42.70%.
A new benchmark reveals AI models fail to complete most real-world knowledge tasks, with top models solving just 3 percent of tasks.
OpenAI researchers found that small amounts of beneficial trait training improved model safety across 44 of 53 benchmarks, according to a blog post.
OpenAI's o3 Deep Research model helped identify 18 diagnoses from 376 previously unsolved cases, achieving a 4.8% additional diagnostic yield.