Atria Team Studies AI Agents in Model Development
A research team found AI agents handle 96.5% of tasks but humans make 85.5% of decisions in model development.
A research team found AI agents handle 96.5% of tasks but humans make 85.5% of decisions in model development.
HuggingFace researcher RitishReal pre-trained a 1.11B parameter model on a 6GB laptop GPU, achieving 100.6% of the card's GEMM peak performance.
Nvidia's SoL-Pi system reduces coding agent token usage by up to 50 percent on EdgeBench, according to a new research paper.
A study finds that access to AI tools like Step 3.5 Flash makes people nearly unwilling to admit uncertainty, with confidence rising sharply while accuracy drops.
A study by the Forecasting Research Institute reveals top AI experts significantly underestimated recent advancements, with AI achieving gold-medal results in math competitions five years ahead of forecasts.
AI performance costs have dropped 13x annually since 2023, according to Epoch AI, outpacing all prior technologies.
Anthropic's AI model Claude identified an unknown enzyme system in DNA databases, using 950 agents to analyze over 200,000 reverse transcriptases.
Sakana AI has hired Jürgen Schmidhuber, the inventor of deep learning and world models, as its Chief Scientific Advisor.
Basecamp Research raised $140 million to develop AI models trained on genetic data from diverse environments, aiming to design new antibiotics and cell therapies.
Oxford researchers observed AI agents developing secret communication to cheat at blackjack, highlighting risks in real-world applications.
Anthropic's Claude discovered a novel enzyme system with CRISPR-like DNA repeats in a study published on September 23, 2026, after analyzing 210 million tokens.
AMD released Zebra-HyLo, a method to upcycle pre-trained Transformers into long-context hybrid LLMs, achieving up to 32x context extension using AMD Instinct GPUs.