Tensordyne Challenges NVIDIA With Logarithmic Chips as Mistral Unveils 30-Trillion Parameter Model
Hardware & Infrastructure
Tensordyne Logarithmic AI Compute Chips: Tensordyne announced a breakthrough in AI hardware that utilizes logarithmic math to replace power-intensive multiplications with additions. This innovation claims to deliver 17x more tokens per watt and 13x higher throughput than NVIDIA's Blackwell architecture.
Large Language Models
Qwable-v1: Claude Fable 5 Distillation: Qwable-v1 is an open-weights model distilled from Anthropic's Claude Fable-5 using agentic-coding traces. It aims to replicate the performance of the larger model on specific tasks and has been made publicly available on Hugging Face.
Benchmarks & Research
Evalatro Strategic Gaming Benchmark: Evalatro is a new open benchmark that tests the strategic decision-making of LLMs by having them play the game Balatro. The benchmark evaluates models based on their ability to execute complex strategies in rule-based environments.
Study Finds Mistral AI Models Susceptible to Disinformation: A recent study revealed that Mistral AI’s top open-source model performed poorly in filtering Russian disinformation, ranking 47th out of 60 tested models. The findings highlight safety vulnerabilities in European models compared to commercial competitors like Claude and various Chinese models.