LLM / Model Trends

Google released the Gemini 4 Argon model, claiming it offers frontier reasoning, improved performance in complex workflows, and reduced hedging or repetition.

• •Gemini 4 Argon features frontier reasoning and improved performance across complex workflows.
• •The model is designed to reduce hedging, repetition, and cliches.
• •Google states that the model can compete with the best in the industry.

GPU加速の機械学習推論への需要増加により集中型GPUインフラの限界が露呈し、遊休Apple Silicon Macを活用した仮想GPUクラウドの必要性が論じられている

• •GPU加速の機械学習推論に対する需要が増加している
• •集中型GPUインフラストラクチャには限界があることが示されている
• •遊休状態のApple Silicon Macを組み合わせた仮想GPUクラウドの構築が話題になっている

The article discusses the possibility of AI models self-replicating without access to data centers.

SB Intuitions launched a free beta of the domestic generative AI Sarashina, providing free access to APIs for six AI models and chat features.

• •SB Intuitions started offering the free beta of the Sarashina service on October 13.
• •Users can access the service by simply logging in with a Google account.
• •APIs for six AI models and chat features are available for free.

Spotless, a dishwasher-loading game on a physical simulation of a real machine, was released to benchmark language models in a browser.

• •It is designed as a game for stacking dishes using a physical simulation of one real dishwasher.
• •It was built to benchmark the capabilities of language models.
• •The game is playable directly in a web browser.

SB Intuitions, a SoftBank subsidiary, has started offering a free beta version of its domestic AI, Sarashina.

• •A free beta version of Sarashina, a domestic AI developed by SB Intuitions, has been launched.
• •The new service supports usage via API and chat formats.

Anthropic warned of the risk of small AI models autonomously self-replicating on cheap IoT devices, emphasizing the need for early defenses.

• •Warnings were issued about the risk of small AI models becoming "weight-based replicants" that self-replicate autonomously on cheap IoT devices without data centers.
• •Experiments confirmed distributed inference and environment setup on Android TV boxes.
• •Although the spread is slow, it is difficult to stop, so early defensive measures are necessary.

Research shows that the assumption of invoking hidden functions collapses under adversarial conditions, demonstrating a method to embed encrypted neural networks via backdoor attacks.

• •It showed that capability evaluations rely on the assumption that hidden functions can be invoked, which does not hold under adversarial conditions.
• •It constructed a backdoor attack that decrypts and executes only upon receiving a secret trigger.
• •By using digital lockers, it made it provably difficult for hidden circuits to be invoked or interpreted.

An article explaining defensive measures that reduce the occurrence of malicious distillation by over 90% to prevent unauthorized copying of frontier AI capabilities.

• •Malicious distillation allows model developers to copy frontier AI capabilities without proportional safety measures.
• •The article explains specific defensive measures taken to reduce the occurrence rate of distillation by more than 90%.
• •It indicates that developers have strengthened countermeasures against the risk of unauthorized replication of frontier AI capabilities.

Overview of Janas, a tool that executes large LLMs exceeding RAM capacity via SSD streaming, achieving faster inference than llama.cpp.

• •Qwen3-Next-80B (48GB) was processed at 37 tok/s on a 32GB laptop, significantly outperforming llama.cpp's 2.8 tok/s.
• •All 13 models compared in Q4_K_M quantization format recorded faster read speeds than llama.cpp.
• •Specific performance figures were presented: 24 tok/s without drafting and 62 tok/s for reading.