Knowledge distillation (KD) is a widely used technique in language model training, but a new study reveals that its application is not one-size-fits-all. KD boosts both reasoning and factual recall during pre-training, but it slows down factual recall during the crucial mid-training phase.
Key points
• KD boosts reasoning and factual recall during pre-training.
• KD slows down factual recall during the mid-training phase.
• The study suggests that KD's application is not one-size-fits-all.
Seedance 2.5, a tool developed by pixio_ai, has been praised for its exceptional performance in creating UGC ads. The tool combines GPT Image 2 and Seedance 2.5 to produce high-quality, realistic content.
Key points
• Seedance 2.5 is a powerful tool for creating UGC ads.
• The tool combines GPT Image 2 and Seedance 2.5 for exceptional performance.
• The results are highly realistic and of high quality.
SaaS AI platforms often fail to deliver in government settings due to a lack of forward-deployed engineering. Agencies typically buy a platform, conduct a 6-month pilot, and then declare success, but the adoption rate is usually low.
Key points
• SaaS AI platforms often fail to deliver in government settings.
• Forward-deployed engineering is key to successful AI adoption.
• Agencies should embed, integrate, and deliver source code for effective AI adoption.
Production AI agents require five layers: interface, orchestration, memory, tools, and governance. Most legal AI platforms only provide the chatbot interface, lacking essential features like memory, audit trails, and RBAC.
Key points
• Production AI agents require five layers.
• Most legal AI platforms only provide the chatbot interface.
• Essential features like memory, audit trails, and RBAC are lacking.
RSLM codecs have been developed to compress billion-scale vector search down to 1-4 bits per dimension without sacrificing recall. The new codecs need zero training and match or beat state-of-the-art engines like Faiss and ScaNN on five benchmarks.
Key points
• RSLM codecs compress vector search down to 1-4 bits per dimension.
• The new codecs need zero training and match state-of-the-art engines.
• The codecs perform well on five benchmarks.
Lily is a local inference engine developed for hybrid compute in Perplexity Computer. The engine is specialized for Qwen3.6-35B-A3B on Apple silicon, ensuring that on-device compute does not bottleneck computer tasks.
Key points
• Lily is a local inference engine for hybrid compute.
• The engine is specialized for Qwen3.6-35B-A3B on Apple silicon.
• The engine ensures that on-device compute does not bottleneck computer tasks.
Qwen3.6-35B-A3B has been benchmarked on an M5 Max MacBook Pro. The results show that Lily is faster than MLX-LM, averaging 1.23× higher prefill throughput and 1.35× higher decode throughput while keeping output quality effectively unchanged.
Key points
• Qwen3.6-35B-A3B has been benchmarked on an M5 Max MacBook Pro.
• Lily is faster than MLX-LM, averaging higher throughput.
• Output quality remains effectively unchanged.
Lily, the local inference engine, has been open-sourced by Perplexity Computer. The engine is specialized for Qwen3.6-35B-A3B on Apple silicon, ensuring that on-device compute does not bottleneck computer tasks.
Key points
• Lily has been open-sourced by Perplexity Computer.
• The engine is specialized for Qwen3.6-35B-A3B on Apple silicon.
• The engine ensures that on-device compute does not bottleneck computer tasks.
Anthropic's cut agent pricing has been reduced by 45%. Kimi K3 has climbed to #3 globally, with a 1/3 cost reduction compared to frontier models. The only strategy that survives is infrastructure that treats models as interchangeable.
Key points
• Anthropic's cut agent pricing has been reduced by 45%.
• Kimi K3 has climbed to #3 globally.
• The cost reduction is 1/3 compared to frontier models.
Claude can now use your computer in the background in Claude Cowork and Claude Code. Give it something to do on your desktop, and Claude clicks, types, and opens apps just like you would, while you work on something else.
Key points
• Claude can use your computer in the background.
• Claude can click, type, and open apps.
• Claude can work on tasks while you work on something else.
