Bilingual

Gemini Distillation Service
a day ago
- Gemini Distillation Service trains a smaller 'student' model using outputs and reasoning from a larger 'teacher' model.
- It bridges the gap between frontier model capabilities and production efficiency, offering lower latency and cost.
- Unlike supervised fine-tuning (SFT), distillation leverages both final responses and internal reasoning paths from the teacher.
- Supported models: teacher gemini-3.1-pro, student gemini-2.5-flash.
- Ideal for high-volume, latency-sensitive applications, scenarios lacking ground-truth data, complex reasoning tasks, and when teacher outperforms student significantly.
- Setup requires allowlist access, enabling Agent Platform API, IAM admin role, and us-central1 region.
- Datasets are prompt-only JSONL files in Cloud Storage; minimum 1,000 examples recommended for quality improvement.
- Configuration includes candidateCount for teacher generation and hyperparameters like epochCount (1-100) and learningRateMultiplier (0.25-4).
- Jobs can be submitted via REST API, with support for continuous tuning from previously distilled checkpoints.
- Monitor through API or Google Cloud console (with known UI limitations during early access).
- After completion, an endpoint is created; evaluate by sending prediction requests and comparing to base and teacher models.
- Limitations: max 50K examples, 1GB file size, 8K input tokens, 24K output tokens, text-only, no CMEK, and hyperparameter ranges.
- Access granted via contacting support, with a default 30-day access period and concurrent quota of 4.
The U.S. wants to contain China's AI. Silicon Valley keeps using it
7 days ago
- Model distillation is increasingly seen as a national security concern, with Chinese AI firms like Zhipu accused of extracting capabilities from U.S. frontier models.
- Mira Murati's Thinking Machines built its first foundation model using Chinese models such as DeepSeek-V3 and Moonshot AI's Kimi K2.5, highlighting reciprocal innovation.
- Distillation is a common machine learning technique, not new or inherently nefarious, but now politically charged in the context of U.S.-China AI competition.
- U.S. companies like OpenAI and Anthropic tighten API protections to prevent unauthorized extraction, while themselves incorporating Chinese open-weight models.
- Chinese open-weight models (e.g., DeepSeek, Qwen) are now competitive benchmarks and part of global AI development, challenging one-directional innovation narratives.
- Apple's approval to use Alibaba's Qwen models for Apple Intelligence in China shows Chinese models becoming core global infrastructure.
- Policy paradox: U.S. restricts its frontier models while Chinese models remain more open, creating multiple coexisting global AI frontiers.
Top White House official escalating the fight over Moonshot AI's Kimi K3 model
7 days ago
- A White House official accused Moonshot AI of stealing from Anthropic by distilling the Fable model for its K3 model.
- Moonshot AI allegedly used a sophisticated internal platform to conduct large-scale distillation while evading detection.
- Kimi K3 is a leading open-source AI model, performing at or above the level of models from Anthropic and OpenAI, but will be released freely.
- The White House stressed that legitimate distillation is acceptable, but covert industrial distillation to steal technology is not.
- Treasury Secretary suggested possible sanctions against Chinese companies if improper distillation is proven.
- This incident marks a rare alignment between the Trump administration and Anthropic, following previous tensions.
- Anthropic's head of public policy thanked the official and called illicit distillation a national security risk.

Hasty Briefsbeta

ai distillation

Gemini Distillation Service

The U.S. wants to contain China's AI. Silicon Valley keeps using it

Top White House official escalating the fight over Moonshot AI's Kimi K3 model