Top stories today:

1. OpenAI claims China’s DeepSeek used its model for training
2. Alibaba brings Qwen 2.5-Max to compete with DeepSeek V3
3. ChatGPT Gov launched for U.S. government agencies
4. ASML Q4 revenue +28% YoY, net income +31%, beating expectations
5. Fed expected to hold rates steady, assess Trump’s impact on economy

0. Data and calendar

Listen to our AI-generated podcast summarizing today’s newsletter (beware of hallucinations):

1. OpenAI claims China’s DeepSeek used its model for training

• Some evidence of “distillation,” OpenAI said amid the growing popularity of DeepSeek in the U.S.

• Tue.: White House AI czar David Sacks alleged “substantial evidence” that DeepSeek “distilled the knowledge out of OpenAI models.”

• Microsoft said to be probing if DeepSeek improperly obtained OpenAI’s data.

• Common practice among startups and academics to use outputs from commercial LLMs to train another model, experts said.

• Separately, the U.S. Navy has banned DeepSeek over “potential security and ethical concerns.”

3. Alibaba releases Qwen 2.5-Max, new large MoE LLM competing with DeepSeek V3

Qwen 2.5-Max outperformed the open-weight DeepSeek V3 and Meta’s Llama 3.1-405B and OpenAI’s GPT-4o across benchmarks shared by Alibaba:

• Trained on over 20T tokens.

• Based on the Mixture-of-Experts (MoE) architecture, which requires fewer compute resources than a traditional LLM.

• Available through Alibaba Cloud’s API services.

• Also, accessible through Qwen Chat and Hugging Face Spaces.

Like other Chinese AI models, Qwen 2.5-Max has censored outputs critical of the Chinese government

4. ChatGPT Gov launched for U.S. government agencies