Top story
Aleph Alpha releases Kolibri, an open-weight German-English model
The company made its 78-billion-parameter mixture-of-experts model available under Apache 2.0 on October 3. It trained the model for German and English and says its context can extend to one million tokens; engineer Tejas Kumar independently tested its tokenizer on the German Basic Law, finding it used fewer tokens than several alternatives on that text. The company's model-quality comparisons remain vendor-run. ·
Why it mattersOrganizations handling German documents can run the weights on their own infrastructure, though the model still needs substantial GPU memory.
Also
White House forms a Super Intelligence Force
The Wall Street Journal interviewed chair Jay Clayton and reviewed its charter: the group has 120 days to propose how Washington should handle AI risks and opportunities.
Why it mattersFederal AI-risk policy now has a coordinating group, though its recommendations are still pending.
Former OpenAI safety-report writer calls for a different safety culture
TechCrunch reports David Robinson resigned and said OpenAI's culture was broken; an OpenAI spokesperson told the outlet the company was strengthening monitoring, security and external evaluation.
Why it mattersHis criticism brings an internal safety veteran's perspective into the public debate over how frontier labs manage risk.
Research worth a look
Microsoft and Hugging Face release ThinkingBox
Their October 3 post describes 507 simulated business workflows, each run 20 times, and grades agents on the resulting records: in their tests, successful-looking tool use often left the wrong database state.
Limited coverage: The Hacker News discovery feed returned a 502 error; The Verge's Robinson article and the Los Angeles Times AI-radio profile were only partly readable.