Drama is unfolding in the US: autonomous drone swarms are changing the face of warfare, states are banning AI-mediated mental health treatment, and scientists are "tweaking" Transformers with a new kind of model. AI is no longer just the future - it's the present with a big question mark.
1. Drone swarms struggle: autonomy + human decision
The military use of AI in the US and Ukraine is pushing the limits of autonomous warfare. Startup Swarmer, with a US background, supplies software for drone swarms that allows small units to decide their own attacks and tactics - but under human supervision. A human operator sets the target, but the drones coordinate the distribution of attacks, avoid collisions and operate even in conditions where communications are limited. This technology is particularly useful in asymmetric conflicts, where the advantages are smaller numbers, flexibility and adaptability.
2. States ban AI therapy without a human expert
Illinois became the second state after Nevada to pass a law banning the use of AI for psychotherapeutic purposes without the direct involvement of a licensed therapist. The "Wellness and Oversight for Psychological Resources Act" provides for fines of up to $10,000 for each violation. AI models without human intervention cannot independently diagnose an emotional state, conduct a therapeutic conversation, or be promoted as therapeutic tools unless an expert is present in the interaction. Critics argue that such bans may jeopardize access to innovative tools that could be particularly helpful in areas with a shortage of therapists.
3. Qwen3-Next: faster and with long context
Alibaba comes up with a new version of the model Qwen3-Next-80B-A3Bwhich combines techniques such as "mixture-of-experts" and effective attention layers. The model handles inputs of up to hundreds of thousands of tokens (in some cases approximately 262 144 tokens) and speeds up inference compared to older versions of Qwen3. All this while maintaining or slightly improving performance on many common tasks. In the US and around the world, this opens the door for use in large contexts - such as law, research, or large-scale documentation.
4. Transformers under new energy: Energy-Based Transformer
Scientists from universities in the USA (Virginia, Illinois, Stanford, Harvard) together with partners from Amazon presented Energy-Based Transformer (EBT). Instead of having the models predict the next token all at once, EBT first offers a rough draft, which it then iteratively "refines" - sort of like a scoring mechanism that evaluates how well the answer fits the context. In early tests, EBTs outperform conventional transformers in reading comprehension and math tasks, though they still have higher computational demands.
The Batch - DeepLearning.AI by Andrew Ng / gnews.cz - GH
English 
Czech 
Spanish 
French 
Italian 
Russian 
Chinese 
German