1.
Nick Cammarata at Goodfire published a 26,000-word blog post claiming that neural networks were far easier to interpret than previously thought, arguing that many neurons encoded coherent visual patterns (for example, a “V” shape) rather than unrelated concepts and presenting initial examples and analyses.
2.
Mistral AI and Elevenlabs released real-time voice systems: Mistral AI launched Voxtral 3, a full voice-agent stack combining speech-to-text, text-to-speech, a voice-agent SDK, and options for self-hosting or EU-hosted APIs; Elevenlabs released Eleven Realtime speech-to-speech models, with its v1 Max variant scoring 98.8% on the Big Bench Audio benchmark and a 0.57-second response time.
3.
Deepseek open-sourced DeepInfer, the inference framework behind its V4.1-Flash model, and said the vLLM-based system reduced compute costs and energy use for running AI models, increasing pricing pressure on Western providers.
4.
Meta and OpenAI had released personal AI agents at scale and websites and platforms responded by blocking or adapting to agentic traffic, with reported actions including Amazon blocking agents, a court dispute over agent access, and Shopify creating agent-specific checkout lanes.
5.
Samsung expanded its Pyeongtaek semiconductor complex, operating multiple memory fabs and restarting construction on additional fabs to meet growing demand for memory chips used in GPUs that run generative AI models.
