New AI Models Push Boundaries
The AI field is seeing rapid advances with the emergence of new multimodal models like Qwen 3.5, which enhances contextual understanding. The developments are enabling real-time, persistent AI companions for gaming and social platforms, while breakthroughs in persistent memory technology could soon merge the speed of RAM with the permanence of storage.
Alibaba's Qwen 3.5 is part of a larger trend of AI models that can process not just text, but also images, audio, and video within a single conversation. This leap is powered by a Mixture-of-Experts (MoE) architecture, which uses a large number of "expert" subnetworks but only activates a fraction of them for any given task, boosting efficiency. The flagship open-source version of Qwen 3.5 boasts 397 billion parameters, yet only activates 17 billion for each token, allowing it to be 19 times faster in long-context tasks compared to its predecessor. This efficiency allows it to compete with other leading models like OpenAI's GPT series and Google's Gemini. A premium version, Qwen 3.5-Plus, extends the context window to one million tokens. The development of such complex models relies on massive datasets, with some models being trained on over 20 trillion tokens of multilingual text and domain-specific data. Techniques like reinforcement learning with human feedback (RLHF) are then used to fine-tune the models' capabilities for tasks like coding and complex reasoning. In the realm of gaming, AI companions are evolving from pre-scripted non-player characters (NPCs) to dynamic partners that learn from player actions and adapt their behavior accordingly. This creates more immersive and personalized gameplay experiences, with some AI companions even able to comment on random in-game events. The global market for AI companion apps has grown to nearly 50 million active users, with revenues exceeding $580 million. The push for more responsive AI is also driving innovation in hardware, particularly in persistent memory technologies like Spin-Transfer-Torque Magnetoresistive Random Access Memory (STT-MRAM). These emerging memory solutions aim to overcome the performance bottlenecks of traditional RAM and storage, which is crucial for the real-time learning and adaptation required by advanced AI models. OpenAI CEO Sam Altman has suggested that the next major breakthrough in AI will likely come from persistent memory, enabling an AI to learn from a lifetime of user data to provide proactive and highly personalized assistance. The goal is to create a seamless integration of memory and processing that allows AI to have a continuous, evolving understanding of its user and the world.
