Microsoft Copilot Cowork Moves to Usage-Based Pricing, Eyes DeepSeek V4 as a Low-Cost Option
Microsoft's AI assistant 'Copilot Cowork' is set to shift from a flat-rate subscription to usage-based pricing. Charles Lamanna, head of the Copilot division, explained that the flat-fee model is not sustainable.
More notably, Microsoft is considering adopting a fine-tuned version of DeepSeek V4 as a low-cost model option. The possibility of Microsoft incorporating an open-weight model developed by China's DeepSeek into its own products could have a significant impact on the balance of power in the AI industry. While Microsoft has built Copilot primarily around OpenAI's GPT models, the company appears to be moving to diversify its model lineup from a cost-optimization standpoint.
The shift to usage-based pricing marks a move toward a more realistic model for AI consumption—from 'fixed monthly fees' to 'pay for what you use.' For enterprise users, this makes it easier to manage costs in proportion to usage, though they will also need to stay mindful of the risk of unexpected charges.
OpenAI Releases 'Deployment Simulation' — Predicting Model Behavior Before Release
OpenAI has announced 'Deployment Simulation,' a method for simulating and predicting a model's behavior using real conversation data before deploying it to production.
The approach makes it possible to evaluate in advance how a model will behave based on actual user interaction patterns, with improvements in both safety and accuracy expected. A key benefit is the ability to catch unexpected behaviors in realistic usage scenarios ahead of time—behaviors that conventional benchmark tests fail to capture.
With concerns over model safety on the rise, pre-deployment prediction technology is drawing attention as a practical approach. That said, the key question will be how faithfully the simulation data can reproduce real-world usage environments.
ALS Patient Works Full-Time with AI and a Brain-Computer Interface
According to a report by The Register, combining AI with a brain-computer interface (BCI) has enabled an ALS patient who has lost the ability to speak to continue working full-time.
The AI decodes brain signals read by the BCI and converts them into text and speech, allowing the patient to manage both communication and work. As an example that highlights AI's potential in medical applications, the case has also been drawing attention in the Hacker News community.
While the technology is not yet ready for widespread adoption, the fact that direct communication from brain signals is becoming a reality offers real hope for people with motor impairments.
The Battle for Physical AI Supremacy: How Tesla, Waymo, and NVIDIA Strategies Differ
An ITmedia article breaks down the differing approaches of Tesla, Waymo, and NVIDIA to generative AI-powered autonomous driving systems.
In 'physical AI'—the fusion of the physical world and AI, which the Japanese government has designated a strategic priority area—autonomous driving is one of the most closely watched application domains. Each company is competing with a distinct strategy:
Tesla**: An approach aiming for full self-driving through end-to-end AI learning
Waymo**: Securing safety margins through integration with high-precision sensors
NVIDIA**: Positioning itself as a provider of autonomous driving platforms
How the evolution of generative AI reshapes the competition among these companies will likely hold the key to the broader adoption of physical AI going forward.
Lexar Proposes Offloading AI Models to SSDs — A Fix for RAM Shortages
With insufficient RAM becoming a serious problem for running AI models locally, Lexar has proposed technology that offloads AI models to SSDs. The approach, which made waves on Hacker News, is an interesting direction for compensating for hardware constraints.
While LLM parameter sizes continue to grow, the amount of RAM in consumer PCs remains limited. Leveraging SSDs as supplementary storage could point the way toward a realistic path for running large models locally. That said, the challenge remains that SSD access speeds are far slower than RAM.
---