The Decoder
Deepseek has made its 75% discount permanent on its V4-Pro model, significantly lowering its pricing. The model now costs $0.435 per million input tokens, making it substantially cheaper than GPT-5.5 for both input and output. This aggressive pricing strategy could pose a challenge to Western AI providers, especially for applications requiring high token usage.
Key Takeaways
- Deepseek's V4-Pro model will permanently feature a 75% discount.
- Deepseek's pricing is significantly more competitive than GPT-5.5, especially for output tokens.
Why it matters:
Deepseek's drastically reduced pricing could disrupt the market for Western AI providers, particularly impacting agentic systems that consume a large volume of tokens.
Read Original →
TechCrunch AI
Scuderia Ferrari HP is partnering with IBM to leverage AI and hybrid cloud technology to enhance the Formula 1 fan experience. This collaboration aims to create more personalized and engaging interactions for fans, making them feel like "superfans." The initiative focuses on delivering deeper insights and tailored content to a global audience.
Key Takeaways
- IBM's AI and hybrid cloud solutions are being used by Scuderia Ferrari HP to transform F1 fan engagement.
- The goal is to create highly personalized and immersive experiences for Formula 1 enthusiasts, fostering 'superfan' status.
Why it matters:
This initiative demonstrates how advanced AI can be applied to create deeper connections between sports teams and their fan bases, potentially setting new standards for fan experience in the sports industry.
Read Original →
The Decoder
UC Berkeley Law will prohibit the use of AI in almost all graded assignments, including outlining, drafting, and proofreading, beginning in summer 2026. The school's decision stems from the belief that aspiring lawyers need to develop their own critical thinking skills before they can effectively leverage AI tools. Research will be the only exception to this AI ban.
Key Takeaways
- UC Berkeley Law is implementing a strict ban on AI for graded work, starting summer 2026.
- The ban covers a broad range of tasks from initial drafting to final proofreading, with research as the sole permissible AI use.
Why it matters:
This policy highlights concerns about foundational legal skills and aims to ensure future lawyers develop independent analytical abilities before integrating AI.
Read Original →
The Decoder
Alibaba's Qwen team has released Qwen3.7-Max, a new AI model designed for extended autonomous agent tasks. This model demonstrates impressive performance, matching Claude Opus 4.6 on benchmarks and outperforming several Chinese competitors. Notably, the AI successfully ran for 35 hours autonomously to optimize code for Alibaba's custom chip and was also used to control a four-legged robot.
Key Takeaways
- Alibaba's Qwen3.7-Max AI model showcases advanced autonomous capabilities and strong benchmark performance, rivaling top-tier models.
- The AI was successfully deployed for a 35-hour autonomous code optimization task on Alibaba's custom chip and demonstrated robotic control.
Why it matters:
This development highlights the growing sophistication of AI in performing complex, long-duration autonomous tasks, including direct application in hardware optimization and robotics.
Read Original →
The Decoder
Google CEO Sundar Pichai's new phrasing of links and sources as a "part" of search signifies a strategic shift. This change aims to keep users within Google's ecosystem by integrating information directly, moving the company from a traffic distributor to an AI publisher. The selection of these sources consequently grants Google significant editorial power over information.
Key Takeaways
- Google is redefining the role of links and sources in its search product, now referring to them as a "part" rather than its foundation.
- This linguistic shift supports Google's strategy to retain users within its ecosystem by integrating information directly through new AI-powered features.
Why it matters:
This redefinition of search as an AI publisher with editorial control over sourced content has significant implications for the web's information landscape and Google's influence.
Read Original →
The Decoder
Anthropic's Claude Mythos Preview, in partnership with 50 organizations, has identified over 10,000 critical vulnerabilities in system-critical software. The AI is discovering bugs at a rate that outpaces developers' ability to fix them, leading to a period of significant risk. Anthropic acknowledges that current safeguards are insufficient to prevent the misuse of these powerful AI models.
Key Takeaways
- AI models like Claude Mythos Preview can discover system-critical software bugs at an unprecedented pace.
- The rate of bug discovery is currently exceeding the capacity of human developers to patch them, creating a security challenge.
- Existing safeguards are deemed inadequate to prevent the misuse of advanced AI models.
Why it matters:
This development highlights a critical gap between AI capabilities and current cybersecurity practices, potentially leading to widespread vulnerabilities if not addressed.
Read Original →
Hugging Face Blog
Nemotron-Labs is developing Diffusion Language Models that aim to achieve "speed-of-light" text generation. These models leverage a novel diffusion process to create text more efficiently than traditional autoregressive methods. The goal is to significantly reduce latency and increase the speed of generating large amounts of text.
Key Takeaways
- Nemotron-Labs is pioneering diffusion models for language generation.
- The new approach promises significantly faster text generation speeds.
Why it matters:
This advancement could revolutionize real-time AI applications, enabling faster chatbots, content creation tools, and other text-intensive services.
Read Original →