PLUS: ByteDance's hyperrealistic AI video, a new AI diagnostic tool, and Stripe's coding minions
Happy reading
A new startup is tackling AI's speed and cost problems by hard-wiring a Llama model directly onto a chip. The specialized silicon from Taalas results in a 10x performance boost at a fraction of the power and cost.
This approach could solve major latency issues that currently limit real-time AI applications. But is creating model-specific hardware the key to more efficient AI, or will its inherent inflexibility become a major obstacle?
In today’s Next in AI:
Taalas’s silicon Llama runs 10x faster
ByteDance’s hyperrealistic AI video draws backlash
An AI tool to diagnose rare diseases
Stripe's coding minions automate development
The 10x Speed-Up
Next in AI: Newcomer Taalas has unveiled a specialized chip that runs a Llama model 10x faster at a fraction of the cost and power, aiming to solve AI's biggest latency and cost bottlenecks.
Explained:
Taalas achieves its performance by creating specialized silicon for each AI model, essentially hard-wiring the Llama 3.1 8B architecture directly onto its chip to eliminate memory access delays.
The first chip hits an impressive 17,000 tokens per second per user, a speed that dwarfs competitors like Nvidia and Groq while consuming 10x less power and costing 20x less to build.
To reach these speeds, the initial version uses aggressive quantization that can slightly degrade quality, but developers can already test its performance via a public chatbot demo and an API.
Why It Matters: This approach could drastically lower the cost and latency barriers that currently limit many real-time AI applications. If successful, it points to a future where powerful AI isn't just for massive data centers but can be deployed more widely and efficiently.
China's Hollywood Disrupter
Next in AI: TikTok’s parent company ByteDance has launched Seedance 2.0, a video generator that creates incredibly realistic clips from simple prompts. Its stunning output quickly went viral, but also drew immediate fire from Hollywood over intellectual property concerns.
Explained:
The model generates hyperrealistic video clips up to 15 seconds long from text, images, and even audio prompts. It offers users fine-grained motion editing control and can create scenes with impressive physics and native audio-video synchronization.
Major studios, including Disney and Paramount, promptly issued cease-and-desist letters, accusing ByteDance of the unauthorized use of copyrighted characters and celebrity likenesses to train its model.
This release intensifies the global AI race, showcasing a powerful new tool in an ecosystem fueled by China's AI youth boom. The situation is also unfolding as Chinese regulators increase enforcement for labeling AI-generated content.
Why It Matters: The realism of Seedance 2.0 forces a major confrontation between tech innovators and established media over the future of content creation and copyright law. This clash accelerates the urgent, global need for clear rules on how AI models can be trained and deployed responsibly.
AI's Diagnostic Leap
Next in AI: A new AI system called DeepRare is outperforming human specialists in diagnosing rare diseases, according to a new study in Nature. The system offers a new path to shortening the years-long "diagnostic odyssey" for millions of patients.
Explained:
In a head-to-head comparison, DeepRare achieved 64.4% accuracy on its first diagnosis, surpassing the 54.6% rate of experienced rare disease physicians on complex cases.
Instead of just matching symptoms, the system uses an agentic workflow with over 40 specialized tools to form and test hypotheses, reasoning through cases much like a human doctor.
The platform is already gaining traction, attracting over 1,000 users from more than 600 medical institutions worldwide since its web platform launched in July 2025.
Why It Matters:
This development could dramatically reduce the time and cost associated with diagnosing over 7,000 rare disorders that affect 300 million people globally. It demonstrates how agentic AI can serve as a powerful collaborator for human experts, tackling complex problems that require deep, traceable reasoning.
Stripe's AI Workforce
Next in AI: Stripe has revealed its homegrown AI coding agents, called “minions,” now autonomously handle over 1,300 pull requests every week. These agents operate end-to-end, producing code that is human-reviewed before being merged.
Explained:
Minions operate within isolated “devboxes,” cloud development environments originally created for human engineers. This allows Stripe to safely run agents in parallel without giving them access to production services or real user data.
The agents are directed by “blueprints,” a custom system that blends deterministic code for predictable tasks (like running linters) with flexible agent loops for complex problem-solving, all built on a forked version of an open-source tool.
This signals a wider industry trend of deploying autonomous agents for software development. For example, the company Ona recently reported that its agents authored 89% of its pull requests over an 80-day period.
Why It Matters: This system shows how AI can reliably automate routine coding tasks at a massive scale, freeing up engineers to focus on more complex, strategic problems. Stripe's success provides a practical roadmap for how other companies can leverage their existing developer infrastructure to deploy effective AI agents.
AI Pulse
Nvidia nears a $30B investment in OpenAI, a more straightforward equity deal that replaces the previously stalled $100B multiyear partnership framework.
Anthropic launched a research preview of Claude Code Security, a new tool designed to reason through codebases to find and suggest patches for complex security vulnerabilities.
The Pentagon is reviewing its relationship with Anthropic as tensions rise over the military's use of its models and a new policy requiring partners to allow "any lawful use" of their AI.
Reddit is testing an AI-powered search feature that converts product recommendations from community posts into shoppable carousels with links to retail partners.