TLDR AI 2025-09-09

Our latest releases for Redis for AI are full of core enhancements, new integrations, and dev tools to help you build fast, reliable apps at scale—especially AI ones. Here's what's new: 🧠 Redis for AI

• Redis LangCache for fully managed semantic caching (Public preview)
• Hybrid search enhancements
• Vector compression
• Vertical scaling (GA)
• New agent framework integrations

Calculate how much you can save with LangCache now.

Some of California's biggest philanthropies, nonprofits, and labor groups are joining forces to push back on OpenAI's plan to restructure into a for-profit company. This has concerned OpenAI executives and resulted in discussions about moving out of the state. OpenAI's proposed plan is being investigated by Attorneys general in California and Delaware. The startup could potentially be sued for breaking nonprofit law and may be required to pay a settlement as a condition for the restructure.

Databricks has confirmed another $1 billion raise at a $100 billion-plus valuation. The company will use the funds to invest in its database for AI agents. The round was led by Thrive and Insight Partners, who also co-led the previous $10 billion raise. Databricks has $4 billion in annual recurring revenue.

Grok is working on a background thinking feature that appears to allow users to continue chatting while the assistant thinks. A screenshot of the toggle for the feature is available.

Google reduced Veo 3 pricing by nearly half to $0.40 per second and added support for vertical 9:16 videos and 1080p HD output.

Shawn Wang, founder of Smol AI, is joining Cognition, the makers of the Devin AI software engineer. The vast majority of Wang's equity/investor capital at Smol AI will be rolling over their investment to Cognition as part of the deal. Wang will continue to work on AI Engineer and Latent Space independently of Cognition. This post explains in detail Wang's reasons for betting on Cognition.

RLaaS (Reinforcement Learning-as-a-Service) has proven performance, lower costs, is more excludable, and is safer. AI companies pivoting away from pursuing AGI towards RLaaS is good news because it promises a switch to a safer mode of AI development. While RLaaS is not without risks, the transition to an RLaaS model would reduce AI risk. The concatenation of specialized models may create or assist general intelligences.

Anthropic details why it implemented ASL-3 safety protections against biological weapons development, noting Claude now exceeds expert performance on virology evaluations, and controlled trials showed significant uplift with Opus 4 for acquiring bioweapons compared to using just the open internet. It's obviously a low-probability risk, but the impact is so high — just one infraction could mean thousands of deaths — that it still requires serious evaluation and restriction.

Effortlessly deploy agentic workflows and generative AI with the Inference Cloud powered by Qualcomm. Designed for today's leading models, it offers exceptional performance and seamless scalability at a fraction of the cost.

Learn more or sign up now!

AI coding agents are great, but they're too wasteful, they don't hold the context of the database, they only have a surface-level understanding, and they're too susceptible to context rot. These problems could be solved if agents had a semantic understanding of all functions and context about how functions are connected to other parts of code. This article looks at practical ways to make coding agents more efficient.

Set Block Decoding enables parallel sampling of multiple future tokens by combining next-token and masked token prediction. This reduces decoding passes up to 5x without sacrificing accuracy.

Reinforcement Learning environments are sandboxes for AI models to learn from. They keep track of an internal state, prompt AIs to take actions to complete a task, and assign scores based on the outcome. More and more startups are offering increasingly challenging environments under the assumption that machine learning is solved and the only remaining work is to provide models with simulations to learn from. Frontier AI labs pay a lot of money for these environments, but the supply side is set to become increasingly crowded - it likely won't be a long-lived business.

A federal judge expressed serious concerns about Anthropic's proposed settlement over the use of 500,000 pirated books to train an internal version of Claude. He questioned how the settlement money would be distributed: “I have an uneasy feeling about hangers on with all this money on the table.”

Ready to build enterprise-ready AI agents that can handle complex tasks without breaking? Get your hands dirty with production-grade frameworks and LLMs in free builder workshops.Snag your spot to build beyond POCs

Perplexity introduced a secure, zero-data-usage AI platform for U.S. government agencies, alongside a $0.25-per-user Enterprise Pro offering tailored to public sector needs.

Cognition raised over $400M at a $10.2B valuation, in a funding round led by Founders Fund with participation from Bain Capital Ventures and others.

Nebius will provide Nvidia GPUs and AI cloud services to Microsoft out of their new New Jersey data center over 5 years.

Anthropic endorses California's SB 53, a bill mandating transparency in safety practices for frontier AI systems.