TLDR AI 2025-09-08

GitHub Universe returns October 28–29 in San Francisco, and the full agenda is now live! Early Bird pricing ends September 17, so this is your last chance to save $400. This year's focus is AI agents, automation, and the tools reshaping how we build software.

• Try GitHub's newest AI features as soon as they're announced.
• Hear from leaders at Red Hat, Logitech, Block, Amazon, Figma, HubSpot, GM, and more.
• Have your resume reviewed at the Career Corner, connect with peers over board games, and meet your open source heroes.

⏰ Early Bird ends September 17. Register now to save $400 →

ASML is committing 1.3 billion euros to Mistral's 1.7 billion euro fundraise, making it set to become Mistral's top shareholder. It is expected to get a board seat at Mistral. The funding round will make Mistral the most valuable AU company in Europe. The cash could help Mistral make Europe less reliant on US and Chinese AI models.

A federal judge ruled in June that it was legal for Anthropic to train on copyrighted materials but that it was illegal to acquire the materials from pirate sites like LibGen. Anthropic has offered to settle the class action suit from authors by offering $3,000 per work.

OpenAI has raised its projected cash burn through 2029 to $115 billion, $80 billion higher than the company previously expected. It will burn more than $8 billion this year, $1.5 billion more than it projected earlier this year. The company plans to develop its own data center server chips and facilities to power its technology and control its soaring costs. Its first chip, developed in partnership with Broadcom, will only be used internally and won't be made available to customers.

The era of rapid compute scaling that drove AI progress since 2020 is over. GPUs from large cloud providers are almost entirely saturated, and frontier labs are now racing to bring new data centers online and reserve as much chip production as they can. Scaling, and therefore investment, is hampered by physical reality and the years required to build more infrastructure, meaning incremental scaling is the only option ahead, despite headlines of massive upfront funding rounds.

Hallucinations persist because current evaluation methods reward wild guessing that might occasionally be correct over admitting a lack of knowledge. The fix is simple: calibrate rewards during training based on certainty, penalize confidently wrong errors more, and give partial credit when models hedge or express doubt.

Building a popular AI safety movement risks undermining existing organic public support by making it appear coordinated rather than grassroots. What has worked so far is authentic opposition to specific policies and elite-focused expert advocacy. The author argues that broad movements inevitably get captured by adjacent political causes like labor or environmental issues, diluting safety-specific goals while exposing credible organizations to guilt-by-association with more extreme protesters.

Use this practical guide to design your AI governance program. Learn how to build on AI principles aligned with global frameworks like the EU AI Act and NIST AI RMF, establish clear policies and oversight, and embed governance in daily workflows. Get inspiration from OneTrust's real-world approach to risk, collaboration, and Agentic AI. Get the eBook

The Agentic Systems Series is a complete guide for building AI coding assistants that actually work in production. It covers everything from fundamental concepts to implementing enterprise-ready collaborative systems. The series provides deep analysis of real production systems, including Amp, Claude Code, and open-source implementations like anon-kode. It is assumed readers are familiar with system design concepts and have a basic understanding of AI/LLM integration and experience with either TypeScript/Node.js or similar backend technologies.

NVIDIA and AMD have been racing to solve data movement at the cluster scale to create a software moat. Voltron Data's Theseus is the first distributed runtime to put movement at the center of scheduling, memory, and I/O, and then prove the payoff with benchmarking results. It is open, composable, and already running on CUDA and ROCm. Theseus runs across both Nvidia and AMD ecosystems for datacenter-scale analytics and AI.

Alibaba has introduced Qwen3-Max-Preview, a 1 trillion-parameter model now accessible via Qwen Chat and Alibaba Cloud. Early benchmarks highlight improvements in instruction following, dialogue, and agentic task performance.

Dozens of open source projects are currently testing different recipes for how to work with AI productively. These efforts show that AI works best when it is given structure. The frameworks are converging on a future where AI becomes a set of teammates to be managed. The more structure you give, the more you get back.

Companies like Anysphere face soaring LLM inference costs, forcing price hikes and prompting exit strategies like reverse acqui-hires. The coding AI market's rapid growth, driven by reasoning models, is now pressured by expensive compute costs, leading vendors to adopt usage-based pricing.

SEO jobs are changing to include the distribution of content. They now focus more broadly on the discoverability of a website. Focus on methods that drive brand awareness growth. Pivot to tracking LLM visibility, structure content for LLM digestibility, change content approaches, and think about the future of conversion to prepare for the next few years.

AI is table stakes - but knowing what the build isn't. Learn how to move the right bets forward faster with product leaders from Miro and AND Digital. Join live on September 16

GPT-5 Thinking leads to far more comprehensive results, but it's a lot slower.

Moonshot AI has released Kimi-K2-Instruct-0905, a mixture-of-experts (MoE) model featuring 32B activated parameters out of 1T total.

Starting January, Medicare will implement an AI pilot program in six states to assist in making coverage decisions through the WISeR Model.

Anthropic's researchers have embraced moderate accelerationism, advocating for measured and responsible AI advancement.

Following several public incidents regarding ChatGPT's personality, OpenAI is absorbing the Model Behavior team into the larger Post Training group, suggesting that personality is now considered a core capability to measure and hone during model development.