TLDR AI 2025-10-20

Securely authorizing access to an MCP server is complex. You need PKCE, scopes, consent flows, and a way to revoke access when needed.This guide from WorkOS explains how to implement OAuth 2.1 in a production-ready setup, with clear steps and examples.WorkOS Connect and AuthKit handle the full flow from authorization to token management.Implement MCP Auth with WorkOS →

Apple engineers testing the early iOS 26.4 build have concerns about Siri's performance despite extended development time. Internal teams are working on two Siri models, one using on-device models and the other powered by Google Gemini. Concerns persist about Apple's AI strategy, potentially leading to reliance on external solutions.

Google Maps integration is now available in the Gemini API, enabling developers to incorporate Maps data into geospatial-aware AI applications. The tool enhances use cases in travel, real estate, retail, and logistics by providing precise, location-based responses. The tool's combination with Google Search further boosts response quality, offering structured data and timely information.

Alibaba Cloud's new Aegaeon system reduces Nvidia GPU usage by 82%, cutting the number of GPUs needed for large AI models from 1,192 to 213. Beta tested in their model marketplace, Aegaeon addresses resource inefficiency by pooling GPU power. This pooling allows one GPU to serve multiple models, optimizing concurrent LLM workload management.

Andrej Karpathy was recently on a podcast where he discussed AGI timelines, Animals vs Ghosts, Reinforcement Learning, AI agents, and job automation. Some of the explanations on the podcast were botched. This post expands on the topics that Karpathy discussed on the podcast to clear things up. It includes links to previous discussions Karpathy has had on these topics.

The biggest mistake people make with paid AI subscriptions is staying in auto mode, where the system picks weaker models instead of manually selecting agentic modes that are far more capable and less error-prone for real work. This guide covers practical choices like when free models suffice versus paying $20 monthly, why prompting techniques that worked on older models no longer matter, and why explicitly telling AI to act as a critic is necessary to avoid talking to a sophisticated yes-man.

In a wide-ranging conversation, Andrej Karpathy discusses AGI's economic impact, the limits of reinforcement learning, lessons from self-driving research, and his vision for the future of education.

Struggling to wrangle your AI agents? Watch hands-on technical sessions to bring agentic AI into your core processes — without the chaos. Don't miss these high-signal talks from Camunda, AWS, Atlassian, and EY SMEs. See how to make agentic AI work for you.

Google's Coral NPU is an open-source platform designed for low-power, always-on AI at the edge. It tackles key issues like fragmentation, performance, and privacy for edge devices and wearables.

NCPNET adapts conformal prediction to dynamic graphs by using a diffusion-based non-conformity score and efficiency-aware optimization.

o4-mini-deep-research costs $2/$8 per million for input/output tokens. It is only available via the Responses API. OpenAI recommends running them using its background mode. This post takes a quick look at the model's abilities.

Discoveries were simple in the early days of science. Many were discovered in the course of pursuing a hobby. To win the 2026 Nobel Prize in physics, you have to be a physicist fully immersed in the world of physics. AI research is still very much in the 'early days of science' category. While many papers look complicated, AI research discoveries are often simpler than they look. Many things we learn about AI capabilities come from informal user experimentation.

Nvidia has revealed the first Blackwell wafer made in TSMC's semiconductor manufacturing facility in Phoenix, Arizona. The Blackwell architecture is now ready for volume production. Nvidia is still working on expanding its manufacturing footprint across the US. It plans to funnel half a trillion dollars towards building AI infrastructure in the US through partnerships with TSMC, Foxconn, and other companies.

This report looks at global sector trends in generative AI, with the data ending on October 10. It covers sectors such as general AI tools, character and chat, design and image generation, writing and content, EdTech AI, customer support, and more. The report looks at total visits at the domain level and does not capture API usage or integrations.

Use Cognizant's open-source library to vibe code a working multi-agent network based on a description; connect APIs and tools; share data between agents without exposing prompts. Read the details

Anthropic is buying TPUs from Broadcom.

Google AI Studio has rolled out new UX improvements to reduce friction and unify the developer workspace.

Meta won't allow AI model providers to distribute their AI assistants on WhatsApp starting on January 15.

Meta's Yann LeCun and DeepMind CEO Demis Hassabis publicly mocked OpenAI after VP Kevin Weil claimed GPT-5 solved 10 previously unsolved Erdős problems, but the mathematician maintaining the list clarified they were only marked “open” because he personally didn't know of solutions.