TLDR AI 2026-04-15

Do you think in keywords? Then why is search still built around them?

Humans don't think like Google, and neither do agents. If you want to offer friction-free customer experiences, you need search tooling that thinks like your customers do.

Algolia's eBook breaks down how agentic AI is reshaping search to better reflect how people naturally seek information.

👉 Read it to learn how to prepare your org for agents by cleaning up your data, fleshing out security rules, and getting the ethics right.

Discover how you can benefit from agentic AI.

OpenAI is scaling up its Trusted Access for Cyber (TAC) program to thousands of verified individual defenders and hundreds of teams responsible for defending critical software. The company has released a variant of GPT-5.4 trained to be cyber-permissive called GPT-5.4-Cyber. OpenAI plans to make its tools as widely available as possible while preventing misuse. It will deploy its technology carefully as it continues to improve. Companies can now authenticate with OpenAI as cybersecurity defenders - those in the highest tiers will get access to GPT-5.4-Cyber.

Google is expanding NotebookLM with Canvas features for creating visual and interactive experiences from existing notebook sources. The new Connectors option suggests integration with other Google services, positioning NotebookLM as a central research layer. Labeling features and auto-categorization could streamline navigation for heavy users managing large source libraries.

Google introduced “Skills” in Chrome, enabling users to save and reuse AI prompts powered by Gemini across websites. The feature extended in-browser AI workflows by allowing one-click execution of common tasks like summarization or content transformation.

Daniel Kokotajlo, the founder of the AI Futures Project, wrote an essay called 'What 2026 Looks Like' in August 2021, before the launch of ChatGPT. His predictions turned out to be pretty accurate. This post features an interview with Kokotajlo where he discusses what he got right, what he got wrong, and how we should think about the pace of AI over the next few years. Kokotajlo recently released a report that predicts the next few years of AI development, which culminates in the rise of superhuman agents capable of wresting control from humanity.

Cursor detailed a multi-agent system that optimized 235 CUDA kernels for NVIDIA Blackwell GPUs, achieving a 38% average speedup, with some cases exceeding 2× improvements.

Google, Microsoft, Meta, Amazon, and Oracle now control two-thirds of the world's compute. Many AI labs depend almost entirely on these hyperscales for access to compute.

Keyless mode now works with TanStack Start, Astro, and React Router, so your coding agent can scaffold working Clerk auth from a single install without touching the dashboard. Core 3 also ships new hooks and agent-friendly APIs for custom sign-in, sign-up, and checkout flows.

Read the full release

Routines in Claude Code allow users to configure a routine once and then run it on a schedule, from an API call, or in response to an event. Routines have their own API endpoint. Webhook routines subscribe to GitHub events. Routines are available on all paid plans with Claude Code on the web enabled.

Agents can build software fast, but developers still need to put the effort into secure environments and code. Cloudflare has introduced scannable tokens to protect credentials, OAuth visibility to manage principals, and resource-scoped RBAC to fine-tune policies. This ensures that users' credentials (tokens) aren't leaked and that the right applications have access via OAuth, and narrows permissions using granular RBAC. Cloudflare customers should review their API tokens, review authorized OAuth apps, and review member and API token permissions.

Diffusion language models (DLMs) offer a compelling promise that parallel token generation could break the sequential bottleneck of autoregressive (AR) decoding. However, DLMs consistently lag behind AR models in quality. This gap stems from a fundamental failure of introspective consistency: AR models agree with what they generate, whereas DLMs often do not. The Introspective Diffusion Language Model (I-DLM) uses introspective strided decoding (ISD) to verify previously generated tokens while advancing new ones in the same forward pass. With gated LoRA, ISD enables bit-for-bit lossless acceleration.

Gemini Robotics-ER 1.6 upgrades robots' embodied reasoning, improving spatial reasoning and multi-view understanding for better autonomy. It enhances visual and spatial understanding, task planning, success detection, and introduces instrument reading capabilities. Available now via the Gemini API and Google AI Studio, the model supports integration with tools like Google Search and VLAs.

Claude prompt caching helps users avoid reprocessing previously used prompts to save on tokens. Caches can either have a five-minute or one-hour time to live. Writing to the five-minute cache costs 25% more in tokens, and writing to the one-hour cache 100% more, but reading from cache is around 10% of the base price. Some users are reporting issues with hitting limits too quickly, and there are reports that Claude's performance has dropped.

Microsoft has reached an agreement to lease high-performance computing capacity at Nscale's data center in Narvik, Norway. The 230 megawatt data center was originally intended for OpenAI's 'Stargate' initiative, but negotiations between OpenAI and Nscale reportedly concluded without an agreement. Microsoft's deal with Nscale involves the rental of 30,000 NVIDIA Vera Rubin GPUs. The expansion will help ensure access to the scalable capacity needed for the next generation of AI workloads for European customers.

Your AI agent can now turn warehouse queries into live, shareable data apps in a single conversation. No BI tool. No SQL. Start free

OpenAI acquired personal finance startup Hiro Finance in an apparent acqui-hire.

Fluidstack is a startup that builds specialized data centers for AI companies.

Meta and Broadcom have extended an existing partnership for the design of Meta's custom in-house AI accelerators through 2029.

Anthropic shared details of its unreleased Mythos model with the Trump administration due to its advanced cybersecurity capabilities, despite an ongoing legal dispute with the Department of Defense.