Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
Sponsored
AI systems fail when agents can't share state or recover from failures. Build multi-agent architectures with LangGraph for orchestration and AWS for durable messaging. Join the AWS technical
Google's Gemini CLI now includes subagents to split coding tasks, enhancing parallel execution by delegating specific roles like frontend updates or testing. This enables multiple tasks to process simultaneously without interference, optimizing workflows for developers. Gemini's setup contrasts with systems like Claude Code, which extends agent coordination across multiple sessions.
Qwen3.5-Omni scales to hundreds of billions of parameters with a hybrid MoE architecture, supporting long-context multimodal inputs across text, audio, and video.
TIPSv2 improves vision-language pretraining by combining distillation, enhanced self-supervised objectives, and richer caption data. The resulting models achieve strong performance across multimodal tasks, with notable gains in zero-shot segmentation.
FlashDrive is an algorithm-system co-design framework that reduces end-to-end latency to 159ms with negligible accuracy loss. VLA inference is a cascade of stages, each hiding a different form of redundancy. Temporal overlap in vision, low entropy in reasoning, velocity smoothness in flow matching, and numerical headroom in weights each yield to a targeted shortcut. The speedups compound to 4.5x with negligible accuracy loss because the redundancies are orthogonal.
The US is in the model of an unprecedented build-out of AI infrastructure. Stargate is a $500 billion endeavor that involves OpenAI, Oracle, and SoftBank. The AI infrastructure project has seven locations across the US that are currently showing active development. They currently add up to over 9 gigawatts of planned capacity, enough to power the equivalent of 20 million Nvidia H100 GPUs - the total amount of AI compute in the world at the end of 2025. This post takes a look at each of the sites and how they are currently being developed.
Microsoft plans to temporarily suspend individual account signups to GitHub Copilot. The weekly cost of running the service has doubled since the start of the year. Microsoft intends to tighten rate limits on individual and business accounts. It will also remove access to certain models for those with the cheapest subscriptions.
TLDR's Applied AI team is tasked with making every process at TLDR legible to code, runnable by anyone, and composable into larger workflows. Join a small, fast moving team using the latest AI tools with an unlimited token budget. Learn more [22].
Live artifacts are dashboards and trackers connected to apps and files with current data.
Anthropic and Amazon have expanded their collaboration to secure up to 5 gigawatts of compute capacity, supporting Claude's training and deployment.
This is a list of the nine best things they said.
Sponsored
As AI agents replace small, iterative commits with massive, asynchronous code payloads, a "verification gap" has emerged. Small human commits told you what changed; large AI commits dare you to find out. To maintain velocity without compromising security, engineering leaders must shift to the Agent Centric Development Cycle (AC/DC).
Claude Design is a tool for creating visual work using the Claude Opus 4.7 vision model. It allows users to design prototypes, pitch decks, and marketing materials, automating brand consistency and enabling collaboration. The tool integrates with Claude Code for seamless transition from prototype to production.
Cursor, an AI coding startup, is close to raising $2 billion in funding, potentially doubling its valuation to $50 billion. Thrive and Andreessen Horowitz are expected to lead the round, with Battery Ventures and Nvidia possibly participating. Cursor aims to triple its annualized revenue to over $6 billion by the end of 2026, driven by its proprietary Composer model and strategic shifts to achieve profitability.
Two of the architects of OpenAI's most ambitious moonshots, Kevin Weil and Bill Peebles, are leaving the company. Weil led the company's science research initiative while Peebles was the researcher behind Sora. Both announced their departures on Friday. OpenAI recently decided to cut back on 'side quests' as it consolidates around enterprise AI and its forthcoming superapp.
The length of tasks AI agents can perform has been growing exponentially over the last seven years. The latest models can (sometimes) do tasks that would take a human a few hours. However, the cost to achieve these time horizons is growing exponentially, and the hourly costs for some models are now close to human costs. There will eventually be a divergence between what time horizon is possible and what is economically feasible.
NEMOTRON OCR V2, developed using synthetic data, is a fast, accurate multilingual OCR model achieving significant accuracy improvements, lowering NED scores to near-zero for non-English languages. Utilizing a synthetic data pipeline with mOSCAR text and diverse fonts, the model trains with pixel-perfect annotations across languages, enabling generalization to real-world documents. The unified architecture reuses feature maps, achieving speeds of 34.7 pages/second on a single A100 GPU, outperforming specialized models in diverse language OCR tasks.
Anthropic is the only major AI lab to publish the system prompts for its user-facing chat systems. The company just released Opus 4.7, which features an updated system prompt. This author used Claude Code to take the Markdown version of its system prompts, break them into separate documents, and then construct a Git history of those files with fake commit dates. The result is in the post.
Peter Steinberger's talks highlighted OpenClaw's contrasting stories: an inspiring public narrative and a more serious engineering perspective on security and scaling challenges. Anthropic launched Claude Design to compete with tools like Figma, expanding beyond chat/coding into design and prototyping with its OPUS 4.7 engine, which received mixed initial user reviews despite strong benchmark performances. Open-source agent stacks like Hermes and advancements in computer-use UX are gaining traction, while applied AI advances in science, medicine, and infrastructure underscore the sector's dynamic growth.
Sponsored
Agents and services calling each other at scale no longer need a network round-trip to validate auth. Clerk's new JWT M2M tokens verify locally using your instance's public key, with significantly lower latency and zero per-verification cost.
Hybrid inference for Android is a new API for Firebase AI Logic that leverages both on-device and Cloud inference. It supports Google's new Gemini models, including the latest image generation Nano Banana models. The new API allows apps to dynamically switch between Gemini Nano running locally on-device and cloud-hosted Gemini models. It is still experimental.
xAI has launched standalone Grok Speech to Text (STT) and Text to Speech (TTS) APIs, enhancing developer options for integrating advanced speech capabilities. These APIs support high accuracy, low latency, word-level timestamps, speaker diarization, and intelligent inverse text normalization across 25+ languages. Grok STT excels in transcription accuracy within phone calls and video/podcasts, providing robust solutions for medical, legal, and financial sectors.
Prefill-as-a-Service (PrfaaS) is a cross-datacenter serving architecture that selectively offloads long-context prefill to standalone, compute-dense prefill clusters and transfers the resulting KVCache over commodity Ethernet to local PD clusters for decode. It combines model-side KV efficiency with system-side selective offloading, bandwidth-aware scheduling, and cache-aware request placement. The design removes the requirement that heterogeneous accelerators share the same low-latency RDMA fabric and enables independent scaling of prefill and decode capacity across loosely coupled clusters. A PrfaaS-augmented heterogeneous deployment achieves higher serving throughput while consuming only modest cross-datacenter bandwidth.
Claude Code is an agentic coding tool that can run shell commands, edit files, and call external services on behalf of the user. This study describes its architecture by analyzing the publicly available TypeScript source code and further comparing it with OpenClaw. The core of the system is a simple while-loop that calls the model, runs tools, and repeats. OpenClaw uses the same recurring design questions but produces different architectural answers when the deployment context changes. The study also identifies six open design directions for future agent systems.
Google is in talks with Marvell Technology to develop a memory processing unit and an inference-optimized TPU. The custom ASIC market is projected to grow 45% in 2026 and reach $118 billion by 2033. Broadcom, Google's primary custom chip partner, recently announced a long-term agreement to design and supply TPUs and networking components through to 2031. Talks with Marvell have yet to produce a signed contract.
Better AI models, like Opus 4.5 and GPT-5.2, increased developer AI usage by 44%, enabling more complex tasks after an initial adjustment period. Industries including media and advertising saw notable usage growth, driven by competitive pressures and new opportunities. Developers shifted toward managing AI output, with significant increases in documentation, architecture, and learning tasks.
Canon defines its search pipeline as a DAG to ensure automatic concurrency across classification, localization, retrieval, and ranking processes. This setup allows independent node control, enabling easy subsystem changes without affecting the overall pipeline. The DAG framework optimizes execution through parallelism, supports durable execution with node-specific retries, and provides clear introspection and execution decoupling.
TLDR's Applied AI team is tasked with making every process at TLDR legible to code, runnable by anyone, and composable into larger workflows. Join a small, fast moving team using the latest AI tools with an unlimited token budget. Learn more [25].
Canva AI 2.0 is available as a research preview starting today to the first one million people who visit the Canva website.
Google is testing a bridge between Gemini subscriptions and AI Studio, offering users subscription-based token access instead of separate API-key billing.
Tencent released a multimodal framework that generated and reconstructed 3D worlds from text, images, and video using a staged pipeline and a unified feed-forward model.
Sponsored
Making sense of the technology landscape has never been more challenging thanks to AI-accelerated change. But understanding what matters and what it means for your organization has never been more important.
Anthropic has released Claude Opus 4.7, offering improved performance on difficult engineering tasks, stronger vision capabilities, and more reliable long-running task execution compared to its predecessor.
OpenAI introduced GPT‑Rosalind, a specialized model designed to support drug discovery and biological research through improved reasoning, tool use, and integration with scientific data sources.
OpenAI updated Codex with background computer control, multi-agent workflows, and deeper integration with developer tools, extending its role across the full software development lifecycle.
Perplexity launched "Personal Computer," an AI platform that shifts the operating system model from manual instruction execution to probabilistic goal completion. The agent utilizes deep web research as its core foundation to autonomously evaluate reasoning paths and drive multi-step workflows. This architecture transforms the computer into an active orchestrator, eliminating the administrative friction of managing fragmented software tools.
Jensen Huang recently appeared on an interview where he discussed mostly familiar topics. However, there were three exchanges where Jensen said more than he probably intended. Jensen even lost his composure when discussing China and whether sales of chips should be restricted in the region. This post details each of the exchanges and notes other observations about the interview.
This post contains rough notes on pretraining parallelisms, whether distillation can be stopped, Mythos and the cybersecurity equilibrium, Pipeline RL, and why pretraining runs fail.
A new Skill and Test Harness help port transformer models to mlx-lm, streamlining contributions and reviews. This tool assists contributors by managing model conversion tasks and provides reviewers with agent-assisted PRs complete with comprehensive reports. It aims to maintain code quality and improve efficiency in an open-source environment where manual review cannot scale.
Sponsored
DigitalOcean Deploy is April 28 in San Francisco. One day focused on production inference at scale. Real customer stories, a fireside with NVIDIA's Kari Briski on agentic AI, and a first look at DO's next-gen inference products. Speakers from Character, Workato, VAST Data, Arcee, and vLLM. In-person seats are limited, so claim yours! REGISTER
Qwen 3.6 introduces stronger repository-level reasoning and front-end workflow handling, along with a thinking preservation feature that maintains context across iterations.
Ternary Bonsai is a 1.58-bit language model family offering improved performance with minimal memory usage. It outperforms 1-bit counterparts, scoring 75.5 on average benchmarks and achieving 3-4x better energy efficiency. The models, available in 8B, 4B, and 1.7B versions, provide flexible deployment across devices like Macs and iPhones under the Apache 2.0 License.
This guide outlines a structured approach to modernizing large codebases using sandboxed agents that operate on isolated tasks, validate changes, and return auditable patches while orchestration remains outside the execution environment.
Anthropic's chief product officer stepped down from Figma's board following reports that upcoming models may include design tools competing with Figma's core product.
OpenAI will pay Cerebras more than $20 billion over the next three years to use servers powered by Cerebras' chips. The deal may also give OpenAI an equity stake in Cerebras. OpenAI has also agreed to provide Cerebras with about $1 billion to help fund the development of data centers. Cerebras is targeting an IPO in the second quarter of this year.