Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
The subagents pattern is now widely supported in coding agents. Subagents are now generally available in OpenAI Codex. There are default subagents called 'explorer', 'worker', and 'default', but it is unclear what differentiates them. Users can define custom agents with custom instructions and specific models.
Leanstral is an open-source coding agent designed for Lean 4, a proof assistant capable of expressing complex mathematical objects. It is designed to be highly efficient and is trained for operating in realistic formal repositories. Leanstral's weights were released under an Apache 2.0 license. It can be accessed in an agent mode within Mistral vibe and also through a free API endpoint.
Apple will invest $14 billion into AI this year, a tiny amount compared to the $700 billion Amazon, Alphabet, Meta, and Microsoft are investing. The company appears to believe that the AI infrastructure build-out will produce inadequate returns. It is spending less due to a conviction that AI models will commoditize and shrink, that existing product lines will absorb the workloads the cloud was built to serve, and that the durable franchise belongs to whoever owns the customer. Apple is betting on its AI-capable devices rather than centralized infrastructure.
Nvidia's focus this year at its GTC event shifted to inference, the type of computing required to run models and allow them to respond to user queries. The AI industry is now less concerned with training AI models, which is what GPUs are best at, and more preoccupied with running them and generating profits from end-users. Inference requires different hardware than chips optimized for training. How far ahead the company remains in the AI-infrastructure race will depend largely on how effectively it is able to pivot its product road map from training to inference.
Sponsored
We benchmarked five MCP architectures against real enterprise queries. Most were accurate 60–75% of the time. CData Connect AI:
Nvidia Dynamo 1.0 accelerates generative AI and reasoning models in large-scale distributed environments by delivering low-latency, high-throughput distributed inference.
The proposed deal could give OpenAI a faster route into corporate adoption while providing PE firms with a potential lifeline for companies in their portfolios that are exposed to AI disruption.
Drew Bent is the education lead at Anthropic.
Open-H-Embodiment is a community‑driven healthcare robotics dataset created for the training and evaluation of AI autonomy and world foundation models for healthcare applications.
Sponsored
This free, editable board report template [4] helps CISOs and security leaders communicate AI risk, posture, and priorities in a way the board understands, using real metrics, risk narratives, and strategic framing.
Claude Opus 4.6 and Sonnet 4.6 now include the full 1M context window at standard pricing on the Claude Platform. This means fewer compactions and more of the conversation kept intact. 1M context is now included in Claude Code for Max, Team, and Enterprise users with Opus 4.6. Standard pricing applies across the full window, and there is no multiplier.
OpenAI's plan to drop its ban on X-rated content sparked vigorous internal debate over the potential risks. The company is forging ahead with its erotica plans despite concerns from council members with backgrounds in fields like psychology and cognitive neuroscience. OpenAI staff have identified several risks, including the potential for compulsive use and emotional overreliance on the chatbot. The company has developed a plan to monitor for a range of potential long-term effects of adult mode. Adult mode's release has been delayed as OpenAI is currently prioritizing other products.
Google's rebranded design tool, Stitch, will transform into a 3D workspace with AI-powered collaborations, replacing the flat canvas design. Enhancements include voice controls, a conversational agent, and the ability to generate functional React applications from designs. These advancements position Stitch as a comprehensive tool from concept to production, likely featuring at Google I/O 2026.
Transformers can execute programs efficiently inside their own inference loop. This opens a path towards AI systems that integrate learned representations with compiled algorithms inside a single computational substrate. Solving humanity's toughest problems will require systems that can both reason flexibly and compute reliably. Future AI systems will have software as part of the model.
The current industry zeitgeist is dialed in on CLIs, just like it was on MCP a few short months ago. Using CLI can result in token savings, but custom CLIs run into the same context problems as MCP, except without structure and many other sacrifices. Individual usage of coding agents looks very different from organizational adoption of coding agents. MCP is the present and future for enterprise and org-level use cases.
AMI Labs and World Labs both raised over $1B on "world models," but the term covers five distinct approaches: JEPA, spatial intelligence, learned simulation, physical AI infrastructure, and active inference. The sharpest result in the piece is V-JEPA 2, achieving zero-shot robot planning after training on just 62 hours of domain-specific data. Each approach solves a different subproblem, and the lines between them are expected to blur fast.
Sponsored
Replay is Temporal's practical conference for devs building real systems. Join for hands-on workshops and leave with prod-ready AI
Z.ai's GLM-5-Turbo offers a streamlined API for generating text responses, catering to tasks like crafting marketing slogans. Key features include customizable roles, real-time streaming, and adjustable creativity through the "temperature" setting.
Residual connections have long relied on fixed uniform accumulation. Attention Residuals replace standard depth-wise recurrence with learned, input-dependent attention over preceding layers. It enables networks to selectively retrieve past representations, naturally mitigating dilution and hidden-state growth. Attention Residuals have been validated on the Kimi Linear architecture, demonstrating consistent downstream performance gains.
Spatial-TTT is a framework that updates a spatial state from streaming visual inputs using test-time training, enabling models to reason over video-based spatial tasks. The method processes visual chunks incrementally and achieves strong results on video spatial reasoning benchmarks.
IndexCache reduces the cost of DeepSeek Sparse Attention by reusing top-k token indices across layers instead of recomputing them every time. The approach removes most indexer computations while maintaining model quality.
AWS is deploying Cerebras CS-3 systems to offer the industry's fastest AI inference via AWS Bedrock, using open-source LLMs and Amazon's Nova models. The collaboration introduces a disaggregated architecture, pairing AWS Trainium for prefill with Cerebras WSE for decode, boosting token throughput by 5x. This setup enhances high-speed inference performance by efficiently utilizing specialized hardware for each computational phase.
Sydney data engineer Paul Conyngham used ChatGPT, AlphaFold protein modeling, and a $3,000 genomic sequencing of his dog Rosie's tumor to produce a half-page mRNA formula that UNSW's RNA Institute turned into a physical vaccine in under two months. One tumor shrank 75% after her first injection in December 2025, marking the first personalized cancer vaccine ever designed for a dog. The pipeline mirrors Moderna and Merck's ongoing human trials. UNSW researchers are now asking why the same approach isn't being applied more broadly to cancer patients.
Sponsored
One API for Google search, Maps, Shopping & more. Clean JSON. No
This research tool visualizes 342 occupations, showing projected growth outlook, median pay, education requirements, and AI exposure.
This page features a collection of architecture figures and fact sheets from The Big LLM Architecture Comparison and A Dream of Spring for Open-Weight LLMs, focusing on the architecture panels only.
claudetop shows users where exactly their tokens and dollars go in real time.
Its rivals, Zhipu and MiniMax, recently traded at valuations between $30 billion and $40 billion in Hong Kong.
ByteDance paused the planned global rollout of its Seedance 2.0 AI video generator after viral clips triggered legal complaints from major Hollywood studios over alleged use of copyrighted characters and likenesses.
AMD is pitching a new device category called the "Agent Computer," always-on local hardware that runs AI agents continuously in the background while you sleep or work on other things, delegating tasks through Slack, WhatsApp, or iMessage.
Sponsored
First impressions matter. With Framer [4], early-stage founders can launch a beautiful, production-ready site in hours. No dev team, no hassle. Join hundreds of YC-backed startups who launched there and never looked back.
* NO CODE, NO DELAYS: Launch a polished site in hours, not weeks, without hiring developers. * BUILT TO GROW: Scale your site from MVP to full product with CMS, analytics, and AI localization. * JOIN YC-BACKED FOUNDERS: Hundreds of top startups are already building on Framer.
Elon Musk just pulled two of Cursor's product leaders into a $1.25 trillion company and told them to build xAI's coding product. AI coding is now a $5 billion+ market. xAI is the only frontier lab without a coding product printing money. AI labs have figured out that developers are the highest-willingness-to-pay users on the planet.
A beta version of Imagine with Claude is coming to Claude Chat. The feature can create custom charts, diagrams, and other visualizations, and then tweak and modify its creations as the conversation develops. The feature will be on by default, with Claude deciding when to build a visual for something. Users can directly request a visual. Videos showing example visualizations are available in the article.
Meta's Avocado model outperforms the company's previous model and the Google Gemini 2.5 model from March, but it doesn't perform as strongly against leading models from Google, OpenAI, and Anthropic. The company has decided to delay the model's release to at least May. Experts believe Meta still has time to catch up to rivals. The company's Llama 4 release fell short of expectations last year.
Axiom raised a $200M Series A at a $1.6B valuation led by Menlo Ventures, one year after founding and with roughly 20 employees, to build what it calls Verified AI: systems that produce formally verified outputs in the programming language Lean, where every reasoning step is machine-checkable, and errors are caught deterministically rather than statistically. The company originally trained on math proofs and then transferred that skill to code verification, a technique that lets it prove not only that code produces correct outputs but also that it won't create unexpected attack surfaces.
Claude's generative UI is a tool call that returns HTML injected into the DOM with incremental parsing as tokens stream. It enables interactive widgets to be rendered inline in Claude conversations. It uses lazy-loading documentation to insert context on demand. The terminal spawns Glimpse windows with bidirectional JSON communication, allowing the terminal to stay as a terminal while the visual content gets a real browser engine.
Commerce has historically been closer to a seller's market. Consumers are largely shown what platforms choose to surface. AI agents will gradually shift the power toward the buyer. Instead of navigating the layer of market and platform incentives, AI will enable consumers to access the best product for their needs directly. This dynamic may be most powerful in B2B commerce, which has historically been opaque, fragmented, and relationship-driven.
Institutional AI could unlock far more value than individual AI by redesigning organizations alongside technology. Unlike individual AI, which boosts productivity with little firm value impact, institutional AI focuses on coordination, finding signals in data, countering bias, and scaling revenue. Successful organizations will be those that integrate AI as a foundational component, similar to the industrial shift to electric assembly lines.
Sponsored
80% of your company's knowledge is trapped in unstructured silos. OpenSearch [13] unlocks it. As the the open-source backbone for AI-powered retrieval and agentic workflows, turning your "data swamp" into an instant answer engine.
Slate is a frontier agent that uses a code environment for swarm orchestration. It can programmatically orchestrate and solve tasks by running a massive number of subagents. Slate supports a range of models and automatically selects the right model for the job. It uses novel context engineering and maximizes caching to lower costs.
A single run(command="...") tool with Unix-style commands outperforms a catalog of typed function calls. Unix made the design decision 50 years ago to make everything a text stream. Large language models made an almost identical decision by making everything tokens. The text-based system that Unix uses is a natural fit for large language models. They can act as terminal operators faster than any human. Developers just need to take what Unix has proven over the last half-century and hand it directly to AI.
Cursor uses a hybrid online-offline eval process to keep its understanding of model quality aligned with what developers actually do. The offline part uses CursorBench, an internal eval suite based on real Cursor sessions from Cursor's engineering team. It measures multiple dimensions of agent performance, including collusion correctness, code quality, efficiency, and interaction behavior. The online part involves a controlled analysis on live traffic, which catches regressions that offline suites miss. Together, this loop keeps Cursor's notion of model quality grounded in production as workflows change.
OpenClaw-RL is an asynchronous reinforcement learning framework for training personalized AI agents from everyday conversations. It supports large-scale environment parallelization for developing more general agents.