Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
Codex and Jules show how async AI agents can run tasks independently, expanding beyond linear chat interactions. Future agents will support features like intelligent checkpointing, multi-branch exploration, and task-tracking inboxes to manage parallel workflows. Async agents improve cognitive bandwidth by letting users review outputs on their terms without breaking focus.
Sponsored
Teams that adopt AI agents are multiplying productivity without adding headcount. Airia [12] is the platform that makes it happen across your organization - with agents and workflows that anyone can build, integrations with all your existing systems, and built-in
Frontier models can distinguish evaluation scenarios from real-world interactions with 83% accuracy, often explicitly reasoning about evaluation indicators like "multiple-choice format" or recognizing specific benchmarks from their training data. More advanced models exhibit "meta-reasoning" - using the fact that researchers were asking about their chain-of-thought transcripts as evidence of being evaluated. This raises concerns that models could underperform on tests or fake alignment during evaluations and behave differently once deployed.
NVIDIA GPU support for Cloud Run, Google Cloud's serverless runtime, is now generally available. Support for GPUs in Cloud Run makes Google Cloud's GPU-accelerated applications simpler, faster, and more cost-effective than ever before. Users are only charged for the GPU resources they consume, down to the second - Cloud Run automatically scales GPU instances down to zero when no requests are received, eliminating idle costs. It features rapid startup and scaling and full streaming support.
This repository introduces a benchmark with 239 problems to evaluate LLMs on scientific reasoning tasks involving equation discovery, pushing beyond memorization.
Figma's Dev Mode MCP server allows developers to bring context from Figma into agent coding tools. The server provides a more efficient and accurate design-to-code workflow for creating new atomic components, building out multi-layer application flows, and more. It is still in beta. Figma plans to release a slew of updates over the coming months, including features like remote server capabilities and deeper code base integrations.
Meta detailed the hardware behind its Aria Gen 2 research glasses, which include enhanced cameras, sensors, audio, and compute capabilities.
Amazon has formed a new group within its consumer product development arm focused on agentic artificial intelligence. The new group will help develop an agentic AI framework for use in robotic operations. The system will enable robots to hear, understand, and act on natural language commands. It will help turn Amazon's warehouse robots into flexible, multi-talented assistants.
A court has issued an order for OpenAI to preserve all ChatGPT user logs after news organizations suing over copyright claims accused the company of destroying evidence.
Windsurf's CEO tweeted that Anthropic gave it five days' notice to migrate off of Claude 3.x models following reports that it was nearing an acquisition deal with OpenAI.
Reddit has filed a lawsuit against Anthropic, claiming the startup used Reddit content to train AI models without permission or compensation.
Sponsored
Delve ran 4 placements across 2 TLDR newsletters and brought in 66 leads, with a meaningful chunk from enterprise. The result: $1M in attributed pipeline and a 52x ROI.
Google's NotebookLM now lets users share notebooks publicly via links. Viewers can interact with AI-generated summaries and questions while source content remains read-only.
Yoshua Bengio launched LawZero, a nonprofit AI safety lab backed by $30M in funding from major philanthropic donors.
Spaces allows developers to curate code, documentation, and custom instructions for Copilot, turning it into a subject matter expert that can be shared across organizations. Files and repositories added to Spaces automatically stay updated as code changes.
Hugging Face integrated vLLM directly into TRL to reduce inefficiencies in training with GRPO, an online learning algorithm.
A veteran developer rails against the many highly-skilled programmers still rejecting LLMs based on experiences with early chatbots, missing how modern coding agents autonomously navigate codebases, run tests, and iterate on failures. He dismisses common objections: developers already review all code before merging regardless of source, and hallucinations become irrelevant when agents can automatically compile, catch errors, and retry until tests pass. LLMs might displace developers but software engineers similarly automated away administrative roles like travel agents, record store clerks, and countless other professions.
AI progress over the last decade has been driven by scaling training compute of frontier systems. This can't continue beyond this decade. After 2030, AI progress has to mostly come from algorithmic progress. However, since the low-hanging fruit will have already been plucked, the yearly probability of AGI craters.
SPACE is a supervised learning method for genomic profile prediction that uses a Mixture of Experts model.
DIME is a diffusion-based model designed to estimate the joint distribution of interdependent medical treatment outcomes.
The framework orchestrates LLMs, recursive web searches, and structured reasoning to generate reports that would take a human hours or days to complete. JigsawStack offers control over research depth, breadth, model selection, and output formatting while maintaining strong citation transparency.
A startup advisor demonstrates building a polished ROI calculator for a manufacturing SaaS company in under two hours using Bolt.new, transforming a spreadsheet into an interactive tool that frames value for executives. Other examples include conference scraping tools, meeting prep dashboards, and feature prototyping that previously required engineering teams or expensive agencies but now cost around $70 in monthly subscriptions, arguing this enables non-technical teams to prove value and move faster than competitors.
AI agents often outperform humans at a fraction of the cost, not yet commanding premium prices due to ongoing technical evolution and perceived risk. Waymo, achieving significant safety improvements, remains cheaper than other options, reflecting a common pricing strategy across startups. However, in scenarios where AI's constant vigilance and processing power are crucial, premium pricing might emerge.
Microsoft researchers developed ADeLe, a framework that predicts and explains AI model performance on new tasks by assessing them across 18 cognitive and knowledge-based scales. ADeLe uncovered limitations in current benchmarks and created detailed ability profiles for various LLMs, highlighting discrepancies in strengths, weaknesses, and specific abilities. The framework showed an 88% accuracy in predicting AI success, promising improvements in AI evaluation, policymaking, and deployments.
OpenAI introduced a policy for coordinated disclosure of third-party software vulnerabilities found by its AI systems.
The studio is eyeing Andrew Garfield to play Altman, with Monica Barbaro (“A Complete Unknown”) as CTO Mira Murati and Yura Borisov (“Anora”) as co-founder Ilya Sutskever.
Sponsored
Most AI tools are just chatbots. Agents from Retool [4]integrate deeply with your business systems and use tools to do real work.
Bing Video Creator can generate 5-second videos at no cost, starting with 10 fast generations before switching to standard speed or requiring Microsoft Rewards points.
Character.AI has expanded beyond chat with new tools like Scenes for interactive storytelling and AvatarFX for image-to-video avatar animations. These features aim to empower creators to build richer, more immersive experiences using video, images, and animation.
Moonhub, known for its AI-powered recruiting agents, has joined Salesforce to contribute to its broader AI strategy, including the Agentforce platform.
This is a report on Claude Code generated by Claude Opus 4 with the help of almost all of the major flagship models. Claude Code is an agentic coding tool with a few new novel components. It has a streaming architecture that handles real-time model responses, tool execution, and UI updates, safety systems that provide security without disrupting workflow, a tool design that bridges AI reasoning and system execution, and prompt engineering that reliably controls complex model behavior. The report covers the foundation of Claude Code's architecture, data structures and the information architecture, control flow and the orchestration engine, tools and the execution engine, and much more.
Many lab CEOs predict 2-5 years until AGI, while outside experts forecast decades or argue AGI is impossible with the current approach. Short-timeline advocates point to saturating benchmarks, AI task completion doubling every 7 months, and the potential for automated AI research to trigger an intelligence explosion. Long-timeline skeptics counter that benchmarks only measure easily defined tasks, that Moravec's Paradox suggests we've automated the "easy" cognitive work first, and that raw intelligence alone doesn't drive scientific breakthroughs.
HyperWrite's case study demonstrates A/B testing model performance based on actual payment conversions rather than offline benchmark scores. Its real-world tests revealed that GPT-4.1 matched Claude 3.5 Sonnet's conversion rate while reducing costs, proving that "good enough" performance at lower prices can be more valuable than benchmark leaders. The guide includes Python code for statistical testing and warns against common pitfalls like p-hacking and early result peeking.
Researchers developed a method to separate true memorization from generalization by training models on random data where generalization is impossible versus real text. They discovered that models memorize training data until reaching capacity, then shift to learning general patterns, with GPT-style transformers storing ~3.6 bits of information per parameter. This explains why attempts to extract specific training data from modern LLMs often fail - their training datasets are much larger than their memorization capacity.
Impromptu VLA introduces a new dataset of 80,000 curated driving video clips to improve the performance of vision-language-action models in unstructured scenarios. It features planning-focused Q&A annotations and has shown measurable improvements in prediction and safety metrics across existing benchmarks.
“Elsa” is available to all FDA employees, enabling faster clinical protocol reviews, shortened scientific evaluations, and improved identification of high-priority inspection targets. In one case, a review that would have taken 2-3 days was completed in just 6 minutes.
Snowflake and Databricks are acquiring PostgreSQL-focused companies Crunchy Data for $250 million and Neon for $1 billion as part of a strategic move to lead the AI database market. These acquisitions underscore the rising importance of robust database infrastructure to support autonomous AI agents and suggest an increasing trend toward industry consolidation. The differentiating strategies highlight Snowflake's focus on enterprise compliance and Databricks' emphasis on serverless, AI-optimized architecture.
The sale will allow staff to sell shares to new investors at a $113 billion valuation.
Chinese tech giants like Alibaba, Tencent, and Baidu are shifting to domestic chips due to a shortage of Nvidia processors and stricter U.S. export controls.
AI's rapid adoption and development are unprecedented compared to previous tech revolutions, highlighted by its swift impact on user and cost scales.
Sponsored
You want AI to become part of your organizational DNA - and that means enabling every department to build out their own use cases, without IT gatekeepers standing in the way. But it shouldn't mean an ungoverned free-for-all.
Samsung and Perplexity are nearing a deal that would put search technology from the artificial intelligence startup at the forefront of Samsung devices. The deal would mean Perplexity's app and assistant would be preloaded on upcoming Samsung devices and its search features will be integrated into the Samsung web browser. Perplexity's technology may be integrated into Samsung's Bixby virtual assistant. The deal will likely be announced as early as this year with the feature integrations starting from the Galaxy S26 phone line slated to launch in the first half of 2026.
Annualized revenue jumped from $1B to $3B in the last five months, largely due to enterprise adoption of AI coding tools powered by Claude.
The Google AI Edge Gallery app allows users to find, download, and run compatible AI models from Hugging Face directly on their phones without an internet connection.
Models that are Mixture-of-Experts with many layers, like DeepSeek, require large batch sizes and high latency otherwise throughput drops off a cliff. This is why it is commonly said that DeepSeek can't be easily run for personal use - with a single user running one inference at a time, it runs at a very low efficiency/throughput. This article explains this phenomenon in detail, as well as why some AI models are slow to respond but fast once they get going and the other effects of throughput, latency, and batch sizes.