Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
The Weighted Perplexity Benchmark offers a tokenizer-normalized evaluation method to compare language models effectively. It adjusts perplexity scores, considering differences in tokenization, to provide a fairer comparison across models. This approach enhances the accuracy of assessing NLP performance.
Perplexity is quietly expanding into India to compete in the next phase of AI adoption. It is rapidly adding millions of users in the world's second-largest internet and smartphone market. Perplexity has partnered with Airtel, India's second-largest telecom operator, to offer a free 12-month Perplexity Pro subscription to all 360 million subscribers. While the company has seen major growth in the country, monetizing its large user base remains a challenge.
Meta has acquired Mark Lee and Tom Gunter from the Apple Foundation Models team. Lee was Ruoming Pang's first hire at Apple, and Gunter was a distinguished engineer at Apple with an eight-year tenure. Meta recently poached Pang with a $200 million sign-on bonus. The company has been on an aggressive hiring spree, pulling top AI talent from across the industry.
Sponsored
Build new functionality, create tests, and fix bugs with Claude Code,
The deep research mode breaks down complex questions, gathers sources from the web, and builds structured reports, while the new Voxtral-powered voice mode enables audio-in.
After classifying ChatGPT Agent as high bio/chemical risk, OpenAI launched a program to pay $25,000 to the first researcher to submit a universal jailbreak that answers all 10 challenge questions.
FineWeb has been updated with English data from CommonCrawl snapshots from January to June 2025.
Windsurf's Wave 11 includes startups across AI-native productivity, dev tools, robotics, and consumer agents.
Sponsored
Looking to get on the AI agents bandwagon? You should go ahead and bookmark Hypermode [4] — and here's why:
Anthropic is in the early stages of planning another investment round that could value the company at more than $100 billion. It is not formally fundraising, but the company has received pre-emptive funding offers from VCs. Annualized revenue for Anthropic's Claude chatbot has climbed from $3 billion to $4 billion in the past month. The company's current investors include Amazon, Alphabet, Menlo Ventures, and Salesforce Ventures.
Google is bringing Gemini 2.5 Pro and Deep Search to Search, offering advanced capabilities like longer queries and follow-ups for AI Pro and Ultra subscribers.
Google has announced updates to the Agent Development Kit (ADK) designed to eliminate friction and supercharge the 'vibe coding' experience when paired with the Gemini CLI. At the heart of the upgrade is a revamped llms-full.txt file that is 50% shorter and easier for large language models to understand. Gemini has a full understanding of ADK, so it won't eat up context windows or fall prey to 'context rot'. The update gives the Gemini CLI a deeper, native understanding of the framework, enabling it to translate high-level plans directly into accurate, idiomatic multi-agent code.
Sponsored
Bigger isn't always better. JetBrains and Hugging Face challenge the scale-first mindset in AI and spotlight Mellum: a task-specific code model [12], now open source on Hugging Face.
Passage of Time is a Model Context Protocol (MCP) server that gives language models temporal awareness and time calculation abilities. With the proper temporal tools, models can uncover surprising insights about conversation patterns, work rhythms, and the human experience of time. The implementation shows the promise of MCP - the protocol doesn't just open up the possibility of smarter tools, but also for AI systems that can learn to perceive and understand the human experience in ways that create genuine mutual understanding.
Stanford's Marin project aims to share not just models, but to make the entire journey accessible, including the code, data set, data methodologies, experiments, hyperparameters, and training logs. The transparency is aimed at fostering a more transparent and accessible future for foundation model research. The project's first models, Marin-8B-Base and Marin-8B-Instruct, have been released under the permissive Apache 2.0 license. This article looks at the engineering challenges the project had to overcome to succeed in creating truly open, scalable, and reproducible foundation models.
Amazon Bedrock AgentCore is a comprehensive set of enterprise-grade services that help developers quickly and securely deploy and operate AI agents at scale using any framework or model. It comprises several services that can be used individually and are optimized to work together so that developers don't need to spend time piecing together components. AgentCore eliminates tedious infrastructure work and operational complexity. Developers can now discover, buy, and run pre-built agents and agent tools from AWS Marketplace with AgentCore Runtime.
In a rare moment of collaboration, researchers from all major AI labs argue that complex tasks requiring serial reasoning must pass through observable language, creating potential chokepoints for detecting malicious intent before harm occurs. However, this monitorability depends on current training paradigms and could degrade through architectural changes, process supervision, or models learning to obfuscate their reasoning when they know they're being monitored.
Anthropic is launching a comprehensive analytics dashboard for its Claude Code AI programming assistant. It will provide engineering managers with detailed metrics on how their teams use Claude Code. Companies are increasingly demanding concrete data to justify their AI spending. The dashboard gives managers visibility into which teams and individuals are benefiting most from these expensive premium tools.
We sometimes do things because that's what the computer told us to do. Large language models are now giving developers their opinion on how APIs should look, and developers have no choice but to follow their advice. This can be a useful testing device - AI can give developers the 'newbie's POV' on how tools should've been made.
Sponsored
Work on challenging problems that matter alongside people who inspire you. At Amazon, we give teams the resources and massive scale to drive meaningful advancements in AI. Meet our AI teams and explore open
Jason Wei and Hyung Won Chung, formerly of OpenAI, are reportedly joining Meta's superintelligence lab as part of its aggressive talent acquisition push.
Scale AI is laying off 200 employees and 500 contractors, citing overexpansion in its data-labeling unit.
Boris Cherny and Cat Wu are reportedly back at Anthropic after departing for Anysphere, the developer of Cursor.
The Google infrastructure will run in the US, Japan, the Netherlands, Norway, and the UK.
Thinking Machines Lab's first product will include a significant open source component and be useful for researchers and startups developing custom models.
Windsurf now has Claude Sonnet 4 again with first party support from Anthropic.
Sponsored
At Amazon, we have a deep conviction that AI will reinvent every customer experience and make possible things we've only imagined. Join us as we develop AI agents and solutions [4] across every corner of Amazon — from AWS and Alexa+ to retail and operations — with over 1,000 generative AI applications built or in progress.
With no revenue or products, the former OpenAI CTO's startup landed one of the largest seed rounds ever from Andreessen Horowitz, Nvidia, and others. Two-thirds of the team are former OpenAI employees.
NVIDIA CEO Jensen Huang confirmed plans to restart H20 GPU sales to China following U.S. government licensing assurances and announced a new fully compliant RTX PRO GPU tailored for industrial AI use in China.
Mistral's has released Voxtral, its first open-source audio model suite. It features a 24B parameter model for large-scale use and a 3B variant for edge deployment.
A former OpenAI employee shares personal reflections on the company's culture and mission, describing it as a uniquely impactful yet complex place to work. The post provides insight into the internal atmosphere during a pivotal time.
xAI had the goal of releasing something that could be called 'the world's smartest artificial intelligence' - and it successfully found benchmarks to enable it to take that claim. However, these benchmarks are misleading. While Grok 4 has a lot of raw intelligence, for most practical purposes, it seems to be inferior to OpenAI's o3. This post takes a more nuanced look at Grok 4's abilities.
Sponsored
Browse pre-built AI agents on Sōkosumi for research, design, analytics, and automation tasks. Create your own agentic team with the click of a button - researchers, digital analysts, SEO specialists, and more. Get 100 free credits for your first agent hire [10]. Building agents? Monetize them on the Masumi blockchain with smart-contract payments [10].
LLM performance degrades significantly as input length increases, even on simple tasks like text retrieval and replication. Multiple controlled experiments revealed that even frontier models don't process context uniformly, and performance becomes increasingly unreliable with longer inputs.
Asymmetry of verification is the idea that some tasks are much easier to verify than to solve. Examples of this are everywhere, for example, Sudokus take a lot of time to solve, but it is trivial to check if any given solution is correct. One of the most important realizations about asymmetry of verification is that it is possible to improve the asymmetry by having privileged information about the task - for example, it is trivial to check test answers with an answer key on hand. AI will likely be much better at verifiable tasks because it's much easier to solve verifiable tasks.
A coding AI agent that supports any LLM backend, including local models, and has both desktop and CLI interfaces. Like typical coding agents, it handles end-to-end development workflows from planning to testing.
Improving AI agents for offensive cybersecurity tasks is cheap and easy: with just $36 of compute time, Princeton researchers increased attack success rates by over 40% through simple techniques like prompt refinement and self-training. Static safety assessments miss this "bubble of risk" where adversaries can cheaply adapt open-source models beyond their original safety profiles, especially in cybersecurity, where clear success signals enable rapid iteration.
The Incentive Flywheel, discovered by Benjamin Franklin, has been at the heart of balancing progress and security for new technology waves since it was created when fires threatened Philadelphia's growth. It won't appear fast enough on its own for AI - it needs to be jump-started. This essay outlines 25 actions that entrepreneurs and policymakers must take by 2030 across agents, foundation models, and data centers. If the West slows down progress in AI, China could dominate the 21st century, but if it accelerates recklessly, accidents will halt progress - like with nuclear power.
Sponsored
You'll probably like our flagship newsletter. It's all about tech, science, and programming.
Anthropic's Financial Analysis Solution can help financial professionals make investment decisions, analyze markets, and conduct research.
Google outlined new AI-powered tools and partnerships designed to enhance cybersecurity, including agentic systems and platform updates.
Amazon S3 Vectors is the first cloud object storage with native support for storing and querying vectors.
Sponsored
Baseten recently outlined their multi-cloud capacity management (MCM) infrastructure for generative AI workloads. Baseten's MCM unifies 10+ clouds into a global GPU pool enabling infinite scaling and 99.99% uptime for generative AI workloads.