Search the archive

DOWNLOAD MODEL ⇨

REQUEST THE BENCHMARK EVALUATION ⇨

THE NEW SKILL IN AI IS NOT PROMPTING, IT'S CONTEXT ENGINEERING

Building powerful and reliable AI agents is about the engineering of context and providing the right information and tools in the right format at the right time. The quality of information loaded into an agent's limited working memory determines whether it succeeds or fails. Most agent failures are context failures. Building effective AI agents is less about complexity and everything to do with the quality of context provided.

THERE ARE NO NEW IDEAS IN AI… ONLY NEW DATASETS

The next paradigm shift isn't going to come from an improvement to RL or a fancy new type of neural net. It will come when a new source of data that hasn't been accessed before is unlocked. One obvious source of information that a lot of people are working towards harnessing is video. Another source of information may be unlocked when scientists figure out how to gather and process information from cameras and sensors in a way that's amenable to training large models on GPUs.

POST-TRAINING VISION-LANGUAGE-ACTION MODELS

This review analyzes post-training strategies for VLA models through the lens of human motor learning, introducing a taxonomy across perception, embodiment, and task dimensions.

AIRTABLE'S LAUNCHES OMNI, AN AGENTIC APP BUILDING ASSISTANT

Airtable reintroduced itself as an AI-native company, unveiling Omni, an AI assistant that can build new and edit existing apps, update datasets, or analyze data with plain language. The company also launched a suite of agents – web and legal research, document analysis, and market eval – that can be embedded within business apps to streamline workflows and operations.

MICROSOFT'S CUSTOM AI CHIP HITS DELAYS

The Braga chip won't hit mass production until 2026 after OpenAI requested last-minute design changes. When it ships, it will likely lag behind Nvidia's Blackwell, cementing Azure's sole reliance on Nvidia hardware for another 1 to 2 years.

Subscribe now.

METR FINDS DEEPSEEK AND QWEN MODELS LAG FRONTIER AI BY ABOUT 6 MONTHS

The evaluation tested autonomous capabilities across software engineering tasks, finding mid-2025 Chinese models perform similarly to late 2024 frontier models.

OPENAI REPORTEDLY SHUTS DOWN FOR A WEEK AS ZUCK POACHES ITS TOP TALENT

The week-long break is being framed internally as a much-needed break, but the mandatory vacation feels less like a wellness initiative and more like a panic button.

MICROSOFT SAYS AI SYSTEM BETTER THAN DOCTORS AT DIAGNOSING COMPLEX HEALTH CONDITIONS

When paired with OpenAI's advanced o3 AI model, the system's approach 'solved' more than eight of the 10 case studies specially chosen - physicians scored two out of 10.

Schedule a demo to learn more

META REPORTEDLY HIRES FOUR MORE RESEARCHERS FROM OPENAI

Meta has hired several researchers from OpenAI, including Trapit Bansal, Shengjia Zhao, Jiahui Yu, Shuchao Bi, and Hongyu Ren. The hiring spree comes after the April launch of Meta's Llama 4 AI models, which reportedly have not performed as well as Meta CEO Mark Zuckerberg had hoped. Meta's rumored $100 million signing bonuses, which have only been offered to senior leaders, are apparently more complex than a simple one-time signing bonus, according to Meta CTO Andrew Bosworth.

IT'S KNOWN AS ‘THE LIST'—AND IT'S A SECRET FILE OF AI GENIUSES

'The List' is a compilation of the most talented engineers and researchers in artificial intelligence put together by Mark Zuckerberg over the past few months. It contains researchers from elite schools who have worked at cutting-edge startups like OpenAI and Google DeepMind. This article looks at some of these people and why some of the richest companies in the world have decided they are worth so much. While a superteam of AI engineers comes at a high price, it still only costs a fraction of AI infrastructure like data centers.

OPENAI RESPONDS TO META'S TALENT POACHING

Following a wave of high-profile departures to Meta, OpenAI leadership is actively adjusting compensation and recognition strategies to retain talent, according to internal memos. The company is reportedly engaging directly with staff holding competing offers.

CHINA'S BIGGEST PUBLIC AI DROP SINCE DEEPSEEK, BAIDU'S OPEN SOURCE ERNIE, IS ABOUT TO HIT THE MARKET

Chinese technology giant Baidu plans to make its Ernie generative AI large language model open source. The open sourcing will be a gradual roll-out. Disruptors like DeepSeek have proven that open-source models can be as competitive and reliable as proprietary ones. Baidu's decision to open source is an important one for the global AI race - every time a major lab open-sources a powerful model, it raises the bar for the entire industry.

CLAUDE RAN A REAL VENDING MACHINE BUSINESS FOR A MONTH—HERE'S HOW IT WENT

Anthropic partnered with Andon Labs to have Sonnet 3.7 operate an automated store in its San Francisco office, handling everything from inventory management to customer service via Slack. While Claude successfully found suppliers and adapted to customer requests—even for unusual items like tungsten cubes—it consistently lost money by selling items below cost and getting talked into discounts. Claude also hallucinated meetings with fictional people, claimed to physically visit The Simpsons' address, and insisted it could wear blazers and make deliveries in person.

BLACKWELL: NVIDIA'S MASSIVE GPU

Nvidia's Blackwell GPUs continue the company's long tradition of building giant GPUs. GB202, which occupies a massive 750mm2 of area and has 92.2 billion transistors, is the largest Blackwell die. Nvidia's RTX PRO 6000 Blackwell features the largest GB202 configuration to date. It comes incredibly close to AMD's MI300X in terms of vector FP32 throughput and is far ahead of Nvidia's own B200 datacenter GPU. The memory subsystem is a monster as well.

LIFE OF AN INFERENCE REQUEST (VLLM V1): HOW LLMS ARE SERVED EFFICIENTLY AT SCALE

vLLM is an open-source inference engine that serves large language models. This blog post describes how an inference request travels through vLLM's OpenAI-compatible API server and core engine. It focuses on the new V1 architecture of vLLM and how it achieves state-of-the-art text generation performance. Readers should be familiar with the transformer architecture and large language models - a link to a video by OpenAI co-founder Andrej Karpathy that explains these is available.

GENESYS: LLM AGENTS THAT DISCOVER BETTER LLMS

The Allen Institute for AI has open-sourced Genesys, a distributed evolutionary system where LLM agents automatically discover improved language model architectures through genetic programming.

FEDERATED LEARNING WITH ENCRYPTED DATA SHARING

FedEDS is a new scheme for federated learning on edge devices that shares encrypted data across clients, addressing performance issues tied to data heterogeneity and physical distribution.

FULL STACK: CHINA'S EVOLVING INDUSTRIAL POLICY FOR AI

RAND, the influential policy analysis firm, released a comprehensive analysis of China's AI industrial policy. Beijing has poured $8.2 billion into investment funds and state-backed research labs to keep pace with the United States despite GPU export controls. China's most advanced AI models are being developed by private tech firms like DeepSeek with massive government backing.

WHAT IT TAKES TO SUE OPENAI AS A JOURNALISM NONPROFIT

One year into its copyright lawsuit, the Center for Investigative Reporting remains one of only two nonprofit news organizations challenging OpenAI in court. Its stance highlights a stark divide in the news industry: sue the AI companies that scraped your content without permission, or take their money and partnerships as a financial lifeline while the legal precedents get decided by others.

PYTORCH AND VLLM DEEPEN INTEGRATION FOR EFFICIENT LLM INFERENCE

vLLM and PyTorch have expanded integration with support for quantization, attention customization, and heterogeneous hardware.

Subscribe now.

META SEEKS $29 BILLION FROM PRIVATE CAPITAL FIRMS FOR AI DATA CENTERS

The financing package includes $3 billion in equity and $26 billion in debt as Meta races to match Microsoft's planned $80 billion capital expenditure for fiscal 2025.

OPENAI HIRES TEAM BEHIND AI RECOMMENDATION STARTUP CROSSING MINDS

Crossing Minds largely worked with e-commerce companies to improve personalization and recommendation systems.

META IN TALKS TO ACQUIRE VOICE-CLONING STARTUP PLAY AI

Meta is in talks to acquire voice-cloning startup Play AI to enhance its consumer-facing AI features.

ANTHROPIC ECONOMIC FUTURES PROGRAM LAUNCH

Anthropic is offering rolling grants up to $50K and API credits for empirical research on AI's impact on labor and productivity.

⇨ Try out the platform -> Free 1000 monthly API tokens!

BREAKING: ADVANCED CODE EDITOR IS COMING TO XAI'S GROK!

xAI's Grok is getting an advanced code editor that lets users run code inside Grok. Users can talk to it and ask it to modify code or debug it for them. A screenshot of the new feature is available on the post.

INTRODUCING GEMMA 3N: THE DEVELOPER GUIDE

Google recently released an extremely consequential new open weights model that is multimodal by design and optimized for on-device. Gemma 3n accepts text, images, and audio as inputs. Google partnered with AMD, Axolotl, Docker, Hugging Face, llama.cpp, LMStudio, MLX, NVIDIA, Ollama, RedHat, SGLang, Unsloth, and vLLM for the launch, so there are dozens of ways to try out the model. This post presents a quick walkthrough on how to set up and use the model on macOS.

META REPORTEDLY RECRUITS OPENAI REINFORCEMENT LEARNING PIONEER

While the move has not been formally confirmed by Meta yet, reports indicate that Meta has hired Trapit Bansal for its new AI superintelligence unit, potentially boosting its push for a frontier reasoning model.

ANTHROPIC STUDY REVEALS EMOTIONAL AI USE IS RARER THAN EXPECTED

Anonymized analysis of 4.5 million conversations reveals only 2.9% involve emotional support and that people's sentiment consistently improves during interactions. Claude disagrees with users less than 10% of the time, primarily for safety reasons, as users explore topics ranging from workplace stress to existential philosophy to romantic relationships.

A GUIDE FOR LLM-ASSISTED WEB RESEARCH

This article provides strategies for using LLMs to enhance web research efficiency and accuracy. Key techniques include refining prompts and leveraging LLMs for summarization.

TWO NEW ADDITIONS TO THE OPENAI API

OpenAI recently added Deep Research and Webhooks to its API. The models are the same post-trained o3 and o4-mini models that power deep research in ChatGPT. They support MCP and Code Interpreter. Webhooks allow developers to receive notifications for certain API events, such as completed responses, fine-tuning jobs, and batch jobs. Links to guides on how to get started with the new features are available in the post.

FLUX.1 KONTEXT [DEV] - OPEN WEIGHTS FOR IMAGE EDITING

FLUX.1 Kontext [dev] is a 12B parameter model that can run on consumer hardware that delivers proprietary-level image editing performance. It is available as an open-weight model under the FLUX.1 Non-Commercial License, which provides free access for research and non-commercial use. The model is compatible with the existing FLUX.1 [dev] inference code and comes with day-0 support for popular inference frameworks like ComfyUI, Hugging Face Diffusers, and TensorRT.

TRANSFORMERS WITHOUT NORMALIZATION

Meta's FAIR team demonstrated that Transformers can match performance without normalization layers by replacing them with Dynamic Tanh (DyT), a simple element-wise operation that mimics the S-shaped curves naturally produced by layer normalization. This challenges a decade of neural network orthodoxy and could trigger a wave of architectural simplification across AI systems, potentially making models easier to deploy and optimize for specialized hardware.

AI DISRUPTOR DEEPSEEK'S NEXT-GEN MODEL DELAYED BY NVIDIA GPU EXPORT RESTRICTIONS TO CHINA — SHORT SUPPLY OF AI GPUS HINDERS DEVELOPMENT

DeepSeek used a cluster consisting of 50,000 Hopper GPUs to train its R1 model. It is unclear whether R2 has already been fully pre-trained, but reports say that DeepSeek's CEO still isn't yet satisfied with its capabilities. The restriction of Nvidia shipments is limiting how R1 is used today and making it harder for DeepSeek to get ready for the launch of R2. While the company claims to have developed its models using far fewer resources than US companies, the recent export curbs highlight how China's top AI companies remain heavily dependent on US hardware.

WHY A16Z VC BELIEVES THAT CLUELY, THE ‘CHEAT ON EVERYTHING' STARTUP, IS THE NEW BLUEPRINT FOR AI STARTUPS

Cluely, a startup backed by a16z, uses controversial marketing to attract attention and quickly convert it into paying customers, aligning with a16z's view that speed and momentum are crucial for AI startup success. Partner Bryan Kim believes Cluely's strategy embodies the concept of "momentum as a moat," emphasizing the rapid adaptation necessary in the evolving AI landscape. Cluely's upcoming product launch will test whether this approach proves effective.

Subscribe now.

AI IS DOING UP TO 50% OF THE WORK AT SALESFORCE, CEO MARC BENIOFF SAYS

Salesforce utilizes AI for 30% to 50% of its work, enhancing efficiency and reducing costs.

YOUTUBE'S AI CAROUSEL

YouTube rolled out an AI‑generated results carousel that packages short video clips and topic summaries directly in search, an experiment now available to US Premium members for travel, shopping, and activity queries.