Search the archive

ACCELERATING THE NEXT PHASE OF PHYSICAL AI

Generalist AI secured $400 million to advance physical AGI, supported by investors like Radical Ventures and NVIDIA.

EVA-BENCH DATA 2.0: 3 DOMAINS, 121 TOOLS, 213 SCENARIOS

EVA-Bench Data 2.0 expands its evaluation to three domains: Airline CSM, Enterprise ITSM, and Healthcare HRSD.

That's it. Get the full recipe and get started

DEEPSEEK SLATED TO DRAW $7 BILLION IN MAIDEN FUNDRAISING

DeepSeek is set to raise around 50 billion yuan in its first funding round. The startup's founder has committed 20 billion yuan of his own money, and the rest of the funding will come from fewer than 10 other investors. The funding round is expected to be completed within the next couple of weeks. DeepSeek has yet to make any statements about whether it has plans for an initial public offering sometime in the future.

META KEEPS DELAYING THE RELEASE OF ITS NEW AI MODEL TO DEVELOPERS

Meta doesn't have a planned date to release its newest AI models to developers. The company is testing its API with partners and had plans to release it this month. The Muse Spark model is reportedly competitive with OpenAI and Anthropic's offerings, but it has yet to be evaluated by outside firms. The delay raises questions about how quickly Meta can monetize its massive investments in building frontier AI models.

MEET DREAMBEANS, AN APP THAT CONNECTS YOU WITH WHAT MATTERS

Google Labs introduces Dreambeans, an app using AI to curate personalized stories based on Google apps data like Gmail and Calendar. It aims to inspire by cutting through digital clutter with content tailored to user interests, such as recommending dog-friendly restaurants based on calendar events.

OPENAI MAKES ITS NEXT HARDWARE MOVE WITH OPAL ELECTRONICS

OpenAI leads a funding round for Opal Electronics, focusing on a new product line extending beyond webcams into AI-native devices for creative work. This aligns with OpenAI's push into hardware, despite delays in its own ambient computing project.

A FUNCTIONAL TAXONOMY OF WORLD MODELS

While language models have given machines an extraordinary command of concepts, vocabulary, and reasoning, the physical world runs on a different substrate of space and time. 'World model' is one of the most important and overloaded terms in AI. It encompasses computer vision, robotics, reinforcement learning, generative AI, and other fields. This article looks at what world models are and what they are composed of.

RUNNING AN AI-NATIVE ENGINEERING ORG

The Claude Code team eliminated outdated processes by adopting just-in-time planning and AI-assisted coding, which required restructuring roadmaps and roles. Code reviews, once comprehensive, now focus on areas demanding human expertise, as automated tools handle style and bug fixes. They emphasize dogfooding their product, maintaining a flat team structure, and questioning processes to improve efficiency and integration with Claude.

agent strategy, and more. WATCH EVERY SESSION ON DEMAND →

I BUILT A VULNERABLE APP AND SPENT $1,500 SEEING IF LLMS COULD HACK IT

This developer created a vulnerable book review app to see if LLMs could find a flag in users' private reviews by reproducing a common class of exploits. GPT-5.5 performed the best, solving the task in seven out of 10 runs. DeepSeek-V4-Pro was the runner-up with only three successful runs. Claude Sonnet 4.6 was the most expensive model to run, and it only solved the task on two runs, but five of the runs stopped because of the max budget. Many models could not complete the task due to security guardrails.

IDEOGRAM 4

Ideogram 4 is an open-weight text-to-image model. It was trained from scratch and not a fine-tune of any existing model. The model introduces a new structured JSON prompting interface. It features best-in-class multilingual text rendering, deep language understanding, explicit bounding-box layout and color-palette controls, and native 2k resolution images.

SLEEP FOR CONTINUAL LEARNING

Google researchers propose a new “Sleep” paradigm that helps models consolidate short-term in-context knowledge into longer-term parameters through distillation and replay. The approach also uses a “Dreaming” stage with reinforcement learning to generate synthetic curricula for self-improvement.

ANTHROPIC BULKS UP ITS ENTERPRISE PARTNER PROGRAM AMID IPO PLANS

Anthropic's Claude Partner Network is a program for third-party sellers of its AI products that helps them move more product. Firms participating in the program must meet a slate of requirements, but joining it gives companies a great deal of credibility when selling Claude to businesses. The move helps Anthropic demonstrate to the market that it is thinking about scale during a time when investors are looking for signs of business maturity. Anthropic recently filed confidentially for an IPO, putting it on a path to go public this fall.

INTELLIGENCE PER DOLLAR

Microsoft introduces "average token usage" on model release cards, emphasizing intelligence per dollar. Models are now benchmarked on performance and the cost of achieving that intelligence. This new metric forces companies to compete on efficiency, aligning pricing with tangible outcomes like completed support cases.

MORGAN STANLEY WILL SOON OPEN ITS TRILLION-DOLLAR WEALTH MANAGEMENT FUNNEL TO AI AGENTS

Morgan Stanley will allow AI agents from thousands of corporations to access its wealth management platforms like ShareWorks and Equity Edge.

BE THERE FOR EVERY CUSTOMER WITH META BUSINESS AGENT

Meta Business Agent, a new AI tool, allows businesses to efficiently manage customer interactions on WhatsApp, Messenger, and Instagram.

INSIDE META'S ATTEMPTS TO PLAY CATCH-UP WITH AI

Alexandr Wang had a rough start at Meta, but he seems to have found his groove.

guide →

CODEX NEW CAPABILITIES

OpenAI released new Codex capabilities and six role-specific plug-ins for data analytics, creative production, sales, product design, equity investing, and investment banking.

BUILDING A HILL-CLIMBING MACHINE: LAUNCHING SEVEN NEW MAI MODELS

Microsoft released seven new MAI models, enabling developers to tune model weights themselves and integrate these into everyday products. The models leverage Frontier Tuning, an approach where AI adapts to specific workflows through reinforcement learning environments. Microsoft also announced a collaboration with Mayo Clinic to develop an advanced AI healthcare model, combining clinical expertise with AI capabilities, initially deploying within Mayo before wider distribution through Azure Foundry.

MINIMAX PROMISES M3 WEIGHTS AFTER 1M-CONTEXT MODEL LAUNCH

MiniMax will release the model weights and a technical report for its M3 model within the next 10 days. The new model is currently available through MiniMax Code, token plans, and an API. It has a 1M-token context window and a guaranteed 512,000-token minimum for API use. The model is the first open-weight model to combine frontier coding, native multimodality, and a 1M-token context window. MiniMax lists standard API pricing up to 512,000 input tokens at $0.60 per million input and $2.40 per million output.

OPEN AND CLOSED MODELS ARE ON DIFFERENT EXPONENTIALS

A lot of businesses want to switch to open models, but these models aren't yet good enough in out-of-distribution tasks. However, they will eventually catch up. The open model ecosystem will be far more diverse and numerous than the closed lab oligopoly. The total market value of the open model economy will dramatically exceed the cumulative value of OpenAI and Anthropic.

THE NEXT FRONTIER OF VISUAL AI IS CODE

Visual AI is shifting from generating final pixel outputs to creating source code for editable artifacts, transforming workflows in design and 3D modeling by enabling continuous iteration and feedback. Code-native generation produces structured representations such as HTML/CSS or Blender scripts, facilitating precise edits and enhancements post-generation. This approach holds promise for industries requiring consistent 3D structures and interactive assets, with visual AI moving towards creating adaptive, editable digital artifacts rather than just static images.

MEMORY IS PURPOSE

Sentra CEO Ashwin Gopinath argues memory is not a sidecar to intelligence but the layer that decides what reality the reasoning operates on, with knowledge being what was present and memory being the subset of the past that should survive because it changes future behavior. The boulder problem illustrates that the same artifact becomes different memory for sales, product, legal, engineering, and CEO views, so freezing ontology at ingestion traps the system inside a frame that is prematurely right.

THE DATA CENTER MOVES TO YOUR MACHINE

Perplexity unveiled a hybrid local-cloud inference system at Computex 2026 that intelligently routes queries between on-device models for lightweight tasks and cloud-based models for complex reasoning, building on the company's earlier Personal Computer agent.

PREVENTING AI INFERENCE THEFT AT SCALE

Vercel outlined how attackers resell stolen AI inference by exploiting exposed endpoints and highlighted why traditional rate limits are often insufficient. The post describes an approach that verifies every AI request using BotID analysis to reduce abuse.

WALL ATTENTION

An attention mechanism that improves long-context reasoning by organizing information around persistent “wall” memory tokens.

GITHUB'S PLAN FOR AGENTS

GitHub has been the home for software for almost two decades. Coding agents have seen massive quantities of code being shipped on the platform, which has grown 1,400% this year. This has put pressure on GitHub's infrastructure, which was designed around human developers moving at human speed. This article features an interview with GitHub COO Kyle Daigle where he discusses how AI has changed things in the company, how he uses AI to work at GitHub, why the AI era is breaking GitHub in new ways, and more.

ANTHROPIC FACES AI SPENDING BACKLASH BEFORE IPO

Anthropic filed for an IPO amid growing corporate scrutiny over high AI costs, which threatens its revenue as companies reevaluate their AI investments. Businesses, Anthropic's key clients, express concerns over spending, with a survey revealing 40% experiencing cost savings below 10%. Despite Anthropic's strong performance, including surpassing OpenAI in business clients, a shift towards cheaper models or open-source alternatives poses a risk.

ANTHROPIC EXPANDS MYTHOS TO 150 ADDITIONAL ORGANIZATIONS IN MORE THAN 15 COUNTRIES

Anthropic has expanded Project Glasswing to an additional 150 partners in more than 15 countries. The new partners will need to meet security requirements before gaining access to the model. Since launch, Project Glasswing partners have discovered more than 10,000 high or critical-level security flaws. Major Project Glasswing partners include Apple, Nvidia, Microsoft, CrowdStrike, and Palo Alto Networks.

TINYFISH BIGSET TURNS TEXT PROMPTS INTO LIVE DATASETS

TinyFish released Bigset, an open-source system that converts text prompts into structured datasets from the live web.

STATE OF MEMORY IN AGENT HARNESS

Mem0 surveyed memory implementations across Claude Code, Codex, Copilot, OpenClaw, Hermes, Bedrock AgentCore, Windsurf, and Devin, and found the same boundary failures everywhere: bounded local storage, mostly keyword retrieval, harness scoping, weak staleness handling, and 57-71% cross-user contamination rates.

ANTHROPIC FILED A CONFIDENTIAL DRAFT IPO REGISTRATION

Anthropic confidentially submitted a draft S-1 registration statement to the US Securities and Exchange Commission for a proposed initial public offering. The filing does not set pricing or share counts and remains subject to regulatory review and market conditions.

OPENAI AND CODEX REACH AWS

OpenAI announced the general availability of its frontier models and Codex on AWS. The integration allows enterprises to access OpenAI capabilities through existing AWS security, governance, procurement, and billing workflows.

NVIDIA JUST ANNOUNCED THE RELEASE OF NEMOTRON 3 ULTRA

Nemotron 3 Ultra features 550B parameters (55B active). It is the most intelligent open weights model from the US. The model will be made available in NVFP4 quantization for higher inference performance. It scores 48 on the Artificial Analysis Intelligence Index, well ahead of the next strongest model, Gemma 4 31B, which scored 39. Nemotron 3 Ultra serves over 300 tokens per second on a pre-release Deep Infra endpoint.

QWEN3.7-PLUS: MULTIMODAL AGENT INTELLIGENCE

Qwen3.7-Plus is a multimodal agent model that unifies vision and language into a single, versatile agent foundation. It can operate as a multimodal interactive hybrid agent, seamlessly blending GUI and CLI interactions within a single agent loop. The model performs consistently across scaffolds and frameworks. Qwen3.7-Plus is now available via Alibaba Cloud Model Studio.

WHY VIDEO AGENT MODELS ARE NEXT — ETHAN HE, XAI GROK IMAGINE

Ethan He was the lead on Nvidia's Cosmos World Model. He then joined xAI and built Grok Image in three months. He has been in the center of some of the most important work in video generation, multimodal models, and real-time world models. This post contains an interview with He where he unpacks what it actually takes to build frontier image and video systems.

OPUS 4.8 PART 2: MODEL WELFARE

Anthropic cares about model welfare, and it attempts to address it through studies. The study of model welfare is difficult, and Anthropic mainly relies on self-reporting by the model. It can be difficult to evaluate whether model responses are really representative of the truth. This article takes a look at Anthropic's findings regarding model welfare with Opus 4.8.

coming from. Get your copy

RUNNING OPENAI MODELS ON AMAZON BEDROCK

OpenAI cookbook walks through building production workflows with OpenAI models hosted on Amazon Bedrock using the Responses API. It covers structured outputs, tool calling, file inputs, state management, prompt caching, and operational best practices.

RETHINKING SEARCH AS CODE GENERATION

Perplexity introduces Search as Code (SaC) to modernize search architectures by allowing models direct control over the search process via an SDK. This approach lets AI models configure search pipelines tailored to specific tasks, improving performance and efficiency over traditional monolithic systems. SaC outperformed competitors in benchmarks, especially in complex tasks like WANDR, demonstrating its robust, cost-effective agentic search capabilities.

NVIDIA LAUNCHES COSMOS 3, THE OPEN FRONTIER FOUNDATION MODEL FOR PHYSICAL AI

Nvidia Cosmos 3 is a new leaderboard-topping open physical AI foundation model. It is a fully open omnimodel with native vision reasoning and multimodal generation across text, image, video, ambient sound, and action. The model is built on a mixture-of-transformer architecture, which pairs a reasoning transformer with an expert generation transformer. It gives developers a powerful pretrained foundation for building physical AI systems with less data and lower training costs.

JETBRAINS'S MELLUM 2

JetBrains introduced Mellum 2, a 12B-parameter MoE language model optimized for coding, reasoning, tool use, and agentic workflows.

US MOVES TO CLOSE THE LOOPHOLE LETTING NVIDIA'S TOP CHIPS REACH CHINESE FIRMS ABROAD

The US Commerce Department has issued guidance extending export license requirements to advance chips sold to any entity headquartered in China, regardless of where that entity is physically located. This closes a loophole that allowed Chinese companies to buy Nvidia chips using a subsidiary in another country. The action is aimed at future sales and not at clawing back hardware already shipped. It doesn't affect the servicing of advanced computing equipment such as servers.