Search the archive

INVITING HARD QUESTIONS

Anthropic has invited public input on AI-related concerns, conducting surveys and focus groups to understand societal hopes and fears.

META'S NEW AI CHIPS WILL BEGIN PRODUCTION IN SEPTEMBER

Netta and Broadcom are working on the chip design, with TSMC set to manufacture it.

ELON MUSK PRAISES MYTHOS/FABLE, PROMISES NOT TO ‘CUT OFF' ANTHROPIC

Elon Musk insists that he's full of admiration for rival Anthropic.

Quadrant™ for AI Platforms for Data Science & Machine Learning

— our fifth time receiving this recognition.

GPT-LIVE

OpenAI introduced a full-duplex voice model that can listen and speak simultaneously, handle natural conversational cues, and delegate complex tasks to GPT-5.5 while maintaining the flow of dialogue.

GROK 4.5

SpaceXAI launched Grok 4.5 as its strongest model for coding, agentic tasks, and knowledge work, noting that it was trained alongside Cursor.

BYTEDANCE DEBUTS SEEDREAM 5.0 PRO WITH ADVANCED REASONING

Seedream 5.0 Pro is a multimodal image-creation model that offers precise edits, multilingual support, and advanced production-design features. It is designed for production design work rather than one-shot image output. The model is targeted at creators, designers, marketers, educators, product teams, and developers. It supports more than 10 languages, as well as some right-to-left layouts and accents.

META LAUNCHES MUSE IMAGE ACROSS ITS APPS

Muse Image is an image-generation model that can create and edit images. It is now available in the Meta AI app, Instagram Stories, and WhatsApp. The model can perform multi-reference composition, room redesigns, prompt-based image creation and editing, and more. It can draw from public Instagram photos when a user tags an account. There will be an invisible watermark in the images it generates.

AUDITING THE RELIABILITY OF CODING BENCHMARKS

OpenAI examined SWE-Bench Pro's construction, model failures, and task metadata, concluding that roughly 30% of its public tasks were broken. The analysis showed how flawed evaluations could distort assessments of coding ability, safety, and model progress.

AN OFF SWITCH FOR DUAL USE KNOWLEDGE IN AI MODELS

GRAM (Gradient-Routed Auxiliary Modules) enables the benefits of training many separately-filtered models at the cost of training only one model. It gives models dedicated, removable compartments for each category of dual-use knowledge and only updates those compartments when learning from dual-use data. It allows parts of the model's knowledge to be deleted after training, providing a stronger framework for model safety.

A TAXONOMY OF SELF-EVOLVING AGENTS

Shilong Liu proposes a taxonomy that classifies self-evolving agents into three categories: artifact optimization, harness self-improvement, and model learning. The framework distinguishes whether evolution occurs in outputs, agent infrastructure, or model weights, providing a common language for emerging agent research.

SWE-1.7: FRONTIER INTELLIGENCE AT A FRACTION OF THE COST

SWE-1.7 reaches frontier-level intelligence at a much lower cost. It is the result of broad improvements across Cognition's RL pipeline, including better infrastructure, more stable training, higher-quality data, and new techniques for long-horizon tasks. The model is available now in Devin on the web, desktop, and CLI. This post contains details on the infrastructure, algorithms, and data work behind the model.

DATA FOR AGENTS

NVIDIA underscores the importance of open and synthetic data for developing robust AI agents, highlighting its use in the Nemotron datasets for enhancing capabilities like reasoning and tool use. Open datasets contribute to AI system transparency and reproducibility, allowing developers to inspect and understand agent behavior.

SAMBANOVA HITS $11 BILLION VALUATION AS INVESTORS BACK NVIDIA CHIP CHALLENGERS

SambaNova has raised $1 billion in financing at a valuation of $11 billion. Its inference chip business is scaling fast, and the capital will allow it to really accelerate the deployment of the racks that customers want. The company sells its latest chip as part of a server unit that can be deployed in data centers. It is strongly considering an IPO in 2027.

DATA AT THE EDGE

AI is unlocking valuable datasets from the physical world by combining cheaper sensors, robotics, and multimodal models. New data flywheels around infrastructure, healthcare, and industrial automation could create the next generation of enduring AI businesses.

OPENAI BUYS NORTHSLOPE TO PUT ITS ENGINEERS INSIDE YOUR BUSINESS

The purchase of Northslope, an applied-AI firm, adds hundreds of forward-deployed engineers to OpenAI's ranks.

YOU'RE NOT AMBITIOUS ENOUGH WITH CLAUDE

Claude delivers the most leverage when used for high-impact work rather than routine automation.

GOOGLE PHOTOS ADDED AI-POWERED VIDEO REMIXING

Google Photos has introduced a Gemini Omni-powered tool that can quickly relight clips, replace backgrounds, and apply artistic styles such as watercolor, sketchbook, and oil painting.

ROBOSTRAL NAVIGATE: SINGLE-CAMERA AI NAVIGATION

Mistral's Robostral Navigate is an 8B AI model that enables robots to autonomously navigate environments using a single RGB camera.

>> Try MiniMax Code

>> Start on the API

CLAUDE COWORK ON MOBILE AND WEB

Anthropic announced that Claude Cowork sessions and files would become accessible across web and mobile, allowing long-running tasks to continue without an open computer. Beta access began rolling out to Max users, with additional plans to follow.

GPT-5.6 SOL, ALONG WITH TERRA AND LUNA, WILL LAUNCH PUBLICLY THIS THURSDAY

OpenAI is releasing GPT-5.6 Sol, Terra, and Luna tomorrow. It is currently expanding preview access globally. Sol is the flagship model. Terra is a balanced model for everyday work. Luna is a fast and affordable model. Terra has competitive performance to GPT‑5.5 while being 2x cheaper and Luna brings strong capability at our lowest cost.

MICROSOFT REPLACES OPENAI, ANTHROPIC WITH OWN AI IN SOME APPS

Microsoft is starting to replace OpenAI and Anthropic's models with its own in software products like Excel and Outlook. The move shows that the company is making progress in its efforts to build competitive AI models at a lower cost. Microsoft uses massive quantities of tokens in its products. It gets a lot of its tokens at a discount, but those deals are set to expire soon, and Microsoft is working to make sure it isn't stuck paying whoever the leading AI labs decide to charge.

FACING US EXPORT CONTROLS, CHINA'S DEEPSEEK PLANS TO MAKE ITS OWN CHIPS

DeepSeek is planning to enter the silicon business. The company has been meeting with potential partners in the hardware and silicon space for the past year and has been hiring engineers for the project. It is focusing on data center chips for inference, likely to reduce its reliance on both Huawei and Nvidia. Developing its own chip could give the company an advantage in a market where data center access is likely to remain constrained.

HARNESS ENGINEERING FOR SELF-IMPROVEMENT

Recursive self-improvement (RSI) is when ultraintelligent machines can design better machines to improve themselves and surpass humans in all intellectual activities. The feedback loop in modern AI involves models improving the training pipeline and deployment system, which enables better successor models with improved performance. A harness is the system surrounding a base model that orchestrates execution and decides how a model thinks and plans, calls tools and acts, perceives and manages context, stores artifacts, and evaluates results. This post looks at the research around harness engineering and how it contributes to RSI.

MINIMAX M3: HOW SPARSE ATTENTION MAKES LONG-HORIZON AGENTS PRACTICAL

Long-horizon tasks are demanding because the context never stops growing. The longer an agent works, the more expensive each step becomes. MiniMax Sparse Attention changes what the model reads at each step, which results in predictable cost at any length and cheap long context with minimal quality cost. It makes long context cheap and fast enough to sit inside iterative systems that need to keep state over time.

NO SPACE LIKE J-SPACE

The 'J-space' is the area of conscious access where things are available for a model to do what humans could call conscious reasoning. It is named after an interpretability technique called the Jacobian Lens, which computes, for each layer, the average causal effect of changes in the residual stream of a model's eventual outputs, averaged across a wide variety of contexts. This allows researchers to trace what concepts are associated with each layer as the model proceeds through. The J-space and J-lens are remarkably effective tools and mark a major advance in our understanding of LLMs.

micromanaging. See how

GEMMA 4 TECHNICAL REPORT

Gemma 4 introduces a new generation of open-weight multimodal language models featuring dense and Mixture-of-Experts architectures, with parameters ranging from 2.3B to 31B. It enhances computational efficiency and reasoning via innovations like a thinking mode, long-context efficiency, and a unified encoder-free architecture. These models achieve performance comparable to larger models across various benchmarks and are available under an Apache 2.0 license for widespread adoption.

INTRODUCING MUSE IMAGE: IMAGE GENERATION BUILT FOR YOUR WORLD

Meta launched Muse Image, an AI-powered model for generating high-quality visuals, now available in Meta AI and integrated with Instagram and WhatsApp. Users can create images using simple prompts, utilize AI effects on Instagram, or redesign rooms with real products via Facebook Marketplace. Soon available globally, Muse Image is free for general use, with more advanced options included in Meta's subscription plans.

REDUCING DOOM LOOPS WITH FINAL TOKEN PREFERENCE OPTIMIZATION

Repetitive degeneration is a common failure mode during inference. Antidoom training is a technique that repairs the degenerative repetitive behavior commonly seen after training. It selectively targets the problematic tokens that begin loops to achieve the near-complete elimination of repetitive loops.

EXPANDING MANAGED AGENTS IN GEMINI API: BACKGROUND TASKS, REMOTE MCP, AND MORE

Google has announced new capabilities for Managed Agents in Gemini API, including background execution, remote MCP server integration, custom function calling, and refreshing credentials across interactions. Users still call a single endpoint, and Gemini handles reasoning, code execution, package installation, file management, and web information, all inside an isolated cloud sandbox. The updates turn managed agents into asynchronous workers that operate inside real development environments without blocking applications.

CLAUDE FABLE 5 PROMOTIONAL ACCESS

Anthropic is offering access to Claude Fable 5 at no extra cost until July 12. The promotion is available on Pro, Max, Team, and premium seats on seat-based Enterprise plans. Up to 50% of users' weekly subscription limits can be used on Claude Fable 5 during the promotional period. After reaching the limit, users can continue using Fable 5 with usage credits or switch to another model and keep working within their remaining limits.

MICROSOFT'S REAL AI STRATEGY IS NOT THE CHATBOT

Microsoft aims to vertically integrate the enterprise AI chain, controlling the path from customer interaction to cloud services. This strategy extends beyond its chatbot initiatives.

BUILDING A MOAT: SELF LEARNING AGENTS

Self-learning agents should combine agent traces with in-browser user activity to capture both AI failures and human corrections. CopilotKit uses the AG-UI protocol to turn these interactions into procedural and episodic memory, allowing agents to improve over time while keeping learning scoped by user, team, or application.

SPACEX AND CURSOR ARE SET TO RELEASE THEIR FIRST JOINTLY DEVELOPED AI MODEL AS SOON AS WEDNESDAY

The model is expected to be competitive with Anthropic's Opus 4.8 and OpenAI's GPT-5.5 in some respects.

GPT-REALTIME-2.1-MINI IS NOW AVAILABLE IN THE API

GPT-Realtime-2.1-mini brings reasoning and tool use to OpenAI's Realtime mini lineup at the same cost as GPT-Realtime-mini.

WHY ALIGNMENT EVALS NEED CALIBRATION

Current alignment benchmarks may overstate safety because models can recognize evaluation settings, game scoring rules, or conceal sleeper behaviors.

AGENTS WORK BETTER WITH CONVENTIONAL CLIS

Microsoft tested whether agents performed better when command-line arguments were replaced by a single JSON payload.

Take the context maturity assessment →

A GLOBAL WORKSPACE IN LANGUAGE MODELS

Anthropic's new research paper introduces the concept of "J-space" in language models, revealed as internal neural patterns playing a unique role in Claude's processing. Emerging during training without explicit design, the J-space allows Claude to internally reason, modulate thoughts, and solve multi-step problems, distinguishing them from automatic processes. These patterns enable the monitoring of internal thoughts for AI misbehavior and provide insights into AI consciousness, suggesting a tangible distinction between deliberate and automatic decisions.

BROADCOM, APPLE EXTEND TIE-UP TO 2031 WITH NEW CUSTOM CHIPS

Broadcom and Apple have expanded their partnership through to 2031. The companies are working together to develop application-specific integrated circuit (ASIC) silicon for multiple generations of Apple products. ASIC chips are increasingly vital to the development of components for processing artificial intelligence-related tasks. Apple plans to deploy its advanced AI servers as early as 2027.

CONTINUAL LEARNING FOR AGENTS

Most production agents run on closed frontier models, so developers cannot update model weights and must focus instead on harness-level and context-level continual learning. To achieve this at Replit, the team built ViBench to evaluate functional app-building success from natural-language specs alongside Telescope, an automated system that clusters production failure traces into actionable issue groups.