Search the archive

CURSOR PREPARES TO LAUNCH ORIGIN PLATFORM FOR CODE REVIEWS

Cursor seems to be preparing to launch Origin beyond the closed partner beta under the name 'Cursor Review'. Cursor Review will enable two new tabs: Codebase covers syncing and managing repositories pulled in from GitHub, and Review contains an automated pull request pipeline that notifies developers when their judgment is needed, allowing humans and agents to work through open PRs across a codebase together. The rollout could land as early as this week.

NVIDIA'S SWITCHYARD ROUTER RESHUFFLES AI MODELS MID-TASK, CUTTING TASK COSTS TO A THIRD IN ITS OWN TESTS

Nvidia has announced Nemotron 3.5 Lightning, a 30-billion-parameter open mixture-of-experts model built for high-volume, specialized agent tasks, and NeMo Switchyard, an open-source library that routes each step of an agent workflow to whichever model fits it best. Lightning delivers up to 4x faster output than comparable models in its class. It can complete agentic tasks roughly 30% faster than Qwen3.6-35B at matching accuracy. Paired with Switchyard, the model retains frontier-level task completion ability while costing roughly a third of running Opus 4.8 alone.

RYAN GREENBLATT – WHAT HAPPENS ONCE AI CAN AUTOMATE AI RESEARCH? (2 HOUR READ)

This post contains a transcript of an interview with Ryan Greenblatt, the chief scientist at Redwood Research, where he focuses on technical AI safety and security work. The interview covers the topic of recursive self-improvement (RSI), one of the most debated in the world of AI right now. One possible outcome of achieving RSI is that the world quickly slingshots towards having tens of billions of superintelligences, each of which is dramatically more competent than human experts across all fields. Greenblatt believes this scenario is possible, and he predicts that humans will automate AI R&D by 2031.

AGI WILL SET OFF AN INDUSTRIAL EXPLOSION

An AI system that can do every remote job will also be able to operate a robot. A system that can master real-time control, spatial reasoning, physical prediction, and continual learning will have what it needs to operate machinery if given actuators to work with. Hardware is not the bottleneck for robotic automation. Once robots can replace human labor, the economy will no longer be bound by a fixed workforce.

THE FRICTIONS THAT MAKE AI FORECASTING HARD

AI forecasting fails when capability gains are mistaken for end-to-end impact, because real-world workflows contain social, institutional, physical, and tacit bottlenecks. Better forecasts require practitioner-informed, cross-framework research that measures whether AI actually changes the limiting step, not just one task.

Reserve your spot today

STEALING REASONING TRACES FROM PROPRIETARY LLM APIS

Proprietary reasoning can be recovered from its encrypted traces. Anthropic, OpenAI, and Google return encrypted chain-of-thought blocks to clients that can be replayed across sessions, users, and models. Researchers took a trace produced by a frontier model, replayed it into a weaker sibling, jailbreaked the weaker model, and recovered the stronger model's hidden reasoning in plaintext, without ever attacking the stronger model directly or triggering its anti-distillation safeguards. The decoded reasoning closely tracks the number of hidden thinking tokens reported by the API, and they contain real secrets and sensitive information.

MEMORY IN VIDEO WORLD MODELS

NVIDIA's WorldTrace is a training-free framework that keeps compressed memories accessible during long autoregressive video rollouts by assigning them stable, in-distribution positional offsets.

MAI-CODE-1.1-FLASH: BETTER, FASTER, AT A QUARTER OF THE COST

MAI-Code-1.1-Flash produces higher quality code at 25% greater token efficiency and at a quarter of the cost compared to the previous model launched in June. This was achieved by optimizing for real-world use across hundreds of thousands of reinforcement-learning environments in GitHub Copilot. MAI-Code-1.1-Flash shows a 22% improvement on Terminal-Bench 2.1 in GitHub Copilot CLI and a 15% improvement on .NET tasks. It is now available in GitHub Copilot.

NEMOTRON 3.5 LIGHTNING

NVIDIA introduced Nemotron 3.5 Lightning, an open 30B mixture-of-experts model with 3B active parameters designed for low-latency, high-volume tasks inside long-running AI agents.

GOOGLE'S NEW AI BOSS INHERITS A RACE TO CATCH OPENAI AND ANTHROPIC

Koray Kavukcuoglu is becoming head of Google DeepMind. He will report directly to Google CEO Sundar Pichai and oversee Gemini model development, frontier AI research, and the Gemini app and developer teams. Kavukcuoglu was previously DeepMind's CTO and Google's chief AI architect. The promotion shows that Google is prioritizing execution over deep research.

THE MODEL PICKER IS A DEAD END

The model picker in AI products assumes one model can handle all tasks optimally, which is flawed. Lovable focuses on model independence, aligning specific models with suitable tasks and adapting as models improve. The control plane monitors builds, switching models to optimize performance, and even incorporates Lovable's self-trained models when they outperform external ones.

the interactive explainer on the ngrok blog

WHERE AN AI WATERMARK CAN HIDE IN PLAIN TEXT

It's possible to hide watermarks in model weights.

RD-SIGNAL-2: FRONTIER CLASSIFICATION AT PRODUCTION SCALE

Raindrop launches Signals 2.0, offering task-specific binary classifiers with the rd-signal-2 model, matching near GPT-5.6 accuracy at a fraction of the cost.

MANUS UNACQUIRED

Some users will need to back up and restore their data as Manus resumes operations as an independent company.

Code sessions, 103 tasks, and 3 models. SEE THE RESULTS.

INTRODUCING GROK BOT

Grok Bot gives each agent its own cloud computer, memory, and access to apps and websites, letting teams delegate multi-step work that continues autonomously.

IMPORT FROM ANOTHER AGENT

The ChatGPT desktop app and Codex CLI now support importing settings, skills, plugins, and projects from another agent.

FOR WEDNESDAY, AUGUST 19TH.

META RELEASED MUSE GLIMMER

Meta has introduced Muse Glimmer, a 30B-parameter open-weight model, under Apache 2.0, optimized for always-on local agents, coding, function calling, and model evaluation.

GPT-5.6-CYBER

OpenAI introduced GPT-5.6-Cyber, a specialized model for vulnerability research, exploit validation, and other advanced cybersecurity tasks. The company also expanded its Daybreak program with Blue and Red access tiers designed to give approved defenders access to increasingly capable AI tools.

LEARNING MORE ABOUT CLAUDE'S MATHEMATICAL CAPABILITIES

Claude, an AI model by Anthropic, improved the lower bound of zeros satisfying the Riemann hypothesis from 41.6% to 67.2%. Using insights from prior research and attempting 650 ideas, Claude coordinated multiple subagents to run numerical checks and re-prove its finding. Two mathematicians and a formal validation confirmed the result, highlighting AI's unexpected potential in advancing mathematical research.

EXPLORING CLAUDE/GPT KNOWLEDGE CUTOFFS & PRE-TRAINING TIMELINES

It's possible to learn hidden facts about how frontier models were trained by probing them with carefully curated requests. The number of parameters models have can be estimated by scoring them on niche facts. Measuring how models break down tokens can reveal facts about the dataset mixtures used to train the model. Training timelines can be estimated by scoring models on date- or self-identification-related questions.

BUILDING AN AI-NATIVE FINANCE TEAM

OpenAI shared five lessons from rebuilding its finance function around AI, with long-term goals including a zero-day close and continuously updated forecasting. The approach emphasized redesigning workflows around decisions, live business context, human accountability, experimentation, and measurable AI-driven output.

ARE AGENTS REALLY KILLING UI?

Agents are shifting software toward hybrid interfaces rather than eliminating UI: products need agent-friendly onboarding, MCP access, and instrumentation alongside human-facing controls. The highest-value screens increasingly handle approval, review, undo, orchestration, and visibility into what agents changed.

Put Your Agent To The Test

QWEN-MM-PLUGINS

This repository contains native multimodal plugins for Qwen models. They enable agent harnesses to be multimodal-native. Each capability includes a skill and an optional MCP server. Each capability's cookbook has a full tool listing, setup, and worked cases.

H3-METAL

h3-metal brings native MiniMax-H3 inference to Apple Silicon. It currently supports prompt-to-video/audio, first/last-frame conditioning, and ordered Ref2VA image/video/audio references. The project is focused on incremental H3-specific Metal performance and memory optimization on the M3 Max and M5 Max.

A CONTROLLED STUDY OF ATTENTION-ONLY TRANSFORMERS

The feed-forward network (FFN) has been a fixed component of the transformer block since its introduction. In modern decoder configurations, it holds roughly two-thirds of non-embedding parameters. A substantial interpretability literature argues that these layers act as the model's parametric memory. This experiment looks at what is actually lost when the FFN is removed entirely.

ANTHROPIC TRIES TO SHORE UP INVESTOR CONFIDENCE AHEAD OF BLOCKBUSTER IPO

Anthropic is meeting up with investors to offer assurances about its rapid pace of growth and insights into the company's strategies to address growing public backlash against AI. The company must address issues like the recent popularity of cheaper AI systems from China, tensions with the Trump administration, and growing backlash to data-center construction ahead of its IPO. The conversation reflects the tremendous uncertainty around who will win the AI race and the financial stability of the businesses that underpin it. Anthropic is targeting a public debut in September or early October.

THE FUTURE IS FOR EVERYONE

Meta plans to democratize superintelligence by creating personal AI agents to enhance individuals' capabilities while ensuring privacy. Emphasizing invention over automation, these tools will empower people to shape their future, boost economic growth through entrepreneurship, and accelerate scientific progress. Meta advocates for a balance of power to prevent singular centralized AI, ensuring AI serves humanity by empowering individuals rather than institutions.

ELECTRICITY PRICING IN THE AGE OF AI

AI's growing demand highlights power as the real bottleneck, with data center electricity needs doubling every two years. The book explores the intricacies of electricity commodities, emphasizing the central role of independent system operators (ISOs) in managing supply and pricing through marginal cost principles. Economic models like "energy-only markets" and spread trading strategies are key for predicting market movements and securing stable economics in the volatile power markets.

CAN AGENTS USE A COMPUTER YET? WE'VE GOT THE DATA

Agents have rapidly advanced in automating computer use, effectively managing repetitive tasks such as ticket processing, data entry, and navigating legacy systems without APIs.

NVIDIA REPORTEDLY TESTING LOWER MEMORY CONFIGS OF RUBIN ULTRA AS MEMORY SHORTAGE BITES BACK

Nvidia has tested at least three lower-memory designs.

NVIDIA TEAMS UP WITH WALL STREET ASSET MANAGERS ON $500 BILLION AI INFRASTRUCTURE PUSH

Nvidia is collaborating with major asset managers like Apollo, Blackstone, and Goldman Sachs to fund $500 billion for AI infrastructure.

MICROSOFT PLANS MAIA 300 AI CHIP UNVEILING IN SEPTEMBER, REPORT SAYS

Microsoft plans to unveil its Maia 300 AI chip in September, aiming to reduce reliance on Nvidia GPUs.

OPENAI REPORTEDLY COMPLETED A $7 BILLION EMPLOYEE TENDER OFFER

The deal, which was part of an effort to provide liquidity to OpenAI's workforce, valued the company at $852 billion, the same as its most recent fundraising round in March.

DYNA-2: A 1-MILLION-HOUR SCALING LAW FOR WORLD-ACTION MODELS

Dyna-2, a world-action model, uses over one million hours of human video data to establish scaling laws that predictably enhance action accuracy on both human and robot tasks.

Launch a cluster on Verda →

OPENAI PAUSES ASTRA OVER CYBERSECURITY RISKS

OpenAI paused work involving its upcoming Astra model after internal evaluations suggested it could approach “Critical” cybersecurity capabilities, including advanced autonomous exploit development.

CLAUDE CODE MAKES AUTO MODE THE DEFAULT

Anthropic said Claude Code's auto mode would become the default for Pro, Max, and Team users on August 14, allowing most actions to proceed without approval prompts.

XAI LAUNCHES IMAGINE IMAGE 2.0 IN GROK QUALITY MODE

xAI launched Imagine Image 2.0 on Grok, enhancing control for creative work with features like Magic Wand and Smart Resize. It ranks second in text-to-image and image editing globally. An API is planned, but currently, the model is available only through Grok's web and app platforms.