Search the archive

REGISTER FOR LIVE + ON-DEMAND ACCESS

CLAUDE SONNET 4.5

Anthropic's latest model boasts the highest score on the SWE-bench Verified (77.2%), alongside major improvements in computer use, reasoning, and math. The release includes major product updates: checkpoints in Claude Code for instant rollback, the Claude for Chrome extension for Max users, and the Claude Agent SDK that makes the infrastructure powering Claude Code available to all developers.

BUY IN CHATGPT WITH INSTANT CHECKOUT

ChatGPT now supports direct purchases from Etsy sellers via Instant Checkout, with Shopify coming soon. The system runs on the open Agentic Commerce Protocol, co-developed with Stripe, and is open for merchant integrations.

ANTHROPIC LAUNCHES CLAUDE AGENT SDK FOR BUILDING VERSATILE AI AGENTS

Anthropic introduced the Claude Agent SDK, expanding its capabilities beyond coding to enable building versatile agents for tasks such as finance management and customer support. This SDK equips agents with tools for context gathering, executing tasks, and iterating based on feedback, optimizing workflows through features like bash scripting and subagents. Developers can leverage the SDK for immediate integration and enhanced agent functionality.

HOW GPU MATMUL KERNELS WORK

The architecture, assembly, and kernel design techniques behind high-performance matrix multiplication on NVIDIA GPUs.

A RESEARCH AGENDA FOR THE ECONOMICS OF TRANSFORMATIVE AI

There are nine "Grand Challenges" for studying the economic effects of AI, spanning issues from income distribution to AI safety to human meaning. Economists need to start researching these issues now due to the speed of AI progress and the unprecedented resource flow towards the AI industry. By the time transformative AI, which they define as driving 3x productivity growth, arrives, it will be too late for economists to aid in the upheaval.

BUILDING AI FOR CYBER DEFENDERS

AI models are now useful for cybersecurity tasks in practice, not just theory. Anthropic has caught hackers using their tools and invested in making Claude Sonnet 4.5 useful for defenders looking to detect, analyze, and remediate vulnerabilities in code and deployed systems. Now, the defenders need to adopt and experiment with AI to keep pace.

[11] or explore additional AI security resources from SANS.

AGENTIC COMMERCE

This site features design flows for embedded commerce in ChatGPT. OpenAI's Agentic Commerce Protocol is an open standard that enables purchases. It enables agents to reason over structured state, invoke tools at each step, and keep customers informed in real time. This page will help developers start with the essentials, deepen their understanding, and prepare for production with focused ACP resources.

DEEPSEEK-V3.2-EXP: SPARSE ATTENTION MODEL

DeepSeek-V3.2-Exp has introduced a sparse attention mechanism aimed at improving training and inference efficiency on long-context sequences.

LORA WITHOUT REGRET - THINKING MACHINES LAB

LoRA, the most popular method for fine-tuning AI models, can match the performance of expensive full fine-tuning when applied correctly, delivering the same results at two-thirds the cost. The key is applying LoRA to all network layers (especially MLPs rather than just attention layers) and ensuring the rank is high enough for the dataset. Thinking Machines found ranks of 256-512 work well for typical post-training scenarios combined with 10× higher learning rates than fine-tuning.

OPENAI'S NEW SORA VIDEO GENERATOR TO REQUIRE COPYRIGHT HOLDERS TO OPT OUT

OpenAI's new Sora video generator can create videos that feature copyright material unless copyright holders opt out of having their work appear. The company began alerting talent agencies and studios about the product and its opt-out process over the past week. The model will be released in the coming days. It won't generate images of recognizable public figures without their permission.

HOW IS IT POSSIBLE THAT CLAUDE SONNET 4.5 IS ABLE TO WORK FOR 30 HOURS TO BUILD AN APP LIKE SLACK?!

Claude Sonnet 4.5's system prompt reveals how it is able to work for 30 hours. This post goes into detail about how it works. A copy of the system prompt is available in the thread.

DO HUMANS REALLY HAVE WORLD MODELS?

There's little difference between the way humans and large language models think.

HUNYUANIMAGE-3.0 RELEASED

Tencent has open-sourced HunyuanImage-3.0, providing both inference code and model weights alongside a detailed technical report.

VIBE CHECK: CLAUDE SONNET 4.5

Claude Sonnet 4.5 is faster than GPT-5 Codex, and smarter and more steerable than Opus 4.1.

LOGICS-PARSING

Logics-Parsing is an end-to-end document parsing model built on a general Vision-Language Model that excels at accurately analyzing and structuring highly complex documents.

↗️

APPLE'S ‘VERITAS' CHATBOT IS REPORTEDLY AN EMPLOYEE-ONLY TEST OF SIRI'S AI UPGRADES

An internal employee-only chatbot, codenamed “Veritas”, is speculated to be Apple's answer to ChatGPT, but it's unknown if or when it'll be released publicly. The company has struggled to get its AI efforts off the ground after a lukewarm reception to its Apple Intelligence features and multiple delays to Siri upgrades.

DEEPSEEK-V3.1-TERMINUS LAUNCHES WITH IMPROVED AGENTIC TOOL USE AND REDUCED LANGUAGE MIXING ERRORS

DeepSeek released V3.1-Terminus, which strengthens its Code Agent and Search Agent and addresses users' feedback about Chinese/English mix-ups. The model shows benchmark gains in agentic tool tasks (BrowseComp, SWE Verified, etc.) and offers both “chat” and “reasoner” modes for different use cases. Terminus remains open-weight under an MIT license, making it viable for customization and deployment both via API and local environments.

HEAD OF CHATGPT ADDRESSES CONTROVERSIAL MODEL ROUTING

OpenAI confirmed it's testing a safety system that automatically switches users from GPT-4o to reasoning models or GPT-5 when conversations touch on "sensitive and emotional topics." Users have complained that the routing triggers too frequently in harmless situations, pulling them away from GPT-4o, which many prefer for personal conversations.

THE AI VILLAGE IN NUMBERS

OpenAI's models lead on various stylistic measures like lexical diversity, positive sentiment, and formality of language, but Anthropic's models lead in goal-directed behavior. While OpenAI's models vary in which stylistic measures they excel at, Anthropic's models all perform well on agentic behavior. OpenAI offers the brightest, most cheerful, and most eloquent model. Anthropic's models tend to excel at getting stuff done.

OPENAI RELEASES WORKPLACE USAGE REPORT SHOWING RAPID CHATGPT ADOPTION

28% of US workers now use ChatGPT at work, with adoption highest among highly educated white collar workers in tech, finance, and manufacturing. It's vague exactly how the tool is being used - OpenAI is primarily for writing, research, programming, and data analysis - but technical teams have emerged as the clear power users.

scale. Save your spot

MODULAR MANIFOLDS - THINKING MACHINES LAB

Normally, neural networks are trained by letting their weights move freely in any mathematical direction. A new method that restricts the weights to specific curved surfaces makes training more stable and predictable. This "manifold Muon" optimizer forces weight matrices to maintain consistent properties (like not stretching or shrinking data too dramatically) by constraining them to geometric shapes called manifolds. While promising, the experiments required computational overhead, which limits real-world adoption.

WE REVERSE-ENGINEERED FLASH ATTENTION 4

Flash Attention 4 is a brand new, highly optimized CUDA kernel that speeds up the attention calculations in transformers, which are the main mathematical operations that bottleneck AI models like ChatGPT. However, the improvements from FA4 are not from clever new math but from improvements in how an incredibly complex async pipeline splits computations across 32-thread ‘warps'.

WHAT PROBLEM IS DSPY SOLVING WITH WHAT ASSUMPTIONS

DSPY's programming model is a good base, but it has issues with defaults and missing affordances. This post identifies these issues and suggests some changes. Adopting these changes would sharpen DSPy's design without fighting it. It would keep the ergonomics that makes teams productive while aligning the optimization layer with the realities of fixed budgets and production constraints.

ALIBABA BETS BIG ON AI WITH NVIDIA TIE-UP, NEW DATA CENTER PLANS

Alibaba has partnered with Nvidia to enhance its cloud platform with advanced AI capabilities, focusing on robotics and autonomous driving. It plans to expand global data centers, opening facilities in Brazil, France, and the Netherlands to meet increasing AI infrastructure demand. The company also launched Qwen3-Max and Qwen3-Omni AI models, emphasizing improved performance in code generation and virtual reality applications.

EXPANDING OUR DATA ENGINE FOR PHYSICAL AI

Scale is expanding its Data Engine for Physical AI, providing robotics companies with high-quality, annotated datasets needed for training foundation models in physical environments. It addresses the robotics data gap by ensuring datasets are abundant, diverse, and enriched, collecting data through dedicated robots and human demonstrations. This aims to accelerate the development of reliable AI systems capable of handling real-world complexity in various applications.

live

A NEW PARADIGM OF PROACTIVE, STEERABLE AI

OpenAI's new CEO for Applications says the long-heralded dream of AI optimists is near - the personalized AI assistant that proactively helps you achieve your goals.

AN AI INDEX FOR ALL OUR CUSTOMERS

AI Index for domains on Cloudflare is a web index that gives content creators the tools to make their data discoverable by AI, giving AI builders access to better data for fair compensation.

RICHARD SUTTON – FATHER OF RL THINKS LLMS ARE A DEAD END

Large language models aren't capable of learning on-the-job, so no matter how much we scale, we will still need a new architecture to enable continual learning.

ANTHROPIC TO TRIPLE INTERNATIONAL WORKFORCE IN GLOBAL AI PUSH

Anthropic's international growth is outpacing the company's most ambitious forecasts.

ANTHROPIC EXPERIMENTS WITH AN AGENT FOR REAL-TIME UI GENERATION

Imagine with Claude, which lets users interact with a classic desktop UI where windows and apps are managed by AI, will be released as a temporary demo for certain plans.

YOUTUBE LABS

YouTube Labs is a new initiative that gives users early access to AI experiments.

🌀 Spin up a Together Instant Cluster

CHATGPT PULSE PREVIEW

ChatGPT Pulse is a new feature for Pro users on mobile that proactively delivers personalized research and updates via visual cards based on user chats, preferences, and connected apps.

PERPLEXITY SEARCH API

Perplexity has released a new Search API that gives developers access to the same large-scale infrastructure behind its public answer engine, enabling high-quality retrieval over hundreds of billions of web pages.

MUSK'S XAI SUES OPENAI, ALLEGING THEFT OF TRADE SECRETS

xAI alleges OpenAI conducted a "coordinated, unfair, and unlawful campaign" to recruit key employees and induce them to bring trade secrets, including xAI's rapid data center deployment methods.

WHAT ARE POPULAR AI CODING BENCHMARKS ACTUALLY MEASURING?

Popular coding benchmarks measure something narrower than what their names suggest. Claude scoring 80% on SWE-bench doesn't translate to it one-shotting 80% of the tasks thrown at it. This post looks at SWE-bench Verified, SWE-bench Pro, Aider Polyglot, LiveCodeBench, and other benchmarks to see what they are actually measuring. Designing a good benchmark is highly labor-intensive. Without human review and annotations, it's nearly impossible to make a good benchmark.

IS OPENAI'S REINFORCEMENT FINE-TUNING (RFT) WORTH IT?

OpenAI's reinforcement fine-tuning (RFT) for o4-mini is supposed to be able to train models to get better at specific tasks using reinforcement learning. It costs up to 700 times more than supervised fine-tuning but only seems to deliver clear wins on agentic coding tasks. The technology gives engineers flexibility in reward design through flexible grader configurations, and the performance gains can be substantial when it works, but it is expensive with limited applicability, making its use difficult to justify.

TOWARD COMPUTATIONAL TASTE: LLMS, AESTHETICS & JUDGMENT

LLMs are being used to model and optimize taste in diverse fields through personalized systems like Taste Engines, Aesthetic LLMs, and Taste Tribes. Methods like LoRe and models like TAPO enable LLMs to adapt to individual aesthetic preferences, potentially revolutionizing recommendation engines and social networks. This shift towards computational taste signifies a new era where machines not only reflect but also shape human preferences, offering both cultural and economic opportunities.