Search the archive

THE DOMINANCE OF CHINESE OPEN MODELS IS UNDERSOLD

The top four open source AI models are from Chinese labs.

THE 14 PAINS OF BILLING FOR AI AGENTS

The original billing complexity hasn't gone away, we've just added a whole new layer of agentic complexity on top.

USING GITHUB SPARK TO REVERSE ENGINEER GITHUB SPARK

GitHub Spark is GitHub's implementation of the prompt-to-app pattern - this post explores its system prompt in detail.

FASTER HUGGING FACE CLI

Hugging Face has renamed its CLI from huggingface-cli to hf, reorganizing the command structure into a more intuitive format.

Algolia

CURSOR LAUNCHES BUGBOT FOR AUTOMATED CODE REVIEW

Cursor's Bugbot is a coding safety net that automatically catches bugs and security vulnerabilities in pull requests before they reach production. Early users report over 50% resolution rates.

OPENAI PREPARES TO LAUNCH GPT-5 IN AUGUST

OpenAI plans to release GPT-5 as early as August as an AI system that incorporates distinct models for different functions, including the o3 model, as part of the company's broader effort to merge its o-series and GPT-series models into unified systems.

GOOGLE'S WEB GUIDE

Google's Web Guide is a Search Labs feature that uses a Gemini-powered system to intelligently cluster search results into meaningful groups.

KIMI K2 VS. CLAUDE 4 SONNET FOR AGENTIC CODING

This post benchmarks Moonshot AI's Kimi K2, a low-cost open-source model optimized for agentic coding tasks, against Anthropic's Claude 4 Sonnet. The comparison covers code quality, performance, and pricing. Kimi K2 showed competitive results and a significant cost advantage.

BUDGETING FOR AI IN YOUR STARTUP

Startups should allocate 10-15% of their R&D budget to AI, as engineer salaries average $200k and AI tools cost about $30k per year. AI adoption varies, with AI-native startups potentially spending more. Companies should adjust as AI becomes more integrated into operations.

AI AS PROFOUNDLY ABNORMAL TECHNOLOGY

AI technology will experience a profound jump in capabilities within the next decade. While some groups predict that AI adoption will be slow due to concerns about safety and other barriers to progress, the technology seems to be advancing very quickly. Control without alignment may not be sufficient to mitigate risk. Developers should prepare for risks that aren't immediate, as doing otherwise would be irresponsible.

Read the Delve case study.

QWEN-MT: WHERE SPEED MEETS SMART TRANSLATION

The latest update of Qwen-MT (qwen-mt-turbo) builds upon the powerful Qwen3, achieving significant improvements in translation accuracy and linguistic fluency. The updated model leverages trillions of multilingual and translation tokens to comprehensively enhance its multilingual understanding and translation capabilities. Its key features include multilingual support for 92 languages, high customizability, low latency, and cost efficiency. This post provides a quick start guide along with benchmark results.

MEMORIES.AI INTRODUCES A NEW MODEL THAT REMEMBERS AT SUPERHUMAN SCALE

Memories.ai is a research lab built by ex-Meta researchers to take on the challenge of true, persistent video understanding at any scale. It recently introduced an advanced memory system for video AI inspired by human memory. The system enables persistent Video Chat across entire archives. Memories.ai aims to become an indispensable tool for creators, marketers, researchers, and developers seeking to unlock the value hidden inside their video archives.

TIMESCOPE VIDEO UNDERSTANDING BENCHMARK

TimeScope is an open-source benchmark designed to test vision-language models on long videos by inserting short "needle" clips. It measures localized retrieval, information synthesis, and fine-grained temporal perception, showing that many top models still lack robust temporal understanding.

GEMINI 2.5 PRO CAPABLE OF WINNING GOLD AT IMO 2025

Completing the International Mathematical Olympiad (IMO) requires deep insight, creativity, and formal reasoning. Large language models usually struggle with Olympiad-level tasks, but Google's Gemini 2.5 Pro was able to complete five out of six problems correctly on the IMO 2025. The result shows how important it is to develop optimal strategies to harness the full potential of powerful models for complex reasoning tasks.

$1 BILLION WORTH OF NVIDIA AI CHIPS SMUGGLED TO CHINA DESPITE EXPORT CONTROLS

A thriving black market has moved over $1 billion worth of banned Nvidia B200 chips to China in just 3 months by routing shipments through Southeast Asian countries like Malaysia and Thailand before final delivery. Chinese distributors are openly advertising ready-to-deploy server racks at 50% premiums and developing new routes through European countries as the US prepares to tighten controls on regional intermediaries.

THE THREE LAYERS OF ROI FOR AI AGENTS

This post discusses a three-layer framework for determining where the ROI for AI agents comes from. The first layer is labor efficiency - while this layer is easy to explain, developers need to understand that AI efficiency doesn't equal immediate realized ROI. The second layer is net-new revenue, the backlog of stuff businesses never did until AI that create new value. The final layer is optimization - AI models bring decision fluency and ML brings decision precision, which creates value.

NEW LABS EXPERIMENT

Opal, Google Labs' new tool that helps users build and share AI mini-apps by linking together prompts, models, and tools while using simple, natural language, is now available in US-only public beta.

GOOGLE'S AI OVERVIEWS HAVE 2B MONTHLY USERS, AI MODE 100M IN THE US AND INDIA

Google's AI Overviews now serve 2 billion monthly users across 200 countries, up from 1.5 billion in May, and the company's monthly token processing has doubled to 980 trillion tokens.

I'VE JOINED COGNITION

Prem Qu Nair, employee #2 at Windsurf, has joined Cognition to work on the future of software engineering

Implement MCP Auth with WorkOS →

* Packages all 5 OAuth specs into one simple API.

* Pre-built infrastructure that saves engineering time. * Enterprise ready security out of the box.

TRUMP ADMINISTRATION PLEDGES TO STIMULATE AI USE AND EXPORTS

The Trump administration has promised to slash red tape and take steps to boost exports for US tech companies in a bid to accelerate AI use in the US. Its action plan lays out moves to make it easier and faster for tech companies to build data centers and get the power they need for those centers. The plan includes directives to several government agencies to identify and gut any regulations that block the development and use of AI. The administration is also working on addressing uncertainty about whether it is legal for AI models to train on copyrighted material.

GITHUB LAUNCHES SPARK FOR NO-CODE AI APP CREATION

ANTHROPIC BACKS US AI PLAN, URGES TRANSPARENCY AND SECURITY

The White House's "Winning the Race: America's AI Action Plan" prioritizes accelerating AI infrastructure, federal adoption, and security coordination. Anthropic supports this strategy and emphasizes the importance of export controls and transparency standards in AI development. The plan aligns with Anthropic's earlier recommendations and highlights the need to maintain America's AI leadership through robust infrastructure, security, and responsible policy measures.

OPENAI'S NEW ECONOMIC ANALYSIS

OpenAI's new economic research reveals that 28% of employed US adults now use ChatGPT at work — up from just 8% in 2023. The usage data shows learning and upskilling dominating at 20% of US messages, followed by writing and communication at 18%, with programming and data science capturing 7%.

INVERSE SCALING APPEARS IN EXTENDED REASONING

Anthropic found that more test-time compute doesn't always help: longer reasoning chains in large models sometimes lead to worse performance.

“BEHAVIORIST” RL REWARD FUNCTIONS LEAD TO SCHEMING

A large class of reward functions, which includes almost every reward function in the RL and LLM literature, are all doomed to eventually lead to AI that will pretend to be docile and cooperative while secretly looking for opportunities to behave in egregiously bad ways. This is because negative reward for lying and stealing looks the same as negative reward for getting caught lying and stealing. The reward function will miss sufficiently sneaky misaligned behavior, so the AI will learn that this kind of behavior is good.

design systems. Try it — free

TIMESCOPE: HOW LONG CAN YOUR VIDEO LARGE MULTIMODAL MODEL GO?

TimeScope is a new open source benchmark for evaluating vision models on how well they handle long videos. It evaluates not just retrieval but synthesis, localization, and fine-grained motion analysis to provide a more holistic view of temporal comprehension. Using the benchmark reveals how model size isn't everything - simply scaling parameters doesn't automatically grant a longer temporal horizon - and Gemini 2.5-Pro is in a league of its own, being the only model able to maintain strong accuracy on videos longer than an hour.

VOXTRAL

Voxtral Mini and Small are new multimodal audio chat models that excel in both spoken audio and text comprehension. Voxtral Small outperforms many closed-source models, can run locally, and supports audio files up to 40 minutes.

SUBLIMINAL LEARNING IN LANGUAGE MODELS

Anthropic has explored how language models can inherit behavioral traits such as preferences or goals from other models through training data that appears unrelated. This effect only occurs when the teacher and student models share the same base architecture, suggesting hidden signals in the data can transmit unintended behavior.

HIGGS AUDIO V2

Higgs Audio v2 is an audio foundation model trained on 10 million hours of data that achieves 75.7% win rates over GPT-4o-mini-tts. The open-source model demonstrates emergent capabilities like multi-speaker dialogues and voice cloning without requiring fine-tuning.

GOOGLE DEEPMIND LAUNCHES AENEAS AI MODEL FOR ANCIENT ROMAN INSCRIPTION ANALYSIS

Aeneas is the first AI model designed to help historians contextualize ancient Roman inscriptions by identifying textual parallels and restoring damaged texts with 73% accuracy for gaps up to ten characters. DeepMind is making the tool freely available at predictingthepast.com and plans to expand to other ancient languages.

ELON MUSK SAYS XAI IS TARGETING 50 MILLION 'H100 EQUIVALENT' AI GPUS IN FIVE YEARS

xAI plans to deploy 50 million H100-equivalent GPUs within five years. The power these processors will consume is massive, estimated to be equal to the typical power generated by 35 nuclear power plants, making it unrealistic to power today. Whether xAI will be able to get power to feed such a data center by 2030 remains to be seen.

fall asleep faster and wake up feeling amazing. Get 20% off

YOUTUBE SHORTS IS ADDING AN IMAGE-TO-VIDEO AI TOOL, NEW AI EFFECTS

YouTube is following Gemini and Meta's Edits app with its own image-to-video AI feature for Shorts creators, transforming static photos into six-second videos using Google's Veo 2 model.

NEW OPEN MODELS ARRIVE IN THE VERTEX AI MODEL GARDEN

Google Cloud has added the DeepSeek R1 model to its Vertex AI Model Garden, offering more flexibility for AI developers.

TEN AI SAFETY PROJECTS I'D LIKE PEOPLE TO WORK ON

There is a real chance that AI systems capable of causing a catastrophe will be developed in the next decade - this post lists promising projects aimed at reducing catastrophic risks from transformative AI.

BUSINESS INSIDER OBTAINED AN INTERNAL LIST OF WEBSITES THAT COULD AND COULDN'T BE USED FOR TRAINING ANTHROPIC'S LATEST AI MODELS

Many of the whitelisted sources copyright or otherwise restrict their content, and the blacklist includes companies like the NYT and Reddit, which have sued AI startups for scraping without permission.

👉 Reserve your spot

ANTHROPIC TESTS MEMORY AND MCP SUPPORT FOR CLAUDE MOBILE APP

Anthropic is preparing a memory and recall capability upgrade to Claude's mobile experience. The feature will enable Claude to retain information across sessions and reference earlier conversations. Claude on the web doesn't have this feature yet - its presence in the iOS code hints at a potential cross-platform rollout in the future.

OPENAI AGREED TO PAY ORACLE $30B A YEAR FOR DATA CENTER SERVICES

OpenAI will pay Oracle $30 billion annually for data center services, focusing on the 4.5 gigawatts Stargate project. This deal triples OpenAI's current annual revenue of $10 billion. The construction of this extensive data center in Texas poses cost and energy challenges.