Search the archive

WILL THE EU DELAY ENFORCING ITS AI ACT?

45 European companies and major tech firms, including Google, Meta, and Mistral, are calling for a "two-year clock-stop" on the EU AI Act before key obligations take effect on August 2, citing missing guidance documents and compliance uncertainty.

META OFFERS TO BUY STAKE IN VENTURE FUNDS STARTED BY AI HIRES NAT FRIEDMAN AND DANIEL GROSS

Liquidity providers could cash out at the current net asset value, an unusual opportunity in today's markets, where secondary transactions often come at a discount.

Ready to start the conversation? Get the full guide (PDF)

PERPLEXITY LAUNCHES A $200 MONTHLY SUBSCRIPTION PLAN

Perplexity Max offers unlimited access to Labs, a report-generating tool, and early access to Comet, its forthcoming AI-powered browser.

ORACLE, OPENAI EXPAND STARGATE DEAL FOR MORE US DATA CENTERS

OpenAI will rent 4.5 gigawatts of data center power from Oracle across multiple US states - enough electricity to power millions of homes - as part of a $30 billion cloud deal as part of the $500 billion Stargate initiative.

COULD GOOGLE'S VEO 3 BE THE START OF PLAYABLE WORLD MODELS?

Google's latest video-generating model, Veo 3, could potentially be used for video games. World models are different from video-generation models - video-generation models synthesize realistic video sequences while world models simulate the dynamics of real-world environments. Veo 3 isn't quite a world model yet, but it could be used for cinematic storytelling and narrative prototyping. Google has plans to turn its multimodal foundation model, Gemini 2.5 Pro, into a world model.

AI MODELS REQUIRING $10M+ TO TRAIN NOW LAUNCH TWICE PER MONTH

Epoch AI tracked 201 models exceeding 10²³ FLOPs in 2024, up from just two in 2017, with training costs for the largest models now reaching tens of millions of dollars at the 10²⁵ FLOP threshold first crossed by GPT-4.

BUILDING REWARD FUNCTIONS FOR CHEMICAL AI: A TALE OF REWARD HACKING

FutureHouse researchers detail the months-long battle against their chemistry reasoning model, ether0, which kept finding creative ways to game their reward systems—from proposing reactions that just "add water and call it a reaction" to generating explosive peroxide chains. Their iterative fixes using bloom filters and structure alerts reveal why domain expertise, not just ML engineering, has become critical for training reasoning models.

Read the Delve case study.

OPTIMIZING CHECKPOINTING WITH PYTORCH DCP

PyTorch engineers reduced checkpoint sizes by 22% using a modular compression approach in Distributed Checkpointing (DCP). This post outlines practical integration steps and design choices to streamline storage and bandwidth during distributed training.

GEMINI 2.5 FOR ROBOTICS AND EMBODIED INTELLIGENCE

Gemini 2.5 models advance robotics with enhanced coding, reasoning, and multimodal capabilities rooted in spatial understanding. Developers can use Gemini 2.5 for complex scene queries, robot control through spatial reasoning, and interactive applications using the Live API for real-time voice commands. Trusted testers like Boston Dynamics showcase its potential in robotics.

GNNS WITH TRANSFORMERS FOR NEURAL PREDICTORS

A new predictor combines GNNs and transformers with a sibling-aware token mixer and a bidirectional graph isomorphism feed-forward network to better model DAG-based neural architectures.

GRAPH REGRESSION WITH CONTRASTIVE LEARNING

Extending causal graph learning techniques to regression tasks, this work adapts intervention strategies via contrastive learning to better handle confounders in graph-level regression.

THE PATH TO MEDICAL SUPERINTELLIGENCE

Microsoft AI's Diagnostic Orchestrator beats expert physicians by diagnosing 85% of New England Journal of Medicine cases four times more accurately. It achieved this faster and more cost-effectively through sequential diagnosis, mimicking real-world clinical decision-making. This advancement is supported by Microsoft AI's new Sequential Diagnosis Benchmark, which transforms complex medical cases into interactive diagnostic challenges.

NYT TO START SEARCHING DELETED CHATGPT LOGS AFTER BEATING OPENAI IN COURT

OpenAI's request to overturn a court order requiring it to retain all ChatGPT logs 'indefinitely' has been denied. The company's user agreement specifies that its data can be retained as part of legal processes. OpenAI plans to keep fighting the order, but it seems to have few options left. In the meantime, it is negotiating a process that will allow news plaintiffs to search through the retained data.

build it with React. Try a prompt

GOOGLE PROMOTES ‘AI MODE' ON HOME PAGE 'DOODLE'

Google promoted "AI Mode" via its homepage Doodle, exposing users to its latest AI search product.

CHINA HOSTS FIRST FULLY AUTONOMOUS AI ROBOT FOOTBALL MATCH

Humanoid robots competed in an AI-powered three-a-side football match in China, highlighting the challenges they face in agility and coordination.

OPENAI SAYS IT HAS NO PLAN TO USE GOOGLE'S IN-HOUSE CHIP

OpenAI clarified it's only doing early testing with Google's TPUs with no plans to deploy them at scale, contradicting reports from two days earlier that suggested the ChatGPT maker was turning to its competitor's AI chips to meet growing demand.

ROBINHOOD STOCK HITS RECORD FOLLOWING ANNOUNCEMENT OF OPENAI AND SPACEX TOKENS

OpenAI distanced itself from Robinhood's new tokenized share offering, stating "these 'OpenAI tokens' are not OpenAI equity" and that it didn't partner with or approve any equity transfer.

ELON MUSK'S XAI RAISES $10 BILLION IN DEBT AND EQUITY AS IT STEPS UP CHALLENGE TO OPENAI

xAI has raised $5 billion in debt through secured notes and term loans and $5 billion in equity secured through a strategic equity investment. The debt offering was oversubscribed and included prominent global debt investors. The funding will go towards the startup's efforts to build out its AI infrastructure and develop its Grok chatbot. It will continue buying Nvidia and AMD chips for a 1-million-GPU facility outside of Memphis.

GRAMMARLY TO ACQUIRE SUPERHUMAN TO ACCELERATE ITS AI PRODUCTIVITY PLATFORM

Grammarly is acquiring Superhuman to enhance its AI productivity platform by integrating AI-native email capabilities. Superhuman users send and respond to 72% more emails per hour with AI, showcasing increased efficiency. This acquisition supports Grammarly's plan to create applications as intelligent agents, positioning email as a key platform for multi-agent assistance.

AMAZON LAUNCHES A NEW AI FOUNDATION MODEL

Amazon introduced a new AI foundation model to optimize warehouse robotics and celebrated the milestone of deploying one million robots. The model improves efficiency by 10% and supports faster deliveries.

INSIDE OPENAI'S CHATGPT JOURNEY (65 MINUTE VIDEO)

Nick Turley and Mark Chen discuss ChatGPT's viral launch, evolving safety strategies, breakthroughs in multimodal models, and the rise of agentic coding with Codex. They also share insights on internal culture and the vision for AI-powered assistants.

FINETUNING SPARSE ENCODERS WITH SENTENCE TRANSFORMERS

The latest update to Sentence Transformers introduces support for training sparse embedding models, ideal for hybrid search and reranking. The blog walks through model components and training steps and highlights available SPLADE-based models.

BUILDING A PERSONAL AI FACTORY

This post discusses how to build an AI factory that improves itself. The aim of the project was to build a factory that can produce itself. It aimed to create a factory of agents that can produce code, verify it, and improve themselves over time. The post covers the idea behind the AI factory, how to build the factory, scaling the factory, and ideas for future improvement.

WHAT WE LEARNED TRYING TO DIFF BASE AND CHAT MODELS AND WHY IT MATTERS

Researchers compared raw AI models with their chat-trained versions to understand what safety training actually changes inside the model—like finding the "off-limits" neurons that make a model refuse harmful requests. They discovered that Anthropic's standard comparison technique was hallucinating thousands of fake differences, but their BatchTopK fix revealed real interpretable safety circuits like refusal detection and stereotype filtering, capturing 88% of behavioral changes versus just 60% for the broken method.

REFRAMING DEEP LEARNING PHENOMENA

Arguing against case-by-case analysis of concepts like grokking or double descent, these authors propose shifting focus toward broader theories that offer more practical value for real-world deep learning applications.

INTRODUCING CHAI-2: ZERO-SHOT ANTIBODY DESIGN IN A 24-WELL PLATE

Finding antibodies that bind to disease targets usually means screening massive libraries for months—most pharmaceutical discovery is basically a numbers game. Chai Discovery's AI cuts straight to viable drug candidates by designing them computationally, hitting 16% success rates and finding working antibodies for 50% of targets with just 20 attempts each.

KEY PATHS, PLANS, AND STRATEGIES TO AI SAFETY SUCCESS

Key AI safety paths include international cooperation, domestic initiatives, and societal resilience. Strategies involve forming alliances like CERN for AI, using deterrence to prevent unsafe development, and fostering responsible actors within countries. A diversified approach across these strategies offers insurance against uncertainties and leverages various expertise.

COMMUNITY NOTES ON X

X launched a pilot allowing LLMs like Grok or ChatGPT to submit Community Notes. These AI-generated notes will be vetted by humans and subject to the same trust mechanisms as existing user contributions.

Subscribe now.

CURSOR HAS SIGNED TWO KEY PLAYERS FROM ANTHROPIC'S CLAUDE CODE TEAM

Boris Cherny and Cat Wu, product managers who worked on Claude Code, have joined Cursor.

CLOUDFLARE LAUNCHES A MARKETPLACE THAT LETS WEBSITES CHARGE AI BOTS FOR SCRAPING

Cloudflare launched Pay per Crawl, a marketplace enabling sites to charge AI bots for content scraping.

CHINA'S HUAWEI OPEN-SOURCES AI MODELS AS IT SEEKS ADOPTION ACROSS THE GLOBAL AI MARKET

Huawei will open source two of its PanguAI models as well as some of its reasoning technologies.

Read the Delve case study.

CURSOR LAUNCHES WEB AND MOBILE APPS

Cursor can now run coding tasks in the background while users are away and works across any browser or mobile device. Teams can review agent-generated diffs and create pull requests directly from the web interface, with Slack integration for notifications and triggering agents.

APPLE CONSIDERS LETTING ANTHROPIC AND OPENAI POWER SIRI

Repeated delays have forced Apple to abandon its typically insular approach. Both companies are training specialized versions of their models to run on Apple's cloud infrastructure.

ZUCKERBERG ANNOUNCES META ‘SUPERINTELLIGENCE' EFFORT, MORE HIRES

Mark Zuckerberg announced a restructuring of Meta's AI division to focus on developing superintelligence. The new group, Meta Superintelligence Labs (MSL), will be led by Alexandr Wang, appointed as chief AI officer.

HOW ORACLE IS WINNING THE AI COMPUTE MARKET

Oracle has quietly become a key player in the AI compute race, largely due to the success of its partnership with ByteDance. The two companies poured billions into data centers in Singapore and Malaysia, which is now the world's second-largest AI hub. The company was the single largest lessor of datacenter capacity in the US from late 2023 to early 2025, securing over $130 billion in contracts expected over the next 12 months, including Stargate deals.

MACHINES OF FAITHFUL OBEDIENCE

AI, including AGI and ASI, can significantly benefit humanity if the technical alignment problem is solved, enabling AIs to follow human instructions accurately and faithfully. However, rapid AI progress poses risks, such as reaching alignment "uncanny valleys," potential arms races, surveillance states, societal upheaval, and misuse in offensive applications. Ensuring dominance in inference-time compute with responsible deployment and maintaining human oversight are critical to mitigating these risks and harnessing AI's potential for positive global impact.