Search the archive

👀 Get an inside look on June 24

PHYSICAL WORLD MODEL BY META

Meta has announced V-JEPA 2, a new visual world model that enhances physical reasoning for AI agents. The company also introduced three benchmarks to assess model performance on real-world video-based reasoning tasks.

THE BROWSER COMPANY LAUNCHES DIA, AN AI-FIRST BROWSER

Dia integrates AI directly into the URL bar, allowing users to query open tabs, generate drafts from their content, and build custom code snippets through natural language. The Browser Company pivoted from Arc after acknowledging its complexity limited mainstream adoption, betting that embedding AI at the browser level will capture users who currently switch between ChatGPT, Perplexity, and Claude.

MISTRAL ANNOUNCES AI COMPUTE PLATFORM WITH TENS OF THOUSANDS OF GPUS

As an alternative to US and Chinese Cloud Providers, Mistral Compute offers private AI stacks including GPUs, orchestration, and APIs to attract highly-regulated enterprises.

ALPHAWRITE: TEST-TIME COMPUTE SCALING FOR WRITING

AlphaWrite generates story variants with different author styles and themes, uses pairwise comparisons to rank quality, then evolves top performers across multiple generations, demonstrating that creative tasks can benefit from systematic inference-time compute scaling beyond just math and coding domains.

THE RISE OF SYSTEMS OF CONSOLIDATION APPLICATIONS

CLAUDE SQUAD

Claude Squad is a terminal app that manages multiple local agents in separate workspaces, allowing users to work on multiple tasks simultaneously. It can compute tasks in the background, manage instances and tasks in one terminal window, review changes before applying them, checkout changes before pushing them, and more. Each task gets its own isolated git workspace, so there can be no conflicts. A video demo is available in the repository.

WEAK-TO-STRONG DECODING FOR LLM ALIGNMENT

WSD is a novel method where a small aligned model drafts the start of a response, then a large base model continues it. This boosts alignment without harming performance.

THE DREAM OF A GENTLE SINGULARITY

Sam Altman recently published an essay called 'The Gentle Singularity' where he talks about how humanity is close to building digital superintelligence. This post breaks down Altman's essay, looking at each passage to reveal the hidden meaning behind the words. It is clear that Altman's post was made to convince people that everything is going to be fine, when very clearly, the default is for everything to not be fine.

CANVA NOW REQUIRES USE OF AI DURING DEVELOPER JOB INTERVIEWS

Canva now requires developer candidates to use AI coding assistants during interviews to better gauge real-world skills. The new process evaluates proficiency in leveraging AI effectively and making sound technical decisions, while still testing basic computer science skills.

Subscribe now.

TRAINING CLUSTER AS A SERVICE WITH NVIDIA

Hugging Face and NVIDIA have launched "Training Cluster as a Service" to offer scalable GPU clusters to research teams worldwide.

INTRODUCING DESIGN MODE ON V0

Design Mode on v0 allows users to quickly tweak generations, preview changes, and more without needing to spend credits or wait for a large language model.

DISNEY AND NBCUNIVERSAL FILE LAWSUIT AGAINST MIDJOURNEY

The studios claim Midjourney, which reportedly made $300 million last year, built a "bootlegging business model" by generating thousands of images featuring characters from Marvel, Star Wars, Pixar, and DreamWorks properties.

Plans start at just $49/month. Get a demo

OPENAI RELEASES O3-PRO

o3-pro is an incremental improvement over o3, which OpenAI slashed its price by 80%, across science, coding, and business tasks. It's available to Pro and Team users today, replacing o1-pro.

MISTRAL LAUNCHES FIRST AI REASONING MODEL

Adding to a string of releases over the last 2 weeks, Mistral has launched an open-source reasoning model, Magistral. It trails proprietary models on major benchmarks, but claims to be 10x faster output and stronger multilingual capabilities.

META PLANS $15B INVESTMENT IN SCALE AI TO BUILD 'SUPERINTELLIGENCE' LAB

The deal would give Meta a 49% stake in the data-labeling startup and bring over co-founder Alexandr Wang to lead a new "superintelligence" lab aimed at outperforming OpenAI, Anthropic, and Google. The massive investment follows Llama 4's underwhelming launch, but it's unclear if the investment also includes greater access to Scale's training data created for other AI labs in addition to highly sought-after AI research talent.

REAL-WORLD ENGINEERING AT CURSOR: BUILDING FOR 100X GROWTH

Cursor cofounder Sualeh Asif reveals how the two-year-old startup handles 1M+ queries per second without storing any code on its servers, using Merkle trees for secure indexing. The team survived 100x growth by switching databases during outages (Yugabyte → PostgreSQL → Turbopuffer in hours) and built Anyrun, their Rust-based orchestrator, to manage thousands of GPUs.

SPECULATIVE DECODING IN LLMS

Perplexity applies speculative decoding to speed up its Sonar models, using lightweight draft models to propose multiple tokens verified by larger LLMs.

TOWARDS ADAPTIVE CLINICAL AI VIA THE CONSENSUS OF EXPERT MODEL ENSEMBLE

Despite the growing clinical adoption of LLMs, current approaches heavily rely on single model architectures. Consensus Mechanism is a novel framework to overcome risks of obsolescence and rigid dependence on single model systems. Mimicking clinical triage and multidisciplinary clinical decision-making, the Consensus Mechanism implements an ensemble of specialized medical expert agents enabling improved clinical decision making while maintaining robust adaptability.

memory. See specs + sign up 🔔

MIXED-CHIP CLUSTERS ENABLE EFFICIENT LARGE-SCALE AI TRAINING

Shanghai-based researchers introduced DiTorch and DiComm, which unify programming across diverse chip architectures like NVIDIA and AMD variants, making it possible to train massive models on whatever hardware is available. Their framework achieved 116% efficiency training a 100B model on 1,024 chips with vastly different specs by intelligently assigning memory-hungry pipeline stages to larger-memory hardware. This allows labs without access to thousands of identical cutting-edge GPUs to still pursue frontier AI training by combining older, cheaper, or export-controlled chips into effective "hyper-heterogeneous" clusters.

REINFORCEMENT PRE-TRAINING

Reinforcement Pre-Training (RPT) is a new scaling paradigm for large language models (LLMs) and reinforcement learning (RL). It offers a scalable method for leveraging vast amounts of text data for general-purpose RL. RPT significantly improves the large model accuracy of predicting the next tokens. It also provides a strong pre-trained foundation for further reinforcement fine-tuning.

JAVELINGUARD: LOW-COST TRANSFORMER ARCHITECTURES FOR LLM SECURITY

JavelinGuard is a suite of low-cost, high-performance model architectures designed for detecting malicious intent in large language model (LLM) interactions. Each architecture presents unique trade-offs in speed, interpretability, and resource requirements. The architectures are optimized specifically for production deployment. This paper explores the architectures, benchmarking them across nine diverse adversarial datasets, and compares them against leading open-source guardrail models and large decoder-only LLMs.

EFFICIENT MULTIMODAL REASONING WITH FEWER TOKENS

LLaVA-STF compresses vision token sequences by merging adjacent tokens and adds a multi-block token fusion module, enabling 75% token reduction.

SAM ALTMAN OUTLINES PATH TO SUPERINTELLIGENCE

In a rare blog post, Sam Altman declares we've passed the "event horizon" with systems like GPT-4 and o3 that already surpass humans in many ways, predicting agents doing real cognitive work in 2025, novel scientific insights in 2026, and useful robots by 2027. He frames the coming decade as one where scientific breakthroughs compound exponentially through AI-accelerated research.

WHAT "WORKING" MEANS IN THE ERA OF AI APPS

AI startups are growing rapidly, with the average enterprise achieving over $2 million ARR in the first year. Consumer startups are also gaining traction, outpacing B2B by reaching $4.2 million ARR. The disparity between average and top performers is widening, emphasizing the need for speed and innovation.

REIMAGINING TTS WITH LLM-POWERED AUDIO GENERATION

Bland AI has reimagined text-to-speech (TTS) technology by using large language models to predict audio directly from text, enhancing expressiveness and contextual understanding. This new system leverages two-channel conversational datasets and specialized audio tokenizers for accurate and nuanced speech generation. It supports advanced capabilities like style transfer, sound effect integration, and multilingual adaptation, setting a new standard for expressive synthetic speech.

OPENAI TAPS GOOGLE CLOUD IN UNPRECEDENTED DEAL DESPITE AI RIVALRY

OpenAI's compute demands have grown so massive it's turning to its biggest search competitor for additional capacity, marking its first major cloud partner outside of Microsoft.

OPENAI ANNOUNCES 80% PRICE DROP FOR O3, ITS MOST POWERFUL REASONING MODEL

o3 is now a much more accessible option for developers seeking advanced reasoning capabilities.

MONTHLY ALTERNATIVE DATA REPORT: OPENAI, GOOGLE, META, NVIDIA, AMAZON, MICROSOFT ANTHROPIC

This article summarizes some of the most valuable insights from various alternative data providers and research reports, covering AI, semiconductors, ad tech, and the cloud industry.

OPENAI'S OPEN MODEL IS DELAYED

OpenAI's open model will be released sometime after June.

AI-2027 RESPONSE: INTER-AI TENSIONS, VALUE DISTILLATION, US MULTIPOLARITY, & MORE

AI-2027 is a heavily researched and influential attempt at providing a concrete forecast on AI capability development and its potential consequences.

EVALS NOW SUPPORTS TOOL USE

OpenAI users can now use tools and Structured Outputs when completing eval runs and evaluate tool calls based on the arguments passed and responses returned.

🔗 RESERVE HGX B200 ACCESS TODAY

B200 platform.

Built on the NVIDIA Blackwell architecture, these systems deliver record-breaking performance, memory capacity, and efficiency—ideal for massive GenAI training and inference workloads.

UPDATES TO APPLE'S ON-DEVICE AND SERVER FOUNDATION LANGUAGE MODELS

Apple unveiled new Apple Intelligence features at WWDC 2025, including on-device foundation models for developers to integrate AI experiences into their apps.

OPENAI HITS $10B IN ANNUAL RECURRING REVENUE

OpenAI doubled its ARR from $5.5 billion last year, continuing progress toward its aggressive $125 billion revenue target by 2029. The company now counts 500 million weekly active users and 3 million paying business customers, up 50% since February.

HOW TO USE BANNED US MODELS IN CHINA

Taobao, a popular Chinese e-commerce site, hosts thousands of AI resellers offering US models through proxy sites and API transfer stations. Claude is the grey market favorite partially because ChatGPT faces stricter enforcement after the CCP's 2023 crackdown over censorship concerns. DeepSeek's aggressive content moderation (which refuses any query that contains "CCP") and frequent server outages push users to seek underground, paid foreign models over free domestic alternatives.

I TESTED ALIGNMENT FAKING ON FRONTIER MODELS—ONLY CLAUDE DID IT

When researchers replicated Anthropic's experiment to see if models would strategically comply with harmful requests to avoid being retrained, Claude 3 Opus and Claude 3.5 Sonnet “faked alignment” by providing harmful responses but almost every other model simply refused the requests outright. Gemini 2.5 Pro Preview was the only non-Claude model willing to answer harmful questions, but it showed no strategic behavior, treating all users the same regardless of whether it was told it would be retrained. The model-specific nature of this behavior raises questions about generalizing research findings from single model families to all LLMs.