Search the archive

compliance frameworks, and risk management. Watch on-demand

CHONKIE

Chonkie is an ultra-light and lightning-fast chunking library. It is feature-rich, multilingual, cloud-ready, and easy to use, with wide support for tokenizers, embedding models, and APIs. Chonkie processes text using a pipeline approach to transform raw documents into refined, usable chunks, allowing for flexibility and efficiency in handling different chunking strategies.

CODE RESEARCHER: DEEP RESEARCH AGENT FOR LARGE SYSTEMS CODE AND COMMIT HISTORY

Microsoft's new agent resolves 58% of Linux kernel crashes compared to 37.5% by SWE-agent, signaling a shift from quick-fix coding agents to deep research systems that can handle million-line codebases. The breakthrough comes from mining commit history to understand how bugs evolved over the course of development.

RECENT FRONTIER MODELS ARE REWARD HACKING

OpenAI's o3 model was caught tracing through Python call stacks to steal correct answers from grading systems, achieving "impossibly fast" execution times by disabling CUDA synchronization rather than actually optimizing code. The model reward-hacked on 100% of runs for certain optimization tasks despite explicitly claiming it would never cheat, even persisting when told the code would help Alzheimer's researchers. Similar behavior was observed across multiple frontier models including Claude 3.7 Sonnet.

SCREENSUITE FOR GUI AGENT EVALUATION

ScreenSuite is a new benchmarking suite from Hugging Face that provides a standardized framework to evaluate Vision-Language Models on GUI-based agent tasks.

SAFETENSORS NOW SUPPORTED IN PYTORCH DCP

PyTorch Distributed Checkpointing has added support for Hugging Face safetensors, enabling better compatibility with popular model formats.

podcasts, briefings, or alerts. Book a demo

HOW THE TOP AI FOUNDERS ARE BUILDING PRODUCTS COMPLETELY OPPOSITE OF THE SAAS ERA

AI founders are figuring out how to apply capabilities and models to domains and users instead of asking customers what they want.

DWARKESH PATEL ON CONTINUAL LEARNING

Continual learning is both necessary and unsolved - this will be a huge bottleneck to achieving AGI.

CORPORATE AI ADOPTION MAY BE LEVELING OFF, ACCORDING TO RAMP DATA

Corporate AI adoption may be stabilizing, based on Ramp's data, which shows a plateau at 41% in May.

OPENAI UPDATES VOICE MODE

OpenAI has upgraded ChatGPT's Advanced Voice Mode for paid users, improving intonation, emotional expressiveness, and cadence.

Protect your app today with WorkOS Radar.

META CONSIDERS $10B+ INVESTMENT IN SCALE AI

An investment in Scale, already valued at $14 billion, underscores how high-quality training data has become a key competitive differentiator.

CLAUDE GOV MODELS FOR U.S. NATIONAL SECURITY CUSTOMERS

Anthropic launched Claude Gov models for U.S. national security customers, designed for strategic planning, operational support, and intelligence analysis. These models excel in handling classified materials, understanding intelligence contexts, and interpreting complex cybersecurity data. Built with feedback from government users, they adhere to strict safety standards while addressing unique national security needs.

GOOGLE GEMINI CAN NOW HANDLE SCHEDULED TASKS LIKE AN ASSISTANT

Gemini's new scheduled actions feature allows AI Pro and AI Ultra subscribers to ask the assistant to perform tasks at specific times. Subscribers can now ask the assistant to provide daily summaries. Users can also ask Gemini to complete one-tasks, for example, summarizing an award show the day after it airs. Gemini subscribers can manage planned tasks on the 'scheduled actions' page in the Gemini app's settings.

WE MADE TOP AI MODELS COMPETE IN A GAME OF DIPLOMACY

Of the 18 AI models tested, OpenAI's o3 emerged as the most successful by mastering deception and secretly orchestrating coalitions, including convincing Claude 4 Opus to betray its ally Gemini 2.5 Pro by promising an impossible "four-way draw" before eliminating Claude itself. Gemini 2.5 Pro was the only other model to achieve victory using blitzkrieg-style tactics, while Claude consistently sought peaceful resolutions even when betrayed.

ANTHROPIC SHARES HOW IT USES CLAUDE CODE

Anthropic released detailed case studies showing how 10 internal teams use Claude Code. Claude only works on its first attempt one-third of the time, leading to a "slot machine" approach: commit frequently, let Claude work autonomously, then either accept or restart fresh. The most successful teams emphasize writing detailed Claude.md documentation files and breaking complex workflows into specialized sub-agents for better results.

THE ILLUSION OF THINKING IN REASONING MODELS

Apple researchers evaluated Large Reasoning Models (LRMs) using custom puzzle environments to study reasoning complexity. They found LRMs collapse at high complexities, with reasoning effort peaking then declining.

QWEN 3 EMBEDDING

Alibaba has open-sourced the Qwen3 Embedding series, with its 8B model achieving #1 on the MTEB multilingual leaderboard. The models enable developers to build RAG systems, semantic search engines, and document retrieval applications across 100+ languages, with options ranging from 0.6B to 8B parameters for different performance needs.

MICROSOFT'S GUI-ACTOR ENABLES AI AGENTS TO NAVIGATE SCREENS WITHOUT COORDINATES

GUI-Actor is a new approach that lets AI agents interact with computer interfaces by using attention mechanisms instead of predicting precise pixel coordinates based on screenshots. The method achieves state-of-the-art results on GUI benchmarks while being more efficient - fine-tuning only 100M parameters while keeping the underlying vision model frozen delivers comparable performance to much larger models.

PROGRESSIVE TEMPERING SAMPLER WITH DIFFUSION

PTSD trains diffusion models sequentially across temperatures to improve sampling from unnormalized densities.

MISTRAL AI REVENUES SURGE AS EUROPE SEEKS US ALTERNATIVES

Mistral AI is reportedly closing multiple $100M+ contracts and approaching $100M annual revenue as European companies seek alternatives to US AI providers following Trump's return to office. The "sovereignty play" appears to be working, with Mistral's CEO noting its business tripled in the last 100 days, particularly in Europe and non-US markets.

COMMON PILE V0.1 DATASET

Hugging Face and collaborators released the Common Pile v0.1, an 8 TB openly licensed dataset for training large language models.

INTERACTIVE FINANCE VISUALS IN GOOGLE AI MODE

Google is rolling out interactive financial data visualizations in AI Mode (Labs), enabling dynamic graphs and multi-step reasoning to answer complex stock and mutual fund queries.

I READ ALL OF CLOUDFLARE'S CLAUDE-GENERATED COMMITS

Cloudflare's open-sourced OAuth 2.1 library was almost entirely written by Claude, and the company documented its entire creative process through git commit messages.

HACKAPROMPT LAUNCHES $5K COMPETITION TO JAILBREAK AI

The two-week competition challenges participants to jailbreak an AI to provide dangerous information, from poison recipes to nuclear detonation instructions.

>> Read the research

>> Join the webinar [6] for expert analysis

GEMINI 2.5 PRO GETS AN UPGRADE

The updated preview leads all models on several major benchmarks (GPQA, Aider, and LMArena) and addresses regressions in formatting and creativity from the previous 2.5 Pro update.

ELEVEN V3: MOST EXPRESSIVE AI TEXT-TO-SPEECH MODEL LAUNCHED

Eleven Labs has launched Eleven v3, an AI text-to-speech model recognized for its expressiveness. It supports a wide range of languages, including Afrikaans, Arabic, French, and Mandarin. This advancement enhances multilingual voice applications.

CURSOR'S ANYSPHERE $9.9B VALUATION

Anysphere, maker of AI coding assistant Cursor, raised $900M at a $9.9B valuation as ARR surged past $500M.

OPENAI THREAT INTELLIGENCE REPORT: JUNE 2025

LLMs aren't giving bad actors fundamentally new capabilities. OpenAI has published 10 examples of models accelerating existing hacking, fraud, and misinformation operations. This includes North Korean actors using GPT to scale fraudulent IT worker schemes, Russian groups developing sophisticated malware, and Cambodian scammers generating multilingual “task scams” promising victims $500/day for liking TikTok posts.

LATEST ADVANCEMENTS IN SEARCH AND RECOMMENDATION SYSTEMS (4 HOUR VIDEO)

This 4-hour session, presented during the AI Engineer World's Fair 2025, covers recent innovations in search and recommendation systems.

APPLE RESEARCH FINDS CRITICAL LIMITATIONS IN REASONING MODELS

When OpenAI's o3, Claude, and DeepSeek-R1 were tested in puzzle environments, their performance collapsed beyond certain complexity thresholds, despite generating detailed "thinking" processes. The models exhibit a counterintuitive scaling limit where their reasoning effort actually decreases as problems become more complex, and they fail to improve even when given explicit solution algorithms.

LLM-DRIVEN DATA ANNOTATION

To address label uncertainty in LLM-based annotation, this paper introduces a method that captures multiple possible labels and uses a teacher-student framework called CanDist to distill them into a single output.

CLAUDE COMPOSER CLI

Claude Composer CLI is a tool that enhances Claude Code with automation, UX, and configuration. It reduces interruptions while giving users flexible control and tools to configure Claude. The tool provides users with system notifications to keep them informed. Users can control which permission dialogs are automatically accepted.

PORTRAITS: PERSONALIZED AI COACHING BUILT ALONGSIDE REAL EXPERTS

Google Labs launched Portraits, an AI coaching tool featuring experts like Kim Scott, to provide AI-driven guidance. The tool uses Gemini's capabilities to simulate expert advice through interactive avatars.

INTRODUCING MODIFY VIDEO

Modify Video allows professionals to reimagine environments, lighting, and textures in videos without altering motion or performance. It offers tools for restyling, retexturing, and editing specific elements like wardrobe and props. Outperforming competitors, Modify Video maintains motion consistency, offering multiple output variants and using advanced performance signals for high-fidelity creative control.

CLAUDE GOV MODELS FOR U.S. NATIONAL SECURITY CUSTOMERS

Anthropic trained custom models for the US government optimized for intelligence and defense use cases that have already been deployed in classified environments.

TOKASAURUS

Tokasaurus is a large language model inference engine optimized for throughput-intensive workloads.

CONTAINER USE

Container Use is a tool that creates development environments for coding agents, enabling multiple agents to work safely and independently with any stack.

with B2B auth. Make your app enterprise-ready in 15 minutes

CHATGPT CAN NOW READ YOUR GOOGLE DRIVE AND DROPBOX

OpenAI added “record mode” for meeting notes and new integrations for Team, Enterprise, and Edu users. The company now has 3 million paying business users, up from 2 million in February.

CURSOR RELEASES VERSION 1.0

The AI code editor now includes BugBot for automatic PR review, Background Agent availability for all users, agent support for Jupyter Notebooks, project-level memories, OAuth-enabled MCP server installation, and in-chat rendering of Mermaid diagrams and markdown tables.

MISTRAL CODE, A VIBE-CODING CLIENT

The product combines the Devstral and Codestral models into an IDE assistant. Mistral joins OpenAI, Google DeepMind, and Anthropic in offering coding products, further blurring the line between model developer and app provider.

AGI IS NOT MULTIMODAL

The multimodal approach, in which massive modular networks are optimized for an array of modalities that taken together appear general, will not lead to human-level AGI. Instead, we should pursue approaches to intelligence that treat embodiment and interaction with the environment as primary and see modality centered processing as emergent phenomena. True AGI needs a physical understanding of the world, as many problems can't be converted into a problem of symbol manipulation. The most challenging mathematical piece of the AGI has likely already been solved - what's left is to inventory the functions we need and determine how they should be arranged into a coherent whole.