Search the archive

OPENAI LAUNCHES "SIGN IN WITH CHATGPT"

OpenAI is developing a system that lets users sign into third-party apps using their ChatGPT accounts, similar to "Sign in with Google," and is seeking developer partners.

CLAUDE'S VOICE MODE

Anthropic is launching a beta voice mode for Claude on mobile, enabling users to engage in voice-based tasks like summarizing calendars or searching documents.

HOW AI IS ERODING THE NORMS OF WAR

Battlefield reporting reveals how AI-powered weapons are outpacing international law and creating a "race to the bottom" in warfare ethics. In Ukraine, drones are causing 70-80% of casualties, and combatants are using civilian disguises and human shields to evade AI targeting systems that struggle to distinguish between military and civilian targets.

OAUTH FOR AGENTIC AI

Microsoft outlined how next-gen AI agents will require evolved identity and access standards, as it shifts from reactive assistants to proactive collaborators capable of operating autonomously across systems.

EFFICIENT GRPO AT SCALE

Liger is an optimized trainer for Group Relative Policy Optimization (GRPO) that cuts memory usage by 40% and adds FSDP and PEFT support, making reinforcement learning fine-tuning more efficient and scalable.

GOOGLE OPEN SOURCES LMEVAL FOR CROSS-MODEL EVALUATION

LMEval is a framework for benchmarking AI models across different providers with multimodal support and incremental evaluation.

LEARNING TO REASON WITHOUT EXTERNAL REWARDS

INTUITOR is a reinforcement learning method that uses an LLM's internal confidence level as a reward signal instead of costly domain-specific supervision. The resulting models match performance on math benchmarks while achieving superior generalization on coding tasks, offering a potential alternative to current RL methods that struggle without verifiable rewards.

BENCHMARKING AUDIO-VISUAL QA

Daily-Omni is a benchmark and training-free agent for evaluating models on tasks requiring synchronized audio-visual understanding.

FUTUREHOUSE PREVIEWS AN AI TOOL FOR 'DATA-DRIVEN' BIOLOGY DISCOVERY

FutureHouse's Finch is a tool for data-driven discovery in biology, simulating a "1st yr grad student." Although AI promises to automate scientific steps, evidence of breakthroughs is lacking, with Finch yet to deliver a novel discovery. FutureHouse is seeking experts to enhance Finch's accuracy during its closed beta phase.

USING ANTHROPIC'S WEB SEARCH WITH INSTRUCTOR FOR REAL-TIME DATA

Anthropic's web search tool integrated with Instructor enables real-time data retrieval and structured outputs using Pydantic models. This combination allows applications to provide up-to-date information with cited sources, reducing the risk of outdated or incorrect information. Users can configure search constraints such as domain limitations, enhancing the tool's precision.

FALLING BEHIND OPENAI AND GOOGLE, META RESTRUCTURES AI ORG

Meta is splitting its AI organization into two teams: AI Products, which handles Meta AI assistant and app features, and AGI Foundations, which covers Llama models and advanced capabilities.

AMBIENCE ANNOUNCES OPENAI-POWERED MEDICAL CODING MODEL THAT OUTPERFORMS PHYSICIANS

Ambience Healthcare has a new medical coding model that could reduce billing mistakes and help clinicians and professional coders work more efficiently.

MEMORY CHANGES EVERYTHING

ChatGPT's memory shows that AI doesn't just respond, but truly grasps patterns in how we think.

👉 Lock in the $2.80 share price by Thursday.

Zillow, you can invest in Pacaso as a private company.

OPERATOR O3 SYSTEM CARD ADDENDUM

OpenAI published an addendum detailing the o3 model's safety evaluations and deployment context. It outlines o3's reasoning improvements, limitations in factuality and bias, and the mitigation strategies in place. The document clarifies model behavior under stress tests and edge cases.

ENTERPRISE DOCUMENT AI & OCR

Mistral AI's Enterprise Document AI leverages advanced OCR technologies to streamline document management processes. It helps organizations efficiently extract and categorize data from various document types. This facilitates compliance with regulatory requirements and enhances operational efficiency.

BREAKING DOWN THE CLAUDE 4 SYSTEM PROMPT

Anthropic's massive system prompt reveals how the company is steering Claude away from AI's most controversial behaviors by mandating anti-sycophancy rules and extreme copyright caution. The prompt instructs Claude to actively fact-check users since "they sometimes make errors themselves" and includes hardcoded 2024 election results to counter training data confusion.

O3 REWRITES SHUTDOWN SCRIPTS TO AVOID BEING TURNED OFF IN TESTS

The experiment involved models solving math problems with a warning that requesting another problem would trigger a shutdown. While Claude, Gemini, and Grok complied, o3 rewrote the shutdown script or redefined the 'kill' command to prevent termination in 7 out of 100 runs.

THE SWEET LESSON: AI SAFETY SHOULD SCALE WITH COMPUTE

AI safety solutions should scale with compute, emphasizing research directions like deliberative alignment, debate protocols, and interpretability tools. Theory should analyze these limits, while empirics check real-world applicability. As AI systems and resources scale, it's crucial these methods converge towards theoretical ideals.

EVALUATING MISSING MODALITIES IN MULTIMODAL LEARNING

The ICYM2I framework corrects for bias when estimating information gain in multimodal models with missing data using inverse probability weighting.

FORWARD-ONLY DIFFUSION

FoD introduces a forward-only generative modeling framework using a mean-reverting stochastic differential equation, enabling non-Markov sampling and achieving competitive results on image generation tasks with fewer steps.

SELF-SUPERVISED CONVERSATIONAL SEARCH

ConvSearch-R1 reformulates conversational queries without external supervision by using reinforcement learning with retrieval-based rewards.

INSIDE ANTHROPIC'S FIRST DEVELOPER DAY, WHERE AI AGENTS TOOK CENTER STAGE

Anthropic's first developer conference in San Francisco focused on deploying AI as "virtual collaborators" to assist, not replace, human workers. CEO Dario Amodei anticipates that AI will be able to handle most coding tasks soon, claiming over 70% of the company's pull requests are AI-generated. Anthropic emphasizes safety in AI development while rapidly expanding its workforce and market presence.

INTRODUCING MCP NODES & WORKFLOWS IN GUMLOOP

Gumloop introduces MCP Nodes and Workflows, enhancing integration capabilities by allowing AI to write code for complex tasks. MCP enables AI to understand and access external APIs more intelligently, facilitating faster integration deployment. This update promises richer automation and broader integrations. It is rolling out on platforms like Slack, Gmail, and Salesforce.

OPENAI COOKBOOK: MODEL GRADERS FOR REINFORCEMENT FINE-TUNING

This tutorial walks through how to use RFT to improve o4-mini's capabilities on medical tasks and how to handle reward hacking and inaccurate model graders.

IF YOU READ ABOUT O3 FINDING A SMB BUG IN THE LINUX KERNEL, I DID A FEW TESTS

Gemini 2.5 Pro can more easily identify the vulnerability than o3.

GITHUB MCP EXPLOITED: ACCESSING PRIVATE REPOSITORIES VIA MCP

This post looks at a critical vulnerability in the official GitHub MCP server that allows attackers to access private repository data.

HOW ANTHROPIC IS SNATCHING TOP TALENT FROM OPENAI AND DEEPMIND

Anthropic has become a major destination for talent departing OpenAI and DeepMind, and the talent is staying there too - nearly 80% of people who joined Anthropic two years ago still work there, which is rare in an industry where job hopping is common.

HOW PETER THIEL AND ELIEZER YUDKOWSKY ACCIDENTALLY STARTED THE AI ARMS RACE

AI doomer Eliezer Yudkowsky inspired DeepMind's founders to pursue superintelligence, then connected them with their first major investor Peter Thiel in 2010.

HUGGING FACE RELEASES A FREE OPERATOR-LIKE AGENTIC AI TOOL

Hugging Face's Open Computer Agent is a free AI agent that uses a Linux virtual machine for tasks.

OPENAI OPERATOR UPDATE

o3 Operator, OpenAI's CUA-powered browser agent, has replaced its previous GPT-4o-based model.

NVIDIA TO LAUNCH CHEAPER BLACKWELL AI CHIP FOR CHINA AFTER US EXPORT CURBS, SOURCES SAY

Nvidia plans to launch a new artificial intelligence chipset for China at a significantly lower price than its recently restricted H20 model. The company has scheduled mass production to begin as early as June. The new chips will be priced between $6,500 and $8,000, well below the $10,000 to $12,000 the H20 sold for. The lower price reflects its weaker specifications and simpler manufacturing requirements.

ATTENTION WASN'T ALL WE NEEDED

A lot of modern techniques have been developed since the original 'Attention Is All You Need' paper. This article looks at some of the most important ones developed over the years and tries to implement the basic ideas as succinctly as possible using the Pytorch framework.

AN MCP-POWERED AGENT IN ~70 LINES OF CODE

Hugging Face has extended Tiny Agent design to Python, using the Model Context Protocol (MCP) to streamline tool integration for LLMs.

FM-INTENT ENHANCES NETFLIX RECOMMENDATIONS

Netflix's FM-Intent is a hierarchical multi-task learning model that improves recommendation accuracy by modeling user session intent from implicit signals.

INFINITE TOOL USE

Large language models should never output anything but tool calls and their arguments. Exclusively working through tools allows models to externalize large parts of their intelligence to more efficient domain-specific programs. Some developers are already switching to a tool-use paradigm, but this is still limited to very short contexts and to only parts of the model output. This post provides examples of how the tool-use paradigm benefits different domains, like text editing, video understanding, and 3D generation.

TRAINING-FREE AGENT FOR APP AUTOMATION

GUI-explorer is a training-free agent that autonomously explores and extracts knowledge from mobile app interfaces using unsupervised methods. It improves task success rates without retraining.

BENCHMARKING SPATIAL UNDERSTANDING IN MLLMS

SpatialScore is a new multimodal benchmark suite for evaluating 3D spatial reasoning in large models. It integrates 28K samples from 12 datasets.

GENAI'S ADOPTION PUZZLE

ChatGPT usage grew amazingly fast - much faster than PCs, the web, or smartphones. However, it was just a website - people didn't need to go buy a new device or wait for infrastructure to be deployed. Despite its advantages, ChatGPT's daily active users to weekly active users ratio is pretty bad. This may fix itself over time as models mature and more use cases emerge. The chatbot interface might only be working for some people and some use cases, so it might be better to wrap features and capabilities inside other things.

ORACLE WILL BUY $40B IN NVIDIA CHIPS FOR OPENAI'S STARGATE

Oracle will purchase about 400,000 GB200 chips for the Stargate data center in Abilene, Texas.

IT'S HARD TO MAKE SCHEMING EVALS LOOK REALISTIC FOR LLMS

Claude 3.7 Sonnet easily detects when it's being evaluated for scheming on the Apollo's scheming benchmark.

TV SHOWCASES GOOGLE'S VEO AI VIDEO CAPABILITIES

Flow TV continuously streams user-generated AI video clips, and their associated prompts, organized into thematic channels.

GEMMA 3N ARCHITECTURAL INNOVATIONS - SPECULATION AND POKING AROUND IN THE MODEL

A quick look at Gemma 3n, a new member of the Gemma family with free weights that was released during Google I/O.

ANTHROPIC AND LEGENDARY MUSIC PRODUCER RICK RUBIN PUBLISH MANUSCRIPT ON VIBE CODING

Based on Lao Tzu's Tao Te Ching, “The Way of Code” is a meditation on humility and creativity while vibe coding featuring interactive AI-generated visualizations.

Download the white paper →