Claude 4 System Prompt 💬, Operator o3 system card 📚, Mistral Document AI 📄

Anthropic's massive system prompt reveals how the company is steering Claude away from AI's most controversial behaviors ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌  ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ 

Sign Up |Advertise|View Online TLDR Together With TLDR AI 2025-05-27 He's already IPO'd once – this time's different (Sponsor) Spencer Rascoff co-founded Zillow, scaling it into a $16B real estate giant. But everyday investors couldn't invest until after the IPO, missing early gains. "I wish we had done a round accessible to retail investors prior to Zillow's IPO," Spencer later said. Now he's doing just that. Spencer teamed up with fellow Zillow exec Austin Allison to launch Pacaso. And they've already surpassed $110M in gross profit and $1B in transactions. They also recently reserved their Nasdaq ticker PCSO. But unlike Zillow, you can invest in Pacaso as a private company. 👉 Lock in the $2.80 share price by Thursday. 🚀 Headlines & Launches Operator o3 system card addendum (7 minute read) 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 (5 minute read) 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. 🧠 Deep Dives & Analysis Breaking Down the Claude 4 System Prompt (20 minute read) 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 (5 minute read) 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 (5 minute read) 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. 🧑‍💻 Engineering & Research Evaluating Missing Modalities in Multimodal Learning (18 minute read) The ICYM2I framework corrects for bias when estimating information gain in multimodal models with missing data using inverse probability weighting. Forward-Only Diffusion (4 minute read) 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 (GitHub Repo) ConvSearch-R1 reformulates conversational queries without external supervision by using reinforcement learning with retrieval-based rewards. 🎁 Miscellaneous Inside Anthropic's First Developer Day, Where AI Agents Took Center Stage (6 minute read) 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 (3 minute read) 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 (25 minute read) 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. ⚡ Quick Links If you read about o3 finding a SMB bug in the Linux Kernel, I did a few tests (1 minute read) Gemini 2.5 Pro can more easily identify the vulnerability than o3. GitHub MCP Exploited: Accessing private repositories via MCP (11 minute read) 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 (7 minute read) 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 (20 minute read) 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 (2 minute read) Hugging Face's Open Computer Agent is a free AI agent that uses a Linux virtual machine for tasks. Love TLDR? Tell your friends and get rewards! Share your referral link below with friends to get free TLDR swag! https://refer.tldr.tech/0b6a6dc1/2 Track your referrals here. Want to advertise in TLDR? 📰 If your company is interested in reaching an audience of AI professionals and decision makers, you may want to advertise with us. Want to work at TLDR? 💼 Apply here or send a friend's resume to jobs@tldr.tech and get $1k if we hire them! If you have any comments or feedback, just respond to this email! Thanks for reading, Andrew Tan, Ali Aminian, Jacob Turner & Sahil Khoja Manage your subscriptions to our other newsletters on tech, startups, and programming. Or if TLDR AI isn't for you, please unsubscribe.