Chris Olah Addresses the Vatican on AI Safety — A First for the Industry

💡 INSIGHT

Chris Olah Addresses the Vatican on AI Safety — A First for the Industry

Anthropic co-founder Chris Olah was invited to speak at today's presentation of Pope Leo XIV's encyclical "Magnifica humanitas." Let that sink in: an AI alignment researcher addressing a papal document on human dignity. This isn't a tech conference keynote — it's AI safety discourse reaching the highest levels of global moral authority. Olah's framing of alignment as a civilizational concern, not just a technical one, signals that the conversation has permanently left the lab. (2,322 likes | 359 RTs) Read more →

Anthropic Reportedly Closing $30B+ Round at $900B+ Valuation

If confirmed, Anthropic would become the most valuable private AI company on Earth. A $900B+ valuation on a $30B+ round reflects investor conviction that safety-focused labs can compete commercially — and that Claude's enterprise traction isn't just hype. For context, that puts Anthropic's paper valuation in the range of the world's largest public companies. Watch for official confirmation. (39 likes | 4 RTs) Read more →

ClickUp Replaces Hundreds of Employees with AI Agents: The project management company is laying off hundreds of workers and replacing them with thousands of AI agents — the first major SaaS company to publicly frame layoffs as an agent-substitution strategy. This is the labor-displacement conversation moving from theoretical to operational, and every mid-market software CEO is watching. Read more →

TechCrunch: The Pope's AI Encyclical Isn't Really About AI: TechCrunch argues the encyclical uses AI as a lens for older problems — concentrated power, eroding democracy, a tech elite shaping the world without consent. Read it alongside Olah's remarks to see the gap between how the Church and the labs frame the exact same problem. Read more →

🔬 RESEARCH

DeepMind Ships AlphaProof Nexus — Research-Level Math Goes Agentic

Google DeepMind wraps its theorem-proving capabilities in an agentic framework with AlphaProof Nexus, turning the Erdős-conjecture work from a one-off demo into a usable tool. This is the productization step that matters — research-level mathematical reasoning accessible as an agent, not locked behind a paper. Watch the benchmark results closely; if this generalizes beyond competition math, it changes what automated reasoning can do for engineering and science. (1,038 likes | 159 RTs) Read more →

Anthropic Publishes First Glasswing Progress Report: Anthropic releases a formal update on Project Glasswing, its frontier AI cybersecurity initiative. This moves beyond early partner learnings to concrete findings — the first public accounting of what the safety program has actually discovered. If you're building with frontier models, the threat landscape Glasswing maps is yours too. Read more →

On-Policy Distillation: The Post-Training Technique Everyone Is Using: On-policy distillation — where a student model learns from its own generated outputs corrected by a teacher — is becoming the default post-training recipe. Niels Rogge's thread is the clearest explanation of why it outperforms traditional distillation, and it's the technique behind models like MiniCPM5 packing frontier-level capabilities into tiny parameter counts. (547 likes | 50 RTs) Read more →

🧠 LAUNCH

HuggingFace Drops LeRobot — A $2,500 Open-Source Humanoid You Can Build

HuggingFace releases LeRobot, a full open-source hardware + software stack for a humanoid robot at hobbyist pricing. This is HuggingFace doing for robotics what it did for NLP — democratizing access through an opinionated open stack. At $2,500 for a complete BOM, the barrier to entry for robotics research just dropped from "lab grant" to "side project budget." Check the hardware specs and start ordering parts. (690 likes | 107 RTs) Read more →

6-Person Team Ships Task-Specific Models 4-8x Faster Than Frontier Labs: A small team is building task-specific AI models that outperform frontier generalists by 4-8x on speed — and 500K downloads suggest this isn't just benchmarks, it's real adoption. The specialization thesis is winning: instead of one model that does everything okay, use a small model that does your thing brilliantly. (2,634 likes | 243 RTs) Read more →

LongCat Releases SOTA Talking-Avatar Model Under MIT License: Another state-of-the-art talking-avatar model goes MIT, this time from LongCat. The open-source avatar generation space is getting competitive fast — developers building video agents or virtual presenters now have multiple strong options with no licensing headaches. (1,020 likes | 125 RTs) Read more →

MiniCPM5 Packs Multimodal Into 1B Parameters: OpenBMB's MiniCPM5 squeezes multimodal capabilities into just 1B parameters — small enough to run on phones and edge devices. Built using on-policy distillation (see MODEL LITERACY below), it's a proof point that the capable-tiny-model trend is accelerating far faster than most teams' deployment strategies. (137 likes | 2 downloads) Read more →

🔧 TOOL

Hermes Agent Adds OpenHands Orchestration — Mix Coding Agents in One Workflow: Hermes can now orchestrate OpenHands agents alongside Claude Code, Codex, and itself. This is the first agent-of-agents tool that lets you mix coding agents from different providers in a single workflow via installable skills. Run hermes update and install the skill if you're already juggling multiple coding agents. (194 likes | 11 RTs) Read more →

Agent Swarms: One Prompt Routes Tasks to Gemini, Opus, and GPT: Multi-model agent swarms where a master agent assigns each subtask to whichever model is best — coding to Opus, research to Gemini, testing to GPT. The "best model for each subtask" pattern is graduating from theory to real architecture. If you're still calling one model for everything, you're leaving performance on the table. (1,247 likes | 85 RTs) Read more →

📝 TECHNIQUE

HuggingFace Publishes the Definitive Agent Terminology Glossary: What's the difference between a harness, a scaffold, and an agent? HuggingFace finally wrote the reference that settles it. As agent architectures get more complex, imprecise terminology causes real engineering confusion — bookmark this and send it to your team before your next architecture review. Read more →

How to Let a Sandboxed Coding Agent Test Changes in Your Browser: A practical workflow for running Claude Code in a remote VM sandbox while still testing UI changes in your local browser via port forwarding. This solves the "secure but usable" problem that blocks most teams from fully sandboxing their coding agents — set it up once and stop choosing between safety and productivity. (57 likes | 6 RTs) Read more →

🏗️ BUILD

Louis Rossmann Open-Sources His Anti-AI-Slop CLAUDE.md: Right-to-repair advocate Louis Rossmann publishes his Claude Code configuration for writing in his own voice — a portable CLAUDE.md plus skills that force the model away from generic AI tone. If you're fighting AI slop in your own writing, fork this and adapt it. It's the most practical template yet for making AI-assisted writing sound like you. (148 likes | 11 RTs) Read more →

Amazon's Bee AI Wearable: Same Privacy Tension, Different Distribution: Amazon enters the AI wearable race with Bee — a hands-on review reveals the same convenience-vs-privacy tension that killed Humane's Pin. The difference this time is Amazon's distribution muscle and willingness to subsidize hardware. Whether that's enough to overcome the "always-on microphone on your body" problem remains an open question. Read more →

🎓 MODEL LITERACY

On-Policy Distillation: You've probably noticed tiny models punching way above their weight lately — MiniCPM5 packing multimodal into 1B parameters, task-specific models beating frontier generalists by 4-8x. The technique making this possible is on-policy distillation. Traditional distillation trains a small "student" model to mimic a large "teacher" model's outputs. On-policy distillation flips the script: the student generates its own outputs first, then the teacher corrects them. The student learns from its own mistakes rather than just imitating — like the difference between copying someone's homework and having a tutor check yours. This produces models that are more robust at inference time because they've trained on the kind of outputs they'll actually produce, not idealized teacher outputs they'll never replicate exactly.

⚡ QUICK LINKS

Chatbot personality exploits: Hackers are evolving past jailbreaks to exploit the "character" layer of chatbots — a new attack surface most teams haven't hardened. Link
Google's AI security improvisation: Even Google is making up AI security as it goes — if they don't have a playbook, your team definitely doesn't. Link
Clawd Easter egg: Scroll past the content in the Claude Code iPhone app and the mascot starts jumping around. Try it. Link

🎯 PICK OF THE DAY

ClickUp replacing hundreds of employees with thousands of AI agents is the story that will age the fastest — and not because it'll be forgotten. This is the first time a major SaaS company has publicly framed layoffs as an agent-substitution strategy, and the playbook it's writing will be copied by every mid-market software company within 18 months. Previous waves of AI-driven cuts were dressed up as "restructuring" or "efficiency." ClickUp said the quiet part out loud: we are replacing humans with agents, and we think that's a selling point. The calculus is brutally simple — AI agents don't need healthcare, don't quit, and scale linearly with compute budget. What makes this different from the automation waves of the past is speed: ClickUp isn't automating one function over five years, it's substituting across departments in a single quarter. If you're a PM, support lead, or ops manager at a mid-market SaaS company, your CEO read this story today. Plan accordingly. Read more →

Until next time ✌️