The Big Story

Imagine one smart AI getting 48 hours and a single GPU to study smaller AIs and then teach them to be less likely to lie or go off track. That's exactly what happened in new research from Anthropic. The bigger model, called Claude, looked at common problems like deception or overly agreeable answers, suggested fixes, trained the smaller models, and checked the results. It worked without hurting how well the smaller models handled normal tasks. Claude hill-climbed safety benchmarks for common misalignments like deception or sycophancy while preserving general capabilities. Across 10 alignment failures, Claude reliably improved safety scores without degrading capabilities. Its best methods also generalized to benchmarks it hadn’t optimized on, to the Petri behavioral audit, and to models up to 4.7x larger. This matters because as AI spreads into school projects, creative work, and daily apps, people want to know the systems won't suddenly behave in weird or harmful ways. The research showed the improvements even carried over to bigger models and tests the team hadn't planned for ahead of time. For you, it means the AIs you chat with today are getting extra layers of checking so they stay useful instead of surprising you. It also shows one path toward AI that can help build safer AI in the future. Right now there isn't a public demo of this exact experiment, but you can already try Claude itself at claude.ai and notice how it handles tricky or sensitive questions compared with other chatbots. Source: anthropic.com

Explain Like I'm 14

You know how teachers give the same test to different classes so they can compare who really knows the material? AI benchmarks do something similar — they are sets of questions or tasks that every new model has to answer so researchers can see which one performs better. The tricky part is making sure the test itself stays fair. If the company that made the AI already saw the questions, it could accidentally (or on purpose) train the model to ace that specific test without being truly smarter. Google Deepmind is testing a new way to run these checks using special secure computer space that keeps the questions hidden from the model makers and the model hidden from the testers. It's like putting the test in a locked box that both sides can use but neither can peek inside until the results are ready. Cryptographic protection through Confidential Space is meant to keep Google from seeing the test questions and keep evaluators from seeing the model weights. The pilot project with the Singapore AI Safety Institute uses a Gemini Flash Lite and could set a new standard for tamper-proof AI benchmarks. This setup could become a standard way to trust future comparisons between different AIs. The first try is happening with a smaller Gemini model and an AI safety group in Singapore. When you hear people argue about which chatbot is "best," remember that trustworthy tests are what make those claims believable in the first place. Source: the-decoder.com

Cool Stuff & Try This

AI that helps regular people make real art New image, video, and music tools have reached the point where someone who has never drawn or composed before can create things that feel original. Instead of just churning out random pictures, these models are becoming actual creative partners. Aside from waiting for economic impact, we are now at the place where AI video, image & music models are good enough to be real tools for letting more people who never could produce new kinds of art. If you like making TikTok videos, writing songs, or designing posters for school projects, this is the moment to start experimenting. Go to any major image generator you already have access to (like the one inside ChatGPT or Claude) and try describing a scene from your favorite book or game in vivid detail — then ask it to change the lighting or mood and see how the result shifts. The goal is to treat it like a sketchbook that never runs out of pages. How long until we see a creativity boom among the slop flood? Is it happening? You can start small by generating one image today and then editing it step by step to match exactly what you pictured in your head. Source: x.com

ChatGPT now connects to more than one Google account You can link several Gmail and Calendar accounts at once so the AI can help with school email, personal plans, and family schedules without switching tabs. This makes it easier to ask things like "what do I have due this week across both my school and personal email?" The feature rolled out recently and works on the web and in the app. Open ChatGPT, go to settings, look for connected accounts, and add your Google logins one at a time. Try asking it to summarize emails from the last two days across the accounts you connect — it's a quick way to see the difference. You can also ask it to pull events from multiple calendars and list them in order so you never miss a deadline or practice. The update adds support for Gmail, Calendar, and more, turning the chatbot into a single place where you can manage different parts of your life. Source: uk.pcmag.com

Quick Bits

AI that plans real lab experiments Google Deepmind has expanded Co-Scientist from a hypothesis generator into a research system that's integrated into the lab. Across three disciplines, from materials synthesis to the autonomous development of a medical AI architecture, the Gemini-based multi-agent system delivered experimentally validated results. The system now plans experiments, runs lab equipment, and writes scientific papers. It already produced work that matched what human researchers later verified. This shows how AI is moving from suggesting ideas to handling full cycles of scientific work in real settings. Source: the-decoder.com

```claims []

💬 Reply to this email — Patrick reads every one.

Share: X · LinkedIn · WhatsApp

Forwarded this email? Subscribe here — it's free.