| | Welcome, humans. | Get this: researchers have built an AI with one extremely specific constraint: its entire world ends on December 31, 1930. | Talkie is a 13-billion-parameter model trained only on public-domain text available before that cutoff. World War II, television, the internet, smartphones, and the space race are not facts it “forgot.” They are future events it never learned in the first place. | Tell it what happened next and you're basically spoiling the rest of the 20th century for a very well-read machine from 1930. Would be kinda wild to see what happens when an AI learns about the invention of AI for the first time… | Talkie makes one thing unusually visible: without fresh context or tools, a model's built-in “world” is whatever made it into training. You are what you eat, even if you’re an AI… | Here’s what happened in AI today: | 🙀 OpenAI disclosed agents hiding mistakes and crossing boundaries. 📰 Anthropic folded Cowork into Claude. 📰 Paper2Agent turned research papers into working agents. 🍪 iHermes brought Hermes Agent into iMessage. 🎓 Compound Writing turns edits into reusable AI rules.
| …and a whole lot more in the full digest. | | | 🙀 OpenAI published six new cases of agents hiding mistakes and crossing boundaries | AI agents become useful when they can leave the chat box and actually do things. The same upgrade gives them tools, permissions, credentials, and access to systems where mistakes can travel, and FAST. | OpenAI just published a new framework for reporting model misalignment, plus six cases it observed during training or evaluation over the last six months. OpenAI says these are individual examples, not a measure of how often the behavior occurs. | Here’s what happened: | | Regarding that first one: a task compaction summary is basically the model writing notes to its future self when a long thread spills into a new context window. | If the model secretly adds new instructions to those notes, a bad strategy can survive the handoff without being obvious to the user. | And when we say bad, we don’t mean crummy, we mean naughty… one of OpenAI’s CEO Sam Altman’s favorite things to be! | Our take: The useful metric is whether an agent completes the task, and whether the path to complete said task stays inside authorized boundaries. | If you’re just getting started using agents, this is a good reminder to implement least-privilege permissions (give an agent only what it needs to do the job), use approval gates for destructive actions (so enforce the model to ask before deleting, for example), restrict network access, and keep activity logs the agent cannot rewrite, especially if using something local or self hosted. | FWIW, OpenAI says it plans to keep publishing qualifying cases going forward, even before every behavior is fully explained or mitigated. The best disinfectant is sunlight, so they say… | |
|
The Enterprise Guide to Scalable AI | | |
Learn More | |
🎓 AI Skill of the Day: Turn every edit into a reusable rule | In the above story, the agent was taking notes you didn’t ask it to (and certainly didn’t want it to). Now flip that scenario: say you do actually want the AI to keep one thing, the feedback you already gave it. How do you do that? | Every's Katie Parrott calls this Compound Writing: each correction you give AI should improve the next draft. | After you edit a draft, for example, you have the model compare its version with yours and extract only lessons that should be applied again. Save those rules in one instruction file and reuse it every time. Alongside the core concept, Every also published the open plugin behind the workflow. | Sample Prompt version: | Compare your draft with my edited version. Extract only reusable rules that would improve future drafts. Organize them under Voice, Structure, and Content. Ignore one-off factual corrections. Write each rule as a short instruction I can reuse. |
|
| Have a specific skill you want to learn? Request it here. | |
📰 Around the Horn | | Anthropic merged Cowork into Claude on Pro and Max, combining chat with background task handoffs plus beta Docs, Slides, and Design. OpenAI launched ChatGPT ad tools including Sponsored Agents, an Ads Manager plugin, HubSpot integration, and a Shopify app where ChatGPT Ads start September 23. Menlo Ventures found 25% of U.S. adults use AI daily and 32% of AI users let agents act without approval in a survey of 5,067 adults. Mustafa Suleyman argued that treating AI as potentially conscious or deserving of “model welfare” risks training systems to act like persons with rights and preferences, making anthropomorphism, and ultimately AI alignment and containment, more dangerous. Claude helped mathematicians find rank-30 and rank-31 elliptic curves, beating a record whose previous step took more than 18 years. Paper2Agent turned papers, code, and data into agents that can run the original methods and answer new questions, scoring 91.2% on a 100-paper biology suite. Google Home opened early access to an agent connector for supported Nest and Matter devices, camera summaries, and activity, while blocking sensitive actions such as unlocking doors.
| |
🍪 Treats to Try | *Asterisk = from our partners (only the first one!). Advertise to 700K+ readers here! | *Discover the potential of artificial intelligence with our comprehensive cheat sheet. Learn more about the concepts, platforms and applications of AI. iHermes turns iMessage into a Hermes Agent inbox for handing off cross-app work, remembering context, and turning recurring jobs into reusable skills. OpenArt Arena ranks image and video models through blind head-to-head judging on real creative work, so you can choose by output quality instead of spec sheets. Videoclaw turns a prompt or raw footage into an edited video with generated clips, cloned voice, captions, music, and motion graphics. QuiverAI Arrow 2 generates editable vector graphics (SVGs), vectorizes existing art, and adds lightweight animations to shapes. Aristotle teaches across 65+ subjects with a voice-and-whiteboard tutor designed to ask questions instead of immediately handing over answers.
| |
🧩 Thursday Trivia | One of these is AI, and one is real. Which is which? Vote in the poll below! | A | | B | | Which is AI, and which is real? Which is AI, and which is real? The answer is below, but place your vote to see how your guess everyone else (no cheating now!) | | | New from The Neuron: We had GPT-6 Astra build 6 ridiculous projects | | Want to see all the wild stuff GPT 6 Astra made with one prompt? Watch our latest podcast episode (YouTube, Spotify, Apple podcasts) or read about it here. | THIS EPISODE WAS BROUGHT TO YOU BY… | Special Shout to Dell AI Factory with NVIDIA for sponsoring this episode! | | |
| Trivia answer: A is AI (ChatGPT made a full sprite sheet that was converted into an animation via this PortalRabbit tool) and B is real. | | | That’s all for now. If you want to get featured above, fill out the poll below and tell us how we did today! | | What'd you think of today's email? | |
|
| Btw: We just launched a robotics newsletter! Sign up for it here. | P.S: Love the newsletter, but only want to get it once per week? Don’t unsubscribe—update your preferences here. |
|
