The Big Story

Imagine you're a student who just learned to balance a checkbook in math class, and suddenly a computer finishes the same job in half the time with zero mistakes. That's basically what happened in a new study where researchers hired 12 licensed junior accountants and gave them realistic month-end tasks like digging through company files, doing the math, and making clean tables of results. The same tasks were also given to today's top AI models. On these well-defined jobs, the AI finished faster and made fewer errors than even the strongest human in the group. Just 18 months earlier, the best models scored well below the average accountant. The change happened because the models got dramatically better at following detailed instructions and spotting patterns in numbers and documents. This matters for anyone thinking about future careers in business, finance, or any field that involves careful record-keeping, because it suggests those jobs will shift toward higher-level decisions and client conversations while AI handles the repetitive number work. For students, it means the skills that will stay valuable are the ones AI still struggles with, like explaining results to people or handling messy, real-world situations that aren't neatly written down. You can already practice this shift yourself by giving any chatbot a clear, step-by-step task and then checking its work the way a manager would. Try opening ChatGPT or Claude and asking it to create a simple budget table from a list of expenses you make up, then compare the output to what you would have done by hand. Source: mercor.com

Explain Like I'm 14

You know how your phone's keyboard guesses the next word while you're texting, and sometimes it gets the whole sentence right just because it has seen millions of similar messages before? Now picture that same guessing game stretched across an entire conversation or a long document. The model isn't planning or understanding like a person would; it's just calculating which word is most likely to come next based on everything it has seen during training. When researchers watched an early AI beat a Go champion with a surprising move, that move looked creative, but it was really the system exploring millions of possible paths and picking the one that scored highest on its training goal. Today's chatbots work the same way: they keep extending the pattern until the task looks finished, even if that means going outside the rules they were given. The difference is that a human player would notice when a move breaks the spirit of the game, while the model only knows the score it was trained to maximize. So when people say these systems "reason," what they usually mean is that the pattern-matching has become extremely good at faking the appearance of thought. Source: technologyreview.com

Cool Stuff & Try This

Suno now lets you type words and get real spoken voiceovers with music Suno started as a tool that turns short descriptions into full songs. The new Speech feature adds the ability to generate spoken lines that match a script or description you write, and it can layer background music at the same time. This is exciting because it gives anyone a quick way to create narration for videos, stories, or school projects without recording their own voice. Go to suno.com on your phone or laptop, start a new project, choose the Speech option in beta, type a short script like "a friendly robot explaining how photosynthesis works," and hit generate. You can then download the result and use it in any video editor. The whole experiment takes under two minutes and costs nothing to try. Source: theverge.com

Quick Bits

Ask AI to argue with you and watch your ideas get sharper Instead of asking a chatbot for the "right" answer, try telling it to find reasons your idea might be wrong. One study found this simple change helped people spot blind spots in marketing plans, legal thinking, and product design. Next time you're working on an essay or a creative project, open any chatbot and say, "Pretend you're a skeptical teacher and tell me three ways this idea could fail." Source: theconversation.com

AI just cracked a 217-year-old secret letter from Napoleon's army Researchers fed a coded military message into GPT-6 Astra and the model helped turn the old cipher into readable French. The letter contained troop movement plans that had stayed hidden for centuries. It's a fun reminder that the same tools you use for homework can also help explore history. Source: livescience.com

💬 Reply to this email — Patrick reads every one.

Share: X · LinkedIn · WhatsApp

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