This week, Deepseek OCR gives us a glimpse into what AI may look like in the future and why visual context compression could reset how models handle long documents. Then, we break down OpenAI’s latest acquisition and what it signals. Let’s dive in ⬇️

In today’s newsletter ↓
🧠 Deepseek OCR and the future of long context
🖥️ OpenAI buys Sky for Mac and what that means
📓 Turbo AI notetaker races to five million users
🌦️ Global weather body backs AI-powered forecasts and warnings
🎯 Weekly Challenge: Try Deepseek OCR

Deepseek is pushing a fresh idea for long context. Instead of feeding walls of text into a model, it arranges the content into a high resolution image, then uses a decoder that can read from that image. You get fewer tokens to process while keeping structure that models usually struggle to hold over many pages.

Think contracts with dense schedules, research packs with charts and footnotes, or spreadsheets pasted into a slide. When the model sees the layout and not just a stream of characters, it can keep relationships that often vanish in plain text, like which number belongs to which clause or which footnote modifies which sentence.

Every long task has a token tax. If you can send one or two visual frames instead of a pile of tokens, cost drops and speed improves. The biggest wins show up on mixed content, for example tables with small fonts and labels, price sheets with footnotes, or PDFs where text and images are tightly packed.

Even simple reads feel more stable, since the image preserves where things live on the page, which helps the model reason about context rather than juggling a flattened stream.

Keep it practical. Pick one real document you already need this week. Pass a single clear page through an OCR to image pipeline, then ask a concrete question, such as list every date and amount, or find the late delivery risk clauses and return the exact lines.

Compare the result to your normal text only method. You are looking for three things, a faster answer, fewer misses on small but important details, and references that point to the right spots on the page.

Expect encoders that protect sensitive fields, decoders that get better at numbers and names, and workflows that store compressed visual context for common references, like vendor terms or house templates. If this approach lands in everyday tools, your long context strategy shifts.

You bring large archives into reach without constant trimming, you ask better questions over structured pages, and you stop wrestling with token limits that used to block real work.

OpenAI acquired the team behind Sky, a native Mac interface that lets an assistant understand what is on your screen and act inside apps with your permission. The plan is to fold the technology and the people into ChatGPT clients so the assistant can operate more naturally on the desktop.

The focus is system level awareness, clear prompts before actions, and a smoother way to take small steps like moving data between documents or clicking through a simple workflow.

This leans into a desktop first future rather than a browser wrapper. A helpful assistant needs to know what is visible, where your cursor is, and which app holds the item you care about.

Sky was built for that kind of context. Pulling it into ChatGPT suggests deeper operating system hooks, along with clearer guardrails around what the assistant can and cannot do. You should expect more visible permission prompts, a slower and safer pace for actions, and a bias toward explain as it goes.

If the Sky approach lands inside ChatGPT on Mac, common chores get easier, like summarizing the visible section of a PDF, turning a selected table into a spreadsheet, or starting a calendar event from text on screen. Timeline and safety are the big questions.

The direction is clear, the assistant is moving closer to your apps, and you will approve each step as it happens.

📓 Turbo AI notetaker races to five million users

🍪 Mondelez turns to generative ads to cut costs

🧑‍⚖️ Leaders call to pause the superintelligence race

🌦️ Global weather body backs AI powered forecasts

🧀 Snack bag triggers false AI gun alert in school

🧰 Canva projects big time savings from agent tools

Here’s what to do:

That is it. One page, one question, one result.

That’s it for this week! Has Deepseek created yet another AI breakthrough to challenge major closed-source models? And, will OpenAI’s acquisitions result in better products or just more investment opportunities for the company? Hit reply and let us know your thoughts.

Zoe from Overclocked