OpenAI has launched ChatGPT for Financial Services, a version of ChatGPT that helps bankers research companies, build financial models and prepare client presentations. Morgan Stanley and Evercore helped shape it. It combines GPT-6 Astra with included datasets from PitchBook, Daloopa, LSEG News and Crunchbase. Teams can trace numbers back to source tables and turn their analysis into documents using their firm’s own Excel, Word and PowerPoint templates. That covers a familiar sequence in banking: find the financials, check the footnotes, update the model, put it into a deck. Connecting those steps could remove a lot of the copying and reformatting between tools. The analyst still has to defend the assumptions behind the valuation. For finance software startups, the commercial question gets harder. A bank evaluating another AI research or presentation tool can now ask how much of that work its ChatGPT workspace already handles. The data providers have a place in this arrangement: OpenAI needs their information. Software companies selling access to that information through a better interface will have more to prove.

Apple’s $1,999 iPhone Duo has a hinge made from more than 100 components. But one of its most consequential design choices is something you might miss in the trailer: both screens have the same aspect ratio. Apple unveiled its first foldable iPhone on September 9. It opens from a 5.4-inch outer screen into a 7.6-inch inner display. Matching their proportions lets content scale between them without changing shape every time you open the phone. The software gives that extra space a purpose. You can run two apps side by side, open two Safari windows and save app pairs. Think of a booking page beside your calendar, or a document beside the conversation discussing it. Apple Pencil support is coming later this year, extending that idea to handwritten notes and document markup. Netflix, Zoom and Slack are already adapting their apps to the folding design. At $1,999, Apple has to persuade buyers that those moments happen often enough to justify the price. A bigger screen earns its place when it lets you finish something you would otherwise postpone until you reach a laptop. Preorders start October 16. Sales begin October 23.

Anthropic’s new economic model shows how AI could make the US economy grow twice as fast while knowledge workers’ wages barely move. The company behind Claude released an interactive explorer of possible economies in 2030. Its “substantial” scenario assumes AI can perform half of knowledge work, much of it autonomously. Adoption takes longer, so most knowledge-work tasks are still done without AI. Even so, output rises while pay for knowledge workers stays essentially flat. The mechanism is worth understanding. Automating a task raises productivity, but it can also reduce demand for the people who performed it. Meanwhile, cheaper knowledge work can create demand elsewhere. Faster building designs and permitting could support more construction projects, increasing demand for construction workers. The extreme scenario makes that split sharper: GDP is 32.4% higher than in the model’s economy without AI, while knowledge-worker wages fall by more than 10% by 2030. It assumes rapid automation and essentially no new knowledge-work tasks for humans. Those assumptions matter. The model leaves out hyper-capable robots, which could change the outlook for physical work too. For workers, the useful question becomes: When AI makes your output cheaper, will customers buy enough more of it to increase demand for your work? Getting faster at a task and gaining bargaining power are two different outcomes.

If this is Apple’s variable-aperture hint, the backstory starts with a $720 Samsung phone from 2018. The Galaxy S9 could physically switch its camera opening between f/1.5 and f/2.4 depending on the light. Those numbers describe a trade-off you’ve probably encountered without knowing its name. A wider opening lets in more light but keeps a narrower range of distances sharp. Close it down and more of the scene comes into focus, at the cost of light. Photographers make that choice whenever they adjust aperture. You can see why it matters when photographing food up close: the middle of the plate looks sharp while the edges blur. Or when one person stands slightly behind another and only one face is properly focused. An adjustable aperture gives the camera another way to handle those situations before the image reaches its processing software. The difficult part is choosing well. Closing the aperture for sharper detail can require a slower shutter, which creates its own problem when someone moves. Apple’s opportunity would be making those decisions automatically, so the person taking the photo gets more keepers without ever learning what f/2.4 means. #AppleEvent

Meta is offering up to $130,000 for successful prompt-injection attacks affecting one Muse user. That number makes more sense when you see what its new personal agent can access: email, calendars, a browser and its own cloud computer. An agent reading your inbox can encounter instructions planted by someone else. If it follows them, access meant to help you can become a way to send information out. Meta’s design assumes the agent may come under attack. Muse runs inside an isolated environment. A separate system called Sentinel controls permissions for connected apps and outbound network requests. The agent receives substitutes for sensitive credentials; real credentials are inserted only after a request is authorized. Meta also says its email connector filters password-reset links, login links and one-time codes. Permission to read an inbox could otherwise become a route into many other accounts. For anyone evaluating personal agents, this is a useful question to ask: what can the system still prevent if the model follows a malicious instruction? Muse’s published architecture gives a concrete answer to examine. Whether those protections hold up in practice is what the bug bounty will help test.