Hire Globally Without Opening a Local Office
Global talent. None of the global hiring headaches.
Why build a legal entity in every country just to hire one great person?
With Oyster’s EOR, you can hire international talent faster while Oyster handles payroll, taxes, contracts, and local compliance for you.
Today In Ai
1 OpenAI launched GPT-Live and cleared GPT-5.6 for public release after weeks of government restriction.
GPT-Live is OpenAI's new family of real-time voice models that can listen and speak simultaneously, removing the turn-taking delay that has made AI voice feel stilted. It is rolling out in ChatGPT now. Listen to the new ChatGPT Voice here.
Alongside it, the GPT-5.6 model family is now publicly available after the US Commerce Department ended a weeks-long export restriction. Both releases arrive on the same day, making this OpenAI's most significant single-day launch since GPT-5.5.
2 Monogram raised $40M to build an AI with a visual interface that generates its own UI on the fly.
Instead of a text chat window, Monogram generates an interactive interface in real time to answer each specific question or request. Looking for a movie? It builds a browsable grid. Planning a trip? It creates a visual itinerary. The bet is that a purpose-built UI for each task is more natural than a conversation. See how it works or try the iOS app.
3 SpaceXAI rebranded from xAI and shipped Grok 4.5, its first frontier model since acquiring Cursor.
The company officially completed its rebrand to SpaceXAI and immediately shipped Grok 4.5, described as its strongest model yet. The release is the first joint output since the Cursor acquisition. Grok 4.5 is comparable to leading frontier models on coding and agentic tasks, and plugs directly into Microsoft Word, PowerPoint, and Excel. Watch it in action.
From The Frontier
The buried number. When Anthropic announced that Claude Cowork was going cross-device, the headline was the platform update. The more interesting detail was tucked into the announcement: 50% of all Cowork usage comes from just two categories of task. Anthropic went further and published a full analysis of 1.2 million anonymized sessions from May to explain what that actually means.
What people actually use it for. Business operations claimed 33% of all sessions. The tasks in that bucket are recognizable to anyone who has worked at a company of any size: building trackers, pulling scattered updates into a single report, writing onboarding checklists, organizing project status across teams. Content creation came second at 16%, covering decks, posts, and proposals. Together they account for half of everything people ask Claude Cowork to do.
Why those two categories. Anthropic's framing is the useful part. They call it "work around the work," the coordination and documentation layer that surrounds actual decision-making. This work exists at nearly every company, it needs to get done regularly, and it almost never belongs clearly to any single person's job description. That combination makes it a natural fit for a background AI worker: nobody is protective of it, everyone needs it done, and the cost of getting it slightly wrong is lower than for core decisions.
The repositioning that matters. Describing Claude Cowork as handling "work around the work" is a deliberate positioning choice. It frames the tool as an agentic background worker, not a replacement for the people doing the actual job. That framing is consistent with a broader pullback from doom talk across major AI labs about the role of AI in the workplace. Whether that reflects a genuine belief about AI's limits or a calculated response to enterprise buyer anxiety is worth watching as the usage data continues to accumulate.
What people are actually watching and sharing
Clone Fable before it goes paid. A viral guide walks through how to extract Fable 5's reasoning patterns and capabilities into a reusable operating manual that cheaper models can follow. The prompts are included. Fable 5 leaves free plans on July 12, so the window to run this is narrow.
Voice AI as a tutor. A viral clip demonstrates OpenAI's new voice model giving real-time pronunciation feedback to someone learning a new language. The loop of speak, hear feedback, correct, repeat is exactly what language tutoring has always needed and rarely delivered consistently.
Multiplayer games against AI, on demand. Mira is a new interactive world model from General Intuition and Kyutai, built in collaboration with Epic Games, that lets you play multiplayer video games against AI opponents without a server, a lobby, or other players. The model generates the game world and the opponents in real time.
A black hole, rendered in code. Computational astrophysicist Chi-kwan Chan is using OpenAI's Codex to produce the first video simulation of a black hole accurate enough for research publication. The clip shows what scientific visualization looks like when the bottleneck shifts from compute time to prompt quality.
How Claude Code got built. Anthropic published The Making of Claude Code, a first-person account from the team that built it. Nearly 1 million views in the first few days. Worth reading if you want to understand the design decisions behind the tool, not just how to use it.
Prompt Station
Turn any design screenshot into working frontend code
This prompt, shared by TawohAwa on X, tells Claude or ChatGPT exactly what done looks like before it writes a single line of code. Attach your design screenshot or file, describe what the product does and where it will be used, specify your target breakpoints, and the model builds clean frontend code that matches the design closely, then checks its own output against the original image before reporting back. The self-check step at the end is what separates this from a standard code request.
Here's a design [attach screenshot or file]. It's for [PRODUCT] and users will mainly [PRIMARY ACTION] on [mobile or desktop].
Build it as clean, working frontend code that matches the design closely: spacing, type, colours, and interactive states.
Done means: it renders correctly at [BREAKPOINTS], every interactive element works, and it matches the design side by side.
Compare your build against the image before reporting. List anything you couldn't reproduce and why.Replace [PRODUCT] with what you are building, for example: "a SaaS onboarding flow", "a mobile checkout screen", or "a landing page for a fitness app". Replace [PRIMARY ACTION] with the main thing users do: "sign up", "browse products", or "complete a form". Replace [BREAKPOINTS] with your target screen sizes: "375px mobile and 1280px desktop" is a good default. The self-check instruction at the end, comparing the build against the image before reporting, is the most important line. It is the difference between a model that guesses it is done and one that verifies.
Global hiring doesn't come with a playbook. Join Oyster's webinars and events to learn how leading companies are hiring, expanding, and staying compliant across borders.