GPT-Image-2 appeared quietly on LMArena this week under three aliases: maskingtape-alpha, gaffertape-alpha, packingtape-alpha. All three were pulled within hours. Three simultaneous code names running in parallel is not a mistake. It is how you A/B test variants before selecting one to ship. Early users reported photorealistic portraits that passed human inspection, text rendering that works correctly inside scenes, and no yellow tint. The yellow-tint problem was one of the core limitations of GPT-Image-1. If the early reports hold, that limitation is gone.

⮕ The photorealism benchmark for image generation is about to reset. GPT-Image-1 was a meaningful step forward when it launched. GPT-Image-2 appears to close most of the remaining gaps: accurate text inside scenes, correct hands, consistent lighting. For anyone producing visual content with AI today, the use cases that were not viable before, marketing visuals with embedded copy, UI mockups that look like real screenshots, branded product photography, are about to become viable.

Buried in Copilot's terms of service, updated October 2025: "Copilot is for entertainment purposes only. It can make mistakes, and it may not work as intended. Don't rely on Copilot for important advice. Use Copilot at your own risk." A Microsoft spokesperson told PCMag the language is "legacy" and will be updated. OpenAI and xAI carry nearly identical disclaimers. These are not edge cases. They are the standard legal position of the three largest AI companies in the world.

⮕ The gap between how AI is sold and how it is legally positioned is now documented in plain language. This will not stay quiet for long. Either enterprise reliability standards will force AI companies to close this gap with real contractual guarantees, or procurement and legal teams at enterprises will start using these terms to push back on AI purchases. If you are buying AI tools for a team or recommending them to clients, this language belongs in your vendor risk assessment now.

DeepSeek confirmed its next model will run on Huawei-designed silicon with no Nvidia GPUs in the supply chain. This is the first frontier AI model built completely outside the US semiconductor stack, and the clearest sign yet that China's AI infrastructure is decoupling in practice, not just in policy.

Anthropic shipped computer use inside Claude Code this week. From the terminal, Claude can now open apps, click through UIs, use the browser, and run end-to-end tests without a human handoff at any step. It was the top launch on Product Hunt this week at 425 upvotes. OpenCode, an open-source equivalent that runs across multiple LLMs, launched in the same window, making the workflow model-agnostic and remixable.

The Japanese government set a target of 30% of the global physical AI market by 2040, driven by a labor shortage severe enough that one investor called it "industrial survival." Physical AI is moving from controlled pilots to real-world national deployment for the first time.

The Ministry of Education released a white paper approving AI tools that monitor expressions, screen for psychological issues, and build behavioral profiles across all primary and secondary schools. Target: universal deployment by 2030.

Most developers paying for API access have at least a few workflows that do not require a cloud connection: prototyping, prompt testing, processing internal documents. Gemma 4 runs locally at 51 words per second on a standard MacBook Pro and costs nothing.

Here is how to set it up:

Step 1. Download LM Studio at lmstudio.ai. Install normally on macOS.

Step 2. Search for Gemma 4 inside LM Studio and download the model. The file is approximately 18 GB. The model has 26 billion parameters but only activates 4 billion at a time, which is why it reaches 51 words per second on consumer hardware.

Step 3. Start the local server inside LM Studio. Any tool that currently calls the OpenAI or Anthropic API can point to this server instead by changing the base URL to http://localhost:1234/v1. No code rewrite required.

Step 4. Test with vision: drag a screenshot into the chat window. The model correctly describes every element on screen. API cost: zero.

Techzip note: This is not a full replacement for production API calls, but it is the right tool for any workflow that requires privacy or offline access. Because the local server is OpenAI-compatible, switching between local and cloud is one parameter change, not a new integration.

Most people saving for retirement have a number in mind. Fewer have a plan for turning that number into actual income.

The Definitive Guide to Retirement Income walks you through the questions that matter: what things will cost, where the money comes from, and how to keep your portfolio aligned with your long-term goals.

If you have $1,000,000 or more saved, download your free guide and start building a retirement income plan that holds up.

Download your free guide.

Andreessen breaks down the architecture behind modern AI agents: LLM plus bash shell plus file system plus markdown plus cron job. Nothing in that stack, except the model, is newer than 50 years old. Because the architecture uses Unix primitives, agents can swap models, migrate between machines, and rewrite their own code. He also predicts the two infrastructure problems that come next: biometric proof-of-human as Turing tests become irrelevant, and a payment layer for agents browsing on behalf of users. The x402 Foundation launched the same week.

Why it made the cut: the clearest available explanation of why agent architecture works, who will control the infrastructure underneath it, and what the two unsolved problems are.