AI Blog
· daily-digest · 5 min read

AI shops, paints, and shakes up the chip market

OpenAI launches a wallet for AI agents, Claude paints with Python, Meituan opens up LongCat-2.5 – plus chip export pressure and the first effects on the job market.

Inhaltsverzeichnis

Today is one of those days when AI doesn’t just talk big, but directly interferes with real-world processes: shopping, drawing, influencing jobs, and triggering geopolitical unease. At the same time, two new model news items show how quickly the technical foundation for agents and generative systems is evolving. In short: the AI world remains a mix of practical magic, regulatory stomach rumbling, and the occasional chaos.

🤖 OpenAI launches ChatGPT Wallet for agents

OpenAI is rolling out ChatGPT Wallet, a payment feature that allows AI agents to shop on behalf of users. Initially, the feature is only available to Pro subscribers in the US. This is an important step from “AI recommends” to “AI acts” — and that’s exactly where things get interesting.

Why does it matter? Once a model not only suggests products but actually pays for them, the boundary between assistance and autonomy shifts. Then it’s no longer just about good prompts, but about questions like: What limits apply? Who is liable for mistaken purchases? And how do you stop an agent from accidentally ordering half of humanity’s shopping cart? For the market, this is a signal: agentic AI is moving from demo topic to product with a payment flow.

🧠 LongCat-2.5: Meituan opens up an agent model

Meituan is releasing the weights of LongCat-2.5, a 1.6-trillion-parameter model for AI agents. According to the report, it was trained entirely on Chinese hardware and achieves 88.4 points on coding benchmarks such as Terminal Bench — just behind Claude Opus 5. So the model is not just large, but specifically designed for agentic tasks.

For the industry, this is interesting in two ways: first, it shows that “open weights” are continuing to gain momentum even in the high-end agent space. Second, it makes visible how strongly the hardware ecosystem is diverging regionally. Anyone who thinks of this as “a bit of open source” is massively underestimating the strategic dimension. This is not just model policy, it’s infrastructure policy with transformers.

🎨 Claude Fable 5.2 paints via Python instead of pixels

Anthropic’s Claude Fable 5.2 takes an unusual approach to image generation: the model does not create images directly as pixels, but writes Python code that places hundreds of thousands of brushstrokes on a virtual canvas. In one example for a photorealistic parking garage, around 700,000 brushstrokes are said to be involved. That’s absurd — but productively absurd.

Why is this exciting? The approach makes image generation more interpretable and potentially more controllable. Instead of a black box that “somehow” spits out an image, you get a traceable process with code as an intermediate step. That’s exciting for creative workflows, research, and monitoring, because sources of error can be analyzed more easily. At the same time, it shows: LLMs are no longer just text machines, but can become generative control centers.

🔬 Differential privacy for EEG: small, private, difficult

A new arXiv study on Differentially Private Federated Class-Incremental Learning for EEG Mental State Classification addresses a fairly specific but relevant problem: how do you train models on sensitive EEG data when only a few clients are involved? Classic federated learning with differential privacy usually works better with large numbers of participants. But with wearables and biosignals, the data situation is often small and private.

This matters beyond EEG. Here lies a core problem of many medical and personal AI applications: the nice privacy promises from ML slide decks become shaky very quickly in small real-world groups. The study is a reminder that privacy does not simply mean “turn on DP, problem solved.” Especially with personal signals, you have to systematically check how much utility is left after privacy protections are applied. A classic case of: well-intentioned does not automatically mean well-generalized.

💼 Australia’s central bank sees first AI effects on the job market

According to The Decoder, Australia’s central bank is observing the first visible traces of AI on the labor market — especially among young entrants to the workforce. Jobs in software development and customer service seem particularly affected, precisely where AI can take over many routine tasks. At the same time, the bank urges caution: not every movement in the labor market is automatically an AI effect.

Still, the finding matters. If job postings and employment among 15- to 24-year-olds come under greater pressure while older groups remain more stable, this could be an early sign of a structural shift. For companies, this means entry-level roles are changing faster than career ladders were built. For applicants, it means AI competence is becoming less of a bonus and more of a must. And yes, the phrase “entry-level jobs are the entry” sounds a lot less reassuring under these conditions.

⚙️ US Congress wants to further tighten AI chip exports

Following the Nvidia scandal, the US Congress wants to further tighten export rules for AI chips to China, reports The Verge. Democrats and Republicans are signaling support for tougher measures, and Nvidia CEO Jensen Huang and other executives are expected to testify before Congress. This is not just about individual deals, but about the question of how much computing power is allowed to circulate freely in geopolitical terms.

For the AI industry, this is a real risk. Export controls affect supply chains, revenue forecasts, and the hardware strategies of global labs. When chips become politically sensitive goods, the competition shifts from models to supply chains. That’s the point where AI news suddenly starts sounding like foreign policy — which, by now, it also is.

🧩 Nvidia, China, and the question of actual revenue

A report from The Decoder suggests that the company Megaspeed may have brokered Nvidia chip orders worth nearly $12 billion for Chinese companies over several years. At the same time, short seller Culper Research estimates Nvidia’s actual share of revenue from China at over 20 percent — far more than the officially stated 7 percent. This matters because the debate over export controls is not only political, but also financially explosive.

If those numbers are true, or even just seem plausible, then one thing becomes clear: despite all restrictions, the China market remains central to the semiconductor and AI economy. For investors, regulators, and competitors, it’s therefore not only about what is sold, but also through which intermediary stations it passes. It’s exactly these gray areas that make the chip market as nervous as a model right before a tool call.


Don’t want to miss any news? Subscribe to the newsletter


Weekly AI news highlights

No spam. No ads. Just the essentials — concisely summarized. Weekly in your inbox.