The AI market is feeling pressure on several fronts today: products are becoming more personal, the web is being overrun by bots, and the hardware supply chain is hitting its limits. At the same time, the topic of biosecurity is moving noticeably closer to political reality — and unfortunately, this is not just a side note for the conference coffee table.

For you, that means the next phase of AI is not just about “better models,” but also more memory, more automation, more infrastructure stress, and more regulation. In short: less demo glitter, more system-level consequences.

🧠 ChatGPT gets a significantly better memory

OpenAI is fundamentally overhauling ChatGPT’s memory system. The new “Dreaming” system processes conversations in the background and updates user information on its own, rather than merely storing it passively. According to OpenAI, the success rate for temporal relevance rose from 52.2 to 75.1 percent. That’s a solid jump — and exactly the kind of improvement you notice in everyday use, even if it doesn’t look spectacular on stage.

Why does this matter? Personalization is a core feature of the next generation of ChatGPT. If the model remembers preferences, projects, or recurring contexts better, it becomes much more useful. At the same time, expectations around privacy, control, and transparency also increase. A memory that “keeps itself up to date” sounds convenient — until it remembers something you had long since forgotten you wanted to forget.
Source: The Decoder

🌐 AI agents are overtaking humans in web traffic

According to Cloudflare CEO Matthew Prince, more bots are now browsing the internet than humans — driven by AI agents. That’s not just a nice statistic for presentations, but a possible turning point for the web: content is increasingly no longer consumed directly by humans, but rather scraped, summarized, and further processed by systems. Cloudflare had expected this point only by the end of 2027.

The likely consequence is a “pay-to-crawl” model, meaning paid or contractually governed access for bots. For publishers, search engines, and platforms, this is a major shift: those who operate content no longer want to provide free training and response material for other companies’ assistants. For you as a user, this may mean more access restrictions, more licensing models, and more disputes over who is actually supposed to keep financing the open web.
Source: The Decoder

🔬 New approach to structure-aware equation discovery

A new package for hypothesis-driven equation discovery, PyCC.id, has appeared on arXiv. The key point: it is not just about fitting any formula from data, but about finding models that are structurally identifiable. In other words: the system should not only calculate nicely, but also be mathematically unambiguous in a meaningful way. That is crucial for differential equations, because many models can explain the data equally well.

Why is this relevant for AI? Such work shows where research in reasoning and training may be heading: away from pure pattern matching and toward systems that learn causal structure and robust parameterization. This is especially exciting for scientific applications, because LLMs and hybrid models could then not only generate text, but also help with modeling. In everyday terms, it sounds dry — until you realize that “ambiguous but fitting” in science is about as helpful as a navigation system without place names.
Source: arXiv

🧬 Tech industry calls for tougher DNA security rules

Several big names in the AI world — including Sam Altman, Dario Amodei, and Demis Hassabis — are urging the US government to screen synthetic DNA orders more strictly by law. The background: AI systems are now said to perform at times better than PhD-level virologists when it comes to lab knowledge, which could significantly increase the misuse risks in the area of bioweapons. This is one of those cases where “capability” and “safety” can no longer be discussed separately in a clean way.

Politically, this matters because voluntary standards often crumble faster than a promise on a tech stage. Mandatory screening for DNA synthesis could become a real precedent: AI safety would be linked here to classic biosecurity. For the industry, that is uncomfortable; for the public, potentially reassuring. And yes: if the industry itself is calling for stricter rules, that should at least make people pause and listen.
Source: The Decoder

🛡️ AI leaders warn of AI-enabled bioweapons

Political pressure is also growing in the US. According to an open letter to Congress, several leading AI companies are pushing for stronger safeguards against AI-aided bioweapons. The demand: close regulatory loopholes before they could, in the worst case, end in a global pandemic. This is no longer an abstract “what if” debate, but a very concrete security and legislative issue.

Its relevance is twofold: first, it shows that the biggest rivals in the industry do in fact share common interests on safety issues. Second, the letter is likely to set the tone for upcoming regulations — for example around access controls, model testing, and misuse monitoring. For the AI industry, this is a balancing act between innovation and responsibility. For policymakers, it is a test of whether they can respond to biosecurity faster than the next model generation.
Source: The Verge

⚙️ TSMC is becoming a bottleneck in the AI era

Despite expansion in the US, TSMC can barely keep up with the enormous demand for AI chips. CEO C.C. Wei said after the annual meeting that they can only support so much. That’s a polite way of saying: the world wants more GPUs than the factories can currently pump out. And this is exactly where it becomes clear that AI is not only a software revolution, but also a physical one.

This is hugely important for the market. If TSMC becomes a bottleneck, it affects not just individual model launches, but the entire supply chain of data centers, cloud providers, and chip designers. More demand with limited supply means higher prices, prioritization of large customers, and possibly delays for new projects. The real lesson: if you want to build AI, you need not only good ideas, but also a lot of silicon and even more patience.
Source: The Verge

🛒 Apple shows the sheer scale of the digital economy

Apple reports transaction volume of US$1.4 trillion for the App Store, of which around 90 percent was commission-free. For AI Radar, this is not just an App Store story, but a look at the platform logic through which AI products are sold and distributed. When digital marketplaces move enormous sums, they are also the places where new AI apps, subscriptions, and agentic services must succeed.

What is especially interesting is the scale of the distinction: much of this economic activity appears to run outside Apple’s classic commission model. That shows how platforms remain gatekeepers, yet do not necessarily take equally large cuts from every transaction. For developers, this matters because distribution channels, fees, and visibility are constantly shifting. For the AI industry, the message is: if you want to win in the app ecosystem, you need not only a good model, but also a viable business setup.
Source: TechCrunch

🛠️ Tool tip of the day

If you want to make AI development, experiments, or model workflows more productive, take a look today at a tool for prompt testing, runs, and evaluation — especially if you’re working with memory, agents, or evaluations. It not only saves time, but also prevents the classic “It worked yesterday” folklore.

---

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