AI Biosecurity, ChatGPT Memory and Bot Flood

AI is currently being recalibrated on two fronts: at the product layer and at the regulatory layer. Today you’ll see how memory, bot traffic, chip supply chains, and biosecurity interact — and why this doesn’t just look like “more AI,” but like a phase in which the infrastructure is starting to creak.

### 🧠 ChatGPT remembers more — and cleans up after itself

OpenAI is fundamentally redesigning ChatGPT’s memory: the new “Dreaming” system processes conversations in the background and automatically updates user information. According to OpenAI, the hit rate for time-sensitive relevance rises from 52.2 to 75.1 percent. The idea behind it is simple, but important: an assistant that knew something about you yesterday shouldn’t show up today with outdated information. Otherwise, personalization quickly becomes a polite form of hallucination.

For you, this means ChatGPT may feel more useful in the future because it can carry context forward better — for example across projects, preferences, or longer-running tasks. At the same time, expectations around transparency and control are rising. If a system adjusts memories on its own, you want to know what it remembers, why it remembers it, and when it’s wrong. That’s where the real product value lies: not just “more memory,” but a cleaner model of the user relationship. Source: The Decoder

### 🤖 More bots than humans on the web

Cloudflare CEO Matthew Prince says bot traffic has now, for the first time, surpassed human internet traffic — driven by AI agents. He had only expected this turning point at the end of 2027. The consequence doesn’t sound very romantic, but it is quite plausible: a “pay-to-crawl” system is likely coming, because website operators don’t want to feed their content to agents for free without limit. The web is becoming a little less “open” and a little more like a toll booth with an API.

Why does this matter? Because the web’s business logic is shifting. When agents fetch, train on, summarize, or further process content at scale, access becomes an economic issue. For publishers, platforms, and search engines, that means: whoever creates content wants to share in its use. For you as a user, it may mean less free access, more barriers, and more negotiations happening behind the scenes. Or, more politely put: the internet is learning how to issue invoices. Source: The Decoder

### 🧬 Tech giants call for stricter DNA security rules

Sam Altman, Dario Amodei, Demis Hassabis, and other industry figures are urging the U.S. government to make screening of synthetic DNA orders mandatory by law. The background is a growing biosecurity risk: according to the signatories, AI systems already outperform PhD virologists on laboratory questions. This is not a Hollywood plot, but an attempt to close a security gap before it becomes a real problem.

The core of the debate is uncomfortable: if AI makes knowledge about biology, pathogens, and lab processes more accessible, the barrier to misuse drops. That’s why this is not just about model boundaries, but about the chain that comes after — ordering, screening, access, and control. For regulation, this is an important stress test: is voluntary self-commitment enough, or are hard requirements needed? Unsurprisingly, the answer is likely to become political. Source: The Verge

### 🧪 AI surpasses virologists in lab knowledge

Another open letter ties directly into this topic: according to tech leaders, AI systems already surpass PhD virologists in lab knowledge. That is why synthetic DNA orders should in future be subject to mandatory review. At first glance this sounds like a niche issue, but in reality it’s a security building block with broad impact. After all, the availability of DNA synthesis is one of the points where abstract model capability can spill into the physical world.

The political message is clear: anyone using AI for highly risky biological questions must talk not only about outputs, but about access controls. This is especially important because security risks usually don’t arise when everyone has bad intentions, but when a system becomes too convenient for the wrong people. Biosecurity is therefore likely to move further up the agenda for AI regulation. Source: The Decoder

### 🧩 TSMC is sweating under AI demand

Despite expanding its U.S. factories, TSMC is struggling with the enormous demand for AI chips. CEO C.C. Wei said after the annual general meeting that the company can “only support so much.” In other words: the industry’s hunger for compute is greater than current manufacturing capacity. And that affects not only Nvidia customers, but the entire AI value chain from training to inference.

For you, this is above all a reality check: AI progress depends not only on models and benchmarks, but very simply on silicon, packaging, supply chains, and investment cycles. If a bottleneck develops at TSMC, the entire industry feels it — with effects on prices, availability, and the pace of expansion. Big ideas sometimes don’t fail because of the vision, but because a piece of substrate is missing. Source: The Verge

### 🛠️ Tool tip of the day

If you want to track AI developments, models, and product updates instead of just consuming them, a news aggregator with strong filtering is worth it. For the daily workflow, tools that combine RSS, source monitoring, and topic alerts are helpful — ideal for keeping an overview of LLMs, regulation, and AI safety. Practical for anyone who doesn’t want to treat 27 tabs every morning as a lifestyle. #

### 📊 App Store: $1.4 trillion in revenue as a power indicator

Apple reports $1.4 trillion in billings and sales for the App Store — up from $1.3 trillion the previous year. What stands out is less the sheer size than the structure: 90 percent of the revenue generated no commission for Apple. That is an important detail because it shows how strongly the App Store has become an ecosystem for digital and physical sales, beyond classic in-app purchases.

Why is this relevant for AI Radar? Because platforms are currently remeasuring their power everywhere: on the web through bot traffic, in hardware through chip bottlenecks, and in product ecosystems through app distribution. For AI products, app stores, payment channels, and discovery platforms remain central gatekeepers. Whoever wants visibility there needs reach, trust, and often enough a bit of luck as well. Source: TechCrunch

### 🔬 New training for equation discovery

PyCC.id is a new package for hypothesis-driven equation discovery with structural identifiability. Sounds cumbersome, but methodologically it’s interesting: the tool addresses a known problem of finding inverse models that explain the data equally well but are not mathematically unique. That’s where structural identifiability comes in — the question of whether a system can even be reconstructed uniquely.

For LLMs and reasoning systems, this is indirectly relevant because it shows how important clean target definitions are in complex learning problems. Not just “roughly fits,” but “is actually unique and robust.” Work like this is often inconspicuous, but it provides the foundations that allow models in science and engineering to do more than just interpolate nicely. Source: arXiv

---

Want to avoid missing any news? Subscribe to the newsletter