AI breaks open CTF contests & Lightweight LLM memory without context - Hacker News (May 16, 2026)

Let’s start with AI and how it’s reshaping two very different worlds: competitive security and long-running assistants. A longtime CTF competitor, Kabir Acharya, argues that open online Capture The Flag competitions are effectively “broken” by frontier AI. The claim isn’t just that models can solve puzzles—it’s that agentic tooling can industrialize the process. In that world, public scoreboards stop reflecting who learned the most, and start reflecting who automated the most. If that’s accurate, it matters beyond bragging rights: CTFs have been a training ground and a recruiting signal, and the community may need new formats that still reward understanding, not just throughput.

On the more constructive side of AI news, there’s a new arXiv paper introducing something called b4-mem: a lightweight online memory mechanism meant to help LLMs retain and reuse information across long interactions—without simply extending the context window. The key idea is pragmatic: keep the underlying model frozen, then add a compact memory state that updates as the conversation progresses. Instead of stuffing more text into the prompt or doing expensive fine-tuning, the memory produces small, targeted adjustments to the model’s attention while it generates. The authors report meaningful gains, especially on benchmarks that stress long-term recall, while mostly keeping general capabilities intact. Why it matters: this is the kind of “cheap persistence” you’d want for assistants and agents that need to remember ongoing projects, preferences, or past decisions—without multiplying compute costs every time the chat gets longer.

Next up: cloud sovereignty, and an uncomfortable gap between policy goals and the hardware reality. One piece argues that Europe’s big investments in “sovereign cloud” programs—meant to reduce exposure to US legal reach—often focus on certifications, operators, and legal structures, while skating past a deeper dependency: the silicon. Many certified setups still rely on mainstream Intel and AMD processors that include highly privileged management subsystems running below the OS and hypervisor. The concern is simple even if the details are thorny: if the most trusted layer is outside your visibility and outside your jurisdictional control, sovereignty becomes a partial promise. The article also points to recent legal changes in the US that could expand who can be compelled in secret. Even if you think the most extreme scenarios are reserved for nation-state threats, the broader takeaway stands—trust isn’t just a cloud contract; it’s a supply chain.

Shifting gears to the open web as a library—and the difference between “free to read” and “free to share.” Charles Stross’ science-fiction novel Accelerando is available as a complete, free online edition under a Creative Commons license that allows redistribution with attribution, but not commercial reuse or modifications. Alongside the text, the page documents how the novel grew out of earlier magazine stories and then became a traditionally published book. Why this is interesting in 2026: it’s an early, durable example of hybrid publishing—print on one track, legal online circulation on another. For readers, it’s access. For authors and publishers, it’s a case study in how licensing choices can keep a work culturally alive without fully giving up control.

Staying with access: Project Gutenberg is highlighting a collection now exceeding 75,000 free eBooks, largely built from works whose US copyrights have expired. Gutenberg’s pitch remains refreshingly plain: no fees, no accounts, and common formats that work on typical devices. The bigger story is longevity—this is digital preservation as a decades-long habit, powered by volunteers and a pipeline that turns scanned pages into usable text. It also points to a growing ecosystem around the texts, including audiobooks. In a moment when information access is often mediated by subscriptions, app stores, or walled gardens, Gutenberg continues to be a reminder that “public domain” can be a living resource, not a historical footnote.

For developers, there’s an updated “Futhark by Example” page—essentially a guided tour through a data-parallel functional language aimed at high-performance array computing. Rather than selling Futhark with lofty claims, the page leans on real, commented programs that gradually move from fundamentals into practical work: performance-minded patterns, numerical tasks, and examples that feel closer to what you’d actually want on a GPU-like workload. It also highlights automatic differentiation examples, which is especially relevant right now as more domains—AI included—depend on efficient math tooling. The significance here is educational leverage: languages like this can be intimidating, and a well-curated example set can be the difference between curiosity and adoption.

Finally, a health story that’s been circulating for years, but continues to draw attention because the potential upside is so large—and the evidence is still developing. Researchers at Arizona State University report sustained improvements in autism-related symptoms in a subset of patients using a microbiota-based approach involving fecal microbiota transplants, particularly among individuals who also have significant gastrointestinal issues. They point to small studies with follow-ups suggesting symptom reductions over time, and mention progress toward more controlled trials. Why it matters—and why caution matters too: if the gut–brain connection turns into a validated clinical pathway for a specific subgroup, that could reshape treatment strategies. But the current base is still limited by study size and the need for larger, rigorous trials to confirm both efficacy and safety.