GameStop bids to buy eBay & ASML EUV machine as Lego - Hacker News (May 4, 2026)

Let’s start with the headline that made everyone double-take: GameStop has made an unsolicited cash-and-stock offer to buy eBay. The proposal puts a big premium on eBay’s share price, and Ryan Cohen says if the board says no, he’ll take the pitch directly to shareholders. GameStop is also promising aggressive cost cuts—particularly in sales and marketing—and says it has a major debt-financing commitment lined up. Why it matters: it’s a high-stakes bet that financial engineering and consolidation can fix two businesses facing real competitive pressure, and markets immediately began pricing in both the upside and the risk of heavy leverage.

Sticking with markets—but moving to chips—there’s a delightful contrast coming out of ASML. The company’s EUV lithography tools are among the most important, hardest-to-build machines in modern industry, enabling the leading edge of smartphone and AI chips. And yet, the most coveted ASML “product” right now is a detailed Lego replica of the EUV machine, originally built as a side project by an employee. It’s scarce, restricted internally, and that scarcity has created a resale frenzy. Beyond the novelty, it’s a reminder that the most complex technologies are often invisible to the public—and sometimes a physical model is the best way to make abstract engineering feel real.

Next, a security-and-trust story that’s depressingly familiar: Notepad++ creator Don Ho is warning about a website that claims to offer an official macOS version of Notepad++. According to Ho, it’s not affiliated with the project, misuses the trademark, and even copies personal bio details to look legitimate—enough to fool casual readers and, apparently, some tech coverage. The key point is simple: there is no official Notepad++ for macOS. Why it matters: impersonation sites don’t just confuse people; they can damage open-source reputations and become distribution points for unwanted or risky software.

On the AI front, there’s a practical lesson about image generation: models are getting better at text, but they still regularly mess up numbers, labels, and sequences—exactly the stuff you need to be correct in a board game, a UI mockup, or an infographic. One approach gaining traction is to split the work: generate a precise ‘underdrawing’ with deterministic tools—think SVG or HTML—so the text is correct, then have a multimodal model re-render it in a chosen style while preserving the layout. The significance is less about art tricks and more about reliability: it’s a workflow that treats AI as a stylist, not a typesetter, which is often the right division of labor.

Related to AI expectations versus reality, Communications of the ACM has a sobering take on huge context windows. Vendors now advertise massive token limits, but the report argues the bottleneck is frequently memory bandwidth and the sheer weight of the KV cache during inference. As the context grows, systems end up compressing, evicting, or paging data to slower memory tiers, raising latency and sometimes degrading output. On top of that, there’s the human-facing issue: attention gets spread thin, and long prompts can turn into noise—what some call ‘context rot.’ Why it matters: the next leap in AI usefulness may come less from bigger context numbers and more from smarter memory, retrieval, and hardware designed for inference-first workloads.

If you’ve been watching coding agents evolve, there’s a notable open-source project called deepclaude that aims to keep the Claude Code-style workflow—where an agent iterates through edits and tool calls—while routing the model requests to cheaper, Anthropic-compatible backends. The pitch is straightforward: agent loops can get expensive fast, so swapping the underlying model provider can dramatically lower the cost of experimentation. The bigger story is ecosystem gravity: once an API style becomes a de facto standard, a whole layer of compatibility tools emerges, and developers start treating model choice like an interchangeable component rather than a locked-in platform decision.

For the storage nerds—and anyone who likes clever measurement—there’s a write-up on inferring the hidden physical geometry of hard drives using read-only timing tests. The author disables caches, times carefully chosen sector reads, and uses those measurements to tease out things like rotation behavior, track boundaries, and even how the drive works around defects. The key takeaway is that modern drives are no longer the neat, cylinder-aligned devices older textbooks imply; firmware-controlled layouts and remapping strategies make them far more complex. Why it matters: even when hardware is a black box, performance fingerprints can reveal design choices—and that can inform everything from benchmarking methodology to forensic analysis.

In education news, NHK WORLD-JAPAN has launched a kids-focused series called “Texico” that teaches core programming concepts without requiring a computer. It’s part of a broader movement toward ‘unplugged’ computational thinking—using stories and everyday reasoning to teach analysis, abstraction, and simulation as mental tools, not just screen skills. Why it matters: access to devices is uneven, and even when kids have screens, foundational thinking can get lost in the apps. Teaching the concepts away from the keyboard can make coding feel less like a club you need gear to join—and more like a way of thinking anyone can practice.

And finally, a lighter story with a serious engineering undertone: a YouTube creator has unveiled an enormous, flyable radio-controlled Airbus A380 replica. This isn’t a desk model—it’s the kind of build that forces you to think about structure, control systems, and ground handling like a real aircraft program, just scaled down and driven by hobbyist obsession. Why it matters: the line between hobby engineering and professional-grade systems keeps blurring, and online platforms are turning niche feats into mass-audience showcases—pulling more people into the mindset of building complex things carefully.