Vintage AI trained pre-1931 & OpenAI cloud exclusivity ends - Hacker News (Apr 28, 2026)

First up, a big shift in the AI business landscape: Microsoft and OpenAI are dialing back one of the most consequential terms of the AI boom—Microsoft’s exclusive right to sell OpenAI’s models. That exclusivity helped make Azure the default home for “the hot models,” because if you wanted OpenAI at scale, you typically came to Microsoft. Under the updated arrangement, OpenAI can pursue distribution with other cloud providers, potentially including AWS. Microsoft, meanwhile, stops paying a revenue share on OpenAI products it resells through its cloud platform. The significance here isn’t just who sells what—it’s that AI model access is becoming more like a competitive market than a single-lane highway. For buyers, that could mean more options and better pricing pressure. For cloud vendors, it’s a reminder that model partnerships are powerful—until they aren’t exclusive anymore.

Sticking with AI, one of the more unusual research releases today is a “vintage” language model called talkie. It’s a 13B-parameter model trained only on English text published before 1931, designed to simulate a pre-modern-knowledge assistant. Why is that interesting? Because modern models are soaked in the web—memes, post-2000 facts, and a lot of repeated benchmark material—so it’s getting harder to tell whether a model is reasoning or just recalling. A historical model gives researchers a cleaner lens: you can ask, “What does the model do when it can’t possibly have seen the modern answer?” The team says talkie lags behind a comparable web-trained “modern twin” on many knowledge questions, as you’d expect, but the gap narrows when you remove anachronistic prompts. It also surfaces very practical hurdles, like OCR noise and “temporal leakage,” where post-1930 facts sneak into datasets. The bigger takeaway: data curation isn’t just paperwork—it shapes what a model can be, and what we can honestly claim it learned.

Now, zooming out from models to the real-world footprint powering them: a WIRED review of state air-permit documents points to a surge in natural-gas projects tied to a small set of US data center campuses serving major AI players. The headline number is huge—permits that could allow emissions on the scale of entire countries. A key detail is the strategy: “behind-the-meter” generation. Instead of waiting years for grid interconnections, developers put gas turbines next to the data centers and largely bypass the grid. The argument is speed and predictability—AI workloads don’t like power uncertainty. The counterpoint is that this can lock in a new wave of fossil infrastructure right when many expected gas and coal to shrink. Even if permitted emissions are upper bounds, the concern from analysts is simple: data centers are steady, always-on demand, which can keep those turbines running hard. For Big Tech, it’s a credibility stress test for climate commitments. For everyone else, it’s a preview of how AI can reshape energy policy by sheer urgency.

That pairs uncomfortably well with an essay making the rounds about what it calls the “Social Edge Paradox.” The idea: today’s AI looks smart partly because it reflects a vast record of human collaboration—debate, apprenticeship, institutional review, and all the messy social friction that produces robust knowledge. But if companies deploy AI in a way that reduces those human processes—cutting entry-level roles, shrinking mentorship, automating away the back-and-forth—then you weaken the very pipeline that creates future expertise and future high-quality training data. The essay also points to research suggesting AI can lift individual productivity while making group output more uniform. In other words, you may get faster drafts but fewer distinct viewpoints. Tie that to warnings about “model collapse,” where AI-generated text floods the ecosystem and future models learn from a blur of averaged content, and you get a strategic question: are organizations optimizing for short-term efficiency at the expense of long-term originality? The practical takeaway is surprisingly managerial, not technical: if you want durable advantage, use AI to amplify high-interaction work, not replace it.

On the hardware side, there’s a detailed look at how ASML—once a scrappy Philips spinoff—became the chokepoint for manufacturing the most advanced semiconductors. The core bet was extreme ultraviolet, or EUV, lithography: turning a brutally hard physics problem into the only viable path for leading-edge chipmaking. What matters isn’t the gadgetry of EUV, impressive as it is. It’s the compounding effect of being the company that solved it while rivals exited. ASML benefited from long-running research ecosystems, deep collaboration with places like IMEC, and a critical moment in the early 2010s when EUV commercialization wobbled—only for ASML to raise money by selling stakes to the very customers who needed EUV to exist, and to pull key suppliers closer. The result is an effective monopoly that now influences the pace of compute progress and the geopolitics around who can build top-tier chips. If you’re watching export controls, supply-chain resilience, or the next leap in AI hardware, this is one of the central leverage points.

In security, a reminder that some of the most dangerous “exploits” aren’t vulnerabilities at all—they’re features used in the wrong context. GTFOBins is a curated reference of common Unix-like executables that can be abused to bypass local restrictions on misconfigured systems. The value here is defensive clarity. It’s easy to think, “This binary is harmless.” GTFOBins shows how ordinary tools can become stepping stones when sudo rules are too broad, SUID bits are set casually, or capabilities are handed out without a threat model. For blue teams, it’s a practical checklist for auditing what’s allowed and under what permissions. For attackers, it’s a playbook for “living off the land.” The important point is the same either way: configuration is security-critical, and convenience defaults can quietly become escalation paths.

For developers, there’s an interesting argument about WebAssembly: people often call Wasm a “stack machine,” but if you actually try to write Wasm by hand, that label can mislead you. The claim is that classic stack-based VMs tend to offer richer stack manipulation, so you can reuse values and restructure computations without naming temporary storage. Wasm, in practice, pushes you toward introducing locals for non-trivial reuse, which makes it feel closer to a register machine with a stack-like encoding style. Why should you care? Because the mental model affects everything from how you reason about performance to how you expect compiler optimizations to show up in the output. If your intuition about “it’s just a stack VM” is off, you’ll be surprised by what’s awkward—or what becomes more verbose—when you leave the comfort of higher-level languages.

On the science and imagery side, NASA’s Astronomy Picture of the Day featured a long-exposure shot of Comet C/2025 R3, with a dense web of satellite trails cutting across the scene. To the naked eye, satellites can look like slow-moving points near twilight. But with a multi-minute exposure, they become bright streaks that can overwhelm faint celestial targets. It’s a vivid illustration of an ongoing tension: satellite constellations are transforming connectivity on Earth while complicating observing conditions in the sky. And in this particular case, the comet itself is also hard to spot right now because it’s close to the Sun from our viewpoint—though it’s expected to be better placed for viewing soon from the Southern Hemisphere. The broader point is that “light pollution” increasingly includes objects in orbit, not just city glow.

Finally, a fun retrocomputing lesson with a real performance punchline: a Quake enthusiast upgraded a 1997-era PC to a hefty 384 MiB of RAM—because the modules were cheap—and then watched frame rates drop dramatically. After swapping parts, reinstalling software, and chasing ghosts, the fix was simply removing RAM. The culprit was an old-school limitation: the motherboard chipset could only cache a certain amount of main memory. Anything above that behaved like “uncached” RAM, slowing things down—especially because the operating system tends to allocate from higher memory addresses. It’s a great reminder that on older systems, “more” can be “less,” and that real-world constraints don’t always match what the spec sheet promised.