Episode Details

Today's topics:

Vintage AI trained pre-1931 - Researchers released “talkie,” a 13B vintage language model trained only on pre-1931 English, raising new questions about data provenance, anachronisms, and evaluation contamination in AI.

OpenAI cloud exclusivity ends - Microsoft and OpenAI revised their partnership to end Microsoft’s exclusive resale rights, reshaping cloud competition as AWS and others may now distribute OpenAI models.

AI data centers’ gas emissions - A WIRED review of air permits suggests behind-the-meter natural-gas plants for AI data centers could enable massive CO2 emissions, challenging Big Tech climate pledges and grid strategy.

Social Edge Paradox in AI - An essay argues heavy AI automation can weaken human apprenticeship and debate, risking creativity loss and “model collapse” as AI-generated text feeds future training data.

ASML EUV lithography chokepoint - ASML’s early bet on EUV lithography turned a Dutch spinoff into the key bottleneck for leading-edge chips, with major geopolitical implications for semiconductor supply chains.

GTFOBins and misconfig risk - GTFOBins catalogs how everyday Unix tools can be abused when systems are misconfigured, helping defenders audit sudo/SUID permissions and “living off the land” attack paths.

WebAssembly: stack vs register - A developer argues WebAssembly isn’t truly a classic stack machine in practice, which matters for how compilers express optimizations and how humans reason about hand-written Wasm.

Satellites streaking comet photos - NASA’s APOD shows Comet C/2025 R3 (PanSTARRS) obscured by satellite trails, highlighting how dense satellite constellations are increasingly affecting astronomy and astrophotography.

More RAM made Quake slower - A retro PC benchmark mystery showed adding RAM reduced Quake performance due to chipset cacheability limits—an old-school reminder that specs don’t always match real behavior.