Amazon Fired Their AI Chief. Here's Why It Took So Long (Plus 5 Newsworthy Moments in AI This Week)

China’s push to break Western control of extreme ultraviolet (EUV) chipmaking moved closer to reality after Reuters reported a six-year, government-coordinated effort to build a domestic EUV machine—an advance that matters because EUV lithography tools are a strategic choke point for AI chip production. The effort, launched in 2019, reportedly reached a milestone where a prototype can generate EUV light, though chip manufacturing still requires additional breakthroughs, especially precision optics. The key bottleneck is lenses: the only lenses that work in these machines are described as Zeiss lenses, with tolerances so tight that stretching a Zeiss lens across North America would vary by just 0.1 millimeter. That dependency turns semiconductor tooling into a geopolitical battleground, not just a supply-chain issue.

The broader implication is that AI is increasingly framed as great-power competition. Western firms have explored blockades on Chinese chip supply, while China aims to reduce dependence on Western “silicon stacks” and the tooling ecosystem that supports them. The same logic extends into mining and rare earths, where control of inputs can become as consequential as control of manufacturing. For watchers, the next signals to track are whether industrial espionage risks rise—especially around Zeiss—and whether China reaches domestic chip prototypes or pilot production runs. At current progress rates, the transcript forecasts that meaningful domestic chip outcomes could land around 2027 or 2028.

A second major theme is the market’s shift from AI “magic” to implementation reality. Reuters’ analysis, based on interviews with CEOs, found that many companies struggled to move beyond writing, coding, and Q&A into more complex, domain-specific work. The sticking points are practical: data pipelines, business-logic encoding, and tool integrations. That gap is expected to force vendor marketing to pivot away from “it just works” promises toward more detailed reference architectures and integration commitments—because buyers are increasingly skeptical.

Search and multimodal models also show a move toward concrete workflows. EXA launched “people search,” positioning AI-powered retrieval for B2B use cases like finding accounts, experts, and candidates, while publishing benchmarks using precision, recall, and ranking quality. Privacy and data-scraping concerns are likely to follow, particularly given how the feature resembles LinkedIn-style scraping. Meta, meanwhile, introduced SAM audio, a unified multimodal audio separation model that can isolate sounds from mixed environments using text prompts and video cues. The practical test will be adoption: whether creative tools (like Adobe’s ecosystem or Final Cut Pro) and accessibility or hearing-aid workflows incorporate it.

Finally, Amazon reorganized its AI operation, consolidating leadership under Peter Dantis and bringing custom silicon and quantum computing under the same umbrella, while Roit Prasad is set to leave. The move signals urgency after lagging rivals in generative AI momentum. In robotics, Physical Intelligence reported emergent learning in its vision-language-action models (PI0, PI0.5, PI0.6): as pre-training scaled, models learned from egocentric human video without explicit instruction to imitate, and fine-tuning with human videos reportedly doubled performance on depicted tasks. The transcript frames this as a potential unlock for robot learning from human POV, with replication across other VLA architectures and faster scaling as the next watchpoints.