Apple trains a secret China AI model, Amazon claims your Twitch clips, and Google drops Gemini 3.7 Flash

If you've ever streamed on Twitch, recorded a video, or written a book - there's a decent chance an AI company already has its eyes on your content. Three stories today all orbit the same uncomfortable question: who actually owns the data that trains these models?

Google's Gemini 3.7 Flash supports a massive context window. How large is it?

The answer is hiding near the bottom of today's newsletter... keep scrolling. 👇

Apple has reportedly developed a custom large language model specifically for its China market, built in partnership with domestic tech giant Alibaba. Reuters cited three unnamed people familiar with the arrangement, describing it as a genuine co-development effort - not just a licensing deal.

The move makes sense strategically. Apple needs a China-compatible AI backbone to power Apple Intelligence features in the world's largest smartphone market, and building entirely from scratch would have been both slow and politically fraught. Partnering with Alibaba lets Apple tap into a domestically approved model while keeping its own hardware and software layer on top.

What makes this striking is the timing. US-China tech tensions are near a peak, with export controls on chips and growing scrutiny of cross-border AI partnerships. Apple threading this needle quietly - with Reuters only now surfacing it - suggests the company knew it was walking a tightrope.

Read the full story →

Apple can't afford to miss the China AI wave, but the Alibaba partnership will draw serious regulatory attention in Washington - expect questions about what data flows where.

Twitch streamers discovered this week that Amazon updated its policies to allow the use of streamer content - videos, clips, chat logs - to train its AI models. The default is opt-in, meaning your content is being used unless you actively go into settings and turn it off.

The backlash was swift. Thousands of users flooded the Twitch subreddit questioning why they weren't asked first, and the discovery prompted Wired to publish a step-by-step guide on opting out. The core frustration is familiar: a platform quietly shifts its terms, benefits from its creator community's labor, and makes the opt-out obscure enough that most people will never find it.

This is part of a broader pattern we're watching across every major platform right now. Whether it's Twitch clips, YouTube videos, or even secondhand books (more on that in a moment), AI companies are aggressively sourcing training data from existing creative work - and the people who made that work are getting nothing in return. If you care about data privacy, this is exactly the kind of story worth following.

Read the full story →

Go to your Twitch settings right now and opt out - it takes 30 seconds, and you almost certainly didn't agree to this knowingly.

Secondhand booksellers across the UK and Ireland are reporting a surge of unusual bulk orders from mystery buyers based in the US, Canada, and continental Europe. The Guardian spoke to multiple bookshop owners who described orders that were large, eclectic, and placed with no apparent regard for condition or topic - exactly the kind of buying pattern that makes sense if you're scanning books for training data rather than reading them.

The suspicion isn't unfounded. Anthropic was previously reported to have spent millions acquiring physical books to scan for AI training purposes - a practice that has drawn legal scrutiny around AI copyright and fair use. The secondhand market is particularly attractive because the books are cheap, out of copyright, and often unavailable in digital form.

For independent booksellers, the orders are actually welcome revenue. But the broader implication is unsettling - that even the dusty, analogue corners of human knowledge are now being systematically converted into AI training sets. Booksellers have no way of knowing who the end buyer is or what the content will be used for.

Read the full story →

The AI data hunger has now reached physical bookshops - and there's essentially no regulation stopping labs from buying and scanning out-of-print books at scale.

Google released Gemini 3.7 Flash this week, a refinement of Gemini 3.6 Flash with targeted improvements to its reasoning core. The headline numbers are impressive: 43.6% on FrontierCode 1.1 Main (up from 34.4%), 65.3% on DeepSWE v1.1, and 1588 Elo on WebDev Arena. That last number puts it firmly in contention for the best coding-focused Flash-tier model available.

The model handles text, images, audio, and video across a 1 million token context window with 64K token output, and supports customizable thinking configurations - meaning you can tune how much reasoning budget it spends on a given task. At $0.75 per million input tokens, it's priced to compete hard against other mid-tier models.

If you're building anything that touches code generation, web dev automation, or agentic workflows, Gemini 3.7 Flash is worth a serious look. The combination of a 1M context window and strong coding benchmark performance at this price point is genuinely rare. Pair it with a fast site builder like 60sec.site - an AI-powered website builder that can spin up a full site in under a minute - and you have a surprisingly capable development stack without breaking the budget.

Read the full story →

Gemini 3.7 Flash is the clearest sign yet that Google is winning the price-performance race at the mid-tier - and developers building coding agents have a genuinely compelling new option today.

Meta this week released Glimmer, an open-weight AI model that anyone can download and run on their own hardware. It arrived alongside a 6,500-word essay from Mark Zuckerberg arguing that AI should be 'for everyone' rather than controlled by a handful of closed labs - a clear shot at OpenAI and Anthropic.

The contrast with Muse Spark, Meta's more powerful model that stays locked behind its own APIs, wasn't lost on critics. TechCrunch's Equity podcast unpacked the contradiction: Zuckerberg talks about democratization, but Meta's best model isn't open. Glimmer is the accessible entry point; the frontier capability stays proprietary. That's not unique to Meta - it's how most labs actually operate - but it complicates the 'AI is for everyone' framing.

Read the full story →

Glimmer is a real, usable open-weight model worth experimenting with - but the 'AI for everyone' narrative gets complicated fast when the best models are still API-only.

The answer is 1 million tokens! Gemini 3.7 Flash supports a 1M-token context window with 64K token output - one of the largest available at its price point. If you want to understand what that actually costs in practice, our Token Calculator can help you run the numbers.

The Amazon/Twitch story got me thinking: have you ever actively opted out of a platform using your content to train AI? Or do you tend to just accept the defaults? Hit reply and let me know - I read every response, and I'm genuinely curious how many people are actually checking these settings.

That's all for today - see you tomorrow with more. For the full archive of everything we've covered, head to Daily Inference.