Twitch will train AI on your stream unless you say no
Issue #014 · 4 min read
Twitch will train AI on your stream unless you say no
Amazon flips the default on 7 years of streamer content. Plus DeepSeek's new model lands everywhere at once.
The big story
Twitch just told its streamers their content has been fair game for Amazon's AI training for years, and the only thing changing now is that there's finally a switch to turn it off.
The setting is opt-out, not opt-in, and Twitch's own CPO said that's the point. Asked why, Mike Minton told users on a livestream: "If this was opt-in, nobody would opt in. That's honestly the answer" — via TechCrunch.
That's a company admitting, on the record, that its own users don't want this.
The new toggle lets streamers opt out going forward. It does not undo what's already happened. Ars Technica reports Twitch content "may be used for future Gen AI model improvements" — language vague enough to cover almost anything.
The Register and The Verge both frame it the same way: the default was quietly set years ago, and the opt-out is the news, not the training itself.
If you run any product with a default nobody would knowingly choose, this is a preview of what happens when someone makes you explain the default out loud.
What shipped
DeepSeek V4 Pro 0813 landed simultaneously on OpenRouter, Vercel's AI Gateway, and artificial analysis benchmarks, and it's already the top story on Hacker News with 945 points. The model is live on OpenRouter, running updated weights on Vercel's AI Gateway, and benchmarked for intelligence, performance, and price on Artificial Analysis. Simon Willison flagged it too, noting the release. If you're evaluating model costs this quarter, this is the one everyone's benchmarking against right now — worth a look before you renew any contract.
GitHub shipped Agent Plugins 1.0, letting you build one plugin and run it across VS Code, Copilot CLI, and the Copilot app. Per the changelog, it launched August 6 with AWS, Anysphere, Microsoft, OpenAI, and Vercel already on board. That's five companies agreeing on a plugin format before most PMs have even standardized their own agent tooling internally — a sign the coding-agent stack is consolidating faster than expected.
Anthropic's new watermarking is catching Claude users at work and in class, and some of them are furious about it. TechCrunch reports users complaining on social media that the watermark does exactly what it's designed to do: reveal Claude was used. This is a product decision worth watching if your own tool touches anything a user might not want traced back.
What I'd actually do this week
Check whether any product you run has a data-use default a user wouldn't knowingly pick — then ask whether you could say your reasoning out loud the way Twitch's CPO just did, and whether you'd want to.
Pull up DeepSeek V4 Pro 0813 on Artificial Analysis and compare it against whatever model you're paying for now. If you're mid-renewal, this is your leverage point.
If your team ships anything through GitHub Copilot, read the Agent Plugins 1.0 changelog before your engineers build a one-off integration you'll have to maintain later.
Reply and tell me which default in your own product you'd least want explained on a livestream.
Tools mentioned
- DeepSeek V4 Pro 0813 OpenRouter
- DeepSeek V4 Pro on AI Gateway
- Artificial Analysis: DeepSeek V4 Pro
- GitHub Agent Plugins 1.0