The Signal — August 13, 2026
Three stories about who's really in control this week: a frontier model that shipped without anyone announcing it, an automation wave absorbing tens of millions of workers, and a platform that trained on its users for years before offering them a say.
DeepSeek shipped a frontier-scale model without telling anyone
DeepSeek's production deepseek-v4-pro endpoint quietly flipped to a new build, DeepSeek-V4-Pro-0813, on August 12. There was no blog post, no changelog entry, no press release. The only trace was a version string on the company's own API pricing page and a listing on OpenRouter. Simon Willison, trying to write it up, couldn't find an announcement page to link to and settled for the OpenRouter page instead.
The move closes a preview window that opened back in April, when the V4 series first appeared as open-weight checkpoints under an MIT license. July's Flash update had promised the official Pro release would "follow soon," and this is it. One caution worth stating plainly: the parameter figures floating around, 1.6 trillion total with roughly 49 billion active, describe the April preview checkpoint, not this build. DeepSeek has disclosed nothing about what's under the hood of 0813, so treat the benchmark and spec claims circulating on YouTube and aggregator blogs as unconfirmed. What we can say is that a frontier-scale Chinese model just graduated to production with zero official communication, which tells you something about how routine these releases have become.
Sources: DeepSeek API Docs · OpenRouter · Simon Willison
China's automation push is colliding with 53 million gig drivers
A think-tank report picked up by the Financial Times finds that more than 53 million people in China worked as food-delivery or ride-hailing drivers as of 2025, up roughly 10 million in just two years. The forces behind that jump are the ones this newsletter tracks constantly: manufacturing automation displacing factory workers, a soft market for white-collar jobs, and gig platforms sitting there ready to absorb the overflow.
The strain already shows in the numbers. Wages fell 1.8 percent for the roughly 37.2 million ride-hailing drivers even as their ranks swelled, and at least four cities including Shenzhen have warned that the ride-hailing market is saturated. This is the human cost that usually gets lost behind the automation headlines. When a factory installs robots or a company thins its engineering team, the displaced workers don't vanish, they migrate to whatever will still pay them, and right now that means driving. The catch is that each new driver splits the same pool of orders a little thinner, and the cities directing traffic are starting to say the sponge is full.
Sources: Financial Times via Techmeme · The Business Times
Twitch trained Amazon's AI for years, and only now offers an opt-out
Twitch updated its terms this week to say user-generated content "may be used for future Gen AI model improvements," and reporting indicates that content has quietly fed Amazon's AI development for years already. The actual news here is the opt-out. Streamers can now toggle off training use, a control they didn't have before.
The framing matters, though. The opt-out applies going forward and is off by default, so there is no way to claw back whatever has already been ingested. It lands as another entry in a familiar pattern: a platform sits on a large library of user content, uses it to train models, and offers consent controls only after the fact. For streamers who spent years building an archive on Twitch, the choice they're being handed now is about the future, not the past.
Sources: Ars Technica · TechCrunch
On the Editor's Desk
We looked at a cluster of arXiv benchmark papers this cycle, but each was a single preprint with no independent corroboration, so they didn't clear our two-source minimum. A batch of US AI policy items, the Trump administration's closed-door framework and the White House's plan to secure AI without new rules, were mostly single-source with no fresh hook, so they're held for now. And a few of the week's bigger stories, the reasoning-trace decryption research and Google's DeepMind leadership handoff, we already covered earlier this week and didn't want to reheat.
