2026-08-26: 6 signals to keep.

Today's signals, kept for later.

2026-08-26

Amazon is unplugging Mechanical Turk, the 21-year human marketplace that trained the AIs

Bezos hid a human inside a chess machine and called it AI. Twenty-one years later, the human is the part they’re deleting.

Amazon said it will close Mechanical Turk on 30 September 2026, ending the 2005 marketplace Jeff Bezos once branded “artificial artificial intelligence.” At its peak more than 500,000 people did penny tasks — labeling, transcription, surveys — that computers could not do. Last month Amazon already froze new customers; a 2023 study found up to 46% of remaining Turkers were using AI to complete the “human” work.

What Happened

CNBC; Quartz; Louise Matsakis on X; Amazon’s own MTurk notice.

Sources

2026-08-26

Debian is holding an eight-way vote on whether Linux should be written by humans

One of the options on the ballot is: Debian is created by humans.

Debian developers are mid-ballot on a 5,000-word General Resolution with eight options, from a Social Contract ban on any LLM-assisted contribution to “let them loose,” plus a climate-dealbreaker option. One proposal is literally titled “Debian is created by humans.” Gentoo and NetBSD already banned AI code; Linus has told kernel AI-haters to fork off. Under 350 ballots in as of the Register piece.

What Happened

The Register (26 Aug); official Debian GR page. Voting extended through 28 Aug.

Sources

2026-08-26

Apple just sold a desk that can host a frontier model

Apple’s most powerful chip ever is not for video. It’s so you can keep the model in the room.

Apple launched the M6 (first 2 nm Mac chip) in a new Mac mini and the M5 Ultra (first quad-die M-series, up to 512 GB unified memory, 1.2 TB/s) in a new Mac Studio. Apple’s own copy says the Ultra is for “running compute-intensive frontier AI models on device” and “huge LLMs with hundreds of billions of parameters entirely on device.” M6 adds a Dual 16-core Neural Engine and ~30% more peak GPU AI compute than M5.

What Happened

Apple Newsroom (25 Aug); TechCrunch; TechSpot. Huge HN (1,200+ / 783+). Preorder now, ships 22 Sept.

Sources

2026-08-26

The camera that was supposed to prove photos are real will happily swear an AI image is a photograph

We built a camera that testifies under oath. It will testify to anything.

C2PA was sold as cryptographic proof that a photo came from a real camera. A researcher showed that the Android version of that promise does not survive contact with rooted or glitched phones: a Pixel can be made to sign AI slop as an unedited camera original, and a YouTube clip can be labeled “captured with a camera.” Google paid a bounty and declined to rearchitect the stack. Independent work this week also argues the C2PA spec fails its own security goals.

What Happened

David Buchanan / retr0id, 25 Aug blog; HN (~169 pts). Pixel Camera is C2PA Assurance Level 2 — the “strongest” mobile implementation — and still signs forgeries. Google closed the report as won’t-fix / infeasible.

Sources

2026-08-26

The government wants data-center air permits to skip the public hearing

They can’t make the data center quiet. So they’re making the neighbors quiet.

EPA is moving to drop the federal rule that states must publicize and take comment on “minor source” air permits — the bucket that covers data-center diesel generators. Communities fighting AI buildout could lose the only notice they get before construction. The comment period on the proposal already closed 21 August; a decision is next. Northern Virginia’s backup generators alone are compared to a mid-sized coal plant.

What Happened

New York Times (25 Aug); Tom’s Hardware (26 Aug); 14 state AGs + 3 cities opposed. HN ~96 pts.

Sources

2026-08-26

OpenAI’s homemade chip just beat Nvidia on the metric that actually pays the bills

ChatGPT’s parent just grew its own stomach. It’s a jalapeño, and it eats less power than Nvidia.

OpenAI published first Jalapeño numbers: 1.5–1.9× more AI work per watt and 1.7–3.6× lower latency than the then-best public Nvidia Blackwell systems on GPT-OSS, DeepSeek R1, and Kimi K2.5. The ASIC was designed with Broadcom, and OpenAI says its own models helped design the chip. It is inference-only; training still leans on partners.

What Happened

Hot Chips results 25 Aug; TechCrunch; The Verge; SemiAnalysis InferenceX. HN ~520 pts. Small-volume deploy late 2026, ramp 2027.

Sources