Morning Edition

The Futurist

Thursday, August 27, 2026
AI & Technology Markets & Crypto Ideas Worth Keeping
S&P 500$7,677.28▲ 0.32%
Dow Jones$53,577.40▲ 0.30%
Nasdaq$26,151.30▲ 0.66%
VIX$15.70▲ 1.62%
BITCOIN$78,454.00▼ 1.68%
RIPPLE$1.42▼ 4.83%
ETHEREUM$2,449.12▼ 1.48%
DOGECOIN$0.09▼ 5.51%
01

AI & Technology

404 Media

Inside the Warehouse Where Amazon Scans and Destroys Books for AI Training

A firsthand account from inside Amazon's AI training facility VGT3, where physical books are scanned and destroyed to feed the company's models. This is the clearest window yet into the physical infrastructure of the AI data economy — and what gets sacrificed for it.

Simon Willison

Qwen3.8-Flash-Next: 125B Parameters, Only 6B Active

Alibaba's Qwen team dropped another open-weights model — a multimodal MoE architecture previewing what's coming in Qwen4. At 125B total but 6B active parameters, it punches well above its weight class in inference speed.

404 Media

Flock's CEO Is Lying to Cops About 404 Media's Reporting

Flock Safety CEO Garrett Langley told law enforcement that 404 Media's reporting on police using license plate readers to track an abortion case was "entirely false." 404 Media has the receipts proving otherwise. Surveillance infrastructure and executive dishonesty are a dangerous combination.

The Neuron Daily

OpenAI's Jalapeño Chip Posts 1,459 Tokens/Sec on GPT-OSS 120B

OpenAI's in-house chip nearly triples the throughput of existing hardware on its flagship open model. If these numbers hold at scale, the era of depending on Nvidia for AI inference economics just got a serious challenge.

a16z

The Case for AI That Improves Itself

Mirendil cofounders — ex-Google and ex-Anthropic researchers — are building systems where AI meaningfully contributes to its own development. Not a toy problem anymore: recursive self-improvement is moving from theoretical to funded.

404 Media

The Plane Spotter Who Was Allegedly Photographing for Beijing

Weiheng Zeng claimed he was just an aviation enthusiast. He allegedly received specific U.S. location assignments from a suspected Chinese government official. Open-source hobbyist culture is now a documented vector for state intelligence collection.

02

X / Twitter Signal

Nvidia added over $400 billion in market cap in a single session after posting record quarterly earnings, with the stock surging more than 8%. That's not a company — that's a country's GDP appearing overnight. The market is pricing Nvidia as the toll booth on the entire AI highway, and so far nothing has convinced it otherwise.

Ethan Mollick issued a blunt warning: organizations are not spending nearly enough on cybersecurity before open-weights "Mythos-class" models and their harnesses become widely available. When frontier-grade capability goes open, the attack surface doesn't expand linearly — it explodes. The window to prepare is closing faster than most security teams realize.

A chart making the rounds shows the vibe shift in ETF inflows in stark terms: 2020's darlings were clean energy, cloud, and emerging markets. In 2026, the top inflow sectors are all AI. Capital has fully re-oriented. This isn't a rotation — it's a consensus that took six years to calcify.

Andrew Ng flagged OpenWorker, an open-source agent that operates locally on your laptop rather than just chatting. The new version adds security and task-completion features. Local, capable, open — this is the architecture that makes enterprise AI gatekeepers nervous.

03

The Thread

The throughput war is the real AI war. OpenAI's Jalapeño chip nearly tripling inference speed on its own hardware, Qwen dropping a 125B-parameter model that runs on 6B active weights, and Nvidia posting numbers that added $400B in a day — these aren't separate stories. They're three chapters of the same arms race over who controls the cost and speed of intelligence at scale. Whoever wins inference economics wins the margin game for the next decade.
When AI wrote 1 million lines of code and then refined them into production software running on millions of machines, the question stopped being whether AI can program — and became who still needs to know that it can't.
Amazon physically destroying books to train its models is the detail that should stop people cold. The digital-physical boundary is gone. The training pipeline now runs through warehouses, through hobbyist communities, through surveillance cameras that may or may not be real. The infrastructure of AI is everywhere and increasingly invisible — which is exactly what makes the Flock CEO's behavior so alarming. Executives lying to law enforcement about their own surveillance product's role in a criminal case is what regulatory capture looks like in practice. The recursive self-improvement story out of Mirendil deserves more attention than it's getting. Two researchers who left Google and Anthropic — organizations at the absolute frontier — decided the most important problem was building AI that improves itself. That career bet tells you something about where they think the ceiling actually is. Meanwhile, Ethan Mollick's cybersecurity warning lands in this context with extra weight: the window to build defenses closes at exactly the moment the tools to breach them go open-source.