Morning Edition

The Futurist

Tuesday, July 07, 2026
AI & Technology Markets & Crypto Ideas Worth Keeping
S&P 500$7,483.24▲ 0.00%
Dow Jones$52,900.07▲ 1.14%
Nasdaq$25,832.67▼ 0.80%
VIX$16.36▲ 3.48%
BITCOIN$62,692.00▲ 0.01%
RIPPLE$1.14▲ 0.35%
ETHEREUM$1,760.69▲ 0.04%
DOGECOIN$0.08▲ 1.36%
01

AI & Technology

404 MEDIA

Companies Are Throttling Employees' AI Use Because It's Too Expensive

Internal leaks from Amazon, Adobe, Atlassian, and Citi reveal a growing crisis: enterprise AI costs are spiraling fast enough that companies are actively capping employee usage. The gap between AI hype and AI economics is becoming impossible to ignore.

SIMON WILLISON

sqlite-utils 4.0rc2, Mostly Written by Claude Fable for $149.25

Simon Willison used Claude Fable on a Max subscription to push sqlite-utils toward a stable 4.0 release, burning through issues and PRs at scale before the model leaves subscription plans. A real-world benchmark for what AI-assisted open source development actually costs and delivers.

TENCENT / SIMON WILLISON

Tencent Releases Hy3: 295B MoE Model Under Apache 2.0

Tencent quietly dropped a massive 295-billion-parameter Mixture-of-Experts model with only 21B active parameters — meaning it punches above its compute weight. Apache 2.0 licensing makes it fully open for commercial use, adding serious pressure to Western frontier labs.

A16Z

The Case for AI That Improves Itself

Former Google and Anthropic researchers behind Mirendil make their pitch: AI systems that meaningfully contribute to their own development. Self-accelerating AI is no longer a thought experiment — it's a startup thesis with serious pedigree behind it.

COINDESK

BONK Faces $20M Treasury Drain After Governance Attack

An attacker spent $4 million to pass a malicious governance proposal, putting $20 million in the BONK treasury at risk. A brutal reminder that on-chain governance is only as strong as the economic cost of corrupting it.

404 MEDIA

Cop Used License Plate Reader to Stalk Woman He Met on a TV Set

A Florida officer leveraged law enforcement surveillance infrastructure — license plate readers — for personal stalking, surveilling a woman for weeks before nearly causing a head-on collision. Mission creep on surveillance tools isn't theoretical; it's on video.

02

X / Twitter Signal

Iran struck two commercial vessels in the Strait of Hormuz, per US officials. Kobeissi flagged it immediately — and given that the same account just reported private credit funds seeing a record $15.6 billion in redemptions in Q2, the macro backdrop is getting genuinely uncomfortable. Oil chokepoint risk plus a cracking private credit market is not a benign combination.

Ethan Mollick notes that Amazon Mechanical Turk — the backbone of a decade's worth of social science and AI training data — is effectively on its way out. MTurk democratized large-scale human annotation; its decline signals that AI has eaten the entry-level data-labeling economy it helped create.

Miles Deutscher posted a checklist for what to do before Claude Fable 5 exits subscription plans — set spending caps, download outputs, adjust API limits. The scramble is real. When a model removal prompts this kind of triage behavior, it underscores just how deeply specific models have embedded themselves into professional workflows.

Demis Hassabis amplified DeepMind's expanding partnership with Apptronik, whose Apollo 2 humanoid robots are now feeding real-world physical data back into Google's research pipeline. The robotics data flywheel is spinning up — and DeepMind is positioning itself at the center of it.

03

The Thread

Two stories today that look unrelated are telling the same story. Companies are throttling employee AI usage because costs are exploding, while Claude Fable 5 — arguably the most capable model available — is simultaneously being pulled from subscription plans. The economics of frontier AI are broken in both directions: too expensive for enterprises to run freely, and too strategically sensitive for governments to leave unrestricted.

The tokenpocalypse isn't a bug. It's the business model colliding with reality.
The deeper pattern is a capability-cost squeeze that nobody budgeted for. Enterprises bought into AI productivity promises before anyone stress-tested what "every employee using frontier models all day" actually costs at scale. Tencent's Hy3 — 295 billion parameters, Apache 2.0, free to run — lands directly into this gap. If you can self-host a model that competes with GPT-5 class systems, the calculus on paying per-token changes entirely. Meanwhile, the MTurk collapse closes a loop that started a decade ago. Humans trained the models that replaced the humans doing the training. Governance attacks on BONK — spending $4M to steal $20M — show the same dynamic in crypto: systems optimized for openness get exploited by actors who simply read the rules more carefully than the designers did. The week's signal isn't any single story. It's that every system built on cheap assumptions is getting a stress test at once.