Morning Edition

The Futurist

Tuesday, August 25, 2026
AI & Technology Markets & Crypto Ideas Worth Keeping
S&P 500$7,674.37▲ 0.43%
Dow Jones$53,277.01▲ 0.98%
Nasdaq$26,180.46▲ 0.43%
VIX$15.87▲ 4.89%
BITCOIN$77,383.00▲ 1.09%
RIPPLE$1.47▼ 0.23%
ETHEREUM$2,458.55▲ 2.00%
DOGECOIN$0.09▲ 0.52%
01

AI & Technology

FINANCIAL TIMES via Simon Willison

Anthropic's Best Model Can't Find Users — Even at $65B Revenue

Anthropic's annualized revenue hit $65B in July, up from $47B in May — but its most capable model is losing ground to cheaper alternatives. The pattern is clear: capability ceiling keeps rising while users route around it on price.

404 Media

Subtlefakes: The Nonconsensual AI Image Problem Just Got Harder to Solve

Slightly altered AI-generated nonconsensual images are flooding X, engineered to slip past detection tools. The attack surface is no longer raw deepfakes — it's deliberate, incremental obfuscation.

Simon Willison

Linus Torvalds Credits AI for Cracking a Debug Session from Hell

Torvalds called an AI his partner in solving a nightmare Linux bug — while also noting it repeatedly declared the problem "impossible." The creator of the kernel using AI for grunt work is not a footnote. It's a signal.

Simon Willison via Drew Breunig

The "Fable Moment" and the End of Coasting on New Model Releases

Drew Breunig argues the arrival of a prohibitively expensive but extraordinary model (Fable) has ended the era of waiting for cheaper models to paper over your bad prompts. Engineering your AI harness now actually matters. The free ride is over.

CoinDesk

Monad Proposes Quantum-Resistant Wallet That Survives Lost Keys

Monad's proposed wallet upgrade addresses two existential crypto problems at once: lost private keys and future quantum attacks. If it works at scale, it's foundational infrastructure for the next decade of onchain activity.

a16z

Mirendil Founders Ask: What Happens When AI Improves Itself?

Former Google and Anthropic researchers launched Mirendil around one question — can AI systems meaningfully contribute to their own development? It's the self-accelerating AI thesis, now with a dedicated lab behind it.

02

X / Twitter Signal

Mollick called out a narrative overcorrection in real time: the discourse has swung from "AI will instantly transform everything" to "impacts will take many years" — and he thinks both are wrong. The truth, as usual, is messier and faster than the pessimists expect and slower than the hype cycle demands. When one of the sharpest AI observers flags a mood shift, it's worth tracking.

Andrew Ng flagged the Marin project as a "precious demonstration of openness" in model training — open code, data, and recipes included. With Anthropic chasing a $2T IPO and OpenAI's training increasingly opaque, Ng is pointing at the alternative path and calling it worth protecting.

Bitcoin crossed $80,000 for the first time since May — up 29% in seven days. Simultaneously, hedge funds are piling into bets against the US Dollar and public debt has hit 100% of GDP. These aren't disconnected data points. The macro case for hard assets is not subtle right now.

A small but important operational note from Teknium: switching models mid-session wipes your prompt cache entirely, forcing a full recompute. Most users have no idea this is happening. In a world where everyone is building agentic workflows, that's a costly and invisible mistake running at scale.

03

The Thread

The central tension this week is pricing versus capability, and it's reshaping everything. Anthropic is generating $65B in annualized revenue while its most powerful model struggles to find users. Franklin Templeton is tokenizing money market funds in Asia. Bitcoin just cleared $80K. Each of these is a story about the same thing: the infrastructure layer is maturing faster than the premium product layer can capture value.
"Prior to Fable, it felt silly to waste too much time improving your coding harness. A new model would arrive at the same price and paper over most of your problems." — Drew Breunig
Breunig's observation cuts to the real shift. For two years, the rational move was to wait — new models made old problems disappear for free. That era appears to be ending. The models that push capability frontiers now cost enough that engineering around them actually matters. Linus Torvalds using AI to crack Linux kernel bugs and then watching it declare the problem "unsolvable" three times is the perfect encapsulation: powerful, useful, unreliable, and here whether you're ready or not. On the subtlefakes story — slightly altered nonconsensual AI images designed to evade detection — the problem is that moderation infrastructure was built for a static threat. It isn't static. Every content policy is now in an arms race against tools that can iterate faster than trust and safety teams can update classifiers. The "Darth Vader" supporting surveillance cameras at a San Diego city council meeting is funny until you map it onto the same dynamic: AI-era persuasion tools showing up in civic spaces, and nobody quite has the policy framework to handle them yet.