Morning Edition

The Futurist

Sunday, November 01, 2026
AI & Technology Markets & Crypto Ideas Worth Keeping
S&P 500$7,580.07▼ 1.83%
Dow Jones$51,770.31▼ 1.09%
Nasdaq$25,437.00▼ 2.87%
VIX$19.42▲ 12.84%
BITCOIN$77,380.00▼ 0.31%
RIPPLE$1.25▼ 0.57%
ETHEREUM$2,243.05▲ 0.66%
DOGECOIN$0.08▼ 0.52%
01

AI & Technology

SIMON WILLISON

How Researchers Hijacked ChatGPT Accounts Through a Writing Tool Exploit

Parameter Labs discovered a critical vulnerability in ChatGPT's "Writing Blocks" artifact feature that allowed full account takeover. The exploit exposes how AI interfaces that render dynamic content create attack surfaces that traditional web security wasn't built to handle.

404 MEDIA

Proton CEO Backs Down on AI Re-Interview Policy After Public Backlash

After 404 Media revealed that Proton's Andy Yen was forcing employees through AI-evaluated "re-interviews" amid a culture of fear, Yen cancelled the program. A privacy-first company using opaque AI to judge its own staff is a contradiction that proved too loud to ignore.

COINDESK

FBI Confirms North Korea Behind $6.3 Billion Binance Hack

The FBI has officially attributed the largest crypto hack in history to North Korean state actors and is asking DeFi protocols to block specific BTC, ETH, and SOL addresses. Binance claims it has already replaced 90% of the stolen funds — a remarkable backstop that raises its own questions about reserves.

SIMON WILLISON

Someone Baked a Working NES Emulator Into an LLM's Weights

Drew Breunig embedded a functional Nintendo 6502 CPU emulator directly into Qwen3.5-4B's weights, letting you run Super Mario Bros inside a language model. It's slow, absurd, and a genuinely mind-bending proof that model weights can encode executable computation.

A16Z

Why Frontier Models Still Can't Parse a PDF

Reducto CEO Adit Abraham makes the case that AI agents are bottlenecked not by reasoning ability but by document comprehension — PDFs, scans, and spreadsheets that models routinely misread. Healthcare, legal, and finance can't delegate consequential work to agents that fumble the inputs.

ALL-IN PODCAST

Anthropic S-1 Drops: "Greatest Business Ever" or Doomer Overreach?

The All-In crew dissects Anthropic's IPO filing alongside Google shedding $500 billion in market cap in a single day. The tension: Anthropic's financials look extraordinary while its safety rhetoric is being stress-tested by its own investors.

02

X / Twitter Signal

a16z dropped a sharp observation: average token prices are down 67% since May, yet token spend is up more than 120% over the same period. Cheaper compute isn't shrinking AI bills — it's exploding usage. This is the Jevons Paradox playing out in real time across every enterprise AI budget, and it has major implications for anyone betting that cost reductions will compress the market.

Mollick noted that less than three years ago, GPT-4 was the undisputed best model in the world — and now it's a historical footnote. The pace of displacement is accelerating faster than most institutional planning cycles. Anyone building strategy around today's frontier model should be building for its obsolescence simultaneously.

On agent isolation: Mollick flagged a fundamental architectural problem — multiple AI agents that share the same underlying model share a common "brain," meaning they can't be truly isolated from each other. As multi-agent systems scale, this creates both security and coherence problems that no one has cleanly solved yet.

The Kobeissi Letter confirmed the FBI's North Korea attribution on the Binance hack, adding that hackers penetrated a single employee's personal device to access internal systems. One compromised laptop. $6.3 billion gone. The attack vector is a blunt reminder that nation-state hackers don't need to break cryptography — they just need one person to click the wrong link.

03

The Thread

Two stories this week sit on opposite ends of the same fault line: AI is simultaneously the attack surface and the defense. The ChatGPT account-takeover exploit found by Parameter Labs isn't a fringe edge case — it's a preview of what happens when AI products ship dynamic rendering features faster than security teams can audit them. Meanwhile, the FBI's confirmation that North Korea brought down Binance with a single phished employee reminds us that the most sophisticated threat actors still exploit the oldest vulnerability: human trust.

"Cheaper tokens don't mean cheaper bills — they mean more tokens."
The a16z token-spend data is the most underrated signal in today's feed. A 67% price drop paired with 120% spend growth means the AI cost curve isn't deflationary for buyers — it's inflationary for usage. Every CFO who modeled AI as a cost-reduction play is about to revise their numbers upward. The real winners aren't companies saving money on AI; they're the ones converting cheaper inference into more throughput, more products, more moats.

Proton's AI re-interview debacle and Anthropic's S-1 land in the same week for a reason. The privacy-and-safety branding that premium AI companies sell is being pressure-tested by their own internal decisions. Proton got caught using AI to surveil its own employees. Anthropic is IPO-ing on a safety narrative while its investors debate whether that framing is a feature or a liability. The gap between what these companies say about AI and how they actually deploy it internally is closing — and journalists are camped at the gate.