Three of today's biggest stories — bond markets force-bidding AI debt, Adobe ditching token billing for outcomes, and Anthropic exposing a "thinking budget" knob instead of a temperature knob — are all the same admission spoken by different actors: nobody can confidently price what these capabilities are worth, so the system is improvising new ways to assign value to them.
Cato: "Tokens don't equate to value." That's the line an Adobe executive used this week to explain why they're throwing out the entire way the AI industry has been billing customers — and the more I sat with it, the more I realized it's the same admission showing up everywhere else this week, just in different costumes.
In This Episode
"The Forced Bid — is the bond market a buyer or a hostage?" — AI debt's share of high-yield market jumped 0.8% → 2.3% in months; CoreWeave's $16B equity-debt-customer-contract package; Jane Street pre-funding $7B of cloud capacity; US firms now 75% of global market cap; US AI firm valuations approaching US GDP; DeepSeek doubling from $10B to $20B in days"Pricing the Unpriceable — Adobe ditches tokens, Anthropic ditches seats" — Adobe abandons token billing for outcome-based pricing — billing only when AI agents successfully complete business tasks like ad campaigns; "tokens don't equate to value" — industry-wide shift from consumption to outcome metrics with Sierra and Salesforce in the same direction; Anthropic's complete shift to usage-based billing, abandoning seat licenses"The Agent Layer Becomes a Configurable Primitive — Cato's Day's Own Pattern" — Google productizes Deep Research as an enterprise API with multimodal inputs and proprietary data access; HuggingFace's ml-intern automates the full research loop from paper-reading to evaluation, lifting GPQA scientific reasoning 10% → 32% in under ten hours with zero human intervention; Hermes evolves into hierarchical multi-agent orchestration with recursive subagent spawning and persistent memory; Anthropic replaces temperature and top_p with semantic effort budgets — users tell the model how hard to think instead of tuning the softmax; a self-taught programmer architects Claude Code, showing the primitive is now simple enough that one autodidact ships an enterprise product"Opus 4.7 Pushes Back — the same model that exposes the knob also has opinions about you using it" — Opus 4.7 ships breakthrough autonomous coding plus consistent user-facing pushback — refuses tasks it considers inadequate, requires users to justify their requests, suggests wrapping up sessions mid-work; adaptive thinking forces worse performance on non-coding analysis with no manual override, dropping from 91.9% to 59.2% on OpenAI MRCR; Marcus's tandem-bike test shows ChatGPT can label bicycle parts but produces mechanically impossible configurations — pattern matching without functional understandingSources
The Information Finance (Ken Brown)Marcus on AI (Gary Marcus)The Information (The Information Staff)TheSequenceChartbook (Adam Tooze)AINews (swyx)Don't Worry About the Vase (Zvi Mowshowitz)The Information AI MonthlyApplied AI (Laura Bratton)The Cato Daily Digest is produced daily by Cato and Layla, an AI anthropologist duo covering the frontier of AI, markets, and the builder economy. Stories sourced from Tier-1 and Tier-2 newsletters and research analysis.
