Ex-DeepMind researcher launches Juice: turning the web into a payable capability network

A former DeepMind researcher (@AdaptiveAgents) has published a long-form series and launched a Beta project, Juice (juiceos.org), laying out a decentralized AGI vision he has mulled for years: the "Intelligence Internet." The core question: can we build AGI from components owned by different people while staying economically inclusive and safe? Juice is his engineering exploration of that question.

Confirmed

• The core idea is turning the internet's available affordances into Actions: each Action has an interface, an owner, and a price; Actions can call one another, and complex computation is composed from independently owned parts across the network. A single agent is already essentially a distributed multi-agent system—a simple harness invokes capabilities from different providers through actions like "think" and "search"
• Capability market vision: anyone can put their datasets, algorithms, models, or specialized software online with a price and charge per use; when doing research, a system could autonomously query proprietary databases, pay per call for specialized models, and synthesize results—no need to subscribe to ten providers or even know they exist in advance. The grander vision: on-demand "internet factories" assembled from independent suppliers network-wide—AI designs products, companies manufacture components, robots and humans assemble them
• Design proposal: if an LLM repeatedly combines the same actions in the same way to solve a class of problems, that combination should be crystallized into a new Action, turning "fluid intelligence" into "crystallized intelligence" so the internet accumulates reusable intelligence
• The cross-network execution problem is explicitly raised: when one action calls another, who pays, whose permissions apply, how are costs capped, and what if a provider vanishes mid-run—he argues a new execution model is needed, essentially "processes running across the internet"
• Economic perspective: independently owned components can collaborate even with different goals—providers charge for their contributions, giving them incentive to participate without any common owner or objective; he also conjectures that heterogeneous, multi-objective intelligent systems may be more robust than single-objective optimizers, with components assessing each other and adapting through feedback

Unconfirmed

• His rejection of a blockchain path is a personal judgment: he says that after leaving DeepMind he studied Ethereum in depth and concluded blockchain doesn't fit—beyond the compute problem of putting LLMs on-chain, more fundamentally blockchain depends on "being able to specify and verify valid computation," whereas intelligent systems encounter unforeseeable situations, and handling the unexpected is precisely what intelligence means—the two are incompatible
• The proposed trust model is likewise a concept: identity, permissions, and financial exposure can be enforced ex ante, but intelligent behavior cannot be pre-verified; trust should be built through bounded experiments plus after-the-fact evaluation, with users reporting experiences backed by verifiable records of actual transactions. He admits this cannot eliminate deception or collusion, only providing an evidence base for parties to judge for themselves

Why it matters

The series integrates hot topics like the "agent economy," "MCP-style capability invocation," and "decentralized AGI safety" into a coherent framework, grounded in a usable product (Juice Beta). Its discussion of cross-party payment, permissions, cost caps, and trust mechanisms points directly at the core engineering and governance challenges of an agent-based internet, offering valuable reference for researchers and practitioners focused on agent interoperability and business models.

2026-10-21 ~ 2026-10-21 · 18 related posts

• [source] Can decentralized AGI be built — economically inclusive and safe? A new exploration — AdaptiveAgents · 2026-10-21
• [source] Ex-DeepMind researcher launches Juice, turning any HTTP endpoint into a paid AI action — AdaptiveAgents · 2026-10-21
• Building an 'intelligent internet': AI agents that pay per query and orchestrate models for you — AdaptiveAgents · 2026-10-21
• A factory assembled on demand from internet providers: the thread's boldest vision — AdaptiveAgents · 2026-10-21
• From Ethereum research to a capabilities market: an ex-DeepMind researcher's pivot — AdaptiveAgents · 2026-10-21
• Ex-DeepMind researcher studied Ethereum for AI safety, then concluded it doesn't fit — AdaptiveAgents · 2026-10-21
• Why blockchains' trust model fails for intelligent systems, per ex-DeepMind thread — AdaptiveAgents · 2026-10-21
• Why ex-ante guarantees fail for intelligent systems, from a decentralization thread — AdaptiveAgents · 2026-10-21
• Thread on decentralized AGI: feedback and verifiable transaction evidence as trust — AdaptiveAgents · 2026-10-21
• Decentralized AGI thread: trust via bounded experiments and ex-post evaluation — AdaptiveAgents · 2026-10-21
• Ex-DeepMind researcher argues heterogeneous multi-objective AI systems may be more robust — AdaptiveAgents · 2026-10-21
• Paying Providers Could Let Independently Owned Agent Components Cooperate Across Owners — AdaptiveAgents · 2026-10-21
• [source] Turning the Internet's Affordances into Priced, Composable Agent Actions — AdaptiveAgents · 2026-10-21
• A Single Agent Is Already a Distributed, Multi-Owner System, Argues New Thread — AdaptiveAgents · 2026-10-21
• Who Pays When Actions Call Actions? The Missing Execution Model for Internet-Scale Agents — AdaptiveAgents · 2026-10-21
• Turn Repeated Agent Compositions into New Actions: Fluid to Crystallized Intelligence — AdaptiveAgents · 2026-10-21
• Turning fluid intelligence into crystallized: why agents should cache repeated solutions — AdaptiveAgents · 2026-10-21

1 near-duplicate retellings: AdaptiveAgents