# Project Fetch: Phase two — Anthropic > [!question] Lingering Questions > - Does "without human assistance" survive scrutiny when a researcher still approves every command and task transition? The article's autonomy claim needs that definition pinned down. > - Do all domains traverse the help→help→do-it-itself arc at similar speed, or do physical/safety-critical domains stall at stage 2 (humans helping models) far longer than digital ones? > [!tip] Article Ideas > - Cross-link the "Mythos-class" model confirmation here into the Fable 5 / Mythos 5 ban thread as independent corroboration — this article confirms the model-class name from a separate, same-week Anthropic source. #addendum > [!link] Cross Reference > - [[3. Resources/Permanent Notes/AI Capability in a Domain Advances Through a Fixed Three-Stage Arc]] — this source's central claim, already captured as a Permanent note with this article as its origin. #xref > - [[3. Resources/Permanent Notes/A Standardized Harness Turns Model Access Into a Second-Sourceable Supply Decision]] — same source-week; that note also grounds itself against this article's confirmation of the Mythos-class model line. #xref > [!quote] Vocabulary > - **Physical agentic AI** — Anthropic's framing for the current moment, where general-purpose model scaling (not robotics-specific tuning) starts producing embodied-task competence. ## Bib capsule Anthropic, "Project Fetch: Phase two," anthropic.com/research, June 2026. https://www.anthropic.com/research/project-fetch-phase-two ## Author's thesis (Adler Q1+Q2) A year after Claude could only help a human operate a robodog, Claude Opus 4.7 performed the same class of tasks alone — about 20x faster than the fastest human team at tasks completed by participants less than a year earlier, using roughly 10x less code. The article frames this as the same three-stage capability arc Anthropic has observed in cybersecurity (models help humans → humans help models → models act alone) now reaching the boundary between AI and the physical world — "the early era of physical agentic AI," produced by general scaling rather than robotics-specific effort. The honest caveat: the model still fails the fine-motor, closed-loop "fetch" itself (precisely moving the ball), which is the actual named capability frontier. ## Key arguments - Claude Opus 4.7 alone was ~20x faster than the fastest human team on the completed tasks (sensor connection, control programming, ball detection — not the literal fetch). - The help→help→do-it-itself progression, first observed in cybersecurity, now shows the same shape at the AI/physical-world boundary. - The capability gains are attributed to general model scaling, not concerted robotics-specific tuning — a recurring Anthropic motif. ## Claim atoms ### Claude Opus 4.7 alone was ~20x faster than the fastest human team > [!quote] ID:rw_pending > "Claude Opus 4.7—operating without human assistance—was about 20 times faster than the fastest human team at all tasks completed by our participants less than a year ago" **Where/why**: Phase Two headline claim; tasks were sensor connection, control programming, and ball detection — not the literal fetch. "Without human assistance" still includes per-command human approval, which needs careful definition. **Tags**: #xref #addendum ### The help→help→do-it-itself pattern now reaches the physical world > [!quote] ID:rw_pending > "first, models are helpful to humans. Then, humans are helpful to models. Finally, models are largely able to do things themselves. We have seen this in cybersecurity and now the same dynamics are starting to take shape at the intersection of AI and the physical world." **Where/why**: The article's core thesis — ports a cross-domain capability progression into embodied AI. Already promoted to a Permanent note. **Tags**: #xref #articleseed ### Models still fail the fine-motor closed-loop "fetch" > [!quote] ID:rw_pending > "The latest Claude models still struggled with using the robot to precisely move the beach ball—the 'fetching' part of Project Fetch." **Where/why**: The named capability frontier — a perceive-error-then-adjust-input closed loop and low-level actuation policy remain untouched. Honest scoping that bounds the headline claim. **Tags**: #xref ### Capability gains came from general scaling, not robotics-specific tuning > [!quote] ID:rw_pending > "this progress is not the result of a concerted effort to improve the robotics capabilities of our models. These improvements, like so many others in the history of LLM development, have emerged from much more general scaling." **Where/why**: Recurring Anthropic motif — domain capability as a byproduct of scaling. Also: Opus 4.7 wrote ~10x less code (1,045 vs. 10,309 lines) with equal-or-better success. **Tags**: #xref #addendum ## Permanent candidates - **The "help → help → do-it-itself" arc is a transferable capability-forecasting lens** — already exists as [[3. Resources/Permanent Notes/AI Capability in a Domain Advances Through a Fixed Three-Stage Arc]], created from this article. Not duplicated here. - The remaining atoms (speed comparison, fine-motor frontier, scaling-not-tuning) are all supporting evidence for that same arc claim rather than distinct reusable claims — not split into separate Permanent notes. ## Processing checklist - [x] All claim atoms have header + verbatim quote + where/why - [x] Triage callouts populated for every tagged highlight - [x] Checked existing Permanents first — found direct coverage, linked rather than duplicated - [x] Connected into MOC — [[3. Resources/MOCs/AI & the Future of Knowledge Work MOC]] ## Links - [[3. Resources/MOCs/AI & the Future of Knowledge Work MOC]] — a forecasting lens for which kinds of work approach the autonomous stage - [[3. Resources/MOCs/AI Agent Safety MOC]] — embodied-agent capability claims bear directly on physical-world agent safety - [[3. Resources/Permanent Notes/AI Capability in a Domain Advances Through a Fixed Three-Stage Arc]] — because evidence-for; this article is that claim's origin source - [[3. Resources/Readwise/Articles/Project Fetch Phase two]] — source document