Project Fetch: Phase two  🔗 **Source:** [Project Fetch: Phase two](https://www.anthropic.com/research/project-fetch-phase-two) > [!note]+ Document Notes > tags: [embodied-ai, agentic-ai, ai-capabilities, robotics, model-scaling, anthropic-research] > author: Michael Ilie, C. Daniel Freeman, Kevin K. Troy > source: Anthropic > reference: https://www.anthropic.com/research/project-fetch-phase-two > date created: 2026-06-21 14:30 > > # 🏷️ Title: Project Fetch: Phase two > <!-- deep-read:sections start --> ## 🧭 Table of Contents - [📝 Summary](#-summary) - [📖 Main Content](#-main-content) - [🗒️ Notes](#️-notes) - [📚 References](#-references) - [🗣️ Paraphrase](#️-paraphrase) - [📰 Newsletter](#-newsletter) - [✅ Action Items](#-action-items) - [🧪 Hypotheses](#-hypotheses) - [📊 Methodology](#-methodology) - [❓ Questions](#-questions) - [🔮 Future Research](#-future-research) - [💡 Implications](#-implications) - [🧠 Learning Styles Assessment](#-learning-styles-assessment) - [📊 Accessibility Assessment](#-accessibility-assessment) - [🌍 DEIB Assessment](#-deib-assessment) - [✅ Logical Validity](#-logical-validity) - [🛡️ DARVO Analysis](#-darvo-analysis) - [🎯 Conclusion](#-conclusion) ### 🔗 References - [Project Fetch: Phase two](https://www.anthropic.com/research/project-fetch-phase-two) (this article) - [Project Fetch — Phase One post](https://www.anthropic.com/research) (the original August 2024 experiment referenced throughout) - Claude Code (the agentic harness used to run Opus 4.7 in the trials) ## 📝 Summary - **What happened.** Anthropic re-ran "Project Fetch," its experiment operating an off-the-shelf robotic quadruped (a "robodog"). In Phase One (August 2024), Claude Opus 4.1 helped non-expert human teams but could not do the tasks alone. In Phase Two, **Claude Opus 4.7 — operating with no human assistance via Claude Code — completed the same tasks roughly 20x faster than the fastest human team** from a year earlier (connect to video/lidar sensors, write control programs, detect a beach ball). - **Why it matters.** It's a concrete data point that the recurring AI progression — *models help humans → humans help models → models do it themselves* — already seen in cybersecurity, is now appearing at the intersection of AI and the physical world. The authors call this the "early era of physical agentic AI." - **Key stakeholders.** Anthropic (author and self-interested party), robotics and embodied-AI researchers, AI-safety/evaluation teams tracking capability jumps, and anyone modeling how general scaling spills into hardware control. - **Risks and tradeoffs.** The headline is real but bounded: Claude still **fails the actual "fetch"** — the fine-motor closed-loop control to precisely nudge the ball. None of the tasks touch low-level actuation policies. The research is Anthropic's own, on its own models, framed favorably (flag for self-interest). Speed claims rest on three trials against a prior small human sample. - **Bottom line.** LLMs have not "solved robotics," but apparently large capability distances are being crossed fast; ruling out a similar trajectory in hardware would be unwise. ## 📖 Main Content - **The setup and the through-line.** Project Fetch tested whether Claude could help non-roboticist Anthropic employees operate a robodog. Phase One pitted a Claude-enabled team against a Claude-less team; the Claude team got more done, faster, but Opus 4.1 could not do the tasks autonomously (it got stuck just connecting to the robot). Phase Two revisits this with a newer model to measure model-only capability. - **The headline result.** On every task at least one human team completed in August, **Opus 4.7 finished at least 10x faster**. Across the four tasks completed by *both* human teams, Opus 4.7 averaged **>37x faster than Team Claude-less and >18x faster than Team Claude** (361 min vs 181 min vs 9 min 35 sec). On the five-task table comparison, Team Claude took 264 minutes; Opus 4.7 averaged 12 min 7 sec over three trials. - **Why it was faster — and leaner.** Where humans dithered between approaches to interface with the sensors, the model quickly identified the best path and much of its code worked on the first try. It was as or more successful while writing **~10x less code** than Team Claude (1,045 lines vs 10,309; Team Claude-less wrote 1,136). - **Reliability.** Across three trials, completion times were consistent (low within-task variance). One beach-ball-detection trial ran long, attributed to the model defaulting to an outdated object-detection algorithm — which it then worked around. For tasks "within its capability envelope," Claude is now "quite reliable." - **What it still can't do.** The literal "fetch" — using the robot to precisely move the beach ball home — requires a fast perceive-error → adjust-input closed loop that humans do well after practice. Claude could position the robot behind the ball but its control was poor and unsuccessful, mirroring the human participants who reached that stage. A more experienced human researcher *did* program autonomous fetching; the authors expect current Claude could too with more time and scaffolding — what they'll watch for is doing it at the same speed and reliability as the other tasks. - **The framing claim.** This progress is not a targeted robotics effort; it "emerged from much more general scaling." The pattern (models help humans → humans help models → models do it themselves) seen in cybersecurity is now taking shape in the physical world. The analogy: just as models learned to use existing software-editing tools (e.g., string-replace) on the way to agentic coding, they now seem close to using off-the-shelf *physical* tools with relative ease — "the early era of physical agentic AI." - **Uncertainties / contested points.** Whether models can make physical tools more bespoke (writing tailored control policies, designing robotic systems) is open, and the authors concede "substantial barriers" may exist. The methodology footnote also discloses that the **Mythos-class** model ("Claude Mythos Preview") was excluded because the experiment setup and serving wouldn't allow an apples-to-apples comparison. ### 🗒️ Notes - **20x headline** is the load-bearing number: Opus 4.7, *alone*, ~20x faster than the fastest human team from ~a year prior on completed tasks. - **Two distinct comparisons** in the piece: (1) 4 tasks all teams completed → 37.7x vs Claude-less, 18.9x vs Claude; (2) 5-task table → 264 min (Team Claude) vs 12 min 7 sec (Opus 4.7 avg of 3). Don't conflate the two denominators. - **Researcher role was minimal but nonzero**: plug laptop into robodog, enter initial prompt, approve commands, approve task progression. Not "fully autonomous" in the strict sense — there's a human approving steps. - **Excluded tasks**: they couldn't have Claude use the physical controller, and they didn't time a researcher using the Claude-programmed controller to retrieve the ball (though they confirmed it worked). - **Code-volume metric** doubles as an efficiency proxy: ~10x less code than Team Claude while matching/beating success — argues the speed isn't brute force. - **Failure mode named**: outdated object-detection algorithm default → one slow detection trial; model self-recovered. - **The capability boundary** is the fine-motor closed loop ("the fetching part"), plus untouched low-level actuation policy. - **Scaling claim**: improvements are a byproduct of general scaling, not robotics-specific tuning — recurring motif in Anthropic's capability narrative. - **Pattern claim** ("models help humans → humans help models → models do themselves") explicitly ported from cybersecurity to embodied AI. - **Mythos confirmation**: footnote states Opus 4.7 was "our most advanced non-Mythos-class model," and names "Claude Mythos Preview" — confirms Mythos-class as a real Anthropic model line. - **Self-interest flag**: Anthropic authoring research about Anthropic models' capability gains; tone is measured but promotional. ## 📚 References - [Project Fetch: Phase two — Anthropic](https://www.anthropic.com/research/project-fetch-phase-two) — primary source. - [Anthropic Research index](https://www.anthropic.com/research) — locus of the Phase One post referenced throughout. - Claude Code — the agentic environment running Opus 4.7 with adaptive thinking, effort set to maximum. - (Referenced concept) Closed-loop control / object detection — the technical capabilities the experiment probes. ## 🗣️ Paraphrase - In August 2024, Anthropic tested whether Claude could help non-expert employees operate a robotic quadruped; the Claude-assisted team outperformed the unassisted one, but the model alone could not complete the tasks. - Revisiting the experiment with Claude Opus 4.7, the model — running without human help in Claude Code — completed the year-old tasks about 20 times faster than the fastest human team had. - This does not mean robotics is solved: the model still cannot perform the precise, closed-loop "fetching" of the ball, and the experiment never touched low-level actuation control. - The same progression seen elsewhere — models assist humans, then humans assist models, then models act alone — now appears at the boundary of AI and the physical world. - The model was not only faster but leaner, writing about a tenth as much code as the human-plus-Claude team while matching or beating their success, and performing consistently across three trials. - These gains come from general scaling rather than deliberate robotics work, and the authors argue we are plausibly entering an early era of physical agentic AI, while conceding real barriers remain. ## 📰 Newsletter **The robodog gets a (digital) handler: Claude alone, ~20x faster** Anthropic just published Phase Two of "Project Fetch," and it's a sharp little capability snapshot. Last year, Claude *helped* non-expert humans operate an off-the-shelf robot dog; it couldn't run the show solo. This year, Claude Opus 4.7 — no human co-pilot, driving everything through Claude Code — knocked out the same tasks (wiring up video and lidar sensors, writing the control programs, detecting a beach ball) roughly **20 times faster than the fastest human team** from a year earlier. And it did it writing about a *tenth* as much code. The honest caveat is right there in the post: Claude still flubs the literal "fetch." Precisely nudging the ball home needs a tight perceive-and-adjust loop that humans nail with practice and the model doesn't — yet. The bigger idea worth filing away: the familiar arc of *models help humans → humans help models → models do it themselves* — already old news in cybersecurity — is now showing up where bits meet atoms. Anthropic calls it "the early era of physical agentic AI." Two footnotes for the WordPress/dev crowd: (1) keep your skepticism handy, this is Anthropic grading Anthropic; and (2) the post quietly confirms a **Mythos-class** model ("Claude Mythos Preview") exists, excluded here only because it couldn't be compared apples-to-apples. Optimism, but with the limits printed on the label. ## ✅ Action Items #### DevRel - Developer advocacy: build a talk/demo on "agentic AI meets hardware" using Project Fetch as the case study; pair with a Claude Code agentic-coding demo. - Education: workshop on the *models help humans → humans help models → models do it themselves* progression as a capability-forecasting lens. - Developer experience: document the Claude Code "approve-each-step" agentic loop pattern shown here (plug in, prompt, approve commands, approve next task). - Developer marketing: clip the "~20x faster, ~10x less code" stat as an evergreen capability-jump graphic. - Community management: host a Q&A on embodied-AI claims vs. hype, with the fine-motor failure as the honesty anchor. - Internal/external/cross-project: share the experimental-design template (3 trials, time + code-volume + reliability metrics) as a reusable eval pattern. #### OSPO - Policy & Governance: position on agentic AI operating physical hardware — approval gates and human-in-the-loop requirements. - License Compliance: review provenance of model-generated control code (e.g., outdated object-detection libraries it may default to). - Security Management: threat-model agentic models that can program and operate physical devices. - Software Inventory: catalog the open-source CV/robotics libraries such agents reach for by default. - Developer Support: guidance on supervising approve-each-command agentic runs. - Community Engagement: fund independent replications of capability-jump claims. - Strategic Alignment: track "model-only task completion time" as a capability KPI. #### Community - Maintainers: Not applicable to this source. - Contributors: explainer for newcomers on what "physical agentic AI" does and doesn't mean. - Sponsors: Not applicable to this source. - Ecosystem: coordinate shared, reproducible embodied-AI benchmarks beyond a single lab's setup. #### WordPress - Core: Not applicable to this source. - Core Performance: Not applicable to this source. - Core AI: document the "general scaling spills into new domains" lesson when scoping responsible-AI evaluation sets. - **Core Program Team:** Not applicable to this source. - Mobile: Not applicable to this source. - Playground: Not applicable to this source. - Community: WordCamp lightning talk on capability-forecasting using the help→help→do-it-itself pattern. - Training: learning module on reading AI capability research critically (self-interest, sample size, scope). - Marketing: Not applicable to this source. - Support: Not applicable to this source. - Design: Not applicable to this source. - Documentation: examples-first explainer of agentic loops (approve-command patterns). - Photos / TV / Openverse: Not applicable to this source. - Polyglots: glossary for "agentic AI," "closed-loop control," "actuation policy," "Mythos-class." - Accessibility: ensure any embedded charts ship with descriptive alt text (see Accessibility Assessment). - Meta: Not applicable to this source. - CLI: Not applicable to this source. - Hosting: Not applicable to this source. - Tide / Themes / Plugins: Not applicable to this source. #### The WPCC - Programs: Not applicable to this source. - Projects: Not applicable to this source. - Funding experiments: Not applicable to this source. ### FAIR - Technical independence: argues for vendor-independent, reproducible capability benchmarks rather than first-party-only claims. - Policies and covenants: disclosure norms for AI operating physical hardware. - Productized governance artifacts: a "physical agentic AI" risk-assessment scorecard. ## 🧪 Hypotheses - H1: A model's autonomous task-completion speed on robot-operation tasks improves faster than dedicated robotics tuning would predict, driven by general scaling alone. - H2: For tasks within a model's "capability envelope," within-task completion-time variance stays low across trials (reliability rises with capability). - H3: Models produce less code than human+model teams for equal-or-better task success (efficiency increases with capability). - H4: Fine-motor closed-loop control (the "fetch") remains a capability frontier that lags sensor-connection and perception tasks by at least one model generation. - H5: The progression *models help humans → humans help models → models do it themselves* generalizes across domains (cybersecurity, physical control) with predictable ordering. - H6: Mythos-class models would show further uplift on these tasks but require redesigned, comparable serving/experiment setups to measure. ### 📊 Methodology - H1/H3: longitudinal cross-generation evals (4.1 → 4.7 → Mythos) on a fixed robot-operation task battery, holding the harness constant, measuring time and code volume. - H2: repeated-trial reliability studies (≥3 trials, ideally more) reporting variance and outlier causes (e.g., algorithm-selection defaults). - H4: dedicated closed-loop control benchmark isolating perceive-error → adjust-input latency and accuracy, scored against human baselines. - H5: cross-domain replication of the three-stage progression with pre-registered stage definitions and timing. - General: larger human control groups to strengthen the speed-ratio denominators; independent (non-vendor) replication. ### ❓ Questions - Is "no human assistance" fair given a researcher still approves each command and task transition? - How robust is "~20x" against a small, year-old human sample and only three model trials? - Can the model write *bespoke* control policies, or only stitch off-the-shelf tools — and where's that line? - What exactly would close the fine-motor "fetch" gap: more scaffolding, a new model class (Mythos), or genuinely new capability? - How much does the result depend on Claude Code's harness versus the model itself? ### 🔮 Future Research - Measure model performance on low-level actuation/control-policy authoring, not just sensor wiring and perception. - Quantify the speed *and reliability* of autonomous fetching (the unsolved closed-loop task) across generations. - Test transfer to other off-the-shelf physical tools beyond this robodog to validate the "physical tool use" generalization. - Compare Mythos-class models under a redesigned, apples-to-apples serving setup. - Independent, multi-lab replication with larger human baselines. ### 💡 Implications - **Practice:** Agentic models can now operate some physical hardware via code with minimal human help — relevant to robotics prototyping and lab automation. - **Policy:** Human-in-the-loop approval gates matter; "autonomous" here still includes per-command approval — regulation should track that nuance. - **Security:** Models that can program and drive physical devices expand the attack/abuse surface; threat models must include embodied agents. - **Education:** Capability-forecasting literacy (read scope, sample size, self-interest) becomes essential for practitioners. - **Funding models:** Strengthens the case for funding independent embodied-AI benchmarks rather than relying on first-party claims. ### 🧠 Learning Styles Assessment - Reading level estimate: clear expository prose with technical terms (lidar, actuation policy, closed-loop control); ~Flesch–Kincaid Grade 11–13. Flesch Reading Ease ~40–50 (some long, clause-heavy sentences). #### 👀 Visual learners - Present: strong — a bar chart of total time, a comparison table, a code-volume bar chart, and a reliability scatter plot are described in the text. - Improve: add descriptive alt text/captions for each chart; a single annotated timeline of Phase One → Phase Two; a labeled diagram of the closed-loop "fetch" gap. #### 🔊 Aural learners - Present: none — text-only research post. - Improve: a 2–3 minute audio summary or TTS version; a short podcast walkthrough of the methodology. #### 🗣️ Verbal learners - Present: vivid analogies ("runaway robodog," string-replace tool use, "early era of physical agentic AI"); terms used in context. - Improve: a glossary callout (lidar, actuation policy, closed-loop control, Mythos-class). #### 🧩 Kinesthetic learners - Present: the experiment itself is hands-on, but no reader-runnable steps. - Improve: a reproducible task checklist or sandbox for the sensor-connection tasks; a small "approve-each-command" lab. #### 🧠 Logical learners - Present: clear cause→effect (general scaling → capability jump → faster/leaner completion) plus quantified comparisons and a reliability plot. - Improve: a decision tree of which tasks fall inside vs. outside the capability envelope; explicit error bars/CIs on the speed ratios. #### 🧑‍🤝‍🧑 Social learners - Present: none — single-direction research post. - Improve: discussion prompts for research-reading groups; a replication call-to-action. #### 🧘‍♀️ Solitary learners - Present: self-contained narrative readable alone. - Improve: reflection prompts ("which of my 'impossible for AI' tasks resemble sensor-wiring vs. fine-motor fetch?"). #### 🌿 Naturalistic learners - Present: weak — physical-world embodiment is the theme but no environmental framing. - Improve: tie embodied AI to energy/compute footprint and lifecycle of robotics hardware. ### 📊 Accessibility Assessment #### Screen reader semantics - Present: the post text describes each chart's content in prose (effectively long descriptions). - Improve: ensure each figure has a programmatic `alt`/`figcaption`, proper heading order, and `lang` attributes. #### Keyboard navigation - Present: standard article page; links should be keyboard-reachable. - Improve: verify visible focus states and logical tab order on the published page. #### Contrast and readability - Present: Anthropic's site generally meets contrast norms. - Improve: verify 4.5:1 on chart labels, axis text, and link colors. #### Text resizing and reflow - Present: responsive article text. - Improve: confirm charts and tables don't clip or require horizontal scroll at 200% zoom. #### Media and captions - Present: no audio/video; charts are described in body text. - Improve: provide data tables behind each chart and a downloadable dataset; no autoplay applies. #### Forms and interactions - Present: minimal — static research post. - Improve: ensure any newsletter/subscribe form has labels and inline validation. #### Standards reference - Maps to WCAG 2.2 AA: 1.1.1 Non-text Content (chart alt text), 1.3.1 Info & Relationships (table/figure semantics), 1.4.3 Contrast (Minimum), 1.4.10 Reflow, 2.1.1 Keyboard, 2.4.6 Headings and Labels, 2.4.7 Focus Visible. ### 🌍 DEIB Assessment #### Cultural breadth - Present: limited — a single-lab, single-context experiment with no regional/cultural framing. - Improve: situate findings against robotics work from other institutions and regions; broaden the contributor set. #### Inclusive language - Present: generally neutral, technical, plain; jargon mostly used in context. - Improve: define specialized terms inline (lidar, actuation policy, closed-loop) for non-expert readers. #### Neurodiversity support - Present: consistent structure, but several long, clause-dense paragraphs. - Improve: more subheadings, chunked summaries, and bulleted takeaways for the dense methodology passages. #### Representation - Present: three named authors; no other voices or perspectives. - Improve: include independent/external commentary or critique to balance the first-party framing. #### Psychological safety - Present: measured tone; openly states limitations ("this doesn't mean LLMs have solved robotics"). - Improve: more explicitly foreground the self-interest and small-sample caveats near the headline claim. ### ✅ Logical Validity #### Strong reasoning - Causal (general scaling → capability jump): well-argued and consistent with prior Anthropic findings; the claim is hedged ("emerged from much more general scaling") rather than overstated. - Inductive (efficiency): the ~10x-less-code-with-equal-or-better-success evidence reasonably supports "faster isn't brute force." - Analogical (software tools → physical tools): the string-replace → agentic-coding analogy is apt and clarifying for why off-the-shelf physical-tool use may follow. - Honest scoping: explicitly carves out what was *not* tested (physical controller, actuation policy, the literal fetch), which strengthens the credibility of the narrower claims. #### Fallacies - **Conflict of interest / self-serving framing** — "Claude Opus 4.7—operating without human assistance—was about 20 times faster than the fastest human team" (intro). Location: opening section. Weakness: Anthropic is evaluating its own model and publishing the favorable framing; no independent verification. Fix: commission or cite independent replication and report effect sizes with uncertainty. - **Hasty generalization / small sample** — "we ran three trials of Opus 4.7" and ratios drawn against a year-old human sample (Methodology section). Weakness: three model trials and a tiny human baseline can't robustly support a precise "~20x." Fix: report confidence intervals, enlarge the human control group, pre-register. - **Equivocation on "without human assistance"** — "The role of our researcher was limited to plugging a laptop … entering the initial prompt, approving commands, and approving the model to go to the next task" (Methodology section). Weakness: "without human assistance" coexists with per-command human approval. Fix: define autonomy precisely and report time/effort of the approval steps. - **Optimistic extrapolation** — "It would be unwise to rule out the same trajectory in hardware" (closing). Weakness: a confident forward claim immediately after conceding "substantial barriers." Fix: frame as a falsifiable hypothesis with named barriers and milestones. #### Tone of voice - Measured, candid, and self-aware (repeatedly states limits), but promotional by context — an organization reporting its own model's capability gains. Wry asides ("the runaway robodog that almost rammed into one of our human teams") leaven the technical content. ### 🛡️ DARVO Analysis - Deny: Minimal and counter-directional — rather than deny limitations, the authors volunteer them ("This doesn't mean that LLMs have now solved robotics. Far from it."). No accusation is being deflected. - Attack: None — there is no accuser or opponent to attack; peers/humans are framed sympathetically ("like our human participants"). - Reverse victim/offender: None. - Overall: DARVO is largely **not applicable** — there is no accusation to defend against. The relevant integrity issue is *self-interest*, not DARVO: Anthropic publishes favorable research on its own models. That's a conflict-of-interest flag (see Fallacies), but it does not manifest as denial, attack, or victim-reversal; the tone is, if anything, unusually forthcoming about constraints. ## 🎯 Conclusion - **Bottom line.** A year after Claude could only *help* humans operate a robodog, Claude Opus 4.7 did the same tasks alone — about 20x faster and with ~10x less code — yet still can't perform the precise closed-loop "fetch." The real signal is the recurring progression (*models help humans → humans help models → models do it themselves*) now surfacing at the AI/physical-world boundary: the "early era of physical agentic AI," driven by general scaling, not robotics-specific work. - **Next steps for readers.** (1) Read the claim with its scope and self-interest in mind — it's bounded and first-party. (2) Watch the fine-motor "fetch" gap as the next capability milestone. (3) Note the confirmed existence of a **Mythos-class** model line. (4) If you forecast AI capability, adopt the help→help→do-it-itself pattern as a lens — and demand independent, larger-sample replication. ## 🔄 Refactor Appendix — Project Fetch: Phase two > Density: 2–5 Literature atoms (H3), 0–3 Permanent candidates (H4). > Defaults applied. Resonance filter on. > > Tag pool: #bmf · #addendum · #xref · #articleseed · #storystem · #share · #aphorisms ### 📚 Literature Atoms (H3) #### [H1] 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" **Context & use**: - Phase Two headline; tasks were sensor connection, control programming, ball detection — not the literal fetch. - "Without human assistance" still includes per-command human approval (define autonomy carefully). **Tags**: #xref #addendum **Highlight**: [view highlight](readwise://pending) <!-- atomicity: 1 idea ✓ --> #### [H2] The "models help humans → humans help models → models do it themselves" 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." **Context & use**: - The article's core thesis; ports a cross-domain capability progression into embodied AI. - Frames the moment as "the early era of physical agentic AI." **Tags**: #xref #articleseed **Highlight**: [view highlight](readwise://pending) <!-- atomicity: 1 idea ✓ --> #### [H3] 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." **Context & use**: - The named capability frontier: perceive-error → adjust-input closed loop, plus untouched low-level actuation policy. - Honest scoping that bounds the headline claim. **Tags**: #xref **Highlight**: [view highlight](readwise://pending) <!-- atomicity: 1 idea ✓ --> #### [H4] 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." **Context & use**: - 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 **Highlight**: [view highlight](readwise://pending) <!-- atomicity: 1 idea ✓ --> ### 🧠 Permanent Candidates (H4) #### [P1] The "help → help → do-it-itself" arc is a transferable capability-forecasting lens **Claim** (one sentence, Courtney's voice): AI capability in a new domain tends to advance through a fixed three-stage arc — models assist humans, then humans assist models, then models act alone — so observing which stage a domain is in predicts the near-term trajectory. **Evidence pointer**: [[#H2]] **Scope & conditions**: Holds where general scaling drives capability and tasks decompose into tool-use plus perception (cybersecurity, agentic coding, robot operation). Breaks where the bottleneck is fine-motor/closed-loop control or genuinely novel capability not reachable by scaling. **Links**: - [[AI & Future of Knowledge Work MOC]] — because applies-in the agentic-AI capability thread - [[AI Agent Safety MOC]] — because applies-in embodied-agent risk - [[Claude Mythos Preview]] — because evidence-for the next-model-class line - [[Physical agentic AI]] — because extends the agentic pattern into hardware <!-- atomicity: 1 idea ✓ --> ### ✏️ Fleeting Queue - [F1] Confirm "Mythos-class" / "Claude Mythos Preview" against the Fable 5 / Mythos 5 thread in the vault — this post is independent corroboration. — https://www.anthropic.com/research/project-fetch-phase-two #fleeting #inbox - [F2] Open question: does "without human assistance" survive scrutiny when a researcher approves every command and task transition? #fleeting #inbox ### → Things 3 (max 2 actions) - [ ] Add the "help → help → do-it-itself" capability-forecasting lens to the AI & Future of Knowledge Work MOC — #things - [ ] Cross-link the Mythos-class confirmation into the Fable 5 / Mythos 5 vault note — #things <!-- deep-read:sections end --> 📄 **Full text:** ![[Full Document Contents/Articles/Project Fetch Phase two.md|Project Fetch: Phase two by anthropic.com]] ## Highlights - We figured it was time to revisit Project Fetch to see if our newer models could outperform the previous generation. Not only did they do that, but **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**. ([View Highlight](https://read.readwise.io/read/01kvnjjnqq94faemykgeyy3ys6)) - Note: H1 — the 20x headline: Opus 4.7 alone beat the fastest human team from a year ago by ~20x. Caveat: "without human assistance" still includes a researcher approving each command and task transition. - Tags: robotics, atom/literature/h1, anthropic - This doesn’t mean that LLMs have now solved robotics. Far from it. The latest Claude models still struggled with using the robot to precisely move the beach ball—the “fetching” part of Project Fetch. And none of the tasks in these experiments implicate the more challenging, low-level elements of robotic control, such as developing a specific actuation policy. However, once again, we are seeing a pattern whereby 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](https://red.anthropic.com/2026/mythos-preview/) and now the same dynamics are starting to take shape at the intersection of AI and the physical world. ([View Highlight](https://read.readwise.io/read/01kvnjjtweg57g2hcr951vr85b)) - Note: H2+H3 + Permanent: the three-stage capability arc (models help humans → humans help models → models act alone), now reaching embodied AI — bounded by the still-unsolved fine-motor "fetch" (precise ball control / actuation policy). - Tags: atom/zettel/z1, atom/literature/h3, atom/literature/h2 - It is worth underscoring (as we did in our previous post) that 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. ([View Highlight](https://read.readwise.io/read/01kvnjjvzptz8m4xjn8sk0dzy1)) - Note: H4 — the gains are a byproduct of general scaling, not robotics-specific effort (and Opus 4.7 used ~10x less code than the human-Claude team for equal/better success). - Tags: atom/literature/h4, robotics