{ "schema_version": "paper_public_manifest_v1", "paper_id": "skild_blog_introducing_s1_in_context_learning_20260817_2026_08", "slug": "skild_blog_introducing_s1_in_context_learning_20260817", "title": "S1: The GPT-3 emergence moment for robots", "authors": [], "source": { "arxiv_id": "", "pdf_url": "", "project_url": "https://www.skild.ai/blogs/s1", "github_url": "", "huggingface_url": "", "original_source": "https://www.skild.ai/blogs/s1" }, "site": { "post_url": "/posts/skild_blog_introducing_s1_in_context_learning_20260817", "canonical_url": "https://haiguangboy.com/posts/skild_blog_introducing_s1_in_context_learning_20260817", "cover_image": "https://static.haiguangboy.com/papers/skild_blog_introducing_s1_in_context_learning_20260817/cover.webp" }, "taxonomy": { "domain": "embodied_ai", "track": "vla", "tasks": [ "embodied_ai", "vla", "action_generation", "robotics", "Embodied intelligence", "Robot foundation models", "In-context learning" ], "related_topics": [ { "paper_id": "decompose_and_reorganize_planning_with_primitives_and_visuomotor_policies_learne_2026_08", "title": "Switching logic should not be left to the policy to learn", "url": "https://haiguangboy.com/posts/dr-lfd", "relation": "contrast", "summary": "S1: A video-demonstration-driven in-context learning (ICL) robot manipulation foundation model—no language instructions, no fine-tuning, and weights unchanged throughout—contradicts the core claim: switching logic should be handed to the planner, not learned by the policy. This is a structural rebuttal to end-to-end scaling.", "strength": "strong" }, { "paper_id": "generalist_blog_gen_1_5_20260818_2026_08", "title": "Robots' GPT-3 moment: plenty of story, not enough evidence", "url": "https://haiguangboy.com/posts/gen-1-5-one-shot", "relation": "contrast", "summary": "The training objective itself is meta-learning: pretraining runs an outer loop on episodic data whose tasks are specified only by in-context demonstrations; at inference, demonstrations drive the inner loop without weight updates—so ICL is not something that emerges accidentally after training a normal policy. This contradicts the core narrative (company's own account, not independently verified) that these capabilities arise directly from scaled pretraining, without any special mechanism design.", "strength": "strong" }, { "paper_id": "dyna_blog_dyna_2_20260811_2026_08", "title": "The first scaling law for human-to-robot zero-shot transfer", "url": "https://haiguangboy.com/posts/dyna-2-scaling-law", "relation": "contrast", "summary": "Data-engine stance: no single data source can simultaneously excel on all three axes—hardware proximity, diversity, and scalability—so it scales multiple sources such as teleoperation and first-person video; every $1 spent on data collection costs an extra $3 for quality control. This contradicts the route bet: human first-person video, not teleoperation or dedicated collection data, is the right scalable substrate for robot pretraining.", "strength": "strong" }, { "paper_id": "learning_a_thousand_tasks_in_a_day_2026_08", "title": "1,000 tasks in a day, thanks to inductive bias", "url": "https://haiguangboy.com/posts/mt3-thousand-tasks", "relation": "same_track", "summary": "Company's self-reported estimate: one in-context demonstration is worth about 380 traditional post-training (fine-tuning) demonstrations—validates the problem quantification: mainstream BC systems average 175–250 demonstrations per task, and bimanual tasks require roughly 8K.", "strength": "strong" }, { "paper_id": "an_open_foundation_model_towards_2026_07", "title": "An_Open_Foundation_Model_Towards", "url": "https://haiguangboy.com/posts/an_open_foundation_model_towards", "relation": "same_track", "summary": "Data-engine stance: no single data source can simultaneously excel on all three axes—hardware proximity, diversity, and scalability—so it scales multiple sources such as teleoperation and first-person video; every $1 spent on data collection costs an extra $3 for quality control—validates Ψ0's extreme data efficiency: 800h + 30h surpasses a 10x-data baseline by 40%+.", "strength": "strong" }, { "paper_id": "jepa_wam_learning_vision_language_action_policies_with_joint_embedding_world_modeling_2026_08", "title": "JEPA-WAM: An innovative approach that makes world models better at understanding state changes", "url": "https://haiguangboy.com/posts/jepa-wam", "relation": "same_track", "summary": "Company's self-reported internal experiment: under perturbations that fundamentally alter the action mode (L5, such as forced hand-switch execution), language-prompted VLA performance degrades up to 3× more than the ICL policy—validates the problem decomposition: the potential WAM now leaves only two design questions—what objective to predict, and how supervision is fed into action generation.", "strength": "strong" } ] }, "analyst_take": { "type": "author_opinion", "text": "The technical judgment of “buying emergence with scale, with no special mechanism design” is almost identical to what another American company, Generalist AI, said in its Gen-1.5 blog. Both companies admit they cannot explain why emergence happens; they only offer a language-model analogy as a hypothesis, with no empirical evidence. The two companies independently serving up the same narrative looks more like a shared belief taking shape in this field than two separate discoveries. The real issue to watch is whether this road of “once scale is reached, emergence naturally follows” will repeat the history of language models—once the emergence threshold is mapped out, the next step is competing over who can pile data across that line more cheaply." }, "ruling": { "importance_score": 3.0, "one_sentence": "S1: The GPT-3 emergence moment for robots" }, "asset_base_url": "https://static.haiguangboy.com/papers/skild_blog_introducing_s1_in_context_learning_20260817", "assets": [ { "type": "cover_image", "object_key": "papers/skild_blog_introducing_s1_in_context_learning_20260817/cover.webp", "content_type": "image/webp", "upload_status": "uploaded", "bucket": "paper-assets", "url": "https://static.haiguangboy.com/papers/skild_blog_introducing_s1_in_context_learning_20260817/cover.webp", "role": "post_cover", "size_bytes": 84034 }, { "type": "public_brief", "object_key": "papers/skild_blog_introducing_s1_in_context_learning_20260817/public_brief.md", "content_type": "text/markdown; charset=utf-8", "upload_status": "uploaded", "bucket": "paper-assets", "url": "https://static.haiguangboy.com/papers/skild_blog_introducing_s1_in_context_learning_20260817/public_brief.md", "role": "public_brief", "size_bytes": 5093 }, { "type": "public_manifest", "object_key": "papers/skild_blog_introducing_s1_in_context_learning_20260817/public_manifest.json", "content_type": "application/json; charset=utf-8", "upload_status": "uploaded", "bucket": "paper-assets", "url": "https://static.haiguangboy.com/papers/skild_blog_introducing_s1_in_context_learning_20260817/public_manifest.json", "role": "public_manifest", "size_bytes": 7050 } ], "published_at": "2026-08-26T22:25:45+08:00", "created_at": "2026-08-26T22:25:45+08:00", "updated_at": "2026-09-02T10:55:06+08:00" }