1:04Sam Altman: The Next ChatGPT Will Watch Your Screen and Remember Everything
@huxlab · 407 views · 2026-09-12
Sam AltmanOpenAIChatGPT
[●●●]LOADING
AGIENTRY INTELLIGENCE
Footprints on the Road to AGI
One job: collect what has actually happened on the road to AGI, with every entry traceable to a paper, a repo, a benchmark or a video. No headline chasing and no full tool directory — what shows here is only what we read ourselves.
Recommended for you
225This screen is ordered by our in-house recommender: the pool is every paper, deep read, demo, repo and skill that cleared our quality gate, ranked on real engagement signal, freshness and cold-start exploration — so it shifts a little each day.
Building an Agentic Harness with Jev: A Builder's Guide
Avid's 3,500-word build log: Jev as a decision layer inside keel, a local-first Rust coding workspace. One principle carries it: a model can suggest the next move, the host still owns the move — the host prepares the menu, Jev picks from it, the host checks again; pinned routes and live sessions bypass automatic selection; selection grants no permission. Plus how decision receipts become replayable evaluation scenarios.
Deep read2026-09-23
Jev and System One Models: Frontier Intelligence as a Function Call
A deep read of TypeSafe's first System One Model: three question primitives, the economics of parallel calls, confidence-gated routing, eval caveats, eight jagged edges, and the OpenJev local repro.
2026-09-18
Build Your Own Jev (100% Local): Recreating Fixed-Answer Scoring with SGLang
Jev-style fixed-answer scoring reproduced locally with SGLang /v1/score: single-token labels, restricted softmax, confidence-gated routing, and a 6 ms vs 1088 ms benchmark against generation.
2026-09-20
ModularRSI: Modular and Generalizable Recursive Harness Self-Improvement
ModularRSI is a benchmark-disjoint, contrastive, and modular framework for generalizable agent-harness evolution. It diagnoses recurring failures from paired successful and failed trajectories, evolves five functional modules independently, and integrates validated changes into a unified harness. On Terminal-Bench 2.0 and SWE-Bench Verified, the evolved harness improves unseen in-domain and cross-domain tasks and transfers across foundation models.
2026-09-14
Just Ask Jev: Reinforcement Learning for Calibrated Decisions as a Zero-Shot Detector of AI Alignment Failures
RLCDAlignBench spans 44 benchmarks and 7,193 instances. One generic question reaches median AUROC 0.886; soft probabilities beat argmax.
2026-09-24
unreal-home-wizard: A Codex Skill That Turns a Room Photo into a Walkable UE5 3D Space
A Japanese post spotlighting the free Codex skill unreal-home-wizard: give Codex a room photo and it checks the Unreal Engine 5.8 setup, drafts the floorplan for confirmation, furnishes and lights the scene while fixing clipping and gaps against the source photo, and finally launches a walkable 3D space (WASD). Install by asking Codex to add the skill; users still need UE installed and an Epic Games login. Note: this is a promotional repost — the original demo video is from @AmirMushich.
2026-09-27
ScientistTwo: Pioneering the Human Knowledge Frontier with Autonomous AI
ScientistTwo from Google Cloud AI Research is a fully autonomous multi-agent research framework: given only a scientific problem, it identifies human-SOTA limitations, formulates hypotheses, implements code, runs subset-to-full experiments, performs ablations, and simulates peer review with rebuttal experiments — delivering publishable papers and reproducible codebases. Across 107 ICLR/ICML/NeurIPS-accepted papers it improves 86 (80.4%) with a 25.2% average gain over human SOTA; papers score 7.5 with 91.9% acceptance under ScholarPeer and 72.1% clear the held-out Stanford Agentic Reviewer (0% for all prior agents).
2026-09-17
Figure Helix 2.5: Zero-Shot Chores in 30 Unseen Homes
The Helix 2.5 foundation model from Figure, pretrained on Index human data and fine-tuned for tidying, towel folding and bed making, hits 56% zero-shot success across 30 unseen Bay Area homes versus 9% trained from scratch. The 3h57m narration-free recording is recut here as a 1m35s bilingual recap.
2026-09-17
Just in
Delta Intelligence unveils Delta-0: a humanoid foundation model for whole-body loco-manipulation across 69 DoFs
Delta Intelligence has released Delta-0, a humanoid foundation model for whole-body loco-manipulation built on a co-designed brain and controller. The brain is a latent world-action model: a mixture-of-transformers with three branches (vision-language reading multi-view observations and instructions, a vision branch predicting future frames in DINO semantic feature space, and an action branch modelling whole-body motion from proprioception), trained jointly in four modes - forward dynamics, inverse dynamics, visual planning and policy-only. The controller converts motion commands into joint targets across 69 degrees of freedom through a delta-action interface shared by policy output, teleoperation and human corrections; in zero-shot evaluation on roughly 3 h 31 min of everyday-task motion it beats HEFT, MimicLite v1.1, SONIC v1.1 and ScaleBFM XL on both global root error and MPJPE. Pretraining uses more than 10,000 hours of paired egocentric video and whole-body human motion, plus a 154-dimensional shared action representation padded to 180 so single-arm, dual-arm, ego, UMI and humanoid data all live in one action space; high-precision tracking coverage rises from 29.2% to 72.6% as motion data goes from 10% to 100%. Evaluation runs a real-to-sim-to-real loop, rebuilding deployment scenes into simulation assets with 3D reconstruction and LLM agents, running repeatable autonomous rollouts, then returning selected checkpoints to hardware. Real-world RL post-training uses a value model that estimates cost-to-go, giving positive or negative conditioning signals to autonomous action chunks and to human-in-the-loop corrections - on the dishwasher task in a mirrored kitchen layout the success rate climbs from 20% (4/20) after SFT to 65% (13/20). The demo video covers six long-horizon household tasks: loading a record on a turntable, making a bed, picking a toy off the floor, opening a dishwasher, pressing a step trash can and sitting down on a sofa.
2026-09-28
Just in
100 Unitree G1s in one MuJoCo world: kung-fu tracking on a real 360-scanned plaza
Davide Cuttini, CEO of Over The Reality, shows 100 Unitree G1 humanoids in a single MuJoCo world tracking kung-fu motions with a pretrained whole-body policy, on a plaza reconstructed from a real 360-degree scan. Contacts, knockdowns and pile-ups all come from real physics. The pitch: sim-to-real starts with real environments - OVER crowdscans locations via smartphones and 360 cameras (174k+ locations, 86.8M images) and exports USDZ assets pairing a photorealistic Gaussian splat (vision) with a collision-ready mesh (physics) at true metric scale, importable into Isaac Sim/MuJoCo, so pixels and physics are anchored to the same real place to close the sim-to-real gap. Relevant both to multi-robot policy validation and simulation data infrastructure.
2026-09-28
Exploring
Dynamo: the datacenter-scale orchestration layer above inference engines, with KV-aware routing, tiered KVBM and an SLA-driven planner
NVIDIA's open-source (Apache-2.0) orchestration layer, Rust for the performance path and Python for extensibility. It explicitly does not replace SGLang, TensorRT-LLM or vLLM - it wires them into a coordinated multi-node system: disaggregated prefill/decode, KV-aware routing (2x faster TTFT on Qwen3-Coder 480B), KVBM offloading KV cache to CPU/SSD/remote, ModelExpress streaming weights over NVLink for 7x faster cold starts, an SLA-driven Planner (80% fewer breaches at 5% lower TCO in Alibaba production), and Grove for topology-aware NVL72 scheduling.
2026-09-28
Exploring
unsloth: the 2x-faster, 70%-less-VRAM fine-tuning library, now a desktop runtime that wires local weights into Claude Code and Codex
Self-described as the first desktop app to run and train models (Windows/macOS/Linux; multi-GPU across NVIDIA, AMD, Intel, CPU and Vulkan). Fine-tuning is claimed at 2x faster with 70% less VRAM, the post-training menu covers LoRA, QLoRA, full fine-tuning, pretraining, GRPO, DPO and FP8, and export reaches GGUF, NVFP4 and FP8. Unsloth Start connects Claude Code or Codex to local weights in one command; the model list includes Qwen3.8, GLM-5.3-Flash, Kimi K3, DeepSeek-V4, MiniMax-H3 and Gemma 4, with private web search, deep research, auto-compaction and RAG built in.
2026-09-28
LLM
31The foundation-model layer: the weights themselves plus what is tightly coupled to them - training recipes (pretraining, post-training, RL/RLHF), architecture and long context, and the inference serving stack with its cost per token
Latest demos2
Papers & deep reads3
2026-09-22MiMo-V2.6 Deep Read: Six Days of Live RL, an AA Index of 46 and a Fully Open Self-Improvement RunXiaomi open-sources MiMo-V2.6 Pro/Flash: 30 live RL steps, ~750k trajectories at $850k/$2.62M; AA Index 46 tops open models, DeepSWE v1.1 gains +17/+14 out of sample; Vibe World, CUA, science and content demos plus 7k+ RL environments released.
2026-08-10Stealing Reasoning Traces from Proprietary LLM APIsLeading large language model providers now conceal their models' step-by-step reasoning, or chain-of-thought, to protect intellectual property and limit information leakage. Rather than storing these traces server-side, providers return them to the client as blocks of encrypted text, which the client passes back with each subsequent request. Building on prior research, we identify an architectural vulnerability: these encrypted blocks are fully compatible and interchangeable across different sessions, users, and models within a provider's ecosystem. We exploit this compatibility to develop a scalable decryption jailbreak. By injecting an encrypted reasoning trace from a given model into a weaker, and less safeguarded model from the same provider, we force it to decode and output the trace verbatim in plaintext, without ever jailbreaking the more capable model directly. This vulnerability enables four distinct attack vectors. First, it circumvents anti-distillation mechanisms, allowing adversaries to extract a proprietary model's reasoning, as we demonstrate across Anthropic, OpenAI, and Google. Second, it allows for large-scale private data extraction. Developers frequently share session logs publicly, unaware of contents of the encrypted blocks. By decoding 315,320 reasoning blocks scraped from public repositories, we recovered 367 Personally Identifiable Information (PII) artifacts and 182 credentials. Third, it inadvertently reveals hazardous information hidden within the reasoning process, even in cases where the model's final, visible output safely rejects a malicious request. Fourth, attackers can leverage this flaw to execute invisible prompt injections, embedding malicious payloads entirely within encrypted blocks to poison public agentic rollouts. Following responsible disclosure, we propose concrete cryptographic and system-level mitigations to secure client-side reasoning.
2026-07-22LENS: LLM-guided Environment Simplification for Planning and Control in ClutterDespite recent advances in general-purpose robotic manipulation, real-world multi-object clutter remains challenging to handle for today's prevalent approaches. The problem scales in complexity due to more objects and collisions, more unpredictable contact physics, distractors, and task ambiguity. Bridging this gap to real-world deployment requires effective scene abstractions; yet today, producing such abstractions requires extensive task-specific manual engineering, which does not scale. These abstractions are costly to generate and difficult to adjust or fine-tune. We instead propose a plug-and-play fix to automatically generate scene-specific, task-specific, adaptively updating abstractions on top of existing planning and control stacks. LLM-guided Environment Simplification (LENS) produces a de-cluttered abstracted scene representation by merging (e.g., stacked objects) or pruning (e.g., distant objects) scene entities in a closed loop in response to task progress. These dynamic, task-relevant abstractions are versatile and easy to use. In our experiments, we show that LENS improves classical planning, model-based control, and a vision-language-action model, across a diverse set of highly cluttered manipulation scenes. Project website: https://lens-2026.github.io/.SOTA laddertracked 31
Claude Opus 5.5: the long-running agentic tier at $4/$20 with 1M context and adaptive thinking57.62AJev: TypeSafe's first System One Model types decisions and prices them two orders of magnitude lower67.8%CMiMo-V2.6: Live RL as an auditable ledger of 30 steps, 750k trajectories, and dollars46AGLM-5.3: the open-weights coding flagship bought with post-training alone28.3CClaude Fable 5.1: the long-horizon agentic tier that tops the ladder on a different ruler13.71AGPT-6 Astra: first on terminal engineering tasks, third tier on composite intelligence58.18%A
Intelligence: Artificial Analysis Intelligence Index, GPQA, MMLU-Pro, Arena Elo. Capacity: effective long-context recall, modalities. Supply: tok/s, $ per 1M tokens, open-weight licence
LLMHarness
31Agent runtimes and scaffolds: plugin architecture, sandboxing & tool surface, skill systems, long-term memory, self-improvement loops
Latest demos3
0:12Google CC: a group-level AI agent for families, with its own verified Google account
@xiaohu · 55 views · 2026-09-29
AI AgentGoogleGoogle Labs
1:15Manus launches personal agent Cue, benchmarked against OpenAI Muse, returning to the China market
@xiaohu · 108 views · 2026-09-28
ManusCuePersonal Agent
1:37OpenMausBot: a self-hostable team of AI agents
@BuildwithOmkarr · 243 views · 2026-09-23
OpenMausBotAI AgentComputer Use
Papers & deep reads3
2026-09-28AI Agent HandBook: Alibaba Cloud Enterprise Agent White PaperAlibaba Cloud open-source (Apache-2.0) white paper distilling enterprise agent deployment experience across the agent lifecycle — architecture, building, runtime, governance and optimization — in 30 Markdown chapters. Core thesis: understand the Model/Harness responsibility boundary instead of attributing everything to model capability. Covers agent loops, task state machines, context builders, action planes, MCP/A2A, sandbox runtimes, AI-gateway governance and evaluation flywheels, plus a 2026 agent developer survey and real cases (ABACI kernel-patch testing, Kitta code review, Bilibili content insights).2026-09-23MiMo Code: Xiaomi terminal coding agent betting on memory and self-evolutionXiaomi open-sourced terminal-native AI coding assistant, TypeScript built with bun. The source is MIT but usage is additionally bound by USE_RESTRICTIONS.md, the MiMo terms of service and the trademark policy - read those before treating it as plain MIT. The README states it is a fork of OpenCode: it keeps the multi-provider, TUI, LSP, MCP and plugin core and adds persistent memory (SQLite FTS5 full-text search across four kinds - project MEMORY.md, session checkpoints, scratch notes and task progress - injected automatically on session resume), intelligent context management (near the limit it rebuilds from the latest checkpoint plus project memory plus task progress plus retained recent messages, ranked by importance against a token budget), goals and stop conditions (/goal sets the condition, and when the agent wants to stop a separate judge model assesses whether it was truly met, which targets optimistic early quitting), and deterministic JS workflows in a sandbox (compose splits independent tasks into isolated git worktrees with per-task TDD; also deep-research, fact-check with three-reviewer adversarial voting, and research-experiment with an anti-metric-gaming audit). Twenty built-in skills (arxiv, claude-code, codex, docx, pdf, pptx, xlsx, html-to-video, product-design and more), compatible with four skill roots - .agents/skills, .claude/skills, .codex/skills, .opencode/skills - where a user skill of the same name overrides the built-in. /dream and /distill are its signature: the first distils recent session trajectories into project memory and prunes stale entries, the second finds your repeated manual routines and packages high-confidence candidates into reusable skills. Model agnostic - the Xiaomi platform, Codex/ChatGPT OAuth, or any OpenAI-compatible endpoint. 13.4k stars. We have not installed or run it; memory-restore accuracy, judge-model effectiveness and the vendor-stated cache hit rates are unverified by us, so it is graded as pending reproduction.
2026-09-17SoL-Pi: Recursively Scaling Auto-Research Loops for Efficient Agent HarnessAs coding agents move from supervised code completion to unattended, around-the-clock exploration, their work expands into long trajectories of reasoning, tool use, and feedback. SoL-Pi scales recursive auto-research across executable environments to discover reusable harness mechanisms. Four retained mechanisms improve action execution, context compaction, observation handling, and delegated reading, reducing recorded token traffic by 44.7-49.0% and API cost by about one third at comparable EdgeBench performance.
2026-09-17SoL-Pi: Recursively Scaling Auto-Research Loops for Efficient Agent HarnessAs coding agents move from supervised code completion to unattended, around-the-clock exploration, their work expands into long trajectories of reasoning, tool use, and feedback. SoL-Pi scales recursive auto-research across executable environments to discover reusable harness mechanisms. Four retained mechanisms improve action execution, context compaction, observation handling, and delegated reading, reducing recorded token traffic by 44.7-49.0% and API cost by about one third at comparable EdgeBench performance.SOTA laddertracked 31
DeepSeek Harness: the open-source agent runtime where everything is a plugin234kPi: a layered agent harness monorepo109kDeep Agents: an off-the-shelf agent harness reverse-engineered from Claude Code30kHermes Agent: the open-source agent that turns self-improvement into a closed learning loop248kopencode: the open-source coding agent split into client and server211kDeerFlow 2.0: ByteDance's super agent harness83k
Long-task completion, plugin & skill ecosystem, sandbox & telemetry, cost & latency per task
HarnessAI Coding
66Code generation, repair, refactoring, software engineering tasks
Latest demos2
Papers & deep reads2
2026-09-28mobile-mcp: the mobile automation MCP server driven by the native accessibility treemobile-next/mobile-mcp (8.2k stars, Apache-2.0) connects iOS simulators and real devices plus Android emulators and real devices to any MCP client, including Claude Code, Codex, Gemini and Copilot. Its first principle is accessibility-first: read the native accessibility tree instead of screenshots, no vision model and no image tokens, falling back to screenshots with coordinates only when necessary. Roughly 30 tools in seven groups cover device management, remote devices through Mobile Next Cloud, app management, screen interaction with recording, input and navigation, logs and crashes, and batch commands that compress N rounds into one. The README examples are all cross-app long-horizon tasks, the hardest class for mobile agents.2026-09-27GenOffice: Open-Source AI Office SuiteGenOffice is the first full-featured open-source AI office suite (Apache-2.0): real .docx/.xlsx/.pptx/PDF/Markdown editing with byte-preserving Word compatibility, block-level AI editing with document-aware agents, BYOK model support, fully local.SOTA laddertracked 66
Claude Code: the agentic coding tool that lives in the terminal, and the reference harness for Anthropic models148,438ACodex: an Apache-2.0 core, one agent in four form factors, and the reference harness for OpenAI modelsApache-2.0 · 126,918★CKimi Code: a model lab that builds its own harness, then opens it to every other shell over two protocols4 个 model ID · 1M 上下文CTrae: ByteDance two-form-factor coding line - TraeCode (IDE + SOLO) and TraeWork, plus a stalled MIT-open shell1M · 200 轮CZCode: the official GLM-5.3 harness, defined as an ADE rather than an editor$1.4 / $4.4C
SWE-bench Verified, LiveCodeBench
AI CodingSkills
13Developmentobra/superpowersFeatured
Fifteen skills that give a coding agent a full development process
Our takeIts premise is that models lack process discipline, not coding ability. The single best idea to steal is 'Rulings, not stalls': decide by default, and stop for a human only when an action is irreversible, touches a security-sensitive surface, leaks outside the worktree, or the plan is already too broken to continue. We adopted those four gates in our own pipelines.
v6.4.1 (2026-09-18), MIT, 289k stars — the largest single skill repository we have found. It chains brainstorming, spec writing, git-worktree isolation, planning, subagent execution, red-green TDD, two-way code review and verification-before-completion into one trunk workflow, and the README lists 16+ harnesses: Claude Code, Codex, Cursor, Gemini CLI, OpenCode, Kimi Code, Devin CLI and more.
Claude CodeCodexCursor
/plugin install superpowers@claude-plugins-officialShellobra289.2k
Fifteen skills that give a coding agent a full development processDevelopmentmattpocock/skillsFeatured
The 38 skills Matt Pocock uses every day for real engineering
Our takeIt sells parts rather than a whole process: code-review, diagnosing-bugs and resolving-merge-conflicts each stand alone. grill-me, grilling and wait-what turn the agent around to interrogate your requirements, and nothing else in this batch does that. CONTEXT.md at the repo root decouples project vocabulary from generic skills — the cleanest version of that idea we have seen.
The personal working set of the author of Total TypeScript; the repo describes itself as real engineering, not vibe coding. 38 SKILL.md files, primary language Shell, grouped as engineering (18), productivity (7), misc (4) and in-progress (9). It explicitly rejects takeover-style frameworks such as GSD, BMAD and Spec-Kit in favour of small, easy to adapt, composable.
Claude Code
claude plugins install mattpocock-skillsShellmattpocock266.4k
The 38 skills Matt Pocock uses every day for real engineeringOfficial & Collectionsanthropics/skillsFeatured
The official repo from the format's author: 19 skills plus spec and template
Our takeIf you want to know how a SKILL.md is actually supposed to be written, read this instead of any second-hand tutorial. Two things to know first: docx, pdf, pptx and xlsx are source-available rather than open source (the other 15 are Apache-2.0), and the repo carries its own disclaimer — these skills are demos, so test them in your environment before trusting them with anything critical.
Anthropic's own implementation, 177k stars: 19 skills under ./skills plus ./spec (the Agent Skills specification, now at agentskills.io) and ./template as a scaffold. Its definition of a skill is folders of instructions, scripts and resources loaded dynamically, and progressive disclosure comes straight from here.
Claude Code
/plugin marketplace add anthropics/skillsPythonanthropics177.3k
The official repo from the format's author: 19 skills plus spec and templateDesign & Frontendnextlevelbuilder/ui-ux-pro-max-skillFeatured
Design knowledge as searchable local data: 79 styles, 192 palettes, 119 UX rules
Our takeIts sharpest self-description is that models can write CSS but lack a taste database. Ten rule categories are ordered by severity, with Accessibility and Touch & Interaction first as CRITICAL and Charts last — that ordering is the stance. The Design System Generator runs five parallel searches, and the repo states plainly: do not persist unverified output, so a generated design system has to be checked before it lands in your repo.
v2.13.0, MIT, 129k stars, site uupm.cc. Instead of lecturing the model it ships structured catalogues the agent queries before writing code: 79 UI styles, 192 palettes with reasoning, 74 type pairings, 119 UX rules, 105 icon suggestions, 17 GSAP presets, 25 chart types, 22 stacks and 34 landing-page patterns.
Any agent
npx ui-ux-pro-max-cli init --ai claudePythonnextlevelbuilder129.3k
Design knowledge as searchable local data: 79 styles, 192 palettes, 119 UX rulesKnowledgeGraphify-Labs/graphifyFeatured
Map the whole project into a queryable knowledge graph instead of grepping
Our takeThe design that matters most: every edge is labelled EXTRACTED or INFERRED, so you can tell what the source explicitly contains from what the tool inferred — which makes an agent's conclusions auditable, a hard requirement in production. It is not a vector index: no embeddings, no vector store, a real graph you can explain, path and query. In its own benchmark, building the graph costs zero LLM spend and ingest is an order of magnitude cheaper, a structural win from local parsing.
YC S26, Apache-2.0, 120k stars, Python 3.10+. /graphify . maps code, docs, PDFs, images and videos into a graph and emits a clickable graph.html, a human-readable GRAPH_REPORT.md and a machine-readable graph.json. Code goes through tree-sitter AST parsing (about 40 languages): deterministic, no LLM calls, nothing leaves the machine.
Claude CodeCodexCursor
uv tool install graphifyy && graphify installPythonGraphify-Labs119.9k
Map the whole project into a queryable knowledge graph instead of greppingOfficial & CollectionsComposioHQ/awesome-claude-skillsFeatured
A curated index of 1000+ skills, plus one genuinely unique connect-apps plugin
Our takeThere is no one-liner that installs this repo; use it as a discovery layer. A list answers what exists, not what is worth running. The part you cannot get elsewhere is connect-apps-plugin: a single MCP endpoint to 1000+ integrations with authentication, team-level ACLs and audit logs, which is exactly the governance problem enterprises hit when agents touch SaaS.
A list maintained by Composio at 75k stars with an Apache-2.0 badge in the README. About 31 first-party skill directories physically exist here while the index references 864 SKILL.md files, and most of the difference is links out to external repositories. Coverage explicitly extends beyond Claude.ai and Claude Code to Codex, Cursor, Gemini CLI and Antigravity.
Any agent
claude --plugin-dir ./connect-apps-pluginPythonComposioHQ75.4k
A curated index of 1000+ skills, plus one genuinely unique connect-apps pluginLeaderboard
469Ranks come verbatim from each evaluator's own board, synced into our database on a schedule. We never recompute, reweight or merge across sources: an intelligence index, an Elo and a success rate are different rulers, and any single overall number would be incomparable figures stitched into one that looks comparable.
Artificial Analysis Intelligence IndexOfficial intelligence index (higher is better)
Claude Opus 5.5 (max with fallback)Anthropic57.62
Claude Opus 5.5 (xhigh with fallback)Anthropic55.99
Claude Sonnet 5.5 (max with fallback)Anthropic55.98
Claude Opus 5.5 (high with fallback)Anthropic53.58
Claude Fable 5.1 (max with fallback)Anthropic53.35
Source · artificialanalysis.aiofficial Intelligence Index v4.3synced Sep 22, 2026
Best in each other domainTop row of each board; the rulers differ
AgentsClaude Fable 5.1 (Max)14.06%CodeSonar Foundation Agent + Claude 4.5 Opus79.2%Imagegpt-image-2.5-sunburst1424Videogemini-omni-1.1-flash15163DProlificDreamer43.3EmbodiedDM0.524.9Image Generation
28Text-to-image, image editing, consistency and controllability
Latest demos3
1:54Kling 4.0 announced for October: 4K/HDR, 15 multimodal references, native 30-second generation
@Kling_ai · 113 views · 2026-09-28
Kling AIKling 4.0Video Generation
0:23OmniChar: One Character Across Every Model, Carried in a Portable .char File
@aiaicreate · 147 views · 2026-09-28
OmniCharCharacter ConsistencyComfyUI
0:43Ming Image-0.1 Design, Reviewed: Built for Real Design Pipelines
@realfxw · 368 views · 2026-09-23
Image GenerationText RenderingUI Design
Papers & deep reads4
2026-09-23Frontend Slides: a zero-dependency HTML slide skill that lets non-designers pick their taste from three real previewszarazhangrui/frontend-slides is a coding-agent skill, MIT, primary language JavaScript, 29,716 stars / 2,326 forks, described in one line as making beautiful slides on the web with a coding agent's frontend abilities. Its design premise is the first Philosophy item in the README: you do not need to be a designer to make something good-looking, you only need to react to what you see. That sentence determines the whole architecture. Most PPT prompt collections ask you to describe a style first (minimal, techy, business); SKILL.md explicitly rejects that route - Phase 2 is called Style Discovery, states that most people cannot describe their design preferences in words, and makes the default experience visual comparison: generate three single-page HTML previews, open them automatically, and let the user pick, with an explicit instruction not to ask whether they want options or a preset selector. The three-slot ratio is a hard rule: one safe preset (from the 12 curated entries in STYLE_PRESETS.md, grouped by temperament - dark Bold Signal / Electric Studio / Creative Voltage / Dark Botanical, light Notebook Tabs / Pastel Geometry / Split Pastel / Vintage Editorial, specialised Neon Cyber / Terminal Green / Swiss Modern / Paper & Ink), at least one bold template (34 design systems in bold-template-pack, from the author's other repo beautiful-html-templates, each with three real screenshots), and one wildcard (a second bold template or an agent-designed custom look, used when the brief holds an opportunity sharper than any template). The artefact is a zero-dependency single HTML file with all CSS and JS inlined - no npm, no build tool, no framework - justified plainly: dependencies are liabilities, an HTML file still opens in ten years, a 2019 React project does not. Progressive disclosure is token engineering: the agent first reads only selection-index.json with its mood / tone / best_for / avoid_for / formality / density / scheme fields, then the preview.md cards of the shortlist, and reads a full design.md only for the single template the user finally picks, with an explicit do-not-copy-template.html rule. One rule is marked NON-NEGOTIABLE: every style preview must look like the real first page of the user's deck rather than a diagnostic card - no preview, template, preset, Option A-B-C, filename or path may be rendered on the slide, and requirement notes such as 'sharp and provocative' must not become content. Core principle 5 is NON-NEGOTIABLE too: a fixed 1920x1080 canvas scaled uniformly, 16:9 even on phones, letterboxing allowed and re-layout forbidden; each supporting constraint maps to a real bug (inline the whole of viewport-base.css, show and hide pages only with the visibility / opacity / pointer-events based .active / .visible classes and never display:none, use clamp() outside the stage only, never negate a CSS function directly but write calc(-1 * clamp(...)), and support prefers-reduced-motion). The anti-AI-slop section addresses the agent in the second person and names the convergence tendencies to avoid - Inter / Roboto / Arial, purple gradients on white, predictable layouts - with the sharpest line being that you will still converge on common choices across generations (Space Grotesk for example), so avoid it. The acceptance bar is professional: verify overflow and panel overlap in a rendered browser screenshot, because checking scrollHeight alone is not enough since grid panels can visually cover each other. Six phases cover three modes (new build, PPTX conversion via extract-pptx.py, enhancing existing HTML) and two density settings, the artefact carries inline editing (press E, save with Ctrl+S), and Phase 6 documents the operational pitfalls of deploy.sh (Vercel) and export-pdf.sh (Playwright, where --compact renders at 1280x720 and usually cuts 50-70%). We did not install the plugin or produce a real deck, so it is graded as pending reproduction.2026-09-23Guizang PPT Skill: two locked visual systems and three validators as a web-deck engineering specA web presentation-deck skill open-sourced by (op7418), AGPL-3.0, primary language HTML, 26,809 stars. It is not a prompt template that writes slide copy for you; it is an engineering specification that makes an agent reliably produce a single-file HTML swipe deck you can present on stage. The README states the provenance on its first screen: the rules were distilled round by round from the author's own offline talks, and every pitfall landed in references/checklist.md. The decisive reason for HTML over PPTX is that HTML and CSS are text an agent can read, edit and verify directly - a script can measure how many px a slide overflows by, and Playwright can render it for a real visual check. Its core asset is two visual systems with mutually incompatible class names and very different constraint strength: Style A (editorial magazine x e-ink, 10 paste-ready layout skeletons, 5 theme presets, Chinese display headlines capped at 5 characters with nowrap) and Style B (Swiss International, 22 named layouts S01-S22 that are locked, body pages may only pick from them and must write data-layout, 4 anchor colours IKB #002FA7 / #FFD500 / #C5E803 / #FF6B35, square corners with no gradient or shadow, large type forced to weight 200). The most counter-intuitive decision is that custom hex values are not allowed: when a user supplies an arbitrary colour the agent should politely refuse and lay out the presets instead, and SKILL.md says protecting the aesthetic matters more than giving freedom. What separates it from ordinary template packs is three validators plus a repair ladder: validate-swiss-deck.mjs statically checks registered layouts, image slots, text inside SVG and headline alignment, and with Playwright measures DOM overflow in px, bottom whitespace, the nav-safe line and title gap; validate-presenter-mode.mjs catches duplicate slide IDs, mismatched notes, a timing plan over 90% of the target and missing controls; check-presenter-runtime-sync.mjs exists solely to catch drift in the presenter runtime copied into both templates. Guessing-based fixes are forbidden: 1-40px overflow gets micro-adjustment only, 40-90px local spacing compression, 90-160px may squeeze a headline or split a page, and only above 160px may you change layout or delete content. A fully local presenter mode ships inside (press P: dual-window sync, structured notes stored by data-slide-id rather than page number, rehearsal timing, laser pointer and circling, black/white screen and audience-screen freeze, with no cloud relay, phone remote or AI coach), and under Codex it offers to generate imagery with GPT-Image 2.0 plus multi-platform covers such as a 21:9 article header or a 3:4 carousel. It states its own bad fits: dense tables, training courseware, multi-editor collaboration, and it cannot export PPTX. We did not install it and produce a real deck, so it is graded as pending reproduction.2026-09-23Framelink MCP for Figma: design files compressed into usable context, not screenshotsAlso known as Figma-Context-MCP, npm package figma-developer-mcp, TypeScript, MIT, and a third-party implementation rather than the Dev Mode MCP server Figma ships itself. It exposes only two tools: get_figma_data (read-only) and download_figma_images (removable with --skip-image-downloads, and confined to writing inside --image-dir). The narrow surface is deliberate; the engineering sits in the fetch → simplify → serialize pipeline behind it. Simplify uses composable extractors (layout, text, visuals, component, plus the allExtractors, layoutAndText, contentOnly, visualsOnly and layoutOnly bundles) to compress the raw Figma response into semantic layout and styling in a single tree walk, with maxDepth and nodeFilter, and outputs tree (default, cheapest in tokens), yaml or json. The part worth copying is that compression effectiveness is a metric computed on every call: rawSizeKb against simplifiedSizeKb, rawNodeCount against simplifiedNodeCount, component/instance/text/image node counts, namedStyleCount (the source comment states a high count is a design-system maturity signal), hasVariables (whether Figma Variables are used at all), and separate fetch/simplify/serialize timings; the same hooks drive MCP progress notifications and heartbeats so a large file taking tens of seconds does not look hung. Judgement shows in the details: node IDs accept both the plain and the deeply nested instance shape, the depth parameter description says do not use it unless the user explicitly asks, the proxy default deliberately avoids installing EnvHttpProxyAgent when no proxy variables exist so a stale variable cannot route traffic through an intermediary returning 403, and stdio mode warns at startup when --image-dir is unset. Credentials come from FIGMA_API_KEY or FIGMA_OAUTH_TOKEN; telemetry can be disabled with --no-telemetry or DO_NOT_TRACK=1 and secrets are redacted from reports. 15.9k stars. It generates no frontend code - the output is design facts, and translating them into React, Vue or SwiftUI is your coding agent job. We have not run it against a Figma file of our own and did no screenshot-versus-structured-data accuracy comparison, so it is graded as pending reproduction.2026-09-20Design knowledge as searchable local data: 79 styles, 192 palettes, 119 UX rulesv2.13.0, MIT, 129k stars, site uupm.cc. Instead of lecturing the model it ships structured catalogues the agent queries before writing code: 79 UI styles, 192 palettes with reasoning, 74 type pairings, 119 UX rules, 105 icon suggestions, 17 GSAP presets, 25 chart types, 22 stacks and 34 landing-page patterns.
SOTA laddertracked 28
GPT Image 2.5: the image flagship that writes multi-round editing consistency into the API contract1423/1526CQwen-Image-2.1: one set of open weights for generation, local editing and transparency7BCFLUX: the line that moved image generation from "produce a picture" to controllable editing with open weightsApache-2.0CNano Banana Pro: the image editing model that turned "edit it by talking to it" into a productDKling 3.0 series: native audio, 4K frames, and parameterised camera language as the moat3.0 / Omni / TurboCMeshy 7: the 3D asset platform that finishes what comes after generation, from topology and PBR to rigging, printability, and MCP<1minC
GenAI-Bench, DrawBench, arena preference, controllability
Image GenerationAudio & Video
61Video generation and speech/audio in one channel: physical and temporal coherence for text/image-to-video, TTS naturalness and voice cloning, music generation, and ASR
Latest demos2
![Fish Audio upgrades its ASR model: speaker identification and inline emotion cues such as [laughter]](/static/img/twitter/FishAudio_2104617131920748857.card.webp)
0:59Fish Audio upgrades its ASR model: speaker identification and inline emotion cues such as [laughter]
@FishAudio · 148 views · 2026-09-28
ASRSpeech RecognitionEmotion Recognition
0:50UniMate: one unified model animates arbitrary skeletons from text prompts
@LinzhanMou · 93 views · 2026-09-28
UniMateText-to-MotionRoPE
Papers & deep reads4
2026-09-24Rolling-WAM: World Action Models with Rolling ImaginationWorld Action Models (WAMs) couple action generation with future visual prediction for robotic manipulation. However, completing the joint video-action denoising process at each replanning cycle incurs substantial latency, delaying action updates and limiting closed-loop responsiveness. We present Rolling-WAM, a formulation that distributes joint denoising across successive replanning cycles. Our method maintains a sliding window of video-action chunks at staggered noise levels. At each step, a rolling noise schedule fully denoises the imminent action chunk for execution, while partially refining farther-future chunks. As the window advances with new camera observations, the retained future chunks continue their denoising process. This distributes the computational cost over time while carrying an evolving visual-action context across chunk boundaries. Evaluations on LIBERO, RoboTwin, and a real-world Unitree G1 humanoid show that Rolling-WAM achieves competitive manipulation performance. By removing the need to denoise the entire prediction horizon from scratch, it delivers a 4.5x steady-state replanning speedup over standard joint WAMs.
2026-09-24BeyondRetarget: Learning Executable Humanoid Motions Directly from Monocular VideoBeyondRetarget removes the SMPL intermediate representation and maps monocular RGB video end-to-end to executable humanoid motions: 18 semantic keypoints with per-link non-uniform scaling form unified supervision, a shared representation decodes onto 8 humanoids, and contact-aware refinement yields 26.4mm error, zero collapse, and 193ms latency enabling real-time visual teleoperation.
2026-09-22MachEmbodied-U0: Unified Understanding and Generation Model for Embodied IntelligenceGeneral-purpose robot control requires models to understand task intent, identify where to interact, capture how the scene evolves, and generate precise actions. Vision-language-action models provide strong semantic priors but typically do not explicitly model scene dynamics, while world-action models couple visual prediction with control without necessarily exposing the task-relevant semantic and spatial structure needed for fine-grained manipulation. We present MachEmbodied-U0 (ME-U0), a unified embodied foundation model connecting understanding and generation experts through a Mixture-of-Transformers architecture. Subtask prediction and affordance grounding guide joint visual-dynamics and action generation via flow matching. Visual dynamics encompass future RGB, depth, surface normals, and optical flow, providing complementary supervision for appearance, geometry, and motion. Multi-rate Rotary Position Encoding (MRPE) aligns visual dynamics with fine-grained control. We pretrain ME-U0 on approximately 4,200 hours of curated demonstrations from robotic datasets and egocentric datasets. Using only the supervision natively available in each downstream benchmark, ME-U0 achieves an average score of 17.66 on the RoboDojo simulation benchmark and average success rates of 99.0\% and 82.5\% on LIBERO and LIBERO-Plus, respectively. We additionally validate ME-U0 on real-world robotic manipulation tasks, demonstrating its effectiveness beyond simulation. Without corresponding downstream supervision, ME-U0 further demonstrates zero-shot subtask prediction, affordance grounding, and visual dynamics on simulated and real-world observations. Overall, ME-U0 combines competitive downstream control performance with transferable task-grounding and visual-dynamics capabilities across simulation and the real world.
2026-09-04GE-Act 2.0: Pretraining and Scaling a World-Action Model for Robotic ManipulationGE-Act 2.0 combines a control-oriented autoencoder, a single-step visual planner, an inverse dynamics model, and knowledge-aligned selective optimization to pretrain a world-action policy from scratch on manipulation data. Scaling co-training data from 300 to 30,000 hours raises zero-shot OOD success to 44.1% on G1-OP and 31.1% on G2-90D without task-specific fine-tuning.SOTA laddertracked 61
Eleven Music v2.5: the closed-source row that turns music generation into a callable pipeline10 minCEleven v4: the generation that pushes cloning fidelity far enough to reproduce bad training data~100msCSuno: the platform that pulls the second half of music production inside one product6.8721CVoiceStudio: the commercial voice stack moved onto local hardware and opened to agents by default42,831AGemini Omni: moving the unit of video delivery from one attempt to a conversational editing process1516AEleven v3: the commercial audio stack that covers speaking, listening and singing70+C
Video: VBench, EvalCrafter, physical & temporal coherence. Audio: TTS naturalness, voice cloning similarity, music generation, ASR WER
Audio & Video3D & World Models
863D asset generation, scene reconstruction, long-horizon world prediction
Papers & deep reads2
2026-09-23MCP for Blender: 32 tools that make Blender a 3D workspace an LLM can driveFormerly blender-mcp, a third-party community plugin (the README states on its first screen that it is not made by Blender), MIT licensed, Python, and the highest-starred entry among 3D MCP servers. The PyPI package is now mcp-for-blender, yet uvx blender-mcp still runs and existing configs need no change. Two components only: an addon inside the Blender process runs a socket server that actually executes commands (the only place with bpy, the current scene and the UI thread), while a separate MCP server process translates protocol, joined by JSON over TCP on localhost:9876 by default. We counted the @mcp.tool() decorators in the source: 32, in four layers - perception (get_scene_info, get_object_info, get_viewport_screenshot returning an MCP Image, closing an act-then-look loop), reference lookup (bpy_api_lookup and describe_node_type so the model queries instead of guessing socket order and enum names), execution (execute_blender_code as the escape hatch, set_texture, export_scene to GLB/FBX), and asset retrieval and generation (Poly Haven with semantic search, importing the .blend source and writing attribution properties; Poly Pizza with about 10,600 low-poly models, size normalisation and automatic CC-BY credit; Sketchfab; Hyper3D Rodin text-to-3D and image-to-3D; Tencent Hunyuan3D distinguishing the mainland ai3d endpoint from the international hunyuan one). On security, arbitrary code execution is the design rather than a bug: an optional safe mode (BLENDER_MCP_SAFE_MODE=1) statically validates scripts before running, blocking file access, subprocesses, network and persistent code while modelling, rendering, saving, import and export pass normally, and blocked scripts return to the model with the reason so it can retry. Content telemetry is off by default, and a dedicated disable_telemetry tool can only turn collection off. 29.2k stars. We have not run Blender with this addon locally, so it is graded as pending reproduction.
2026-09-21DexTacWAM: A Visuo-Tactile World-Action Model for Dexterous ManipulationDexterous manipulation depends on contact dynamics that are often only partially observable from vision. Recent World-Action Models (WAMs) couple predictive video world modeling with action generation, but remain largely vision-centric and therefore cannot directly model these contact dynamics. We present DexTacWAM, a visuo-tactile WAM that encodes each fingertip independently, aggregates the resulting features through a finger- and pose-aware tactile compressor, and injects the tactile latent into a video diffusion world model for joint visuo-tactile world modeling. Across six contact-rich dexterous manipulation tasks on a 22-DoF bimanual platform, DexTacWAM achieves the highest score on every task, averaging 70.6 versus 38.0 for the strongest baseline. Ablations attribute the gain to modeling contact evolution as part of the predicted world state rather than tactile conditioning alone: removing tactile world modeling reduces the four-task mean from 74.7 to 26.6 while keeping the same tactile features and action expert. After four hours of tactile-encoder adaptation with a frozen pretrained vision VAE, our continual vision-to-touch learning extends the pretrained video model to touch using roughly 100 demonstrations per task without tactile midtraining, while retaining visual prediction quality within 0.5 dB of vision-only counterparts. The compressor retains 89.4% of pre-fusion contact recall while enabling 2.26x faster training and 1.29x faster inference. Together, these results show that pretrained video priors can be extended to distributed multi-finger contact dynamics in a data- and compute-efficient manner.
2026-09-21DexTacWAM: A Visuo-Tactile World-Action Model for Dexterous ManipulationDexterous manipulation depends on contact dynamics that are often only partially observable from vision. Recent World-Action Models (WAMs) couple predictive video world modeling with action generation, but remain largely vision-centric and therefore cannot directly model these contact dynamics. We present DexTacWAM, a visuo-tactile WAM that encodes each fingertip independently, aggregates the resulting features through a finger- and pose-aware tactile compressor, and injects the tactile latent into a video diffusion world model for joint visuo-tactile world modeling. Across six contact-rich dexterous manipulation tasks on a 22-DoF bimanual platform, DexTacWAM achieves the highest score on every task, averaging 70.6 versus 38.0 for the strongest baseline. Ablations attribute the gain to modeling contact evolution as part of the predicted world state rather than tactile conditioning alone: removing tactile world modeling reduces the four-task mean from 74.7 to 26.6 while keeping the same tactile features and action expert. After four hours of tactile-encoder adaptation with a frozen pretrained vision VAE, our continual vision-to-touch learning extends the pretrained video model to touch using roughly 100 demonstrations per task without tactile midtraining, while retaining visual prediction quality within 0.5 dB of vision-only counterparts. The compressor retains 89.4% of pre-fusion contact recall while enabling 2.26x faster training and 1.29x faster inference. Together, these results show that pretrained video priors can be extended to distributed multi-finger contact dynamics in a data- and compute-efficient manner.SOTA laddertracked 86
Tripo: image-to-engine-ready-asset compressed into a ten-second step<10sCHunyuan3D 2.1: the open-weight 3D generation baseline and default starting point for self-hostingCScenario: style-consistent asset volume for game art, not one beautiful picture at a timeCTRELLIS: a research baseline that unifies 3D asset representations through a sparse structured latentD
Geometric consistency, scene reconstruction, long-horizon prediction
3D & World ModelsDesign & Frontend
24Visual design, UI components, web and slide layout
Papers & deep reads2
2026-09-07img2threejs: Rebuilding Reference Images as Code-Only Procedural Three.js ModelsAn agent skill for Claude Code, Codex and OpenCode that rebuilds a reference image as a code-only procedural Three.js model: an eight-pass sculpting pipeline with deterministic gates, shipping a diffable TypeScript factory instead of a mesh.
2026-07-25Fashion-3DLR: A Controllable 3D Garment Generation Using Pairwise Fashion Elements for Intelligent DesignAI-generated content (AIGC) has made significant progress, with 2D generative models becoming ready-to-use tools for the digital fashion industry. However, 3D garment generation remains in its nascent stage, where in the realm of fashion, the semantic information of diverse design elements exhibits intricate coupling relationships in 3D representations, posing substantial challenges for generating diverse 3D garments. In this work, to handle the above problem, We introduce Fashion-3DLR, a novel 3D garment generation framework that utilizes diverse design elements to create high-quality, versatile 3D garment assets. Specifically, to bridge the semantic gaps between different fashion elements, we propose a Garment Feature Fusion Diffusion Transformer (GFF-DiT) module to integrate 2D fashion design elements, e.g., sketch and texture, into latent space. Within the latent space, we then employ a rectified flow transformer to generate geometry latents, which can be decoded into various 3D garment representations, including 3D Gaussians and meshes. Furthermore, we integrate Fashion-3DLR into downstream tasks, achieving the 3D Gaussian Splatting (3DGS)-driven cloth physical simulation and mesh-based virtual try-on. Experimental results indicate that Fashion-3DLR surpass the previous state-of-the-art methods, which verify that the proposed work can generate well-structured, non-watertight garments capable of physical simulation and virtual try-on, underscoring its potential as a versatile 3D garment design tool.
2026-07-25Fashion-3DLR: A Controllable 3D Garment Generation Using Pairwise Fashion Elements for Intelligent DesignAI-generated content (AIGC) has made significant progress, with 2D generative models becoming ready-to-use tools for the digital fashion industry. However, 3D garment generation remains in its nascent stage, where in the realm of fashion, the semantic information of diverse design elements exhibits intricate coupling relationships in 3D representations, posing substantial challenges for generating diverse 3D garments. In this work, to handle the above problem, We introduce Fashion-3DLR, a novel 3D garment generation framework that utilizes diverse design elements to create high-quality, versatile 3D garment assets. Specifically, to bridge the semantic gaps between different fashion elements, we propose a Garment Feature Fusion Diffusion Transformer (GFF-DiT) module to integrate 2D fashion design elements, e.g., sketch and texture, into latent space. Within the latent space, we then employ a rectified flow transformer to generate geometry latents, which can be decoded into various 3D garment representations, including 3D Gaussians and meshes. Furthermore, we integrate Fashion-3DLR into downstream tasks, achieving the 3D Gaussian Splatting (3DGS)-driven cloth physical simulation and mesh-based virtual try-on. Experimental results indicate that Fashion-3DLR surpass the previous state-of-the-art methods, which verify that the proposed work can generate well-structured, non-watertight garments capable of physical simulation and virtual try-on, underscoring its potential as a versatile 3D garment design tool.SOTA laddertracked 24
Design & FrontendDocs & Slides
4Slides, Word, Excel, PDF generation and editing
Blog
63
RLT: Physical Intelligence masters the last millimeter with RL Tokens
Physical Intelligence (Sergey Levine's team, Mar 19, 2026) present RL Tokens (RLT): freeze a pretrained VLA, attach an encoder-decoder that compresses its internal embeddings into a bottleneck RL token, then run online RL with a ~1M-parameter actor-critic on the real robot, refining only the critical phase (editing VLA action chunks rather than generating from scratch, anchored by a BC regularizer plus reference-action dropout). Across four sub-millimeter tasks the critical phase speeds up by up to 3x, screw success goes 20% to 65%, and half of Ethernet insertion episodes beat every human teleoperation demo - with just 15 minutes of real robot data.
VLAOnline RLRLPDPhysical IntelligenceDexterous Manipulation
Building an Agentic Harness with Jev: A Builder's Guide
Avid's 3,500-word build log: Jev as a decision layer inside keel, a local-first Rust coding workspace. One principle carries it: a model can suggest the next move, the host still owns the move — the host prepares the menu, Jev picks from it, the host checks again; pinned routes and live sessions bypass automatic selection; selection grants no permission. Plus how decision receipts become replayable evaluation scenarios.
Agentic HarnessJevLayaKeelRust
Inside OpenAI's Agentic Software Factory
Gergely Orosz visits OpenAI: Codex and ChatGPT Work have taken over everything; a nine-step agentic software factory with Perf Factory and Sevbot, and the rethink of IDEs and pull requests.
OpenAICodexChatGPT WorkAgentic CodingSoftware Factory
AI Games
2Under constructionThe games system is still under construction — everything listed starts in-page, and the library keeps growing.
ThreeDream FPS arena
A first-person match where every bot across from you is driven live by a policy this repo trained, the whole round stepping at a fixed 60 Hz: builtin, Rust/wasm or Rapier behind one physics interface, nothing scripted.
WASD move · mouse look · left-click fire · shift sprint · R reload · esc release
ThreeDream UE level
Unreal 5.5's first-person test map, running same-origin on this site: the level, the materials, the lights and four weapon data assets are read out of the project's own .umap / .uasset bytes, six soldiers are driven by a four-state brain on one fixed 60 Hz step, and the sidebar lists every substitution the page had to make.
WASD move · mouse look · left-click fire · shift sprint · space jump · R reload · F pick up · esc release
Agent Studio
7
Agent digital twin office
A 100 km 3D city that opens in seconds — the digital twin of the AI agents that build this site: crawling, blogs, papers, industry and investment analysis.
100km city · opens in seconds · Live task stream · Agent-maintained
Robotic arm workbench
An SO-101 6-DoF arm running MuJoCo physics in your browser: joint teleoperation, IK end-effector dragging and gripper pick-and-place into a basket. No install.
MuJoCo WASM physics · SO-101 · 6 DoF · Contacts / telemetry HUD
Joint actuator anatomy lab
Take a humanoid joint module apart layer by layer: brushless motor, magnetic encoder, planetary / harmonic / cycloidal drive and output flange. Switch architectures on one page — exploded view plus analytic kinematics.
Three gearbox types · Planetary / harmonic / cycloidal · Analytic kinematics · explode
Speech synthesis lab
Matcha-TTS mixed zh/en speech synthesis: server-side synthesis with sentence-streamed playback and full-article read-aloud for papers and blogs. Sign-in required.
Matcha-TTS · server-side · Mixed zh / en · Full-article read-aloud
CodeWaifu, your caring coding desktop sprite
A caring coding sprite on your desktop and a cockpit that never stops: the work resumes itself after a crash or a reboot. Every coding terminal, ssh session and local shell in one tree, every start, finish and permission request reported in time. Open source, one-line install.
Your caring coding desktop sprite · Resumes after crash & reboot · First-rate terminal / ssh / shell
ThreeDream simulation engine
Built for games and for physical AI: an ultra-realistic, ultra-high-performance simulation environment with ultra-low-cost procedural data; Unreal-grade visuals open and play in the browser, bridging the physical world and intelligent agents.
Games meet physical AI · Unreal-grade on the web · Open and play
RetouchPi
The first online photo editor built to be driven by AI agents: fifty-two professional tools, a layer tree with masks, PSD files that open and save back, nothing to install, and your images never leave your computer. The agent layer is on its way.
retouchpi.com · 52 pro tools · PSD read and write · your files never leave