←
m0h59,562 posts
m0h avatar

m0h

@exploraX_

exploring AI models & tools | github.com/Moh4696

◷ Joined September 2019▧ 4,191 Photos and videos
m0h avatar
m0h@exploraX_·6m

I open-sourced 8 skills for claude/codex to build 3D projects and motion animations. • FREE & MIT licensed • one-command install npx skills add Moh4696/open-3d-skills 🔗: github.com/Moh4696/open-3d-s…

Post attachment
m0h avatar
m0h@exploraX_·11h

btw I think opus result is more better, I’m running more test on website creation, games etc. But gpt6.2 sol is a cheap as opus

m0h avatar
m0h@exploraX_·11h

🚨 GPT-6.2 Sol vs Claude Opus 5.5. i gave both models the exact same prompt to recreate the White House in 3D, and here are the results. after a long wait, GPT-6.2 Sol is finally here to challenge Opus 5.5, with both models competing in the same category. which model wins? source-code in the comment.

0:57
OpenAI@OpenAINov 17

Some of you have been asking very politely. Some of you have just been asking. GPT-6.2 Astra and Sol are now available. Super efficient and built to perform where it counts—the real world. openai.com/index/introducing…

1:59
m0h avatar
m0h@exploraX_·23h

i think this model are fire but i still love opus 5.5

m0h avatar
m0h@exploraX_·23h

🚨 OpenAI finally launched GPT 6.2 Sol & Astra. and GPT 6.2 Astra built this race car entirely from code. we finally have OpenAI's response to Anthropic's Opus 5.5 and Fable 5.5. the big question is: are these models actually better? I'm running more test to answer this.

0:21
OpenAI@OpenAINov 17

Some of you have been asking very politely. Some of you have just been asking. GPT-6.2 Astra and Sol are now available. Super efficient and built to perform where it counts—the real world. openai.com/index/introducing…

1:59
m0h avatar
m0h@exploraX_·Nov 17

🚨Opus 5.5 is still the most efficient model for building animated websites. this website was built with just a single prompt + ref image. I open-sourced all the skills i used to make this possible. links in the comments. if you want to build websites like this, check out my guide below.

0:40
m0h@exploraX_Oct 30

m0h @exploraX_ Oct 30 Article How to Build a 3D Animated Website with Opus 5.5 (Full Guide) even grandma can now build a cool animated website like this. in this article, i’m going to walk you through exactly how I built this animated website, from start to finish. every message I sent,

m0h avatar
m0h@exploraX_·Nov 17

sam altman: “we would go public at some point.” sam said this yesterday during an interview when asked about Anthropic public sale going live this week. he also argued that access to private markets shouldn’t be limited to investors alone. he believes giving the public exposure to these opportunities is a good thing.

5:45
m0h avatar
m0h@exploraX_·Nov 16

here are four (4) skills to turn claude/codex into a full video editing studio: > video-use: github.com/browser-use/video… > ffmpeg skill: github.com/kajisho5/ffmpeg-s… > palmier pro: github.com/palmier-io/palmie… > videocut skills: github.com/Agentchengfeng/ch… all FREE and opensource.

Post attachmentPost attachmentPost attachmentPost attachment
m0h@exploraX_Nov 11

here are four (4) skills to turn claude/codex into a full motion graphic studio: > remotion: github.com/remotion-dev/remo… > hyperframe: github.com/heygen-com/hyperf… > manim: github.com/3b1b/manim > motion skills: github.com/iart-ai/motion-sk… all FREE and opensource.

Post attachmentPost attachmentPost attachmentPost attachment
m0h avatar
m0h@exploraX_·Nov 16

The Recreate Prompt: Build a photoreal, real-time Mount Fuji as a static, no-build-step three.js r169 page (ES modules + import map, three.js vendored locally) served by a tiny serve.py on port 9013 (no-cache headers, threaded, listen backlog 256 because the page loads ~40 modules at once). Every asset is generated in code: no downloaded models, image files, HDRIs or AI textures. Textures are painted on 2D canvases at load or computed in shaders. Heavy precomputation runs in Web Workers behind a minimal loading bar and is cached in IndexedDB so the second load opens in about a second. The goal: someone who knows Fuji recognises the exact mountain, not a generic volcano. Real geography, real scale, real sun. Real place names are fine; no rail operator, airline or brand names or logos. 1. World frame Metres, y up, origin at the summit's ground position, sea level y = 0, +x east, -z north. 80 x 80 km world (Fuji Five Lakes, Misaka range, Suruga Bay all inside). Far plane 120 km. Logarithmic depth buffer. Every custom ShaderMaterial includes the logdepthbuf chunks. For the terrain, write log depth in the vertex stage instead (gl_Position.z = (log2(max(1e-6, 1.0 + w)) * logDepthBufFC - 1.0) * w) so early-Z still culls hidden ridges; three's per-fragment log depth disables it and costs 15-47 % of frame time. Earth curvature in every vertex shader: y -= |dxz|^2 * uCurv. Real sun and moon: NOAA solar position for 35.36 N, 138.73 E from a JST date and time. 2. The mountain Float32 heightfields in a worker: 4096 x 4096 over 80 km plus 2048 x 2048 over the central 16 x 16 km cone, blended with smoothstep(80, 500, distance to cone edge). Mip chains for both. The same data feeds the GPU (float textures) and a JS sampler that matches the GPU bilinear exactly (used for raycasts, camera clamping and object placement). Concave-up radial profile, relative to a ~850 m base: coneRel(r_km) = 2615.14 * exp(-(r / 6.3762)^1.231) + 591.92 * exp(-r / 1.0659), giving ~3,776 m at the rim, ~3,000 m at 2 km, ~2,300 m at 4 km, ~1,500 m at 8 km, ~1,000 m at 14 km. 8 % stretch on the NW-SE axis; the north apron stays higher than the south. Ouchi-in crater: ~700 m across, ~230 m deep, eight rim peaks, Kengamine highest on the SW rim. From afar the top must read flat and slightly notched, never rounded. Hoei crater on the SE flank (~1.3 km, rim ~2,700 m, open downslope) with the Hoei-zan hump below it: a bulge on the right skyline from the south, on the left from Gotemba, hidden from Lake Kawaguchi. Osawa-kuzure on the WSW flank down to ~2,200 m, up to 500 m wide and 150 m deep, jagged red-brown headwall. Radial gullies from ridged noise in polar coordinates at periodic angular levels (30 / 60 / 120 / 240 per turn, increasing with radius), then a 300,000-droplet hydraulic erosion pass with deposition; smooth lava aprons below ~2,000 m. ~60 flank cones mostly on the NW-SE axis, lava ridges heading north. Surroundings as capsule chains with real water levels: Kawaguchi (830 m, bent, narrow waist with a bridge), Yamanaka (980 m), Sai, Shoji, Motosu; Aokigahara forest; Fujiyoshida; the Misaka, Tenshi and Tanzawa ranges; Mount Ashitaka; Hakone; Suruga Bay coast and the Fuji River. A land-cover map (forest, fields, town, tea) computed in the same worker. Far skyline beyond the world: a ring mesh 58-98 km out with the Southern Alps (W), Yatsugatake (NW), Chichibu (N), Tanzawa (E) and Izu (SE), open sea to the SW, seasonal snow. Past the 80 km edge the terrain shader fades to a low plain so nothing looks cut off. 3. Terrain rendering Camera-centred geometry clipmap: 10 rings of 384 x 384 quads, 2 m cells doubling per ring, CDLOD-style morphing, inner-ring discard with a 2-cell overlap so cracks never open. 128-quad rings are too coarse: at 18 km they turn the notched summit into a hump. When the camera is more than ~8.8 km from the summit, add a fixed 1024 x 1024 "hero" mesh (±7,600 m, 15 m spacing, with a skirt) for the cone so the silhouette is exact; sink the rings 6 m just outside it so the seam never shows. Normals from central differences, fbm micro-displacement and detail normals near the camera. Clamp the camera above the ground everywhere. Materials blended by height, slope, curvature and noise: scoria with rust bands, grey ash fans, treeline band, conifer and broadleaf forest (Aokigahara darkest), fields, towns (with night window lights), red-brown rock on steep walls. Skip the snow, forest and detail work for fragments that cannot see them. Seasonal snow: winter cover to 1,500-2,000 m; late-spring streaks driven by curvature (gullies keep snow, ridges lose it); snow-free summer; autumn dusting on the top 300-500 m with autumn colours below. Subsurface tint, sparkle, sharper normals. 4. Light, sky and atmosphere Hillaire-style atmosphere: transmittance LUT 256 x 64, multiple-scattering LUT 32 x 32, sky-view LUT 256 x 128 with absolute azimuth and two lights (sun and moon). Rayleigh, Mie, ozone and a low valley haze. Ambient: SH9 projected from the sky-view LUT every time the sun moves. Aerial perspective integrated per fragment through the same medium with adaptive steps (1 near, up to 8 far), so it works in mirror passes and on every built-in material. Patch built-in three.js materials via onBeforeCompile for curvature, aerial perspective, terrain sun visibility and cloud shadow, so houses, trees, the train and the pagoda sit in the same light. Terrain self-shadowing from a GPU sun-visibility map (2048 x 2048, a horizon march over the height textures with Earth curvature and a distance-growing penumbra), rebuilt in ~2 ms whenever the sun moves. It covers all 80 km and draws the mountain's own long shadow. Object shadows: one shadow cascade following the camera (2048 at medium, 4096 at high); the terrain shader samples the same map. Exposure computed from sun / moon elevation (deterministic, no adaptation history) and folded into light intensities so night keeps precision in half-float buffers. ACES filmic with toneMappingExposure = 0.6 to cancel three's /0.6. Sky dome: sun disk with limb darkening, twilight, Earth shadow, Belt of Venus, sparse stars, a faint Milky Way, the moon, city glow. 5. Clouds Raymarched from a 128^3 Perlin-Worley texture built in a worker from precomputed lattices (~1 s). Density presets: cumulus; unkai sea of clouds at ~1,800 m; the kasa-gumo cap cloud as a hat-shaped dome with sharp tops, soft undersides, powder and silver lining plus two lens layers; an optional lenticular downwind; summit spindrift plumes. Live mode: half resolution, 32 view steps, 3 light steps, temporal reprojection. Capture mode: full resolution, 84 view steps, 5 light steps, no history. Adaptive steps with empty-space skipping. A top-down cloud shadow map. 6. Water Hand-written planar mirror per lake height (oblique near-plane clip, HalfFloat target with depth texture, scale 0.42 at medium and 0.75 at high or capture, MSAA 4 / 2), active only for lakes on screen. Hide petals, mist, birds and near trees in mirror passes. Wind slider from mirror-still to breezy; Schlick fresnel, dark green-blue body, shoreline darkening, wet band. Suruga Bay as a 240 x 110 km sea plane with real waves, sky reflection and sun glitter. Two drifting mist sheets on Kawaguchi at dawn. 7. Vegetation and life Instanced code-built trees: stacked noisy cones for conifers, smooth jittered icosahedron clusters for broadleaf and maples, cherries with falling petals. LOD tiers: full meshes within ~270 m, simplified to ~1 km, impostors from an atlas rendered at load to ~3.4 km, terrain colour beyond. Wind sway driven by scene time. Black kites circling over the lakes; swan pedal boats and a sightseeing boat on Kawaguchi. 8. Human details Chureito five-storey pagoda: stone base, five roofs with upturned corners, white walls, vermilion posts and brackets, railings, bronze finial with nine rings, a long stone stairway, cherry trees, a clear glade around it. Fujiyoshida: ~7,000 instanced houses, a main street, roads, utility poles with sagging wires, a large vermilion torii aligned with the mountain, the Kawaguchi bridge, the 5th-station toll road and lake roads. Yoshida trail zigzagging from 2,300 m with ~14 stone-walled huts set beside the switchbacks, a rim path, a shrine and torii on the rim, the summit station on Kengamine; glowing huts and a chain of moving headlamps at night. A long white unbranded high-speed train with a blue stripe (16 cars) on a viaduct across the coastal plain, a girder bridge over the gravel bed of the Fuji River, and tea fields as curved hedge rows (instanced, capped ~110,000). 9. Post and performance EffectComposer on a HalfFloat 4x MSAA target, very subtle per-shot bloom, a fade pass for dip-to-black cuts. No chromatic aberration, lens dirt or heavy vignette. ?q=low|med|high: low drops mirrors, MSAA and shadows and halves forest radii; high raises tree radii, mirror resolution and the shadow map. Target 60 fps at 1080p on an M-series Mac; the mirror and sea-of-clouds shots are the expensive ones. 10. Shots (~9 s each, slow eased moves, quick dip-to-black cuts) Sakasa Fuji, Lake Kawaguchi: the upside-down mountain at dawn, autumn maples, mist. 18 Nov 06:40. Chureito pagoda above Fujiyoshida, cherry blossoms, striped late-spring snow, descending. 12 Apr 08:20. Red Fuji from the east, first light on the snow-free cone, kites. 28 Aug 05:24. Kasa-gumo cap cloud from the WSW, Osawa-kuzure, spindrift, winter afternoon. 22 Jan 15:25. Ouchi-in summit crater, drone over the rim, sea of clouds. 6 Aug 07:05. Night climb: headlamps zigzagging up the Yoshida trail, huts, stars, Milky Way, town lights, fading to goraiko. 10 Aug 03:30-04:18. High-speed train on the coastal plain, tea fields, the Hoei hump on the right skyline. 15 Jan 12:10. Diamond Fuji from Lake Yamanaka, winter: solve the camera by back-projecting the summit along the real sun vector (sun elevation 0.5-18 degrees; at Yamanaka it lands near 10 degrees), kites. Plus a free camera: drag to orbit, scroll to zoom, WASD fly, Q / E down / up, Shift for speed, never below ground. 11. HUD and URL Minimal HUD: title "富士山 Mount Fuji 3,776 m", shot chips (keys 1-8), play / pause (space), date and time sliders (JST), season, cloud preset (clear, cumulus, cap cloud, sea of clouds), lake wind, fps; H hides it. URL: ?shot=N&hold&hud=0&t=0&k=0.5&season=spring&time=05:40&q=low|med|high&capture&nocache&pr=1. 12. Tooling window.__fuji.renderAt(t, shot, k, overrides) renders a deterministic frame: every animation (water, clouds, trees, birds, boats, train, headlamps, petals) runs on scene time, never the wall clock. tools/record.py (Playwright headless Chromium with Metal ANGLE, 127.0.0.1 URLs, poll readiness with evaluate instead of wait_for_function): an EDIT list of the 8 shots totalling 54 s with 0.3 s dips, writing a silent 1080p60 H.264 X clip (video/mount-fuji.mp4, ~16 Mb/s) plus a CRF 11 master. tools/shoot.py: stills to captures/, including named reference views (Kawaguchi, Fuji city, Gotemba). 13. Realism checklist Kawaguchi silhouette (symmetric concave slopes, flat notched top, no Hoei hump); Hoei on the right from the south; Osawa-kuzure from the west; eroded, branched gullies; late-spring streaks; aerial perspective (hazy blue-grey base, crisp summit); shadows including the mountain's own; real sun positions and an aligned Diamond Fuji; correct reflections; no cracks, popping, shimmer or banding; life in every shot; clean console; frame-exact clip. 14. Process PLAN.md first. Build the heightfield alone and check hillshades in node, then the clipmap, then sky and aerial perspective, then silhouette checks from Kawaguchi, Fuji city and Gotemba until the profile, notched top and Hoei hump read correctly. Then snow, clouds, lakes, vegetation, human details, shots, tooling and the clip, with a preview and screenshot after each stage. Ship README.md, PROMPT.md and an MIT LICENSE. Keep the git repo private until told otherwise.

m0h avatar
m0h@exploraX_·Nov 16

🚨 Qwen 4 Max just launched, so I put it to the test against Opus 5.5 to build Mount Fuji in 3D. I gave both models the same prompt, with no 3D assets. Alibaba claims the model excels at 3D modeling and even outperforms Opus 5.5 on some LBs. what do you think of the results?

0:54

📢 Meet Qwen4-Max, your new AI teammate for work, creativity, and discovery! Qwen4-Max is built on a new MoE architecture with native vision and a 1M-token context window. Highlights: - Outperforms Qwen3.8-Max with up to ~50% fewer tokens, while average cost per task drops over 60% in our internal evaluation. - Advances real-world productivity through greater proficiency with professional tools, stronger long-horizon execution, and higher-quality deliverables. - Handles real tasks, including 3D modeling, game development, manufacturing, finance, law, healthcare, and frontier scientific discovery. - $1.5/M tokens for input, $4.5/M tokens for output, and $0.075/M tokens for implicit caching — all lower than Qwen3.8-Max, with implicit caching down by 70%. The API is live on QwenCloud. You can also try Qwen4-Max on Qwen Studio, QwenWork and Qoder. Start building today! 🚀 - Blog: qwen.ai/blog?id=qwen4-max - API: qwencloud.com/models/qwen4-m… - Token Plan: qwencloud.com/pricing/token-… More to come! The open weights of smaller Qwen4 models will be released. Stay tuned. 👀

Post attachment
m0h avatar
m0h@exploraX_·Nov 14

🚨Fable 5.2 built this seaport in just 30 minutes, entirely from code. the result is incredibly realistic, and no 3D assets were used. what do you think about the output? live demo and source code in the comments.

0:35
m0h avatar
m0h@exploraX_·Nov 13

here are 4 plugins that can reduce Claude Code/Codex token usage by up to 70%. all are FREE and open-source. 1️⃣ Claude Code Setup: recommends hooks, skills, and MCP servers for your project. 🔗: github.com/anthropics/claude… 2️⃣ Claude Mem: saves context across sessions to reduce repeated tokens. 🔗: github.com/thedotmack/claude… 3️⃣ Headroom: compresses tool outputs to save tokens. 🔗: github.com/headroomlabs-ai/h… 4️⃣ OmniRoute: routes across 1,400+ models and compresses prompts. 🔗: github.com/diegosouzapw/Omni… to install, just paste the repo link to your agent.

Post attachmentPost attachmentPost attachmentPost attachment
m0h@exploraX_Nov 11

here are four (4) skills to turn claude/codex into a full motion graphic studio: > remotion: github.com/remotion-dev/remo… > hyperframe: github.com/heygen-com/hyperf… > manim: github.com/3b1b/manim > motion skills: github.com/iart-ai/motion-sk… all FREE and opensource.

Post attachmentPost attachmentPost attachmentPost attachment