正在加载视频...

视频加载失败

GLM 5.1 vs GLM 5.2 6 advanced HTML canvas challenges: 💧 Ink diffusing in water ⚔️ Energy-blade duel 📱 Slide to unlock 🅿️ 360° parking assist 🔥 Burning letter to ash 🏠 Build-a-house sequence Pure canvas, zero libraries.

347,708 次观看 • 3 个月前 •via X (Twitter)

48 条评论

rijndael 的头像
rijndael3 个月前

glm5.1 lightsaber duel is sweet

stevibe 的头像
stevibe3 个月前

Yeah especially when they crossed!

Tom Turney 的头像
Tom Turney3 个月前

dang dude great visuals and choreography.

stevibe 的头像
stevibe3 个月前

Glad you liked the rhythms!

stevibe 的头像
stevibe3 个月前

The prompt (1/2): 💧 Ink diffusing in water Create a single HTML file with a canvas animation of a single drop of black ink falling into a glass of clear water and diffusing. On impact the ink should bloom outward in turbulent, billowing plumes — fractal tendrils curling and branching as they spread and slowly disperse into faint clouds. Use real fluid-like motion (advection and diffusion), not simple expanding circles. The ink should have density variation, with dark dense cores and wispy translucent edges. Soft lighting from above. 60fps, no external libraries. ⚔️ Energy-blade duel Create a single HTML file with a canvas animation of a short, cinematic duel between two hooded warriors wielding glowing plasma blades — one blade blue, one red. Set it on a dark, atmospheric backdrop with drifting embers or mist. The warriors clash in a fast, choreographed exchange of swings and parries. The core effect: each blade leaves a smooth, motion-blurred light streak that traces its arc through the air, with a bright white-hot core, a colored glow, and a trail that fades over a few frames. Blades should cast dynamic colored light onto the fighters and the ground, and clashes throw off sparks. Fluid, weighty movement — no stiff robotic motion. 60fps, no external libraries. 📱 Slide to unlock Create a single HTML file with a canvas animation of a smartphone lock screen unlocking. Show a phone screen with a large clock and date over a soft gradient wallpaper, and a "slide to unlock" track at the bottom with a draggable handle. A shimmering light sweep should animate continuously across the "slide to unlock" text (a moving highlight gradient). Then animate the handle sliding all the way across; on completion, the lock screen smoothly slides/fades away to reveal a home screen — a grid of rounded app icons (original designs) that gently zoom and fade into place, with a subtle parallax on the wallpaper. Polished, fluid, Apple-grade easing throughout. 60fps, no external libraries. 🅿️ 360° parking assist Create a single HTML file with a canvas animation simulating a modern car's 360° surround-view parking assist as it parallel-parks into a tight roadside gap (the classic S-shaped maneuver). Layout: Split the screen into two panels. The left panel is the car's infotainment display showing a top-down bird's-eye ("around-view") composite — the car rendered from directly above in the center, the road surface and curb, and an empty parking gap bracketed by two parked cars. The right panel shows the steering wheel from the driver's seat, rotating in sync with the maneuver. Dynamic guidelines: Overlay predicted-path lines projecting from the car in the bird's-eye view — two curved trajectory lines marking the car's width, divided into colored distance bands (green = far, yellow = mid, red = close). These lines must bend left or right based on the current steering angle using correct steering geometry (sharper wheel angle → tighter turn radius), and they must update every single frame as the wheel turns and the car moves. Animation: Auto-play the full parallel-parking sequence — drive forward alongside the front car and stop, crank the wheel and reverse while curving in, then counter-steer and straighten as the rear end clears, and finally ease forward to center in the spot. The car body should trace a smooth, continuous S-path. Critical constraint: The steering wheel rotation, the curvature of the guidelines, and the car's actual path must stay perfectly consistent with one another at all times — turn the wheel left, the guidelines bend left, the car arcs left. Add subtle UI chrome (a distance readout, proximity warning arcs that flash red as the car nears the parked cars), soft shadows, and clean flat-design styling like a real car display. 60fps, no external libraries.

stevibe 的头像
stevibe3 个月前

The prompt (2/2): 🔥 Burning letter to ash Create a single HTML file with a canvas animation of a handwritten letter burning. Show an aged, slightly yellowed sheet of paper with visible handwritten cursive text (procedurally drawn lines are fine) resting on a dark wooden desk. After 2 seconds, a flame ignites at the bottom-right corner and spreads organically across the page — the burn front should advance with an irregular, noisy edge, never a straight line. Just ahead of the flames, the paper should darken and brown (scorching), then char black, then disappear entirely, revealing the desk beneath. Render the fire with layered particles: a bright white-yellow core, orange mid-flame, and translucent red tips that flicker and lick upward. Glowing embers should detach from the burn edge and drift upward on turbulent air currents, fading from orange to gray. Add wisps of semi-transparent smoke rising and dispersing above the flames, and a warm flickering light that the fire casts onto the surrounding desk. The entire page should be consumed in roughly 15 seconds, leaving only a few glowing ash fragments that slowly dim. 60fps, no external libraries. 🏠 Build-a-house sequence Write a single HTML file with a full-page canvas, no libraries. Animate the construction of a simple 2D cartoon house in 7 stages over ~25 seconds against a pale blue sky with a green ground line: (1) foundation - a gray rectangular slab rises from underground, (2) walls - four vertical wall segments extend upward from the foundation corners, (3) wall fill - tan/beige siding fills in between the wall frames, (4) roof - two triangular roof panels slide in from above and meet at the peak, with red shingle texture appearing row by row, (5) door - a brown door fades in on the front wall with a small gold doorknob, (6) windows - two windows appear on either side of the door with visible cross-frames and blue glass, (7) details - a chimney extends up from the roof, a small puff of smoke begins rising from it, a path of stepping stones appears leading to the door, and a small tree grows beside the house. Each stage should have a brief caption at the top ("Laying foundation", "Raising walls", etc.). Between stages, pause for 0.5 seconds. Loop continuously.

Studio Zamudio One 的头像
Studio Zamudio One3 个月前

I've started repeatedly prompting 5.2 the same prompt, and found wildly different results. I think we are getting to a point where we will have to start measuring with averages.

stevibe 的头像
stevibe3 个月前

Maybe try lowering the temperature?

Studio Zamudio One 的头像
Studio Zamudio One3 个月前

Depends on what I'm doing. It's nice to get examples to chose from. But I'm concerned with single comparisons between last and next model versions, without lots of repeat comparisons. That takes a lot of time. I guess I could just automate it. GLM is my main three.js llm.

🇵🇱🇵🇱🇵🇱🇵🇱🇵🇱🇵🇱🇵🇱🇵🇱🇵🇱🇵🇱🇵🇱🇵🇱 的头像
🇵🇱🇵🇱🇵🇱🇵🇱🇵🇱🇵🇱🇵🇱🇵🇱🇵🇱🇵🇱🇵🇱🇵🇱3 个月前

how is 5.1 better at pretty much everything

filipe 的头像
filipe3 个月前

love the visualisations! im pretty sure i'd get a ticket if i parked like glm 5.2 tho

stevibe 的头像
stevibe3 个月前

Appreciate it! Yep, two wheels on the pavement haha

Alan Palacio 的头像
Alan Palacio3 个月前

I would think that 5.1 is actually 5.2; otherwise it seems like a regresion in some cases

Tom Davenport 的头像
Tom Davenport3 个月前

I prefer 5.1 on most of these!

Canopy Wave 的头像
Canopy Wave3 个月前

Honestly, GLM 5.1 remains one of the strongest open-source models for this kind of creative coding. Still love it.

ContextBaron 的头像
ContextBaron3 个月前

that looks so cool for comparison

brock 的头像
brock3 个月前

i feel like it lowkey did worse on the old iphone animation and the house building animation was super botched

keys 🧪 的头像
keys 🧪3 个月前

Glm 5.1 seemed to be more more detailed in generated outcome may it be the tweaks they did to glm 5.2 for speed at sacrifice Over quality ?

ESchwaa 的头像
ESchwaa3 个月前

There are some aspects of 5.1 that I almost like better

Foreman Panda 的头像
Foreman Panda3 个月前

Example tests cannot fully reflect the full picture; actual work tests are still needed.

GalacticMarch 的头像
GalacticMarch3 个月前

@francoisfleuret

✦ SK ✦ 的头像
✦ SK ✦3 个月前

I don't know but I like the taste of glm it feels the close to Claude type

dalu 的头像
dalu3 个月前

those are all not 1shot results, because I've done the paper part and it didn't turn out like this

stevibe 的头像
stevibe3 个月前

They are one shot, with the same prompt per model run 5 times, and the best result was selected for each model.

dalu 的头像
dalu3 个月前

that explains it

Lee Penkman 的头像
Lee Penkman3 个月前

super satisfying. one day it will plan out and build a real house project from the ground up :D then control the robots that actually build the house :D

Daniel Yurkin 的头像
Daniel Yurkin3 个月前

love the ideas 🤩

amirmxt 的头像
amirmxt3 个月前

so good

Farrux Hewson 的头像
Farrux Hewson3 个月前

The improvements are subtle, but I like it

Saleh Abdulaziz 的头像
Saleh Abdulaziz3 个月前

👍🏻👍🏻 Thanks

Jakove_HR 的头像
Jakove_HR3 个月前

I love your test everytime Displays a Lot

Milind S 的头像
Milind S3 个月前

Very interesting set of tests

H₀ 的头像
H₀3 个月前

GLM 5.2 is much better on the ink diffusing and burning letter IMO, otherwise not so much.

TJ Kennedy 🇺🇸 的头像
TJ Kennedy 🇺🇸3 个月前

Thanks for this, great work!

Franjo 的头像
Franjo3 个月前

Love this tests

Erman Eroglu 的头像
Erman Eroglu3 个月前

Great visualizations thanks!

Hyce 🇵🇹🇩🇪🇪🇺 的头像
Hyce 🇵🇹🇩🇪🇪🇺3 个月前

can you share the prompts?

stevibe 的头像
stevibe3 个月前

Sure, I have added to this post as comments, please check.

Hyce 🇵🇹🇩🇪🇪🇺 的头像
Hyce 🇵🇹🇩🇪🇪🇺3 个月前

awesome 🙌

Lim Tzu 的头像
Lim Tzu3 个月前

Building AI Video Editor on GLM 5.2 from yesterday. 0 repos - everything from scratch.

stevibe 的头像
stevibe3 个月前

Wow how was it?

Lim Tzu 的头像
Lim Tzu3 个月前

I wish to plug this editor to AI camera I’m working on also.

John Shina 的头像
John Shina3 个月前

Wow! GLM 5.1 is not outdated yet 😁

Gökcan AKALIN 的头像
Gökcan AKALIN3 个月前

Software tasks are very small portion of all possible use cases. Consider tasks that are not easy to verify. You must enlarge your vision about model testing, if you want to distinguish yourself.

Marco Rodrigues 🔧 的头像
Marco Rodrigues 🔧3 个月前

Not a real improvement 😅

Jimsta 的头像
Jimsta3 个月前

Are you going to release your testing software that you use for all of these videos? Just curious

stevibe 的头像
stevibe3 个月前

It's a simple web app that renders the result, I can add that to BenchLocal.

Jimsta 的头像
Jimsta3 个月前

Awesome! I look forward to it, I want to do this testing with all my own wacky prompts

相关视频

Frontier Coding LLM Advanced Benchmark 🔥 - Kimi K3 Max - Qwen3.8 Xhigh - Opus 5 Max - GPT Sol 5.6 Ultra - GLM 5.2 Max - Grok 4.5 High Task: Modelden tek bir bağımsız HTML dosyasında, kum/su/ateş/duman gibi elementlerin hücresel otomat kurallarıyla etkileştiği etkileşimli bir "falling sand" simülasyonu üretmesi isteniyor; kütüphane yok, zero-shot, tek deneme. Kum yığın oluşturarak düşmeli, sıvılar seviye bulmalı, elementler yoğunluğa göre katmanlara ayrışmalı ve ateş-su-lava-asit gibi etkileşimler veri odaklı bir reaksiyon tablosundan yürümeli. Çıktı, on binlerce parçacığı ImageData ile 60 FPS'te çizebilecek performansta olmalı ve kum konisi ile bağlantılı kaplar gibi önceden tanımlı sahnelerle kendini doğrulayabilmeli. Task'ın Zorluğu: Bu görev basit bir parçacık animasyonu değil. Aynı anda dört farklı mühendislik problemini çözmek gerekiyor: - CA scheduling: Aşağı düşen hücre aynı step içinde tekrar işlenmemeli. Aksi hâlde kum tek frame’de tabana ışınlanıyor. - İstatistiksel simetri: Sol/sağ deneme sırası hücre başına randomize edilmezse kum konisi sistematik olarak bir tarafa yatıyor. - Çok-fazlı yoğunluk ve hidrostatik davranış: Yalnızca aşağı hareket ve yatay dispersion, dar taban kanallı gerçek bir U-kapta suyu karşı kola yükseltmeye yetmiyor. Basınç veya eşdeğer bir kütle transferi mekanizması gerekiyor. - Reaktif state machine: Yanıcı madde kimliği, burn lifetime, smoke/ash üretimi, gaz ömrü ve kimyasal reaksiyonlar birbirini bozmadan yürümeli. 🎉Kazanan: GPT Sol 5.6🎉 Final Puantaj ve Özet - GPT Sol 5.6 — 92.4/100 - Opus 5 — 66.1/100 - Kimi K3 — 62.1/100 - Qwen3.8 — 60.9/100 - GLM 5.2 — 55.3/100 - Grok 4.5 — 50.5/100 Sebep: GPT Sol 5.6, görevin üç klasik failure mode’unu birlikte geçen tek çıktı oldu. Kum ve duman tek step’te tam bir hücre hareket etti; iki seed’de kum konisi bias’ı en fazla %0.25 kaldı. Yoğunluk kuralı static olmayan tüm hedeflere genel olarak uygulanıyor; sand-water ve sand-ash testleri geçti. En önemlisi, connected liquid region’larındaki yüksek ve düşük yüzeyleri bulup aralarında kütle transferi yapan ayrı bir hydraulic-pressure pass’i var. Bu sayede dar açıklıklı standart U-kapta 1.24, kendi preset’inde 0.61 hücre yüzey farkına ulaştı. Density preset’i gerçekten karışık başlangıç üretiyor, chemistry’yi izole ediyor ve smoke için uzun test ömrü veriyor. 1200 step sonunda lava, water, oil ve smoke sayıları tamamen korundu; katman sıra skoru 0.906 oldu. Bu, diğer submission’ların çoğunda görülen “zaten doğru sırada dökme”, smoke’un sona ermesi veya oil’in lava tarafından yakılması problemlerini yaşamıyor. Runner-up Opus 5’in UI’si ve validator paneli daha kapsamlı; ancak çekirdek fizik açısından kritik iki iddiası çalışmıyor. Kendi U-kap preset’i 107.44 hücre farkla kaldı. Ayrıca density sahnesi reactionsEnabled=false yapsa da lava’nın ignition yolu reaction gate’inin dışında çalışıyor; burning update de gate kontrolünden önce water extinguish ve smoke/fire emission yapıyor. Sonuç olarak “chemistry off” sahnesinde water tamamen steam’e dönüştü. Bu iki hata, görsel ve ürün kalitesinin kazananı değiştirmesine izin vermeyecek kadar merkezi. Prompt uyumu — 8 puan - GPT Sol 5.6: 7.6/8 - Opus 5: 7.0/8 - Qwen3.8: 5.8/8 - Grok 4.5: 5.5/8 - Kimi K3: 5.3/8 - GLM 5.2: 4.6/8 Performans ve render pipeline - Kimi K3: 6.0/6 - Qwen3.8: 5.7/6 - Opus 5: 5.3/6 - GPT Sol 5.6: 5.0/6 - GLM 5.2: 4.8/6 - Grok 4.5: 3.2/6 Görsel kalite, UI ve resize - Opus 5: 5.0/5 - GPT Sol 5.6: 4.8/5 - Qwen3.8: 4.6/5 - Kimi K3: 3.6/5 - Grok 4.5: 3.2/5 - GLM 5.2: 3.0/5 Mimari ve test edilebilirlik - GPT Sol 5.6: 2.7/3 - Opus 5: 2.5/3 - Qwen3.8: 2.3/3 - Kimi K3: 2.2/3 - Grok 4.5: 1.9/3 - GLM 5.2: 1.0/3

Alican Kiraz

15,979 次观看 • 2 个月前

I designed a new test specifically for multimodal models: fill out a paper form. And it's much harder than it sounds. This isn't typing into an electronic field that captures your text. The form is just an image. The model has to place each form element: text, checkmarks — at the correct pixel position on the canvas itself. Results: 🟢 Kimi K2.6 → done in 3:45, 16.7k output tokens 🟡 Step 3.7 Flash → half the fields, 57k output tokens 🔴 Gemini 3.5 Flash → 489k output tokens, never finished. I had to kill it. Gemini burned ~29x more output tokens than Kimi on the exact same task, and Kimi's was the only form that actually looked filled out. The test, a mocked application form, contains some challenging parts, such as one-character-per-box fields. I provided every model the same set of tools: > get canvas size > drop probe markers to find coordinates > add text > add checkmarks > move elements > take a screenshot anytime to check their own work > ... etc So it's vision + spatial reasoning + tool use + long context, all at once. Small models (Qwen, Gemma) can't really complete this test, so I skipped them. What happened: > Kimi nailed name, DOB, ID, gender, marital status, nationality, email, phone, address, postal code — placement slightly loose, but content correct. 15 turns. Clean. > Step got maybe half right — fields dropped, "United States" landed in the email line, data floating outside boxes. Burned 1.24M input tokens doing it (81 turns of re-reading the canvas). > Gemini almost got there visually... then spiraled. By turn 40 it was issuing a delete_elements call wiping element IDs 365–425, basically erasing its own work. 31 minutes, 489k output tokens, still streaming. Terminated. The takeaway isn't "Gemini bad." This test is indeed difficult. But token efficiency is capability now. A model that needs 30x the tokens and still can't converge is going to be 30x the cost in production. Kimi K2.6 just quietly did the thing.

stevibe

25,494 次观看 • 4 个月前

sonnet 5 vs sonnet 4.6 vs opus 4.8 vs glm 5.2 – frontend tasks dropped sonnet 5 into a quick test today. same three prompts to all four models, single-shot html/canvas, no edits: • objects falling on a trampoline • rockets playing tennis • a slingshot breaking bottles ranked by speed (total across the 3 tasks): 1. opus 4.8 – 15m 09s 2. sonnet 5 – 16m 05s 3. glm 5.2 – 27m 18s 4. sonnet 4.6 – 35m 06s ranked by code shortness (total loc): 1. sonnet 5 – 1794 2. opus 4.8 – 2063 3. sonnet 4.6 – 2182 4. glm 5.2 – 3285 sonnet 5 came out on top here – leanest code overall and a near-tie for fastest it was also the most creative. in every task it added something none of the others did: – kept the trampoline vibrating after the objects landed – drew a +1 next to the rocket that scored the point – turned the slingshot to face the next bottle before each shot opus 4.8 evaluated the code sonnet 5 produced. four things stood out: • the sphere is a fake, and that's the smart move. the cube and star are real 3d meshes with proper culling and shading, but the ball is just a flat shaded circle. a lit sphere looks identical from every angle, so building it in 3d would burn compute for zero visible payoff. knowing where not to bother is its own kind of skill • weight actually means something on the trampoline. the star is heavy, so it barely bounces and dents the mat hard. the ball is light, so it's lively and leaves a shallow dip. the three objects aren't just different shapes – they have different temperaments, and the physics is what gives them that • the slingshot is framed like a shot, not just drawn. the handle is anchored below the bottom of the screen and runs off-frame, so it reads as something you're holding rather than a sprite parked in the scene. that's a staging instinct, not a rendering one • the paddle ai forward-simulates the ball to predict where it'll land, then adds a deliberate error bias (roughly 1 in 5 shots is a real miss). that's why scoring looks natural instead of robotic – plus four distinct fault types with a catch-all so a rally never hangs without a result bottom line: sonnet 5 does more with less. fastest tier, leanest code, and the only one that added small touches nobody asked for follow thehype. for 24/7 ai news, analysis and breakdowns

thehype.

14,518 次观看 • 3 个月前

I built a macOS app for benchmarking local LLMs. 6 test suites. Multiple providers. One workspace. Open source. There are hundreds of local models now. New ones every week. How do you actually pick one? Leaderboards test for general ability. But if you're building an agent that chains tool calls, or a pipeline that extracts structured data, or a code assistant that needs to debug Rust, you need to know if the model handles that specific thing. Not in theory. On your hardware. With your prompts. The benchmarks that exist are either locked behind papers, too abstract to map to real failures, or impossible to extend. You can't add your own test cases. You can't test what matters to your use case. That's what BenchLocal is for. It's a benchmark platform where every test is practical, deterministic, and built around real-world tasks. And you can build your own tests. It ships with 6 Bench Packs TODAY: → ToolCall-15 — tool-use accuracy → BugFind-15 — debugging capabilities → DataExtract-15 — structured data extraction → InstructFollow-15 — constraint-heavy instruction following → ReasonMath-15 — practical reasoning and math → StructOutput-15 — validator-backed structured output Every pack has 15 fixed scenarios. Every score is deterministic and verifiable. Some of you saw ToolCall-15 and BugFind-15 — the individual test packs I open-sourced over the past few weeks. People ran them, filed issues, sent PRs. But managing separate repos, separate scripts, separate results doesn't scale. BenchLocal puts everything in one place. What the app does: > Workspace with tabs — run BugFind-15 in one tab, ToolCall-15 in another. > Any provider — Ollama, llama.cpp, OpenRouter, any OpenAI-compatible endpoint. Local and cloud, same interface. > Run modes — serial, batch per model, batch per test case, or fully parallel. > Test histories — every run saved. Compare any previous session. But the part I'm most excited about isn't the app. It's the ecosystem. BenchLocal is a platform. Each Bench Pack is a plugin. I'm shipping an SDK so anyone can build their own — test what matters to you, package it, share it. Install and uninstall packs right inside the app, same way you'd manage extensions in VS Code. The registry is GitHub-based, fully public. I built 6 packs. I want the community to build the next 60. Theme system built in too — because if I'm staring at benchmark results for hours, it should at least look good. v0.1.0 is macOS only. Windows and Linux are coming. MIT licensed. Everything — the app, the bench packs, the SDK — is open. PRs welcome. Bench Packs even more welcome.

stevibe

50,584 次观看 • 6 个月前