DeepSeek vs OpenAi vs Claude
(35.44KB,
716x407)
Tiap hari ada aja gebrakan dari china, bilangnya side project tapi pas di benchmark orang2 peformance skornya sama dengan claude sonnet 3.5 sama o1 reasoning openai. Hasil benchmark dari https://artificialanalysis.ai/models/deepseek-v3 , dari Quality test cuma beda 1 level dari best model OpenAi sekarang (o1) dengan skor 89, dengan perbedaan skor cuma beda 1 angka, tentu DeepSeek menang jauh dengan perbedaan harga yang jauh lebih murah (skor price o1 model 26 sedangkan deepthink r1 cuma 0.9), tapi ada yang dikorbanin yaitu speed, skor output token per second (ots) o1 openai 210 sedangkan deepthink r1 cuma 15.
nb :
- Price per token, represented as USD per million Tokens. Price is a blend of Input & Output token prices (3:1 ratio)
- Output Tokens per second received while the model is generating tokens (ie. after first chunk has been received from the API for models which support streaming).
Test prompt :
write a script for a bouncing yellow ball within a Rhombicosidodecahedron, make sure to handle collision detection properly. make the Rhombicosidodecahedron slowly rotate. make sure ball stays within the Rhombicosidodecahedron. implement it in p5.js
- Model Deepseek : https://chat.deepseek.com/ ( Deepthink (R1) ) setau gw ini model paling advance dari DeepSeek
Hasilnya :
1. DeepSeek : Run di https://editor.p5js.org/ langsung jalan tanpa error (◍•ᴗ•◍)❤ inget ye ini first try.
2. Claude : Gw cobain di claude model 3.5 Haiku first try error, percobaan kedua jalan tapi bentuknya ga jelas.
- DeepSeek DeepThink (R1) : https://pastebin.com/raw/uPDm0zUf (1 kali prompt)
- Claude 3.5 Haiku model : https://pastebin.com/raw/nr6anzdL (2 kali prompt)
Dari hasil test sebenernya ga adil sih bandingin claude 3.5 haiku ini sama deepthink, harusnya gw pake claude 3.5 sonnet tapi karena trafficnya claude lagi penuh modelnya ga bisa dipake, jadi buat perbandingan gw pake model yang available. Tapi tetep menurut gw deepseek menang banyak karena model nya opensource under MIT license, jadi bisa dipake bebas publik, bisa di enhance sendiri, bisa di training sendiri (the power of open source)
Gw pernah liat di twitter ada yg jalanin modelnya lokal di iPhone 16 (setau gw pake deepseek model yang 8billion, gw lupa nama nama persis modelnya apa) dan ya ini bikin gw excited, model ai yang makin advance dan bisa jalan di hp lokal (tanpa internet) what a dream right? semoga kedepannya banyak model yang JAUH LEBIH ADVANCE dan bisa run dilokal tanpa makan resource lebih.
kalo ada yang mau test run bisa pake editor online, copy codenya yang di pastebin terus paste di https://editor.p5js.org/ terus run aja
oh ya i forget to mention deepseek r1 model masih gagal karena bentuknya itu square bukan Rhombicosidodecahedron still need 2x prompt buat make sure bentuk shapenya bener, but still buat first try not that bad
i do re-prompt and success at the FIRST TRY
let ballPos;
let ballVel;
let boundaryRadius = 190;
let containerSize = 200;
function setup() {
createCanvas(800, 600, WEBGL);
ballPos = createVector(random(-50, 50), random(-50, 50), random(-50, 50));
ballVel = p5.Vector.random3D().mult(5);
angleMode(DEGREES);
}
function draw() {
background(0);
lights();
// Enable camera controls
orbitControl();
// Draw rotating container
push();
rotateX(frameCount * 0.5);
rotateY(frameCount * 0.5);
stroke(150, 150, 255);
noFill();
strokeWeight(1);
// Note: p5.js doesn't have a built-in Rhombicosidodecahedron
// This is a placeholder sphere showing the collision boundary
// Replace with actual Rhombicosidodecahedron geometry if available
sphere(containerSize);
pop();
// Update ball physics
updateBall();
// Draw bouncing ball
push();
translate(ballPos.x, ballPos.y, ballPos.z);
fill(255, 255, 0);
noStroke();
sphere(15);
pop();
}
function updateBall() {
ballPos.add(ballVel);
// Sphere-based collision detection
let distance = ballPos.mag();
if (distance > boundaryRadius) {
// Calculate collision normal
let normal = p5.Vector.div(ballPos, distance);
// Reflect velocity
let dot = ballVel.dot(normal);
ballVel.sub(normal.mult(2 * dot));
// Add position correction
ballPos.normalize().mult(boundaryRadius * 0.95);
}
// Dampening to prevent perfect perpetual motion
ballVel.mult(0.995);
}
@1168 with deepthink r1 model? impressive right? china is number 1 (´∀｀)♡
1955 : Space race (US vs USSR)
2025 : AI race (US vs China)
@1171 china is good at improving while west is good at invention, they both good
(116.17KB,
1895x695)
The one that’s better than o1 is 70B one
@1174 but still in math problem lightweight model 1.5b deepseek is still better than o1, that's still very impressive tho
@1173
kalau opini gw sendiri sih
barat : condong ke imajinasi
china : condong ke kerja keras dan ketelitian
@1176
maka dari itu penemuan lebih masif di eropa dan amerika tapi asia timur (china, jepang, korea) banyak bisa improve dari penemuan eropa
(62.07KB,
800x461)
@1180
“It’s a side project. I call it DeepSeek.”
breaking news deepseek release another model for image generation called Janus Pro 7B
https://huggingface.co/spaces/deepseek-ai/Janus-Pro-7B