What's going on, guys? Josh Pokok here. Now, recently Meta just released their new Llama 3.3 open-source AI model, and this

new state-of-the-art 70B model is supposed to have similar performance to the Llama 2 70B model. Open-source models

are progressing rapidly, and we're seeing them getting closer and closer to actual closed-source models. We're going to dive

into it today, show you how you can use it, run some tests, and see what it's made of. Let's dive right into it.

All right, guys. So all the links in this video will be linked down below, so if you want to take a deeper look at this

model, you can definitely check that out. I'm not going to be able to go through each and every little thing today, but

I'll start by giving you an overview. If you haven't heard of this model, all right, so the main models that Meta

currently has is Llama 2, which they had the 7B, the 13B, and they had Llama

2, which is a lightweight model, which is like 7B and 13B, and then 70B and 70B.

And now they have Llama 3.3, which is a 70B model, and they've been pushing

something which they call the Llama Stack. Now, if we look at the Llama 3.3 eval

values, we can see that they are very impressive. Obviously, I'm not going to go through each and every little one here,

as you can see, Llama 3.3 70B, compared to Llama 2 70B, it is beating that uh, that

model in many different benchmarks, as you can see here, as well as the 70B model. It is um, just below it in certain

areas, and it's even above it in other areas. So we can see human eval, it's just one below. Uh, math, we can see that it's

actually, four ahead or three ahead of the 405B model; reasoning: it's at 50.5

when the 405B model is at 49.  You can also see it compared to Gemini Pro 1.5,

uh, GPT-4, and Claude 3.5 here. All right, so benchmarks definitely don't mean

everything by any means. You know, when you actually start testing it, sometimes you'll get a better sense, which we're

going to get into in just a second. But as you can see, these are definitely some pretty impressive benchmarks. You know, in

certain areas, it's actually sometimes a little bit above Claude 3.5 or below Claude 3.5 in other areas, so it's

impressive to see these open-source models catching up. Now, there are a few main different places you can actually

access this model. If you want to download this model, you could go to llama.com/downloads. Right here, I'll leave a link

down below, and you can request access to this model like so and get it right away. You could go to Hugging Face right here,

and we can see the LLaMA 3.3 instruct model right here. We can also go

to Ollama right here. We can see LLaMA 3.3 right here. We can see all the different information on it right here:

evals, new capabilities, tool use, etc. So pretty cool. All you would do is simply just run LLaMA. Now, you would need some

pretty good hardware to actually run this locally, but you can actually access it on either Perplexity.ai or

gpf.chat right here, which we're going to be using to actually run some tests. All right, so here are some of the test

questions we're going to go ahead and ask LLaMA 3.3. Let's dive right into it.

Write a Python function to generate the Fibonacci sequence up to the 10th number. All right, so it generated this really

fast. I'm going to go ahead and run this code. All right, so here we have the code. I'm going to run this and we can see

that it got this correct. Next, I'm going to say: generate a batch script to create 10 files with the names a1.txt

through a10.txt. All right, so now we got that script loaded up that it generated. I'm going to run this and boom, it

created those files, so that is a pass. Next is a question: a factory produces widgets at a rate that doubles every 3 days. If it produces 100 widgets on day one, how many widgets will it produce on day 19? Explain your reasoning step by

step. And boom, it got the answer correct, which is 6,400 widgets. And you can see all the different reasoning steps

right here, which is actually really good. Next is: generate an SVG of a tree. Okay, I pasted the code here and as you can see,

this is a pass. All right. Next, I'm going to ask it a pretty hard question that I haven't seen really any LLM pass just

because they're not really that great at counting specifically: generate 13 sentences that have 10 words and end in

the word "monkey". Okay, so this is kind of interesting. This is like one of the only ones that I, at least remember, um, that

actually generated. It seems, at least for my account, each and every phrase or sentence has about nine words in it and

they do end in the word "monkey," but then I said, okay, these have nine words, it should have 10, and then they said, "Here

are 13 new sentences with exactly 10 words that end in the word 'monkey.'" And from what I can see, it literally just

gave me the exact same output. So, this is a fail, but I mean it was almost kind

It's impressive that it got nine, like it did. Each one having nine. Most of the models I've done this test on will have

like 10, 11, even 13 or 8, right? So it's a fail, though. Next is: generate me the pawn

game using Python. All right, so now I got the code in here. I'm going to run it and

boom! We got the pawn game here. It has a scoreboard at the top. This is definitely one of the better pawn games that um, one

of the models has generated. It pretty much has everything working well out of the box. The scoreboard is working, so this is

definitely a pass. All right, next is a question: A man stands on the river's side; his dog is on the other. The man calls his

dog, who immediately crosses the river without getting wet and without using a bridge or a boat. How did the dog do it?

The answer is that the river was frozen. That's correct. Next is: create a landing page using CSS, JS, HTML. It should be a

website for a beauty store that has a header, banner, features, testimonials, and checkout section. Make it look professional and very modern. All right,

so I got the code here. Let's go ahead and check this out. Okay, so here is the beauty store. We got the Home, Home,

Products, About Us, Contact Us. We got a banner, a featured section right here, a customer testimonial section, a checkout

section. So, all in all, it's definitely not beautiful; it's, you know, not the greatest store by any means, and I

personally wouldn't use it, but it did get this question correct, so that is a pass. All right, next question: You are in

a room that has three switches and a closed door. The switches control three light bulbs on the other side of the

door. Once you open the door, you may never touch the switches again. How can you definitively tell which switch is

Connected to each of the light bulbs all right. So it said, turn switch one to the on position for 5 minutes. This is to

heat up the corresponding light bulb. Turn switch one to the off position and turn switch two to the on position.

Immediately open the door and observe the light bulbs. The light bulb that is on is controlled by switch two; the light

bulb that is off but is warm is controlled by switch one; and the light bulb that is cool is controlled by switch three. That

is correct. Next is: I left my campsite and hiked south 3 miles, then I turned east and hiked east for 3 miles,

then I turned north and hiked for 3 miles, at which point I came across a bear inside my tent eating food. What color was the

bear? All right, so it realized that the only way you can actually hike um, down, right, and then up and be at the same

spot is the North Pole, which would mean it is a polar bear, which would mean that it is white. So that is correct. All right,

so next is an answer that most LLMs have gotten wrong: It is, how many words are in

your answer here? Let's see what it gets. Okay, so it says the conversation has just begun, and I'm going to try again

and say, how many words are in your response here? All right, so it said in my response there is one word that is

incorrect. There are 1, 2, 3, 4, 5, 6, 7. So that is another fail. All right, next question:

In a room there are two men: one lies and one always tells the truth. You don't know which is which. You may ask one

question to one of the men to determine which door leads to freedom and which leads to death. Which

question do you ask? To solve this, I would ask one man, if I were to ask the other man which door leads to freedom,

What would he say if I asked the truth-teller? He would tell me that the liar would point to the door that leads to

death. If I asked the liar, he would tell me what the truth-teller would say and point to the door that leads to

death. So, regardless of who I ask, the answer will steer me away from the door that leads to freedom. I can safely

assume that the opposite door leads to freedom. So that is correct. All right, next question. You have three boxes

labeled apples, oranges, and apples and oranges. Each box is labeled incorrectly. You can open one box and look inside. How

can you label all boxes correctly by opening only one box? All right, so it said since all the labels are incorrect,

the box labeled 'apples and oranges' cannot actually contain both apples and oranges; it must contain either apples or

oranges. Let's say I open the box and find only apples inside. This means the box labeled 'apples' cannot contain only

apples, since it's labeled incorrectly; it must contain oranges. The box labeled 'oranges' must contain both apples and

oranges since the other two options are taken. If I open the box and find only oranges inside, the logic is similar: the

box labeled 'oranges' contains apples and oranges, and the box labeled 'apples' contains oranges. All right, so this is

correct. All right, so next is: find the derivative of f(x) = x³ - 3x²

+ 2x - 1. All right, so it said that the answer is f'(x) = 3x² -

6x + 2, which is correct. Next is: solve the system of equations 2x + y = 5,

x - y = 1. All right, so this answer is x = 2 and y = 1, which is correct. Next,

If a car travels 100 m in 5 seconds, what is the average speed in m/second and

km/hour? It got 20 m/second and 72 km/hour, which is correct?

Next, the heights of a group of students follow a normal distribution with a mean of 170 cm and a standard

deviation of 10 cm. What percentage of students are taller than 180 cm? It got approximately

15.87% of students are taller than 180 cm, which is correct. Next, write a Python one-liner that creates a

list of all numbers from 1 to 100 that are divisible by seven. All right, so I got the code here.

I'm going to run it, and you can see that we got all the different numbers that are divisible by seven here. So that's

another pass. Next, create a responsive grid layout with HTML and CSS that displays four items per row on large

screens, two items per row on medium screens, and one item per row on small screens. Okay, and here is our code live.

Let's go ahead and shrink the screen here. So we got four items on a medium-size screen, we got two, and then on a

small screen, we got one. So that is correct again. All right, so there are all the questions. Uh, honestly, I definitely

NEW Lama 3.3 70B Meta AI Best Open Source LLM!?🤖 [TESTED] Beats Lama 3.1 405B?!

Josh Pocock

Llama 3.3 70B 모델 소개 및 성능 비교

Meta에서 새롭게 출시한 오픈소스 AI 모델 Llama 3.3 70B 소개

Llama 2 70B 모델과 유사한 성능을 가짐

기존 클로즈드소스 모델과의 성능 격차가 줄어들고 있음

Llama 3.3의 평가 지표(eval values)는 매우 인상적임

Llama 2 70B 모델 및 다른 모델들과의 다양한 벤치마크 비교 결과 제시

벤치마크 결과는 모든 것을 의미하지 않으며 실제 테스트가 중요함을 강조

모델 다운로드 방법(llama.com/downloads) 소개

Hugging Face를 통한 접근 방법 설명

Ollama를 통한 접근 및 정보 확인 방법 설명

로컬 실행을 위한 고성능 하드웨어 필요성 언급

Perplexity.ai 또는 gpf.chat를 이용한 접근 방법 설명

피보나치 수열 생성, 배치 스크립트 생성, 공장 생산량 계산 등 다양한 질문에 대한 테스트 진행

특정 단어로 끝나는 문장 생성 테스트 (성공적이지 못함)

Python을 이용한 장기 게임 생성 테스트 (성공)

CSS, JS, HTML을 이용한 랜딩 페이지 생성 테스트 (성공)

응답의 단어 수 세기 테스트 (실패)

미적분 문제 풀이, 방정식 풀이, 속도 계산 등 다양한 수학 문제 풀이 테스트 (성공)

Llama 3.3 모델의 전반적인 성능 평가 및 다른 모델과의 비교

2025년 오픈소스 모델의 급성장 전망 언급

채널의 다른 콘텐츠 소개 (마케팅, 세일즈, AI 등)

비즈니스 관련 마케팅, 세일즈, AI 구현 상담 안내