LLaMA 3 Is HERE and SMASHES Benchmarks (Open-Source)

preview_player
Показать описание
Meta finally dropped LLaMA 3, and it’s a banger! Let’s review the announcement and see why this changes the face of AI. Many more videos about LLaMA 3 coming soon!

Join My Newsletter for Regular AI Updates 👇🏼

Need AI Consulting? 📈

My Links 🔗

Media/Sponsorship Inquiries ✅

Links:
Рекомендации по теме
Комментарии
Автор

LLAMA 3 (70B) is their middle version, that's why they didn't compare it to Claude 3 (Opus). Meta still has an unreleased (±400B) version that is currently still training, according to Mark Z. 👍🏻

GavinS
Автор

Cannot believe we have llama 3 Before GTA 6🎉😮😮😮

japneetsingh
Автор

0:00 - Introduction & excitement for Llama 3 launch
0:32 - Overview of Llama 3 & Meta AI platform
1:01 - History of Llama & Open-Source AI impact
2:14 - Testing Llama 3 with code generation (Snake game)
2:36 - Enhanced Performance & Capabilities of Llama 3
3:52 - Focus on Multi-Step Tasks & Agent Potential
4:25 - Benchmarks & Comparisons with Other Models
7:32 - Trust & Safety Measures: Llama Guard & Cybersec Eval
8:15 - Making Safety Tools Accessible
9:16 - Meta AI as a New Assistant, Features & Global Rollout
11:33 - Faster Image Generation & Creative Applications
12:59 - Llama 3 Integration in Search & Recipe Example
13:10 - Meta AI in Facebook Feed
14:05 - Meta Llama GitHub Page & Code Access
14:37 - Llama 3 Model Card & Specifications
14:58 - Benchmark Comparisons: Llama 3 vs Llama 2
15:21 - Conclusion & Upcoming Testing Video

dmitrymatora
Автор

it’s crazy how it beats claude sonnet. the model isn’t even free to some people anymore since atrophic switched their free model to haiku. in comparison, meta 3 70 b is not only open source, it’s also free ! (limited only to available countries tho). what a freaking time to be alive

Kutsushita_yukino
Автор

I'm waiting for Llama 4 outperforming GPT-5

TheRealUsername
Автор

This sucks: "Meta AI isn't available yet in your country". Yes i can use a VPN but from EU it still sucks.

Автор

Matthew! The 70B one "IS" the middle one so the comparison is correct. The high end one is 405B dense model and is still in training. Once that is released, then they can properly compare that high-end model with GPT4-TURBO and OPUS, etc.

senju
Автор

Looks great. Already works in ollama. Looking forward to their 405B parameter model...though I'm not looking forward to renting something to run it.

nathanbanks
Автор

I asked LLaMA 3 a VFX question and a simple math question for a daily use case and it did better than Claude 3 Opus. It recognized the order of questions and answered them respecively whereas Claude 3 Opus just melded them into one.

berkertaskiran
Автор

Been running the local model...pretty impressive for an 8B. Can't wait for the fine tuned uncensored models.

braticuss
Автор

Suddenly, the 70b model is on huggingface

WayneMetcalf
Автор

I asked llama 3 a question gpt4 and claude opus needed multiple tries to answer correctly and it got it right in one try

zeMasterRuseman
Автор

Unbelievable! I asked GPT-4 and Meta to troubleshoot a past issue I had with my VMware and a Linux host. Interestingly, I already knew the solution. GPT-4 provided a lengthy troubleshooting suggestion that didn't fix the issue, whereas Meta quickly identified the problem and offered multiple solutions, one of which was the correct answer ! Great first impression so far !!!

kamelsf
Автор

This is a great model. I have installed it locally on LMStudio with the 8B version and tried "write the game snake in python", and it did it greatly in one shot. Even with colors, and we lose when crossing a wall. Wow !

jacquesmaltais
Автор

Math question: Write an equation of the line passing through the point (20, 10) with an undefined slope.
Answer: x=20

JohnLewis-old
Автор

you didn't talked about the +400 billion parameter model they said they'll release, I don't think that there's a 35 billion parameter models and the 80 billion is the middle size

felipe
Автор

I work in cybersecurity and your videos are extremely helpful. I’d love to see you do a video on llama guard and cyber security eval 2.

daniellee
Автор

Matt, Wooohooo!!! 🎉🎉 Can‘t wait for the default tests and I hope to see it in LM Studio soon!

MeinDeutschkurs
Автор

for the math question i think something like a convolution would be interesting especially with a graph that shows it correctly especially.

borisverhaar
Автор

Hope to see something like codellama 3 and also see it in groq

enekxtw