AI News: A Massive Week For AI Advancement!

Here's the AI news that you might have missed this week. I recorded this on Thursday so any news from Friday didn't make this video.

Discover More From Me:

Resources From Today's Video:

#AINews #AITools #ArtificialIntelligence

Time Stamps:
0:00 Intro
0:27 Llama 3.1 Release
6:35 Mistral Large 2
8:40 Apple's Open-Source AI
9:06 Gemini 1.5 Flash
10:22 OpenAI Free Fine-Tuning
10:56 OpenAI SearchGPT
12:20 ChatGPT Voice Update
12:42 Anthropic Scraping Too Much
13:15 Grok 2 and Grok 3
15:19 Tesla Optimus Updates
15:51 Bing AI Search Update
16:31 Luma Video Loop Feature
18:00 Kling AI Video Widely Available
19:49 Runway Scraping YouTube
23:29 Stable Video 4D
24:51 Adobe Illustrator AI Updates
25:38 Leonardo For Teams
26:27 Sakana AI - Ukiyo-e Artwork
26:56 Suno AI Update
27:16 Fake Biden AI Hoax
28:23 EA Uses AI For College Football Game
29:26 Final Thoughts

[SeattleShelby]

My seven-year-old kid was trained on thousands of hours of YouTube without permission.

[matthew_berman]

Matt: Rule #1 -> If you go on vacation, something amazing will drop 😂

[TheLastShow]

what set your channel apart is that you REALLY are passionate about the topics...
It never feels like your sharing boring news.. feels like you're sharing a new toy or a game ..

Keep it up man .. I really missed you this week!

[JohnTackman]

Matt: the billion parameter number doesnt reflect the amount of data the model is trained on, rather it's the number of parameters that get tuned with the data.. basically, it's a larger set of knobs and dials for the data to push around while being trained

[Joseph_Roffey]

I think we can all agree on the most important rule for when scraping videos for training AI models — if they contain ad breaks or breaks to talk about sponsors then the AI has to be trained on that too. I expect all of my AI videos to try to get me to use Skillshare and to play League of Legends or it won’t feel real.

[ayaiai]

Nice to see Perplexity already added Llama 3.1 405B as an option for the generative engine!

[qgqcqbr]

Introduction and Overview - 00:00:00
Meta's Llama 3.1 Release - 00:00:33
Llama 3.1 Model Sizes and Capabilities - 00:01:04
Comparison with Other State-of-the-Art Models - 00:01:38
Open Source Accessibility and Licensing Terms - 00:02:10
Usage and Accessibility of Llama 3.1 Models - 00:03:11
Other Open Source AI Developments - 00:06:36
Google's Gemini Model Update - 00:09:20
OpenAI's Fine-Tuning for GPT-4 - 00:10:17
Anthropic's Scraping Controversy - 00:13:12
Elon's Grok AI and Predictions - 00:14:05
Tesla's Optimus Robots - 00:15:33
Bing AI Redesign - 00:16:01
Luma AI's New Loop Feature - 00:16:28
Singularity's Text-to-Video Advances - 00:18:19
Runway's Training Data Controversy - 00:20:09
Public Perception and AI Scraping Debate - 00:22:12
Stability AI's Stable Video 4D - 00:23:43
Adobe's AI Features in Illustrator and Photoshop - 00:24:42
Leonardo's Team Collaboration Feature - 00:25:38
Sakana AI's Ukio-e Artwork Model - 00:26:08
Sunno's Stems Pro Feature - 00:27:08
Fake Joe Biden Audio Clip Incident - 00:27:38
AI in College Football Games - 00:28:38
Conclusion and Future Plans - 00:29:12

[NianLisa]

I’m 52 hoping to end the rat race by 60 with above $1M. I know money is a liability to be exchanged for assets with real value like real estate (properties for rent) stocks (dividends) bonds (interest) But, what is it with bitcoin? I hear a lot about it and I'd love to diversify my portfolio.

[tomconroy]

Hey Matt, Could you do a video (if you havnt already) on how someone can train an AI to do niche things? I play alot of TTRPGs like pathfinder and DND5e and would live an AI assistant to help clarify rules and that kind of stuff. But i have no idea how to go about making that kind of AI or training data.

[jvlbme]

The parameter count isn't really a measure of the amount of data it trained on, but rather the amount of node connections.

[HonorMacDonald]

We should probably keep in mind that "this AI trained on X without permission" has approximately the same legal weight as "I was walking down the street and X looked at me without permission." (as you kind of mentioned).

It's not that these companies are doing anything "wrong" at least in a legal sense, it's that we need laws that address this new use case that our existing laws never imagined.

And thank you for the good thumbnail. :)

[High-Tech-Geek]

Note: Groq is not Elon Musk's Grok

[aiforculture]

What a time to be ali-- um, on vacation!

[NolanKhalaf]

Advice for my best YouTuber

Do separate videos
- One for AI News only.
- One for trying & experimenting on models & new stuff.

[panpiper]

Tesla's Optimus bots are ALREADY being used in the factory "doing useful tasks", except that it is very low numbers. What they are saying is that next year they will have thousands.

[curiosus]

30:57 whatever happens with AI in the end, whatever that means, Matt is the true embodiment of doing what you love.

[rononeil]

Matt: Thank you so much for taking time out of your vacation to keep us up to date! Your the most informative AI Youtuber that brings AI to the forefront of us. Enjoy the rest of vacation. Nerd out!

[suzannecarter]

You are such a hard worker, heaven knows, you deserved a vacation and we all know cool stuff will drop when you do

[zeddy]

On the training data and ethical concerns,
I understand that this is a different perspective, but let's consider this: as humans, we learn by reading, watching, or being taught how to do things in life and become independent. Similarly, these machines and their transistors are essentially learning to walk and become effective tools that we can use, employing the same learning concepts. If information is freely available on the web and accessible to the public, why can't the same principle apply to LLMs? I find it perplexing that people immediately object, claiming ownership of their data and not wanting LLMs to learn from it. In the same vein, it’s like humans going to a store or library to learn from a book.

[rs-dms]

A question about youtube videos.
If we view AI training as similar to what humans do with websites and other media, how is AI "viewing" a youtube video different from what a human does? After all, if I understand correctly, the training does not store the video, it merely "views" it and builds a neural net from the data.