AI News: The AI Arms Race is Getting Insane!

Matt Wolfe
12 Apr 202428:10

Summary

TLDRThis week in AI saw major announcements from Google and OpenAI, with Google's Gemini 1.5 becoming available in 180 countries and OpenAI's GPT-4 Turbo model improving at coding and math. New large language models were released, including Stability AI's non-commercial Stable LM2 and ML's Mixr 8X 22b. Meta is also on the verge of releasing Llama 3, an open-source model. Additionally, companies are developing AI chips to reduce reliance on Nvidia GPUs, and there's a push for AI companies to disclose their training data. In the world of AI music, Udio is gaining support from musicians, and Spotify is testing AI-generated playlists.

Takeaways

  • πŸ“’ Google's Cloud Next event in Las Vegas featured numerous AI-related announcements, with a focus on enterprise and developer tools.
  • 🌐 Google made Gemini 1.5 available in over 180 countries, offering a 1 million token context window for advanced language understanding and audio processing.
  • πŸŽ₯ An example of Gemini 1.5's capabilities includes analyzing an hour-long audio file to provide key takeaways and generate YouTube video titles and thumbnails.
  • πŸ€– OpenAI announced a significantly improved GPT-4 Turbo model, which is now available through the API and has shown better performance in coding and math tasks.
  • πŸš€ Stability AI released Stable LM2, a 12 billion parameter model that can be used both non-commercially and commercially with a membership.
  • 🌟 ML released a new large language model using a mixture of experts architecture, featuring 176 billion parameters and a 65,000 token context window.
  • πŸ”₯ Google introduced new versions of their open-source large language model, Gemma, tailored for coding and efficient research purposes.
  • 🐫 Meta is close to releasing Llama 3, an open-source model expected to be as capable as GPT-4, with multiple versions for different use cases.
  • πŸ’° Tech companies like Google, Intel, and Meta are developing their own AI chips to reduce reliance on Nvidia's GPUs, which currently dominate the AI training market.
  • 🎢 AI music generators like Udio are gaining popularity and support from musicians, offering a platform for creating music with AI assistance.

Q & A

  • What is the main focus of the AI news in the spring of last year?

    -The main focus of the AI news in the spring of last year was the ramping up of announcements related to new large language models becoming available, as well as the developments in AI technology by major companies like Google and OpenAI.

  • What significant announcement did Google make at their Google Cloud Next event in Las Vegas?

    -At the Google Cloud Next event, Google announced the availability of Gemini 1.5 in over 180 countries with native audio understanding system instructions and JSON mode, among other features.

  • What is the context window of Gemini 1.5 and how does it compare to other models?

    -Gemini 1.5 has a context window of 1 million tokens, which is significant because it allows for a much larger amount of input and output for the model to work with, providing a combined 750,000 words for interaction.

  • How did Bill use Gemini 1.5 to enhance his YouTube content creation?

    -Bill used Gemini 1.5 to analyze an hour-long audio file from a video interview, generate key takeaways, suggest high click-through rate YouTube titles based on principles of Daryl EES and top YouTube creators, and even provide feedback on which thumbnail to use for the video.

  • What is the difference between GPT-4 Turbo and the previous model in terms of capabilities?

    -GPT-4 Turbo is an improvement over the previous model with enhanced capabilities in coding and math, and it is updated through December 2023. It is also considered the strongest and most powerful model according to the chatbot Arena.

  • What is the open-source model released by Stability AI and how does it compare to Mixl 8X 7B model?

    -Stability AI released Stable LM2, a 12 billion parameter model that slightly underperforms the Mixl 8X 7B model. It can be used both non-commercially and commercially, but commercial use requires a Stability AI membership.

  • How did Mistol release their new large language model and what are its specifications?

    -Mistol released their new model, Mixr 8X 22b, through a torrent link. The model features a 65,000 token context window and a total of 176 billion parameters, with eight experts each having 22 billion parameters.

  • What is the significance of the new large language models released by Google and Meta?

    -Google released new versions of Gemma, their open-source large language models, focused on coding and efficient research. Meta is close to releasing Llama 3, an open-source model expected to be as good as GPT 4, available for anyone to use and build upon.

  • What are the implications of the new AI chips introduced by Google, Intel, and Meta in relation to Nvidia's dominance in the GPU market?

    -Google, Intel, and Meta are all developing their own AI chips to reduce reliance on Nvidia's GPUs. While Nvidia currently leads the market, these companies are attempting to catch up and provide alternative options for AI model training and development.

  • What is the controversy surrounding AI companies and their use of copyrighted material for training?

    -There is a debate over the use of copyrighted material for training AI models. A new bill has been introduced to Congress that would force AI companies to reveal the copyrighted material used in training their generative AI models. This is due to concerns over data privacy and the ethical use of content.

  • How is Adobe approaching the acquisition of training data for their AI models?

    -Adobe is taking a different approach by offering to purchase video content from creators for their AI training data. They are willing to pay between $3 to $7 per minute for everyday video footage, which is a shift from the traditional method of data acquisition.

Outlines

00:00

πŸš€ Google Cloud Next and AI Announcements

The paragraph discusses the recent Google Cloud Next event in Las Vegas, highlighting new AI-related announcements. Google introduced Gemini 1.5, which is now available in over 180 countries with advanced features like a 1 million token context window. The video also mentions the use of Gemini 1.5 by a content creator for packaging a long video interview for YouTube, showcasing its capabilities in generating key takeaways and suggesting high click-through rate titles. Additionally, OpenAI's GPT-4 Turbo model received a mention, which is an improvement over the previous models and is now considered the most powerful by the Chatbot Arena community. The paragraph emphasizes the importance of these AI models and their potential applications in various fields.

05:01

🌐 New Large Language Models and AI Developments

This paragraph delves into the release of new large language models, including Stability AI's stable lm2, which is a 12 billion parameter model, and Mixture of Experts models from ML. The open-source community is also making strides with the release of a new model via a torrent link, Mixr 8X 22b, which features a 65,000 token context window and a total of 176 billion parameters. Furthermore, Google has released new versions of their open-source large language model, Gemma, including one tailored for coding and another for efficient research. The paragraph also mentions Meta's upcoming release of Llama 3, an open-source model expected to rival GPT-4 in capabilities.

10:03

πŸ’‘ AI Chip Innovations and Video Generation Models

The focus of this paragraph is on the efforts of major tech companies to reduce their reliance on Nvidia GPUs for AI training. Google, Intel, and Meta have all introduced their own AI chips, such as Google's Axion processors, Intel's gouty 3 AI chip, and Meta's MTI accelerator. These developments aim to improve performance and efficiency while reducing costs. The paragraph also touches on Google's new image generation model, Imagen 2, which can create short animations and GIFs, and other AI-generated video innovations like Magic Time, which specializes in timelapse videos. The advancements in AI chip technology and video generation models underscore the rapid progress in the AI field and the increasing competition among tech giants.

15:04

🎢 AI in Music and the Future of Content Creation

This paragraph covers the emergence of AI in music generation, with platforms like Udio gaining popularity. Udio allows users to create music by providing prompts, style suggestions, and even AI-generated lyrics. The paragraph also mentions Spotify's new AI-driven playlist feature. In the realm of content creation, Adobe's initiative to purchase video content for training their AI models is highlighted, as well as Meta's efforts to identify AI-generated images using their own AI detection system. The discussion emphasizes the growing role of AI in creative industries and the potential for AI to transform content creation and consumption.

20:05

πŸ€– AGI Predictions, AI Ethics, and AI-assisted Art

The paragraph begins with Elon Musk's prediction of achieving AGI within the next year or two, contrasting with Yan LeCun's view that current large language models will not reach human-level intelligence. It then shifts focus to the potential of Meta's self-supervised learning architecture, V Jeppa, to achieve AGI in the future. The Humane pin, a device designed to replace smartphones, receives critical reviews for its impracticality and high cost. The paragraph also discusses the use of AI in art, specifically an AI-assisted artist who was paid $90,000 for generating card art, and the importance of human intervention in refining AI-generated concepts. The discussion concludes with the launch of the Next Wave podcast, which aims to delve deeper into AI topics and features various guests to provide insights into the AI world.

25:06

πŸŽ‰ Launch of the Next Wave Podcast and AI Tools

The final paragraph announces the launch of the Next Wave podcast, a platform for deeper discussions on AI topics, ethics, and implications. The podcast, produced by HubSpot, offers a more in-depth conversational format compared to the video content. The paragraph also promotes the podcast's competition, which includes giveaways such as Apple Vision Pros, and encourages viewers to subscribe, like, and review the podcast. Additionally, the speaker highlights his own AI-focused newsletter and website, futur tools, which curate the latest AI tools and news, and provides an AI income database for subscribers. The paragraph concludes with a call to action for viewers to engage with the content and stay updated with AI developments.

Mindmap

Keywords

πŸ’‘AI News

AI News refers to the latest developments and announcements in the field of artificial intelligence. In the video, the host discusses various AI-related news, highlighting the significant advancements and updates from companies like Google and OpenAI, which are central to the theme of AI progress and innovation.

πŸ’‘Google Cloud Next Event

Google Cloud Next Event is a conference where Google makes new announcements related to its cloud services and technologies. The video mentions this event as a platform where Google announced updates on AI models and other cloud-related technologies, emphasizing its relevance to enterprise and developers working with AI.

πŸ’‘Large Language Models

Large Language Models are advanced AI systems designed to process and understand large volumes of language data. The video focuses on the release of new and improved language models like Gemini 1.5 and GPT-4 Turbo, which are significant for their enhanced capabilities in processing text and generating responses.

πŸ’‘Gemini 1.5

Gemini 1.5 is a specific AI model mentioned in the video that has been made available in over 180 countries with features like native audio understanding and a 1 million token context window. It represents a leap in AI's ability to process and generate extensive amounts of text, which is crucial for various applications like content creation and data analysis.

πŸ’‘OpenAI

OpenAI is a research organization that develops and releases AI models with the aim of promoting the understanding and ethical use of AI. In the context of the video, OpenAI's announcements and models, such as the improved GP4 Turbo, are compared with Google's offerings, showcasing competition and progress in the AI industry.

πŸ’‘Stable Diffusion Model

The Stable Diffusion Model is an AI model released by Stability AI, which is designed to generate images from textual descriptions. The video discusses the release of Stable LM2, a 12 billion parameter model, which is significant in the open-source AI community and represents ongoing advancements in generative AI.

πŸ’‘Meera Model

The Meera Model, or Mixr 8X 22B as mentioned in the video, is a large language model from Mistral that features a 65,000 token context window and a total of 176 billion parameters. It is highlighted as a potentially powerful open-source model, indicating the rapid growth and capability expansion of AI models.

πŸ’‘AI Music Generator

An AI Music Generator is a tool that uses AI to compose music based on given prompts or styles. The video mentions Udio, a platform supported by musicians and investors, which generates high-quality music that is often indistinguishable from human-composed music, demonstrating the creative potential of AI.

πŸ’‘Humane Pin

The Humane Pin is a wearable device that projects a display onto the user's hand and responds to voice commands, intended as an alternative to smartphones. The video discusses the mixed reviews of the device, noting its current impracticality and the challenges it faces in replacing traditional smartphones.

πŸ’‘AGI (Artificial General Intelligence)

AGI, or Artificial General Intelligence, refers to highly autonomous systems that possess the ability to perform any intellectual task that a human being can do. The video cites differing opinions on when AGI might be achieved, with some predicting it to be within a year and others suggesting that current large language models will not reach human-level intelligence.

πŸ’‘AI Image Generation

AI Image Generation is the process of creating images using AI algorithms. The video discusses Meta's efforts to identify AI-generated photos using their own detection systems, as well as Adobe's approach to purchasing video content to train their AI models, highlighting the ethical considerations and practical applications of AI in creative fields.

Highlights

Google's Cloud next event in Las Vegas featured numerous AI-related announcements, emphasizing the growing importance of AI in enterprise and development sectors.

Gemini 1.5 was launched in over 180 countries with enhanced capabilities such as a native audio understanding system and JSON mode instructions.

The Gemini 1.5 model boasts a 1 million token context window, allowing for extensive input and output interactions, equating to 750,000 words in total.

An example of Gemini 1.5's practical application includes analyzing an hour-long audio file to generate key takeaways and suggest YouTube titles based on content analysis.

OpenAI's announcement of the significantly improved GPT-4 Turbo model, available through the API, has sparked interest despite being somewhat overshadowed by Google's announcements.

Stability AI's release of the 12 billion parameter model, Stable LM2, demonstrates the ongoing growth in the open-source AI realm.

ML's release of a new large language model using a mixture of experts architecture, Mixr 8X 22b, features a 65,000 token context window and a total of 176 billion parameters.

Google's release of new versions of their open-source large language models, Gemma, includes one tailored for coding and another for more efficient research purposes.

Meta is reportedly close to releasing Llama 3, an open-source model expected to rival GPT 4 in capability and be publicly available for use and development.

The competition between companies to develop AI chips is heating up, with Google, Intel, and Meta all seeking to reduce reliance on Nvidia's market-leading GPUs.

Google's Imagen 2 model represents their foray into AI image generation, capable of producing animations and GIFs from text prompts.

Adobe's unique approach to AI involves purchasing video content from creators to train their large language models, offering a potential revenue stream for content creators.

The introduction of a bill to Congress aiming to force AI companies to reveal the copyrighted material used in training their generative AI models could have significant implications for the industry.

Udio, an AI music generator, is gaining support from musicians and investors alike, showcasing the potential for AI in creative fields.

Spotify's new AI-driven feature allows users to generate playlists based on prompts, further demonstrating AI's infiltration into everyday applications.

Elon Musk's prediction of achieving AGI within the next year and a half contrasts with Yan LeCun's skepticism about large language models reaching human-level intelligence.

The Humane pin, a device designed to replace smartphones, has received unfavorable reviews, highlighting the challenges in creating practical and user-friendly AI technology.

A card game developer's use of an AI artist to generate card art for $90,000 illustrates the growing potential and commercial viability of AI-assisted creativity.

The launch of the Next Wave podcast offers a platform for in-depth discussions on AI, providing valuable insights and perspectives on the latest developments in the field.

Transcripts

play00:00

so just like spring of last year AI news

play00:02

is really ramping up there has been a

play00:04

ton of announcements this week I'm

play00:06

really having to figure out what stuff

play00:08

to filter down that I think you'll find

play00:10

important because at the end of the day

play00:12

nobody really cares that like Walmart

play00:14

got a new AI chat bot or something so

play00:16

I'm going to break down the stuff that I

play00:17

found important interesting or just

play00:19

downright fun that I think you're going

play00:21

to enjoy so let's get right into

play00:23

[Music]

play00:26

it this week Google had their Google

play00:28

Cloud next event out in Las Vegas where

play00:31

they made a ton of new announcements

play00:34

plenty of them related to AI more of the

play00:36

announcements were more relevant to like

play00:38

Enterprise and developers that are

play00:41

building with AI models but there were

play00:43

some pretty interesting and fun

play00:45

announcements that I think you'll enjoy

play00:47

as well so I'll be kind of sprinkling

play00:49

them throughout this video the real

play00:51

story of this week is all of the news

play00:54

about new large language models becoming

play00:57

available or soon coming available ing

play01:00

Google's event this week they announced

play01:02

that Gemini 1.5 is now available in 180

play01:06

plus countries with Native audio

play01:08

understanding system instructions Json

play01:10

mode and more now we have talked about

play01:13

Gemini 1.5 in the past but most people

play01:16

haven't had access to it until now of

play01:18

course the biggest Factor about Gemini

play01:21

that people are most impressed by is the

play01:23

fact that it's got a 1 million token

play01:25

context window now one token just to

play01:28

refresh your memory is about 75% of a

play01:32

word so 1 million tokens means that

play01:35

between the input that you can give the

play01:37

model and the output it'll give back you

play01:39

have a combined

play01:42

750,000 words to work with Gemini 1.5 is

play01:46

now available via the API so if you're a

play01:49

developer and you want to build with

play01:50

this model it's now available for you my

play01:53

buddy bill of Al do here has probably

play01:54

one of the best examples that I've seen

play01:57

of somebody actually using Gemini .5 he

play02:00

shows an example here where he says I

play02:02

just dropped in an audio file of an

play02:04

hourong video interview and now it's

play02:07

helping me package it for YouTube we can

play02:09

see in his screenshots here he actually

play02:10

uploaded the MP3 file told it to analyze

play02:13

this audio recording for his interview

play02:16

he then asked it to give the key

play02:17

takeaways and come up with 10 high

play02:19

click-through rate YouTube titles based

play02:21

on the principles of Daryl EES and top

play02:24

YouTube creators keep each title to 50

play02:26

characters or less so from this audio

play02:28

file it then gave them the key takeaways

play02:30

he also offered it two thumbnails and

play02:32

said which of these thumbnails is better

play02:34

suited for this YouTube video it

play02:36

analyzed the two thumbnails and gave

play02:38

feedback on which thumbnail to use and

play02:40

then once he picked the thumbnail he had

play02:42

it suggest the 10 titles and it gave

play02:44

some pretty decent titles now that's all

play02:47

pretty cool we can kind of do that with

play02:49

Claude right now the difference is he

play02:51

uploaded an audio file and it did this

play02:54

from the audio file with Claude you

play02:56

would actually have to get the text

play02:57

transcript upload it and you would

play02:59

pretty much get the same end result most

play03:01

likely where I thought this was the most

play03:03

impressive was when he asked it to

play03:04

generate timestamps and you can see it

play03:07

actually recommended these timestamps

play03:10

with an explanation of each of the time

play03:12

stamps and then there is a shorter

play03:15

version of the timestamps that it also

play03:16

generated this to me is really

play03:18

impressive because I've tried to use

play03:19

Claude and I've tried to use chat jpt to

play03:21

generate timestamps for these videos

play03:24

that you're watching right now while it

play03:25

gets the various sections right and it

play03:28

knows what I'm talking about in the

play03:29

video

play03:30

it just completely gets all the timings

play03:32

wrong and struggles to give an accurate

play03:35

timestamp to the actual chapter this

play03:37

example that billal shared is really

play03:40

really useful in my opinion but as open

play03:42

AI does every single time Google makes

play03:45

an announcement they come out with their

play03:46

own announcement in most of the past

play03:48

scenarios open ai's announcement really

play03:51

overshadowed Google's announcement but

play03:53

this time we just got a vague

play03:54

announcement of a majorly improved gp4

play03:58

turbo model is now available in the API

play04:01

and rolling out inside of chat GPT we

play04:03

don't have a ton more details than that

play04:06

but if we take a look here at the openai

play04:09

documentation we can see we've got the

play04:11

newest model here gp4 turbo and this is

play04:14

the April 9th edition of it Vision

play04:16

request can now use Json mode and

play04:18

function calling it's got the same

play04:20

128,000 tokens we've been working with

play04:23

and it's updated through December 2023

play04:26

which is also what the previous model

play04:28

was updated through supposedly this new

play04:30

model is a lot better at coding and also

play04:33

a lot better at math and for a while L 3

play04:36

Opus was the cream of the crop the best

play04:39

model out there but it seems now that

play04:42

according to the chatbot Arena here the

play04:45

newest version of GPT 4 Turbo the April

play04:49

9th Edition now took over Claude 3 Opus

play04:52

again as the strongest most powerful

play04:55

model as voted on by the people that

play04:59

rank this system here now they do show

play05:01

Gemini Pro down here as being below

play05:04

Claud and GPT 4 Turbo but I don't

play05:06

believe I'm not 100% sure but I don't

play05:09

believe that this is taking into account

play05:12

the newest 1.5 model but that's not the

play05:14

only news we've gotten in the world of

play05:17

new large language models the open

play05:19

source world is continuing to heat up as

play05:22

well in fact stability AI released

play05:24

stable lm2 which is a 12 billion

play05:28

parameter model and according to most

play05:30

benchmarks it just underperforms the

play05:33

mixl 8X 7B model and although they kind

play05:36

of make it out to be like an open-source

play05:39

product it does say it can be used

play05:41

non-commercially as well as commercially

play05:43

but if you have a stability AI

play05:45

membership so if you do want to use it

play05:47

commercially you got to pay which

play05:49

doesn't feel very open source to me well

play05:51

ml said hold my beer and watch this

play05:53

releasing a new large language model

play05:56

using the mixture of experts

play05:58

architecture but they released it in an

play06:00

interesting way they released it as a

play06:02

torrent link directly on X with almost

play06:06

no context in order to actually download

play06:08

the weights for this model you would

play06:10

need a torrent downloader something like

play06:12

views and then if you paste this URL

play06:15

into your address bar it will start the

play06:17

download over inside of your torrent

play06:19

downloader however be aware it is a 281

play06:23

gigabyte file now I don't have a ton of

play06:26

information about this new model from

play06:28

mistol however in the rundown newsletter

play06:31

they gave us a little bit more details

play06:33

this week according to Rowan over at the

play06:35

rundown the new model is Mixr 8X 22b so

play06:40

the previous model was 8X 7B so it was

play06:43

eight separate models that the router

play06:46

called upon to get the prompt answered

play06:49

and each of the models that it was

play06:51

calling upon was a 7 billion parameter

play06:54

model each this new one has eight

play06:56

experts but each expert instead of being

play06:59

a billion parameter model is now a 22

play07:01

billion parameter model so it was just

play07:03

trained on a lot more data according to

play07:06

this breakdown the new model features a

play07:09

65,000 token context window and a

play07:11

combined total of 176 billion parameters

play07:15

I haven't used this model yet myself

play07:17

channels like Matthew burman's channel

play07:19

does a really really good job of testing

play07:21

these large language models I have a

play07:23

feeling that this one is going to be the

play07:26

strongest open-source model once some

play07:29

more of the tests come out around it but

play07:31

we have even more large language Model

play07:33

news even more coming from Google Google

play07:36

released Gemini 1.5 their closed Source

play07:38

model but they also rolled out new

play07:40

versions of Gemma which is Google's

play07:43

open- Source large language models these

play07:45

two new models are code Gemma a model

play07:48

fine-tuned for using with coding and

play07:51

recurrent Gemma which is designed for

play07:53

more efficient research purposes now all

play07:56

of the Articles and resources I make in

play07:58

this video I will share in the

play08:00

description below so if you do want to

play08:01

dive deeper into how these Gemma models

play08:04

compare on the benchmarks against other

play08:06

models check out the links below but

play08:08

overall Gemma appears to be pretty on

play08:11

par with the other open-source coding

play08:14

specific large language models and in

play08:17

the final bit of large language Model

play08:19

news we also learned this week that meta

play08:22

is really close to releasing llama 3

play08:25

llama 3 is expected to be roughly as

play08:27

good as GPT 4 but open source and made

play08:30

publicly available for anybody to use

play08:33

and fine-tune and build on top of now

play08:35

they did say with llama 3 they were

play08:37

going to release several different

play08:39

versions of the model it kind of sounds

play08:41

like what Claude did with Haiku Sonet

play08:43

and Opus there will be different models

play08:46

sort of more fine-tuned for different

play08:49

reasons that's my takeaway from this as

play08:51

well according to Tech crunch they

play08:53

announced that this is coming out

play08:54

sometime within the next month hopefully

play08:56

sooner so really looking forward to that

play08:59

because I truly am rooting for both the

play09:01

open source and the closed source side

play09:03

to keep on pushing the boundaries and

play09:05

each side is just making things better

play09:07

and better for us the consumers now I

play09:10

know a lot of people that watch this

play09:11

channel have small to medium businesses

play09:13

and they're looking to use AI to improve

play09:15

their marketing and their business

play09:16

that's why for this video I partnered

play09:18

with HubSpot so that I can get your

play09:19

eyeballs on their completely free report

play09:22

all about how AI is completely

play09:24

redefining startup go to market strategy

play09:27

I'm going to put the link in the

play09:28

description to you can download this

play09:30

report but this is a must read for any

play09:33

small to medium business that plans on

play09:35

using AI as part of their startup

play09:38

strategy you're going to learn about the

play09:39

various strategies that startups are

play09:41

using to bring their products to Market

play09:43

you also learn about the most popular AI

play09:45

tools and best practices for scaling and

play09:48

if you know me I love me some AI tools

play09:51

the free report also covers how AI is

play09:53

driving startup scalability and drawing

play09:56

the attention of investors as well as

play09:58

the future of a guy within goto Market

play10:01

strategies of course my favorite section

play10:03

as a guy who made a business around

play10:05

curating amazing tools is this section

play10:07

all about the tools and best practices

play10:10

that HubSpot recommends for your goto

play10:13

Market strategy again this free report

play10:15

was provided by HubSpot who is

play10:18

sponsoring this video so thank you so

play10:20

much again to HubSpot for sponsoring

play10:22

this and once again the link should be

play10:24

right at the top of the description to

play10:25

make it easy for you to find now if the

play10:27

main story of the week was all of the

play10:29

large language models that have come out

play10:31

this week the sort of B story of the

play10:33

week the Side Story is it seems that all

play10:36

of these companies that are building

play10:38

these large language models are all

play10:40

trying to release the Reliance on Nvidia

play10:43

gpus at the moment Invidia owns the

play10:46

market on gpus being trained for AI but

play10:49

Google Intel and even meta are all

play10:52

trying to bring that chip generation

play10:55

inhouse and stop giving so much money to

play10:58

Nvidia at the Google Cloud next event

play11:00

this week Google introduced their Axion

play11:03

processors Intel introduced their gouty

play11:06

3 AI chip which is apparently a 40%

play11:09

better power efficiency than nvidia's

play11:12

h100 gpus and meta announced a chip as

play11:15

well this new chip is called an

play11:17

MTI or meta training and inference

play11:21

accelerator this is the second

play11:22

generation of the Chip and according to

play11:25

meta article here it is three times

play11:27

improved performance over the first gen

play11:30

chip now again I will link to all these

play11:32

articles that talk about all of these

play11:33

chips that were announced this week

play11:35

because a lot of this sort of technical

play11:37

stuff here is a little bit over my head

play11:39

I don't personally understand how these

play11:41

chips work but if you're somebody that

play11:43

wants to Deep dive and truly understand

play11:45

what makes these chips better than

play11:47

what's available feel free to read the

play11:49

articles in the description however back

play11:51

at Nvidia GTC earlier this year Nvidia

play11:54

announced their next iteration the

play11:56

Nvidia Blackwell which is supposedly

play11:58

already four times more powerful than

play12:01

the h100s again the chip that is sort of

play12:04

the industry standard for training AI

play12:06

models right now so while all these

play12:08

companies are making their own chips to

play12:10

release Reliance on Nvidia Nvidia is

play12:13

still way out ahead with their latest

play12:16

generation of chips making it extremely

play12:19

hard for these other companies to catch

play12:20

up with the compute power that Nvidia is

play12:22

producing right now also during the

play12:24

Google next event Google revealed image

play12:27

in 2 this is sort of like Google's

play12:29

answer to open ai's dolly or adobe's

play12:33

Firefly it's their own internal AI image

play12:36

generation model however what makes

play12:38

image in 2 different than tools like

play12:41

Dolly and Firefly is that it can

play12:43

actually generate animations it can

play12:46

generate GIF files or GIF files however

play12:49

you like to say it here's some examples

play12:51

of the types of animations that it will

play12:52

make you know similar to what we get out

play12:55

of like Pika or Runway but these are

play12:58

very short to 3 second clips and they

play13:00

seem to be designed to just make like

play13:03

short Loops or like little GIF GIF files

play13:06

they described it as text to live image

play13:09

so if you use an iPhone you've got that

play13:11

live photo feature where when you take a

play13:14

photo it sort of saves like two seconds

play13:16

of video as well so that you can find

play13:19

that right spot in the photo it

play13:21

seemingly is designed to generate that

play13:23

kind of little tetiny short clip that

play13:25

wasn't the only video announcement that

play13:27

they made at this event they also

play13:30

revealed Google vids now this is

play13:32

something that we don't have a lot of

play13:34

information about they put out a little

play13:36

like teaser video it's a minute and 27

play13:39

seconds here and it appears to make

play13:42

videos that look almost like PowerPoint

play13:44

Style videos using AI says let's choose

play13:47

a style and then you can see the styles

play13:50

of the videos Almost look like you know

play13:52

something you'd get out of canva or

play13:53

PowerPoint or keynote or something like

play13:55

that you pick your style you can give it

play13:57

a script or let AI generate a script and

play13:59

then it creates like a slide

play14:01

presentation Style video very

play14:03

reminiscent of like a PowerPoint or

play14:05

Google slides video again not available

play14:08

yet but it does say coming soon to

play14:10

Gemini for Google workspace but here's

play14:12

some research that came out this week

play14:14

about a new video generator it's called

play14:17

Magic time and it makes timelapse videos

play14:21

this one is very specific to time lapses

play14:24

so you can see some differences here

play14:26

where a prompt given to a normal

play14:29

animated video like bean sprouts grow

play14:31

and mature from seeds generate something

play14:33

like this where this magic time style

play14:36

generates this like timelapse version or

play14:39

Construction in a Minecraft virtual

play14:41

environment this one shows almost like a

play14:42

drone rotating view around it where

play14:45

magic time is making this time lapse and

play14:48

you can see a couple other examples here

play14:50

again it makes a very specific type of

play14:53

video which is amazing for people like

play14:55

me who love to make videos cuz I can use

play14:57

stuff like this for b-roll their GitHub

play14:59

page which will be linked below has a

play15:02

lot of examples of the type of

play15:03

animations it can do but the coolest

play15:06

part is the code is available on GitHub

play15:08

so if you know what you're doing you

play15:09

could run it locally or on a cloud but

play15:11

they also have a hugging face demo that

play15:13

you can play around with right now but

play15:15

for now let's take a peek at one of

play15:17

their like cached models here so I'll

play15:18

click on this and this was a prompt of

play15:21

cherry blossoms transitioning from

play15:23

tightly closed buds Etc and if I play it

play15:26

that's what that looks like and since

play15:28

we're talking about video if you

play15:30

remember last week I talked about how

play15:32

Neil Mohan the CEO of YouTube said that

play15:35

if open AI trained on their videos that

play15:38

would be a clear violation of their

play15:41

policies their terms well according to a

play15:43

report from The New York Times open AI

play15:45

transcribed over a million hours of

play15:48

YouTube videos to actually train gp4 now

play15:51

there hasn't been a ton of confirmation

play15:52

if you read the article it's a lot of

play15:54

hearsay and Google claims to have looked

play15:56

at the robots. text file on you YouTube

play15:59

and saw that open AI probably was

play16:01

scraping data but there was no real

play16:03

actual confirmation but pretty soon we

play16:06

could have a law that forces the AI

play16:08

companies to reveal what the models were

play16:11

actually trained on so there was a new

play16:13

bill introduced to the Congress on

play16:15

Tuesday that intends to force artificial

play16:17

intelligence companies to reveal the

play16:19

copyrighted material they use to make

play16:20

their generative AI models the bill

play16:23

would actually Force companies to file a

play16:25

report about what copyrighted material

play16:27

they used at least 30 days before

play16:30

actually releasing that AI model no clue

play16:32

whether or not something like this will

play16:34

get past but seeing as how some of the

play16:36

biggest companies in the world namely

play16:38

Google Microsoft and meta may not want

play16:42

to reveal what data they actually

play16:45

trained on and being as powerful as they

play16:47

are I would imagine there could be some

play16:49

lobbying going on behind the scenes to

play16:52

keep bills like this from getting past

play16:54

but I'm just speculating here Adobe on

play16:57

the other hand is taking a completely

play16:59

different approach they're actually

play17:00

willing to buy data off of creators to

play17:04

train on for their large language models

play17:06

Adobe wants to create their own version

play17:08

of Sora and in order to do that they

play17:10

need a large amount of video training

play17:13

data well in order to get that video

play17:15

training data Adobe is offering to pay

play17:18

between $3 per minute and $7 per minute

play17:22

to purchase video content off of

play17:24

creators Adobe is looking for everyday

play17:26

things like people riding a bike or

play17:28

walking down the street or the types of

play17:30

things you'd see in normal stock video

play17:33

so if you like to go out into the world

play17:34

and film stock video footage you might

play17:37

be able to make some extra money by

play17:39

submitting it to Google and letting them

play17:41

pay you a few bucks per minute of video

play17:43

I think if a bill like we just looked at

play17:46

does end up getting passed this is

play17:48

probably the way of the future for how

play17:50

these large language models will get

play17:51

trained meta also announced this week

play17:53

that they're going to take stronger

play17:54

measures to identify AI generated photos

play17:58

basically when people upload photos to

play18:00

places like Facebook Instagram or

play18:02

threads meta is going to use their own

play18:04

AI detector to look for things that are

play18:07

commonly present in AI photos and try to

play18:10

identify when the photo or image was

play18:12

made with AI previously meta was able to

play18:15

label photos that was made with their

play18:17

own emu AI image generator as well as

play18:21

identify AI photos that the uploader

play18:24

specifically marked as AI but now it

play18:26

appears that they're actually going to

play18:28

use AI to try to spot AI another big

play18:32

story of the week was udio udio is a

play18:35

really really good AI music generator

play18:38

you give it a prompt you can suggest

play18:40

Styles you can decide whether you want

play18:42

to write your own lyrics have ai

play18:43

generate lyrics omit lyrics Al together

play18:46

and the outputs are really good like a

play18:49

lot of songs I've heard I would not have

play18:51

known that they were AI unless somebody

play18:53

told me that they were AI they are that

play18:55

good now I'm not going to go too deep

play18:57

into udio in this video because the

play19:00

video that I released the day before

play19:02

this video went live was all about AI

play19:04

music generators with a very heavy focus

play19:07

on udio so if you really want to learn

play19:09

about udio check out this video here but

play19:12

one really interesting thing about udio

play19:14

before I move on is that it's actually

play19:16

being supported by other musicians

play19:19

musicians like will IM and common and it

play19:21

has backing from people like the

play19:23

co-founder and CTO of Instagram and the

play19:26

head of Gemini at Google and a16 is

play19:30

backing it so lots of money getting

play19:32

pumped into this and musicians actually

play19:34

seem to be supporting this platform so

play19:37

something really interesting to watch

play19:38

here and since we're on the topic of AI

play19:40

music Spotify is rolling out a new

play19:42

feature where you can have it generate a

play19:45

playlist for you using AI you just give

play19:47

it a prompt like I need pump up music

play19:50

for the gym and it'll make a playlist of

play19:52

songs it thinks is pump up music for the

play19:54

gym or I need sad music because I'm

play19:57

painting a sad picture and I want to

play19:59

hear sad music and it will generate a

play20:02

list of sad music for you nothing super

play20:04

mind-blowing or groundbreaking but I

play20:06

thought I'd share since we're on the

play20:08

topic of Music anyway now let's talk

play20:09

about AGI for a second Elon Musk thinks

play20:12

it's coming within the next year in an

play20:15

interview on an xas Elon Musk said that

play20:18

artificial intelligence will be smarter

play20:20

than the smartest human probably by next

play20:22

year or by

play20:23

2026 so sometime within the next year

play20:26

and a half he believes we're going to

play20:28

essentially have AGI Yan laon one of the

play20:31

Godfathers of AI and meta's cheap AI

play20:35

scientist thinks otherwise he believes

play20:37

that large language models will never

play20:39

actually reach human level intelligence

play20:41

in his article he doesn't actually say

play20:43

that he doesn't believe that AI can't

play20:45

reach human level intelligence he just

play20:47

doesn't believe that large language

play20:49

models the current standard that we're

play20:52

using for AI right now isn't going to

play20:55

reach human level intelligence back in

play20:57

February we talked about how meta is

play20:59

working on V jeppa this is meta's new

play21:02

architecture for self-supervised

play21:04

learning this is the technology that Yan

play21:07

laon actually believes will at some

play21:10

point hit human level intelligence this

play21:13

week the much anticipated Humane pin

play21:15

started getting into the hands of

play21:16

consumers if you don't remember it's a

play21:19

little pin device that goes on your

play21:20

shirt it has a little projector that can

play21:22

project things onto your hand it has a

play21:24

camera it has a microphone so it listens

play21:26

to voice commands and it's designed TR

play21:28

to be sort of a replacement for a

play21:31

smartphone and so far the reviews

play21:34

haven't been super favorable here's a

play21:37

quick recap as of right now the Humane

play21:39

pin is an incredibly poor proposition so

play21:42

no I can't recommend the AI pin in the

play21:44

form in which I received it I can't

play21:46

imagine a world or a use case where

play21:48

someone would prefer this over what

play21:50

already exists the pin is not worth the

play21:52

money not yet and probably not anytime

play21:55

soon you definitely should not buy

play21:58

planning to replace your phone now the

play22:00

biggest complaints about the Humane pin

play22:03

were that it really doesn't do anything

play22:05

beneficial over your smartphone people

play22:07

complained about holding up your hand

play22:09

like this to get through menus and stuff

play22:12

actually gets fairly tiring on your

play22:14

hands they also said that in bright

play22:16

light you can barely see the projection

play22:18

so it's hard to use people complained

play22:20

that the gestures were confusing and

play22:22

complicated there's no privacy if you

play22:25

are actually trying to prompt it you

play22:27

have to speak speak out your prompts so

play22:30

in public it just feels awkward to be

play22:32

talking to a little computer on your

play22:34

chest but the biggest complaint that

play22:36

everybody made about it was that it is a

play22:38

$700 product with a $24 per month fee

play22:43

and then if you ever cancel that $24 per

play22:46

month fee the product just stops working

play22:49

it's a paper weight saying that every

play22:51

single person that talked about the

play22:53

Humane AI Pin said that they thought

play22:55

that the technology was really really

play22:58

cool just not very practical or usable

play23:00

yet and that it still might get to

play23:02

something that people actually want that

play23:05

could replace a phone it's just nowhere

play23:07

close yet finally I thought this one was

play23:10

a kind of fun article about somebody

play23:12

making a lot of money using AI a card

play23:16

game developer paid an AI artist

play23:19

$90,000 to generate card art because no

play23:22

one was able to come close to the

play23:23

Quality that he was delivering now I

play23:25

have a little bit of an issue with them

play23:27

saying they're an AI artist cuz I think

play23:30

the better term is AI assisted artist I

play23:34

don't actually know if like an AI artist

play23:36

is a thing because if you're just

play23:38

pressing a button and letting AI

play23:40

generate the output I don't really feel

play23:42

like you are much of the artist however

play23:44

I do really appreciate AI assisted art

play23:48

where these cards here somebody

play23:50

generated the images for every single

play23:52

card using AI but then they went and

play23:55

touched it up using Photoshop or

play23:57

whatever image editing tool they use to

play24:00

make sure that the colors the

play24:02

consistency of the characters all the

play24:04

Styles matched there wasn't extra

play24:05

fingers all the weird stuff that we get

play24:07

with AI they actually had to manually go

play24:10

in and fix so AI generated the initial

play24:14

sort of rough draft of the image but

play24:16

then the artist actually then went and

play24:19

made it into what the card designer

play24:22

wanted AI helped them make a lot more

play24:25

images at scale but the artist still had

play24:27

to do work on all of the images so it's

play24:30

more AI assisted art not really AI

play24:33

generated art I don't know that's a

play24:34

little bit of a soap box I'll step off

play24:36

it now I just thought it was really cool

play24:38

that somebody was using an AI art

play24:40

generator to make the initial concept

play24:42

then they cleaned it up with Photoshop

play24:44

they did this at scale and earned

play24:46

$90,000 from the company that hired them

play24:49

to do this and finally the last

play24:51

announcement this is more of a selfish

play24:52

announcement but our podcast the next

play24:55

wave is officially launched you can find

play24:59

the first episode that we did with our

play25:01

vend from perplexity it is available on

play25:03

YouTube Spotify Apple podcast anywhere

play25:05

you listen to podcast I highly recommend

play25:07

the YouTube version because the editors

play25:09

did an amazing job of putting some cool

play25:12

graphics and overlays on it it's a

play25:14

really fun video to watch but if you're

play25:17

driving going for a workout doing

play25:18

whatever you do and you like to listen

play25:20

to audio podcasts it is available in

play25:22

pure audio form as well HubSpot who yes

play25:26

is the sponsor of today's video but is

play25:28

also the producer of this podcast is

play25:31

doing a really cool competition they're

play25:33

giving away Apple Vision pros and all

play25:35

sorts of cool stuff for subscribing

play25:38

liking reviewing doing all of that kind

play25:40

of stuff with the podcast I don't have

play25:42

all the details on the competition yet

play25:44

make sure you subscribe to the podcast

play25:46

and like it and maybe leave a review and

play25:49

the good people at HubSpot could

play25:50

potentially hook you up it's not to

play25:52

mention I might be a little biased but I

play25:54

also think the podcast is really good

play25:55

with videos like this I spend a very

play25:57

short amount of time time per topic with

play25:59

something like the next wave podcast me

play26:01

and my co-host Nathan Lans we get to

play26:04

Deep dive have longer form conversations

play26:07

talk more about the ethics and the

play26:08

implications and the long-term timelines

play26:11

that we see with this technology so it

play26:13

really gives us that platform to go much

play26:15

longer much deeper and bring on really

play26:18

amazing guests who could help us and you

play26:21

better understand the stuff that we're

play26:23

talking about in this crazy fast-paced

play26:26

AI world that we're in that's my pitch

play26:28

for the podcast check it out I'll make

play26:30

sure again it's linked up in the

play26:31

description it's called the next wave

play26:33

podcast I really think you're going to

play26:35

enjoy it and that's all I got for you

play26:37

today if you haven't already check out

play26:39

futur tools. where I curate all the

play26:41

latest and coolest AI tools all of the

play26:44

most interesting AI news that I come

play26:45

across and I have a free newsletter

play26:48

which will keep you in the loop with

play26:49

just the coolest tools and the most

play26:51

important AI news that I come across and

play26:53

if you sign up you get free access to

play26:55

the AI income database a cool database

play26:57

of interesting ways to make money with

play26:59

all of these various AI tools I'm going

play27:01

to have to add recording video for AI

play27:04

training for Adobe to the list pretty

play27:06

soon but check it all out it's over at

play27:08

futur tools. if you like videos like

play27:11

this you want to stay in the loop with

play27:12

the News latest tutorials latest tools

play27:14

all that good stuff make sure you like

play27:16

this video subscribe to this Channel and

play27:18

I'll make sure it keeps on showing up in

play27:19

your YouTube feed thank you so much for

play27:21

tuning in thanks again to HubSpot for

play27:23

sponsoring this video you all rock I

play27:26

appreciate you letting me nerd out over

play27:28

on YouTube and actually enjoying

play27:30

watching it for some reason I don't get

play27:32

it but I'm having fun I hope you're

play27:34

having fun let's do it again I'll see

play27:36

you in the next one

play27:38

byebye

play27:40

[Music]

play27:43

CSF has you cover with AI TOS

play27:48

too

play27:51

m i press the Subscribe button if you

play27:55

like

play27:58

where's the light for you

play28:01

to now w flow it out and he loves you

play28:08

[Music]

Rate This
β˜…
β˜…
β˜…
β˜…
β˜…

5.0 / 5 (0 votes)

Related Tags
AI NewsLarge Language ModelsGoogle CloudOpenAIGPT-4AI MusicAI EthicsTechnology AdvancementsAI in BusinessAI Art