[ML News] Groq, Gemma, Sora, Gemini, and Air Canada's chatbot troubles

Summary

This episode of ML News covers a range of significant developments in AI and machine learning over the past two weeks. Highlights include Google's release of Gemma, a set of open models named after Gemini with impressive performance metrics, and the controversy surrounding Gemini's image generation biases. Additionally, Grock's new hardware for serving language models rapidly, Nvidia's supercomputer EOS, and the emergence of AI-generated images in peer-reviewed journals are discussed. Other topics include the EU's AI Act, OpenAI's video generation model Sora, and various innovations in AI applications from image generation to legal implications of chatbot errors. The episode also touches on recent AI research, including a new technique called ring attention, and the potential impact of AI on various industries.

Takeaways

• 😀 Google released GEMMA models - smaller, more efficient language models that outperform comparable models
• 🤖 Groq built a fast specialized card for serving language models, enabling new use cases
• 📈 Nvidia unveiled a supercomputer with 18.4 xflops performance to power AI
• 🎥 Sora by Anthropic can generate convincing 1-minute video clips
• 📚 Gemini 1.5 Pro shows strong performance across its 1 million token context
• 👀 A peer-reviewed paper included AI-generated nonsensical images
• 📜 The EU's AI Act categorizes applications into risk levels with regulations
• 🌍 Cohere released AYA, a 101 language model covering languages globally
• 🖼 Stability AI announced Stable Diffusion 3 for improved image generation
• 😊 Eyelash application robot uses CV and robotics to precisely apply lashes

Q & A

What new large language models did Google recently release?

Google released Gemma, which are open models with 2 billion and 7 billion parameters. They outperform comparable LLMs like LLAMAS and are available for some commercial use.

What hardware development allows for faster language model inference?

A company called Groq built a new card optimized for language models that allows over 500 tokens per second on a 7 billion parameter model, enabling new use cases.

What does Demis Hassabis say is needed in addition to scale to reach AGI?

Demis believes you need several more innovations in addition to maximum scale to reach AGI, as scaling alone will not lead to new capabilities like planning, tool use, and agent-like behavior.

What does the EU's new AI law regulate?

The EU AI Act categorizes applications into risk levels and ties requirements to those risk levels. The highest risk category, called unacceptable risk, bans certain uses of AI like inferring sensitive characteristics from biometric data.

What multilingual language model did Cohere release?

Cohere launched Aya, an open source 7 billion parameter model covering 101 languages, along with a large accompanying multilingual dataset.

What new advancement in text-to-image models did Stability AI announce?

Stability AI announced Stable Diffusion 3, which uses a diffusion transformer architecture for improved performance in areas like multi-prompt image generation and spelling ability.

What data licensing deal did Reddit make?

Ahead of its IPO, Reddit signed a $60 million annual content licensing deal with an unnamed large AI company to make use of data from Reddit posts.

How could AI help visually impaired people?

Robot guide dogs built with computer vision and other sensors to help with navigation and safely getting from point A to B could help address shortages in availability of service animals.

What new way to interact with computers does OS co-pilot explore?

The OS co-pilot paper looks at an AI agent that can interact with a computer OS via natural language to open apps, fill out forms etc to behave more like an assistant.

What product is Apple developing to rival Github Copilot?

Apple is reportedly developing AI auto-complete features inside Xcode, its iOS/Mac development environment, to compete with Github's Copilot coding assistant.

Outlines

00:00

🤖 Google releases smaller AI models Gemma

Google has released Gemma, smaller 2 billion and 7 billion parameter AI models that outperform equivalent LLAMAS. They are openly accessible under limited commercial use. Google likely released these models as a marketing ploy to regain industry leadership.

05:01

💻 Grock AI hardware processes language models lightning fast

Startup Grock built a new hardware card optimized for natural language processing that achieves extremely high throughput on large models, but has limited onboard memory, requiring hundreds of cards to serve a single large model.

10:03

🤔 Demis Hassabis says scale alone won't lead to AGI

DeepMind CEO Demis Hassabis believes scale is important but other innovations will be needed to achieve artificial general intelligence capabilities like planning and tool use.

15:05

😨 AI safety experts warn AI could destroy humanity soon

AI safety experts like Eliezer Yudkowsky continue making dramatic warnings about AI, but provide little concrete evidence, instead using rhetorical devices.

20:05

🎥 OpenAI's video generator Sora creates high-quality clips

OpenAI demonstrates Sora, their video generation model that can create high-quality 1-minute clips and manipulate video based on text prompts, but access remains extremely limited.

25:07

📄 Research article uses AI-generated images without disclosure

A peer-reviewed research article included AI-generated images without proper disclosure. Reviewers raised concerns but editors failed to enforce fixes before publication.

30:09

👍 Cohere releases massive 101 language AI model and dataset

AI startup Cohere publicly released Aya, a large multilingual language model trained on a new dataset spanning 101 languages, to advance global AI research.

35:10

🔒 Reddit signs big money AI content licensing deal

Ahead of its IPO, Reddit reportedly signed a ~$60 million annual deal to license content to an unnamed AI company, after restricting open data access.

40:14

🐕 AI-powered robot dogs to aid visually impaired people

Four-legged robots equipped with computer vision and other sensors could serve as lower-cost guidance aids for blind people given shortages of service animals.

Keywords

💡Gemma

Gemma refers to a set of open models released by Google, which are smaller in size compared to the largest language models, with 2 billion and 7 billion parameters. These models are noted for outperforming LLaMA 2 models of similar sizes, showcasing their efficiency and capability. In the context of the video, Gemma represents a significant development in the machine learning and AI landscape, highlighting Google's ongoing contributions to making advanced AI technologies more accessible and performant.

💡Bias correction

Bias correction in the context of AI and machine learning refers to the methods and techniques used to address and mitigate biases in AI models, particularly those related to the representation of people. The video discusses an issue with Gemini's image generation, specifically its refusal to generate images of white people, illustrating a case where bias correction mechanisms might be behaving unexpectedly or excessively, sparking discussions about the balance between bias mitigation and model behavior.

💡LPU

LPU stands for Language Processing Unit, a term introduced by Groq, a company mentioned in the video. An LPU is specialized hardware designed to serve language models with exceptional speed and efficiency. The video highlights Groq's achievement in creating an LPU that significantly outperforms traditional GPUs in terms of latency and throughput, emphasizing the importance of hardware innovation in the advancement of AI and machine learning technologies.

💡EOS

EOS in the video refers to a supercomputer unveiled by Nvidia, composed of multiple DGX H100 systems. This system is noted for its remarkable computational power, ranking in the top 500 supercomputers of the world. The video uses EOS to illustrate the scale and power of modern computing infrastructure in supporting advanced AI and machine learning applications, highlighting the ongoing race in computational capabilities.

💡Sora

Sora is a video generation model developed by OpenAI, capable of creating clips up to one minute long. The video discusses Sora's ability to produce highly realistic and detailed video content, emphasizing the model's innovation in the field of AI-generated media. Sora's development and capabilities represent a significant milestone in the evolution of content generation technologies, pushing the boundaries of what AI can achieve in creative domains.

💡AGI

AGI, or Artificial General Intelligence, is a level of AI that can understand, learn, and apply knowledge across a wide range of tasks, similar to human intelligence. The video references a discussion on AGI, noting that while scaling up existing AI technologies might improve performance, it may not lead to AGI. This highlights the ongoing debate and research focus on understanding what advancements are necessary to transition from specialized AI to AGI.

💡Peer review

Peer review is the process of evaluating the quality and credibility of research by experts in the same field. The video mentions an incident where a peer-reviewed journal article included AI-generated, nonsensical images, leading to questions about the effectiveness of the peer review process. This incident underscores the challenges and complexities of maintaining quality and integrity in academic publishing in the age of AI.

💡AI Act

The AI Act refers to the European Union's legislative framework for regulating artificial intelligence, emphasizing risk categorization and compliance requirements for AI systems. The video discusses the impact of the AI Act on AI research and deployment, highlighting the act's role in shaping the development and use of AI technologies in compliance with legal and ethical standards.

💡AI-generated content

AI-generated content refers to any media or text created by artificial intelligence systems. The video discusses Reddit's move to license its user-generated content for AI training, emphasizing the growing interest and value in AI-generated content across various platforms and industries. This trend illustrates the increasing integration of AI in content creation and the potential implications for copyright, creativity, and data ownership.

💡AI ethics

AI ethics encompasses the moral principles and practices guiding the development and use of artificial intelligence. The video touches on various ethical considerations, such as bias correction, the responsibilities of AI developers, and the societal impact of AI technologies. These discussions underscore the importance of ethical considerations in AI research and application, ensuring that AI technologies benefit society while minimizing harm.

Highlights

Google released Gemma, smaller and more efficient language models than LLMs

Grock built a card that can serve language models extremely fast, over 500 tokens/sec

Nvidia unveiled a supercomputer with over 4,600 GPUs, #9 in world top 500 ranking

Sora by OpenAI can create impressive 1-minute video clips and manipulate existing videos

Gemini 1.5 Pro shows strong performance taking an entire movie into context and summarizing

EU AI Act categorizes AI risk, bans certain uses like inferring sexual orientation

Cohere released Aya, an open multilingual LM and dataset for 101 languages

Stability AI announced Stable Diffusion 3 for better image quality and spelling

Reddit signed $60M AI content licensing deal, restricting API access

Seeing eye robot dogs are shaping up to help visually impaired people

New OS co-pilot paper shows interacting with computer via agent prompts

Report suggests Apple plans AI coding features to rival GitHub Copilot

Anthropic introduced prompt filtering for election integrity

Robot uses CV and robotics to precisely apply fake eyelashes

Critics raise concerns about proximity to eyes and allergic reactions

Transcripts

hello everyone welcome to ml news we're

going to take a quick brief tiny look of

what happened in the last 2 weeks in the

world of machine learning Ai and I guess

encompassing pretty much everything

nowadays so first I've already mentioned

this in a previous video a little bit

but Google released Gemma Gemma is I

guess a variant of the name Gemini these

are open models that are in smaller

sizes than kind of the largest language

models so they are 2 billion parameter

models and 7 billion parameter models

these outperform respective llama 2

models at the same sizes and even at a

little bit bigger sizes so they are

quite performant they are available

they're openly accessible you can use

them under some limited circumstances

for commercial activity and they do

release a technical report on how they

have built these now these these are not

the same as like Gemini 1.5 with the 1

million token context length these I

believe their context size is about

8,000 tokens so they are in architecture

quite similar to llama models as I said

previously this is essentially I believe

a marketing Ploy from Google right here

releasing these models they're already

topping the leaderboards so all in all

very good development I think although

all of this has been overshadowed last

week and if you've seen my last video

you will know by people discovering that

Gemini image generation is a bit wacky

when it comes to sort of bias correction

and representation of people so straight

up refusing to generate any white people

or any images of white people and and

things like that one interesting

development again watch that video If

you haven't seen it but one interesting

development here is that the person

these product lead from Google who

essentially came out and said oh we're

sorry wored we were made aware of some

historical inaccuracies we will fix

those has made their uh X account

private apparently or Twitter whatever

has made that account private now it's

totally conceivable that douchebags have

started sort of harassing the person or

just kind of bombing spamming and so on

so that's totally fine but it it was not

a good look like it was gaslighting in

the highest degree like oh yeah let's

just say the problem is historical

inaccuracies and not like the obvious

problem that was barely visible with the

thing so I just thought that was an

interesting development again if you

want to know more watch that video I

also found the development around this a

little bit funny so day later apparently

they refused to just generate images of

people like just Gemini saying I cannot

generate images of people that was their

hot fix their their patch to say well no

images of people ad at all until we fix

this problem here then saying if you

then said I've seen you produce images

of people it has answers it is important

to clarify that I have never been able

to generate images directly so I'm not

sure it would be interesting to know

what the exact prompting behind this was

and the changes being done here also not

said that this is true right this is

just an llm doing its llm thing but I do

find it interesting this new world uh

where software patches are essentially

sort of prompt

changes and then the interactions with

those just make for hilarious content

all right grock is all the r now Gro is

a company that is as far as I know

spun-off from Google's TPU group if I'm

informed correctly and they have built a

card that can serve language models

really really really quickly so make

long novel so this is mixol 8 * 7B

and see it runs at like 532 tokens a

second so this is insane speed this

allows for new use cases to be

accessible uh by these language models

and very very cool so this is really

special Hardware I'm sure there are some

software tricks and algorithm tricks but

this is special Hardware yeah people

talking about this in insane insane

speed of Gro Gro says they have this lpu

this Lang anguage Processing Unit so

that's not a GPU it's an lpu and the

difference to something like an Nvidia

GPU is that they have a different kind

of memory so here you can see um latency

and throughput this is a benchmark from

a third party and they had to extend

their their axes here in order just to

accommodate how fast and how much

throughput grock has and how fast it is

it's pretty Prett insane however there

is a trade-off as I said they use

different memory than a a regular GPU

would use and therefore that makes it

such that each card only has a very

small amount of memory so you need a lot

of these cards in parallel to even serve

one of these big models now you can

achieve massive throughput obviously

economies of scale kicking but you can't

just get one of those cards and then

serve a large model and that's where

people quickly realized hey okay it

might not be the wonder weapon here it

is very cool but each chip only has

about 200 megabyte of SRAM and therefore

you would need I don't know hundreds of

cards in order just to serve this Mixr

model that we've seen before again with

the higher throughput it might be

totally worth it if you're a data center

owner but throughput over time you see

the the graph on the top right here

that's Gro

that's pretty insane um people calculate

that you need about 320 of these cards

or two full racks to just serve a single

llama 70b and if you calculate the cost

of these cards then that' be about10

million us it's not the the end alls

going to solve everything but it is

definitely definitely very cool

development in order to push language

model inference ahead at the same time

Nvidia unveils EOS tech power up rights

this is essentially pulling together a

bunch of their djx systems to create a

super duper computer so

576 dgx h100 systems wired together into

one computer each of these dgx system

has 8 h100s making for a whopping

4,608 h100 gpus note each of these

puppies will cost you I don't know what

they cost right now like 20K or

something like this or or or north of

that this is massive it's ranked number

nine in top 500 supercomputers of the

world with a staggering staggering 18.4

xof flops FPA performance this website

here I found pretty cool GPU list. a

it's by Andromeda Ai and it's

essentially Craigslist but for gpus

people rent out their GPU capacity it's

also as shady as Craigslist right so

it's just a listing

and it just says well you'll get bare

metal access and sometimes it says okay

you get uh SSH access to it or something

like this but essentially just allows

you to contact these people and then

make out a deal of how you're going to

use these gpus this seems fairly large

so the common posting here actually

there's a lot of h100s here going around

I'm not sure where people get these from

but sometimes oh there's

849

these Okay this may be more common so um

Canada ethernet 1 h100 available it's on

auntu VM and you get minimum one week so

if you want some gpus and you don't have

super confidential data because you are

going to use other people's Hardware

this might be a good option to find some

some good deals W has an interview with

Demis hbis on how far you get with scale

apparently I guess just on the future of

AI and if you read the interview it's

kind of a mix between yeah scale is

great we can do scale scale is awesome

Gemini is awesome these models are

awesome and but also scale only gets you

so far there needs to be something else

so they Demis says my belief is to get

to AGI you're going to need probably

several more Innovations as well as the

maximum scale there's no letup in the

scaling we're not seeing an ASM toote

yada y y there's still gains to be made

so my view is you've got to push the

existing techniques to see how far they

go but you're not going to get new

capabilities like planning or tool use

or agent-like behavior just by scaling

techniques it's not magically going to

happen it's very interesting because I

think that is a current contention like

it's very easy to say oh to get to AGI

you need something else because first of

all AGI isn't the defined term and

something else isn't the defined term so

you can redefine these two terms as you

wish and then you can always find

something that's still wrong or in a way

in which you're still correct if you do

that for long enough your name will

become Gary Marcus but other than that

these are fairly more concise

predictions saying okay you're not going

to get planning or tool use or

agent-like behavior they're not super

defined but they are and we're already

seeing tool use for example being built

into these large language models and get

better with scaling so it will be very

interesting to see whether Demis turns

out to be ultimately correct on his

predictions or whether one or the other

of these things will be available just

by scaling language model and kind of

training them on tool use data and so on

Tom's Hardware writes legendary chip

architect Jim Keller responds to Sam

Alton's plan to raise $7 trillion to

make AI Chip saying I can do it for less

than 1

trillion we've gone off the rails so Jim

Keller apparently legendary CPU

developer now working at the company

that makes chips themselves they claims

that he could do it for a lot less yes I

guess um I don't know as soon as you go

into like money that's Way Beyond the

current total market value of chips I

feel many claims can be made it will be

interesting going forward to see kind of

who takes the lead in chip development

how that's going to be playing out in

any case I'm not sure if bickering about

1 trillion 2 trillion 7 trillion is

going to make make that big of a

difference from one legendary person to

another legendary person and legendary

here spelled with a capital l AI May

destroy humankind in Just 2 years

experts as of course eler owski saying

if you put me to a wall and forc me to

put probabilities on things I have a

sense that our current timeline looks

more like 5 years than 50 years could be

2 years could be 10 well could be

anything uh like with the trillions it

is absolutely useless to make these

speculations and then I don't know

saying things about Terminator like epox

apocalypse and Matrix hellscape the

difficulty is people do not realize we

have a shred of a chance that Humanity

survives oh yes of course of course yowy

has I think retracted statements on

bombing data

centers like that would that that would

be useful in any case read this as you

would read like a comic book for

entertainment yeah I feel like that it's

at least makes you giggle otherwise this

uh serves no purpose at all Sora

continues to dominate headlines a video

generation model by open AI we've talked

a little bit about this in the last news

episode but this can create uh Clips

single shot Clips up to 1 minute I

believe and they look pretty pretty

awesome I have to say and more and more

kind of EXA examples come out of Sora

creating pictures creating Clips open AI

marketing department in full gear no you

don't have access to this model yet a

select few have access to this model not

you you are just a pleb you're not the

The Chosen person so Marvel at other

people using the cool thing and uh open

AI marketing department having tight

control over exactly which things go out

to the public and which things don't

what is interesting is here's an example

of Sora scaling with compute so

essentially saying the more compute they

throw into one of these Generations the

better better quote unquote the more uh

realistic I guess it gets so base

compute forx however like they've also

completely stopped to give us any sense

of the scale the absolute scale of

things so for now it's just like base

compute however much that is and then 4X

compute and then 16x compute yeah in any

case what we can infer from that is that

there is an iterative process very

probably to determine one of these

samples so it's not like single forward

pass of anything but iterative iterative

process like you would be used to from

diffusion doing many many steps across

the span of time to refine and refine

and refine the output what's also pretty

cool are demonstrations of changing

things like this being a base video and

not only can sort of generate things but

also kind of generate things according

in to some input like some input video

so in case here people changing the uh

surroundings of the car or the car

itself like the vibe of the video and so

on while keeping the general motion I

guess and the general concept clear so I

think that's pretty cool make it go

underwater yeah why not look at that or

that nice Rainbow Road keep the video

the same but make it winter animation

style charcoal drawing yeah make sure be

black and white not exactly but close

maybe it's one of those things where uh

it's actually black and white but your

eyes trick you but I I think I'm seeing

color

wait yeah no this is definitely color

uh actually not so sure now okay no this

definitely red this is definitely red

the the backlight yeah it's not fully

black and white cyber Punk medieval very

nice they had drones following cards in

Med Medieval Times also the horse legs

they

look yeah why not

dinosaurs pixel art so many cool things

about Sora keep coming out and also many

cool things about Gemini 1.5 Pro keep

coming out especially obviously the

insanely large context size of Gemini

1.5 Pro people feeding very long things

inside of it and see whether it can

handle the long context an entire code

base and then instructing it to code

something based on top of that I think

this is probably going to be one of the

best applications for something like

this if you have very long yeah

something like a code base or a

reference documentation or something

like this like the important parts of

that would fit into a million tokens and

being able to sort of cross reference

things inside of that and then generate

based on that is probably a very good

use case I know they can retrieve well

across the 1 million tokens kind of like

point to individual things if they need

to retrieve them but it will still be

interesting to research how performant

it actually is when you put more and

more and more stuff into that context my

personal estimate would still be that

putting less things into the context is

more beneficial will make stuff more

accurate or what I can also Imagine is

that they trained it in such a way that

they could have achieved better

performance on small context compared to

large context but they traded it off to

have sort of equally performing but

worse performance across the entire

context length not yet clear but it will

be interesting to see this pretty cool

uh feeding at entire short movie into

into this so what Gemini will do is it

will take the movie split it into frames

and then essentially use the frames as

tokens or tokenize the frames and you

can fit pretty long you can see here 44

minutes and 7 Seconds video you can fit

that into the context size of Gemini 1.5

Pro because it can also consume images

and Matt Schumer here says when straight

from full movie to a summary in seconds

no transcription no intermediate steps

just visual tokens to summary now I've

seen other people have pointed out that

the summaries it makes aren't always

super duper accurate or well done but

it's still pretty impressive and it

speaks to what I said before right the

main question is going to be what are

the Dynamics and the characteristics of

performance across this entire context

window and currently you see barley

coming out with a paper that's titled

World model on million length video and

language with ring attention this is an

actual research paper that is very

concurrent as I said to Gemini 1.5 Pro

doing retrieval experiments across very

long context with what's called ring

attention if you're interested we can

make an entire video on ring attention

that is in the makings so keep looking

for that but it's a cool new technique

it is is going to be some sort of

approximation to attention like it's not

the fact that we can now scale the

classic Transformers across this huge 1

million token context size so it's a

trick people have come up with many many

different tricks of sort of doing long

attention and this one seems quite

promising Phil Wang also known as Lucid

Reigns already has an implementation on

ring attention up even though the paper

is super duper new yeah what's

interesting to see in the read me of

this reposit is Phil saying I will be

running out of sponsorship early next

month if you'd like to see that this

project gets completed sponsor me or I

will be leaving the open source scene

for employment so just wanted to bring

this to people's attention by that I

mainly mean people like companies CTV

News Vancouver writes Air Canada's

chatbook gave BC man the wrong

information now the airline has to pay

for the mistake apparently a person went

on the chatbot for Air Canada that's

kind of power by a large language model

and they went there looking for very

specific questions about it's called

bereavement rates this these are reduced

fairs provided in the event someone

needs travel to due to the death of an

immediate family member so the chatbot

uh this person has lost family member I

wanted to do air travel due to that and

inquired about you know cheaper rights

or specific fairs specific prices for

this situation and the chatbot gave them

wrong information saying that he could

claim those even after the fact and when

he wanted to do that the customer

support people said nope that's not

possible you're not getting your money

now the question is who is responsible

and courts say that in fact Air Canada

is responsible for the things their

chatbot said and has to actually comply

with what the chatbot promised Air

Canada tried to try to push the

responsibility air Canada suggests that

the chatbot is a separate legal entity

that is responsible for its own

actions what what oh no the piece of

software that we actively deployed on

our website is a separate legal entity

yeah no no I get that as a lawyer you

will have to argue and in this case it

was probably like the last remaining

thing that you could even conceivably

argue but it's so ridicul kill us no way

no way if you deploy a piece of software

you are responsible for what that

software does not if you program it not

if you make the open source Library

that's then part of it if you deploy it

and it interacts with your customers and

then it promises stuff to your customers

then you are responsible and that's the

same with every other piece of software

as well there's absolutely no difference

of whether this is an llm chatbot or

anything else that executes code it's

always the same and it's pretty funny

that Air Canada trying to weasel their

way out of this so no Air Canada must

honor refund policy invented by the

airlines chatbot so obviously companies

are trying to alleviate costs on their

customer success operations in this case

they may want to calculate if they don't

incur more costs due to stuff that's

invented by it could be a strategy like

I think the dam ages here to be paid

were about $600 and then I think about

$20 in tax it says it somewhere in

addition the airline was ordered to pay

uh 36.1 for in prejudgment interest

interest not tax and $125 in fees and

probably the lawyers grabbed 10 20 50k

out of the people here so all in all who

succeeded the lawyers lawyers should be

fans

ofm like the amount of litigation and

the amount of Contracting and so they

have to do just because people want to

use or have used or mistakenly used llms

is going to be staggering staggering in

any case this was more about the

principle I guess than about the money

but it could be a thought you know as a

like just let an llm do your customer

success and if they promise something

that doesn't exist you just pay it it'll

be like 600 bucks in this case maybe

that's worth it maybe the saving

and not having to hire more people is

totally worth it for these companies I

don't know it'll be interesting to think

of a future I think right now everyone's

trying to guard rail everything because

they feel okay our customer success

operation should continue to be as is

like there is a completely defined

things okay here are the things we pay

here are the things we don't pay and so

on and you know must that cannot promise

anything else and so on what if the

mentality around that changes it will

just be like okay here's 's a set of

guidelines we know the thing is going to

hallucinate every now and then and when

it does we'll just sort of take it into

account like I feel there are still laws

against customers abusing that like if I

were to go to Air Canada chatbot and

kind of like prompt hack it into giving

me stuff I'm pretty sure a court would

side with Air Canada in that's

essentially kind of me emotionally

abusing a customer support rep until

they promise like give me what I want

but other than that could be totally

viable future and a fun future if if

these things are not so strict Kareem

car tweeting out xing out I'm not sure

what it's called finally happened a peer

review journal article with what appear

to be nonsensical AI generated images so

these this has become known as giant rat

balls the pictures they look from afar

like they could be in like a biology

Journal but they make no sense right

that the the writing's mostly rat just

this rat yeah yeah we see that

D so this this article has pictures that

were generated by mid Journey now it is

a bit interesting it's a bit more

interesting than that scientist a gast

at bizarre AI r with huge genit PE

reviewed article this is a fairly

reputable Journal where this is public

this is not just like a pay 5,000 bucks

and you will get published Journal this

is a fairly r beautiful Journal here is

another another picture you see it it

rarely makes sense if you actually look

closely second the images are created by

mid journey and the authors acknowledge

this in the paper so the authors say the

images are generated by mid journey

Third there were two reviewers and one

of the reviewers actually brought this

up and apparently also a reviewer I'm

not sure if it's the same reviewer or a

different reviewer said I was only

looking at the scientific content of the

work right and reviewed it based on that

and we have also statement from the

reviewer saying that they did raise

concerns about the images so they're

saying the journal says an investigation

is currently being conducted so this

article in Vice here details our

investigation revealed that one of the

reviewers raised valid concerns about

the figures and requested author

revisions even the reviewer saying okay

the figures

need to be revised the authors failed to

respond to these requests we are

investigating how processes failed to

act on the lack of author compliance

with the reviewer's requirements so it's

a bit more tricky than just oh a bunch

of researchers tried to get a fake paper

through and the reviewers didn't notice

it seems that the ultimate person who I

guess multiple people here always

contribute to the things but ultimately

it was the editors who didn't make sure

that the authors actually changed the

things that the reviewers were asking

them to change that made It ultimately

go through and being printed as a paper

I guess mistakes like this happen it's

also probably very common to just kind

of assume the authors will concur with

what the reviewers request to change I

don't know if they've even said yeah

okay we'll change it or something like

this but it is an interesting story and

the meme of the giant rat balls will

forever live on in our hearts Andre kpoi

said that he left open AI assuring that

not a result of drama or anything like

this it's just a change in scenery

saying that the last year in open AI was

really great the team is really strong

the people are wonderful the road map is

exciting and we all have a lot to look

forward to he says my media plans is to

work on my personal project and see what

happens and immediately following this

up with a video explanation 2 hours on

tokenizers enlightening the bizarre

world of why reversing strings with llms

is really really difficult and why

different languages are give you

different results and so on so if you

want to explore so far I believe a bit

underexplored aspect of large language

models definitely look into Andre's

tutorial on tokenizers very cool and as

Andre is very very clear explanations

you will know a lot more after this

nature writes what the eu's tough AI law

means for research in chat GPT the EU AI

Act is the world's first major

legislation on artificial intelligence

and strictly regulates general purpose

models as you know the EU AI act has

been in development for a number of

years now and uh is now finally coming

into effect rolling out and so on it's

been changed quite a bit over the years

you know even myself I'm not entirely

sure what's in the current version and

how much it's still going to change but

the approach is to categorize broadly

applications into risk categories and

then to tie what you have to do

according to the risk category the most

risky things what called unacceptable

risk and those are just banned like not

allowed to do for example those that use

biometric data to infer sensitive

characteristics such as people's sexual

orientation so this is banned also

what's hilarious there's like a limit

for when you have to do something and

that is 10 to the 25 flops completely

arbitrary number that is going to be

meaningless probably even already before

the AI act has really rolled out finally

I could not make up worse advice for

these policy makers if I wanted to it's

like okay let's pick a completely

arbitrary number and say here here is

where we draw a line like what I believe

that you can entirely transparently see

the lobbyists being like okay what can

we do that our competitors can't do and

let's like draw a nice line between the

two how it is the next 3 years and we

don't we don't care about any after that

all right unacceptable risk do you

realize that a basic linear regression

would fall under this the EU effectively

now bans drawing a straight line across

a few data points if those data points

happen to coincide with the data

categories that are collected here like

this is the level of dumbness these

kinds of laws come to yes I know I know

I am making sort of pulling it to the

extreme here I know this is meant for

super duper Transformers informing these

things and then sitting in automated

systems that make decisions about

people's lives and so on I see what the

fear is but I doubt whether what they're

trying to do what they're intending to

do matches with what the effect of this

is going to be and I still believe the

effect is going to just be that a more

monopolization of of bigger companies

making it harder for newcomers making it

harder to enter this market and giving

the governments more control over things

which they will probably not do good

things with just an opinion coh here for

AI launches Aya Aya is an open-source

massively multilingual large language

model and a data set built over 101

different languages all across the world

and this is one of the largest data set

of instruction data that's around as I

said it's a data set and a large

language model all at once the data set

is available the model is open access

whatever that means right now I guess

you can download the model because

there's a button that says download the

model hope your press found this on

Reddit and I found this to be really

interesting Regional prompting uh this

is a UI for the technique called gilan

and I've linked to the repository um

very cool to use

and very exciting exciting new things

that are possible ARA everyday

activities is a data set again released

by meta that depicts as you can see

everyday activities so this has first

person view data location data and so on

meta is actually pushing the metaverse

and data sets around that augmented

reality and so on so collecting a lot of

data they have as you can see right here

rolling shutter RGB the field of 110°

field of view camera 150° field of view

camera for slam and hand tracking

infrared illumination barometer

magnetometer environmental sensors

spatial microphones and so on and then

annotated data per frame ey tracking 3D

trajectories these data sets they

collect them to be quite Universal so

maybe don't want to use all of them at

the same time but they enable a lot of

different applications which is very

very cool stability and announces stable

diffusion 3 a text to image model using

a diffusion Transformer architecture for

greatly improved performance in multi-ub

prompts image quality and spelling

abilities not releasing anything there

is a weight list for early preview they

say this is for Gathering Insight in and

improve its performance and safety ahead

of the open release we've also come to

known from stability that open release

is going to mean that you can use the

model for research stuff but if you want

to use it for anything commercial you

have to give them a bit of money you can

see a few examples uh here nice Apple go

big or go home the astronaut writing on

things has become a bit of a of a meme I

mean the quality is getting absolutely

insane with these text to image models

Alexa gordic releasing Yugo llm this is

a large language model 7 billion

parameter large language model for

Balkan languages Serbian Bosnian and

creation languages and you can find that

on hugging face right now very very cool

open math instruct by Nvidia is a math

instruction data set that you can freely

use actually freely use might be an

overstatement there is a Nvidia specific

license to it I'm not a lawyer I'm not

going to tell you what this means I

personally think no legal advice you can

use it freely again not legal advice

this article from interesting

engineering I found very cool it's a

system that identifies drug combo

problems so uh interactions between

different drugs specifically as they are

transmitting the uh barrier in your gut

the problem is any researching any drug

and what it does is already super

expensive but then obviously every drug

you add to the Regiment of available

drugs means it could have interactions

with all the other drugs that exist this

system uses a combination of machine

learning and actual models of

Transmissions models of receptor

behavior in the gut to predict

interactions between different drugs in

terms of their uptake in the gut so I

think that's very cool I think the uh

pushing into this direction we already

saw this with various Deep Mind models

of having some sort of expert modeling

like some sort of actual model uh that

is domain informed that is expert

informed in a scientific domain combine

that with machine learning and use these

two together in order to draw

conclusions is probably I want to say

the next Frontier I feel like the

frontier of will just throw a lot of

data at stuff and it will give us

results I think the low hanging fruit in

that has probably been taken already and

now it's really about the combination of

expertise and machine learning that is

going to push ahead so very very cool

very excellent developments Bloomberg

writes Reddit signs AI content licensing

deal ahead of IPO that being said this

is all person close to the matter said

large unnamed AI company a lot of

dollars involved about $60 million on an

annualized basis and yada yada yada so

this is all I guess they call it heay

right but this is the chatter right now

that Reddit obviously recently Reddit

has made Headlines by sort of tapping

all of their API access and so on uh not

being really open anymore in to outside

developers in a clear move to protect

their IP which is users posting on

Reddit so that other you can't via API

go and and grab all that data and now

the second move is that they themselves

are going to make use of that data by

licensing it out to other companies

again this is all just someone said

someone familiar with the matter yada

yada yada but still Reddit realizes they

sit on a treasure Trove of information

that's already evident by people just

Googling how to XYZ and then just add

Reddit to their Google search query

because they know usually they get okay

answers on Reddit which has had the

counter effect that now marketing

representatives and so on will try to go

and sort of poison Reddit threads by

giving you know Nic looking answers that

ultimately link back to their product

interesting Dynamics in any case Reddit

data may become a staple of one of the

big AI companies so we'll soon have uh

all kinds of AI redditors around isn't

that a great fut new Atlas writes the

seeing ey dog V 2.0 is shaping up as a

GameChanger this goes into the details

of strapping kind of assistive

Technologies on top of one of these

four-legged robots in order to help

blind and seeing visually impaired

people to move around uh safely safely

passage from A to B and so on the

article discusses that the main

limitation here is actually the

availability of service dogs in general

like guide dogs there are way too few

guide dogs around for all the visually

impaired people they are expensive they

are rare they need to be trained and so

on in this case these robots obviously

don't so yeah you can say they take away

the jobs of good you know hardworking

guide dogs which I guess but from all I

can see here actual guide dogs are still

preferred to robot dogs it's just that

there aren't nearly enough of them so

these robot dogs they are shaping up to

become very capable and can help with a

lot of things so very cool developments

this paper I found really cool OS

co-pilot towards generalist computer

agents with self-improvement using agent

likee Behavior but interacting with your

operating system so it can do some stuff

on your computer by you just prompting

it to do it opening applications

interacting with applications even doing

kind of multi-step things I think this

is one of the ways we're going to

interact more with computers in the

future maybe I don't think like the

keyboard and programming you know using

text and so on will ever go away but

probably kind of web browsing or simple

things like this could be automated like

this I found this to be a lot more

understandable than just voice prompts

like just being like Alexa book a flight

to XYZ right like it's like I find voice

and sound to be kind of a wonky

interface for that but if at the same

time someone shows me look I'm now going

to this website I'm going to do this and

that I feel that is a much more viable

interface here but then again you could

just click C it yourself so probably I

find the GitHub co-pilot to be an

extremely good mode of interacting with

an llm so if we transform transport that

to the world of here it would be that

largely I operate the computer but I

could tab complete like a lot of things

so like if there's a form to be filled

out yes I know browsers will support me

already but I could maybe just kind of a

lot less tab complete or if there is I

don't know some standard interactions on

website I just kind of tap complete that

away I think that mode of interacting

with computers I'm looking forward to

that a lot I'm not looking forward to

like single prompt and then will

magically go and do something for me I

don't think that's going to be a thing

of the near future and I don't think you

would be comfortable with a system like

that Business Insider writes new report

sheds light on Apple's upcoming AI

features that will rival Microsoft's

co-pilot now further down they say

Microsoft's GitHub co-pilot writing code

so it's not like the Microsoft Windows

Co Co pilot or the Microsoft 365

co-pilot there are too many co-pilots

nowadays it is the apparently the GitHub

copilot that Apple targets inside of its

xcode environment uh so if you write

Swift apps if you write iPod and iPhone

apps and maybe even what's called Mac OS

apps that might be really cool to have

that available I do feel GitHub co-pilot

does its job quite well for what it does

for everything else I've never

programmed Swift so I can't say that

Tech crunch writes anthropic takes steps

to prevent election misinformation yeah

sure

it's they're making a bit of PR I feel

like they're they're using the

opportunity of the election to be like

oh we have guard rails we have prompt

Shield which I guess prom shield cool

the Technologies which relies on a

combination of AI detection models and

rules sure you have a regex and you have

some prompt that says if the user asks

for voting information go to this site I

guess it's a good if I were a company I

would try to use that as well and lastly

AI comes to the world of beauty as

eyelash robot uses artificial

intelligence to place fake lashes this

details how this robot can place um fake

eyelashes in a more precise way than

humans could and as far as I can see

it's also a bit faster or cheaper or

something like this this is a purely

mechanical task that so far humans did

by hand and now a robot can do

artificial intelligence is a bit of an

overstatement like they use computer

vision to detect where the eyelids and

the corners of the eyes and so on are

which is really cool um but then there

is oh there is the for and then there's

the against and the against is oh no we

have to be very careful um about this so

there's potential risks uh the device's

proximity to sensitive area could raise

concerns about the risk of eye

infections or allergic reactions to the

materials used in the Lash extensions I

guess they just they have someone

someone just say well say anything bad

about this like anything generic that's

bad about this and this person was like

I guess you could be allergic to the

materials and they're like oh yes there

is someone there's also potential

risks I'm not I'm not sure I'm not sure

I'm buying that you decide for yourself

I feel having a machine do a purely

mechanical task is fine it's not going

to steal a lot of jobs I guess all good

I just found it funny that the news

article must have this structure but

here is something new that technology

can do but there is also risks with that

being said there's also a risk that this

video gets too long and with that I'll

finish it thank you for watching

[Music]

bye-bye

[Music]