0:00

Microsoft has released a new research

0:02

paper on Orca a revolutionary AI model

0:05

that learns from complex explanations of

0:07

gpt4 the most powerful language model in

0:10

the world this is a huge deal and I'm

0:13

going to tell you why in this video so

0:15

what is orca and why is it so important

0:18

well Orca is a 13 billion parameter

0:21

model that learns from complex

0:22

explanation traces of gpt4 which is a

0:25

much bigger model that can generate

0:27

almost any kind of text you can imagine

0:29

now why would Microsoft want to create a

0:31

smaller model that learns from a bigger

0:33

model isn't bigger always better when it

0:36

comes to AI well not exactly bigger

0:38

models are more powerful but they also

0:40

have some drawbacks they are very

0:42

expensive to train and run they require

0:45

a lot of computing resources and energy

0:47

and they are not very accessible to most

0:49

researchers and developers that's why

0:52

there has been a lot of interest in

0:53

creating smaller models that can still

0:55

perform well on various tasks such as

0:58

answering questions summarizing texts

1:00

generating captions and so on these

1:03

smaller models are usually fine-tuned on

1:05

specific data sets or instructions to

1:08

make them more specialized and efficient

1:11

however there is a problem with this

1:13

approach smaller models tend to have

1:15

poor reasoning and comprehension skills

1:17

compared to bigger models they often

1:19

make mistakes or give irrelevant answers

1:21

when faced with complex or ambiguous

1:23

queries they also lack the ability to

1:26

explain how they arrived at their

1:27

answers or what steps they took to solve

1:30

a problem but Orca is not just another

1:32

smaller model that imitates a bigger

1:34

model Orca is a smaller model that

1:36

learns from the reasoning process of a

1:38

bigger model it learns from the

1:40

explanations that gpt4 gives when it

1:43

generates its answers these explanations

1:45

are not just simple sentences or phrases

1:47

they are detailed traces of how gpt4

1:50

thinks step by step how it uses logic

1:53

and Common Sense how it connects

1:55

different pieces of information and how

1:57

it simplifies complex concepts by

1:59

learning from these explanations Orca

2:01

becomes much more capable and

2:04

intelligent than other models it can

2:06

handle more diverse and challenging

2:07

tasks it can give more accurate and

2:09

relevant answers and it can also explain

2:11

its own reasoning process to humans this

2:14

is a huge breakthrough for open source

2:16

AI Orca is set to be open source soon

2:18

which means anyone will be able to use

2:20

it and build upon it it will enable more

2:22

people to access the power of gpt4

2:25

without having to pay for it or deal

2:27

with its limitations Orca will also open

2:30

up new possibilities for AI research and

2:32

development especially in areas that

2:34

require more reasoning and understanding

2:37

skills to understand how Orca works we

2:39

need to First understand how gpt4 works

2:42

so gpt4 is more than a text generator it

2:46

performs tasks requiring reasoning like

2:49

answering factual questions summarizing

2:51

lengthy texts generating captions

2:54

writing essays and more interestingly

2:57

gpt4 can provide explanations for its

3:00

outputs these are found in the model's

3:01

internal States essentially its thoughts

3:04

or memories which hold the logic and

3:07

information used to generate outputs by

3:09

using specific prompts we can unveil

3:12

these internal explanations giving a

3:14

detailed view of how gpt4 thinks solves

3:17

problems and uses diverse sources of

3:19

information including its own memory the

3:22

web and Common Sense these explanations

3:24

are very valuable for smaller models

3:26

that want to learn from gpt4 they

3:29

provide more signals and guidance for

3:31

how to perform various tasks and how to

3:33

reason and understand different concepts

3:35

they also make the learning process more

3:37

transparent and interpretable for humans

3:39

this is what Orca does Orca learns from

3:42

these explanations that gpt4 generates

3:45

when it performs different tasks it uses

3:47

these explanations as its training data

3:49

and tries to imitate them as closely as

3:51

possible Orca also tries to generate its

3:54

own explanations when it performs

3:55

similar tasks and Compares them with

3:58

gpt4's explanations to improve itself so

4:01

Orca is actually based on vicuna a

4:04

previous open source model that was

4:06

fine-tuned on question answer pairs from

4:08

GPT 3.5 Orca extends by kuna by adding a

4:12

new technique called explanation tuning

4:14

which allows it to learn from complex

4:16

explanation traces of gpt4 explanation

4:19

tuning is a Fresh Approach that enhances

4:21

gpt4's skill to follow specific

4:24

directives by refining this AI with

4:26

prompts like summarize this in a

4:28

sentence or create a love Haiku we make

4:31

it more Adept at particular tasks but

4:33

explanation tuning goes beyond it hones

4:35

gpt4 to reveal its thought process using

4:38

prompts like think sequentially or

4:40

explain like I'm a child this way gpt4's

4:44

reasoning becomes more transparent this

4:46

technique involves standard and

4:47

explanation prompts former our usual

4:50

tasks like who leads France or craft a

4:52

winter poem the latter instruct gpt4 to

4:56

clarify its logic like think in steps or

4:58

show how you did it using both prompt

5:01

types together gpt4 produces complex

5:04

explanation traces for instance with the

5:06

standard prompt who leads France and the

5:09

explanation prompt think in steps gpt4

5:12

might provide a step-by-step explanation

5:14

this comprehensive response not only

5:16

tells us who the president is but also

5:18

illustrates gpt4's problem-solving

5:20

strategy and information sources

5:22

offering more insight than a simple

5:24

answer Orca leverages explanation traces

5:26

as learning material striving to mimic

5:29

them and generate its own for

5:31

improvement but where do these traces

5:33

come from Orca Taps into flan 2022 a

5:36

massive collection of over 1 000 tasks

5:39

and 10 000 instructions covering a

5:41

spectrum of subjects by sampling from

5:43

flan 2022 Orca gets a variety of tasks

5:46

and uses them to query gpt4 for

5:49

explanation traces it also creates

5:51

complex prompts from the data set to

5:53

test gpt4's reasoning like mashing two

5:56

tasks into one this way Orca learns from

5:59

diverse and intricate tasks fostering

6:01

many aspects of human intelligence Orca

6:03

is evaluated on a number of benchmarks

6:06

that test its generative reasoning and

6:08

comprehension abilities these benchmarks

6:10

include multiple choice questions

6:12

natural language inference text

6:15

summarization text Generation image

6:17

captioning and so on Orca is compared

6:20

with other models of similar size or

6:22

larger size such as vikuna 13B text

6:25

DaVinci 003 a free version of gpt3 chat

6:29

GPT 3.5 and gpt4 orca's performance is

6:33

Stellar topping all other open source

6:35

models in most benchmarks particularly

6:38

those needing deeper reasoning despite

6:40

its smaller size it matches or beats

6:42

chat GPT in many areas even competing

6:45

with gpt4 in tasks like natural language

6:48

inference or image captioning

6:50

here's a quick look at orca's Benchmark

6:52

performances on big bench hard BBH it

6:56

scores a 64 accuracy more than double of

6:58

vicuna 13bs 30 and surpassing chat gpts

7:03

59 and gpt4s 62 on super glue SG it

7:09

achieves an 86 average beating vicuna

7:12

13B 81 Tex DaVinci 003 83 chat GPT 84

7:19

and nearly matching GPT 4 88 on CNN

7:24

daily mail CDM Orca earns a rugel score

7:27

of 41 outperforming vicuna 13B 38 text

7:32

DaVinci 003 39 chat GPT 40 and closing

7:38

in on GPT 4 42 on Coco captions CC it

7:43

scores a cider of 120 higher than vicuna

7:47

13B 113 text DaVinci 0.003 115 chat GPT

7:54

117 and GPT 4 119 so as you can see Orca

8:00

is a highly versatile efficient model

8:02

performing well across tasks and domains

8:05

and soon to be open source it also works

8:08

on a single GPU orca's success reveals

8:11

multiple insights about ai's future

8:12

firstly it indicates that learning from

8:15

explanations as opposed to just answers

8:17

notably boosts AI intelligence and

8:20

performance by studying gpt4's

8:23

explanations Orca not only gains

8:25

Superior reasoning skills but also

8:27

provides a transparent look into its

8:29

problem-solving process secondly Orca

8:31

proves that despite their size smaller

8:34

models can match or outperform larger

8:36

ones learning from gpt4 Orca side steps

8:40

size related drawbacks showing that

8:42

smaller models can be more approachable

8:44

and efficient needing fewer resources

8:47

and energy and thirdly orca exemplifies

8:50

how open source AI through inventive

8:53

methods can match proprietary Ai and

8:56

demonstrates how open source ai's wider

8:58

accessibility can benefit more people

9:01

and spur more applications concerning

9:03

its positioning Orca isn't just a mini

9:06

gpt4 or another open source model while

9:09

it doesn't match gpt4's broad capacity

9:12

or knowledge base it harnesses gpt4's

9:15

reasoning making it smarter than other

9:17

small models it also surpasses gpt4 and

9:20

transparency by generating its own

9:22

explanation traces unlike other open

9:24

source models Orca learns from a varied

9:27

range of tasks and complex explanations

9:29

making it more intelligent and versatile

9:31

therefore Orca occupies a unique

9:34

position in the AI sphere combining

9:37

gpt4's prowess with open source ai's

9:39

accessibility and demonstrating the

9:41

potential of explanation-based learning

9:43

alright that's it for this video thank

9:46

you so much for watching if you like

9:47

this video please give it a thumbs up

9:49

and share it with your friends and if

9:51

you haven't already please subscribe to

9:53

my channel and turn on the notification

9:54

Bell so you don't miss any of my future

9:56

videos I'll see you in the next one

0:03

hey what is up guys welcome back to

0:05

another YouTube video at the world of AI

0:07

in today's video we're going to be

0:09

showcasing a new open source project by

0:11

Microsoft research and that is called

0:13

Orca it's a progressive learning project

0:15

from the complex explanation traces of

0:18

gpt4 now in this video we'll delve into

0:21

the new research paper which focuses and

0:24

introduces an Innovative technique for

0:26

training open source models now this is

0:28

quite remarkable guys because we're able

0:31

to see that there's open source projects

0:33

that are slowly but surely getting

0:35

closer to GPT 3.5 as well as gpt4 types

0:40

of results we can see in the suite right

0:42

here that it's surprisingly capable of

0:44

reaching GPT 3.5 turbo results in terms

0:48

of its performance capabilities and it's

0:50

basically able to do this by Computing

0:52

with no internet actual connection as

0:55

well as about eight gigabytes of hard

0:58

drive space and this is quite

1:00

astonishing because we're able to surely

1:02

get to this open source model that can

1:05

compete with chat gbc's 3.5 models as

1:08

well as gpt4 models and slowly but

1:11

surely we're able to see that there's

1:12

going to be a time where we're getting

1:14

closer to this point where open source

1:17

models can be able to actually compete

1:19

with these different types of GPT 3.5

1:21

models now this technique presents a

1:24

significant advancement by enabling

1:26

small models with a size of 13 billion

1:28

parameters to go beyond merely imitating

1:32

large foundational models such as GPT

1:34

3.5 and this is why we're going to be

1:37

showcasing this project as well as this

1:39

research paper

1:40

now the paper introduces a novella

1:43

approach that empowers these smaller

1:45

models to acquire reasoning capabilities

1:48

as well as well as which helps them

1:51

enable them to understand and

1:53

demonstrate the underlying thought

1:54

process behind arriving towards a

1:57

solution now this breakthrough is a

2:00

notable step forward in the field of

2:01

machine learning as it showcases the

2:04

enhancements of models capabilities that

2:06

can go beyond the pattern's recognition

2:08

Now by incorporating this new training

2:11

technique the researchers at Microsoft

2:13

research have been able to bridge the

2:15

gap between large foundational models

2:17

and smaller models which enables the

2:20

latter reasoning capabilities similar to

2:22

large counterparts and basically with

2:24

this development we're able to see that

2:26

it opens up a new possibility for

2:28

smaller models to tackle complex tasks

2:31

and provide more insightful responses

2:33

for different types of generative

2:35

content and this is something that will

2:37

slowly but surely Showcase in this video

2:40

we're going to talk a little bit about

2:41

the data that is used to create this

2:43

project as well as taking a look more as

2:46

to what you can do with this as well as

2:48

taking a look at the experiments of this

2:50

project and with that thought guys

2:52

before we delve deep into this project

2:54

it would mean the whole world to me if

2:55

you guys can go give Twitter page of

2:57

world of AI a follow I'm going to be

3:00

posting the latest AI news over here now

3:02

if you guys haven't subscribed please do

3:04

so turn on the notification Bell like

3:06

this video and firstly I just want to

3:08

say I am so sorry for not uploading for

3:11

the last few days you can see there's

3:14

been a trend where I haven't been

3:15

uploading daily after this two-week Mark

3:19

it's been really busy in June guys I've

3:22

been really hectic with a lot of

3:23

different projects as well as a lot of

3:25

different things so I'm gonna try my

3:27

best to upload daily but at this rate I

3:29

don't think I'll be able to do so in

3:31

within this month I just want to tell

3:33

you guys and be honest because I'm not

3:35

going to be able to work towards putting

3:37

out videos because I have a lot of

3:38

important stuff on my end that need to

3:41

complete but once that is done usually

3:44

around in July I'll be more free to

3:45

upload back to my regular schedule where

3:48

I upload two videos a day and with that

3:50

thought guys let's get right into the

3:52

video

3:54

now let's take a look at this figure in

3:57

the introduction now in figure one

3:59

there's the context of orca's research

4:01

project which demonstrates the superior

4:03

performance of orca now this is just

4:06

simply a 13 billion parameter model of

4:09

orca being compared to various

4:11

Foundation models including open ai's

4:13

chat gbt 3.5 as well as gpt4 now this

4:18

evaluation is conducted using by kona's

4:20

evaluation set now the results indicate

4:23

that orca's outperforming these

4:26

foundational models in terms of its

4:27

performance and accuracy on specific

4:30

evaluation sets now additionally to this

4:32

what the paper is showcasing is that

4:35

there's a similar favorable outcome when

4:37

testing Orca against evaluation sets

4:40

used in other research works it

4:42

indicates that there's General

4:43

Effectiveness across different scenarios

4:46

now moving on to the second figure we

4:49

can see that it focuses more on the

4:51

concept of explanation tuning with Orca

4:53

now again it utilizes a model with 13

4:56

billion parameter like the figure one we

4:59

showcased and basically this objective

5:01

is to close the performance gap between

5:04

Orca as well as the for foundational

5:06

models like text DaVinci or like

5:10

different ones such as orc uh Chaturbate

5:13

and mikona 13B and basically we're able

5:17

to see that's closing the Gap with these

5:19

different types of model now the

5:21

valuation is carried out with a diverse

5:23

range of professional and academic exams

5:25

such as lsats as well as SATs GMAT Etc

5:29

and notably this is the evaluation that

5:32

is conducted in a zero shot setting

5:34

meaning that there's no type of

5:36

conversation on the topic data that is

5:39

used to Aid the model's performance now

5:41

the results presented in this figure are

5:43

indicating that Orca achieves a

5:45

significant progression in closing the

5:48

performance Gap with the foundational

5:49

models all these different types of LMS

5:52

over here not specifically in comparison

5:54

the chat GPT Orca was able to

5:57

demonstrate a reduction in the

5:59

performance Gap by 5 Points and

6:01

furthermore it should be noted that the

6:03

Gap further diminishes when the system

6:05

message are optimized and this suggests

6:08

that the orca's explanation tuning

6:10

capabilities are able to enhance its

6:12

performance on professional and academic

6:14

exams which is kind of showcasing that

6:18

it can enable it to achieve competitive

6:20

results without the need for explicit

6:22

training as well as specific exam result

6:24

data now this is quite remarkable guys

6:28

because you can see that you're able to

6:29

train smaller models on Lower data sets

6:33

which can help you outperform or become

6:36

like you're able to be competitive with

6:40

these larger models which is quite

6:41

amazing as it showcases that you're

6:44

going to be able to efficiently and

6:46

effectively train different types of

6:48

models and this is something that will

6:50

surely showcase later on in the video

6:54

now before we actually go more deep in

6:57

depth with the other stuff in the video

6:58

where we talk a little bit more about

7:00

the actual analysis as well as

7:02

experiments I wanted to talk about the

7:04

challenges with existing models now for

7:07

now there's a problem with which we can

7:10

see with the existing models which is

7:12

interpretability and explainability in

7:15

complex machine learning models and

7:17

these are some of the several challenges

7:18

that are faced with these existing

7:20

models and these challenges include the

7:23

increasing complexity of models which

7:26

show that there's a lack of transparency

7:28

in decision making processes and the

7:31

difficulty in understanding the

7:32

underlying relationships between the

7:34

inputs and the outputs now additionally

7:36

to this there's a limited

7:37

interpretability which hinders the

7:39

identification of biases errors and

7:42

errors of uncertainty within the model

7:45

now the need for this method can be

7:47

addressed with challenges that are

7:49

crucial to enhance the trust the

7:52

transparency and accountability in the

7:54

deployment a complex machine learning

7:56

models across various domains and this

7:59

is something that they've noted down

8:00

with the Microsoft researchers which are

8:03

working on this project as some of the

8:05

challenges with existing parameters of

8:08

the actual models that exist

8:11

in this research paper there are

8:14

primarily four focuses that are

8:16

addressing the challenges that are

8:18

mentioned above and firstly we're going

8:21

to talk about the progressive learning

8:22

framework now we introduced this novel

8:26

Progressive learning method at the start

8:28

which is called Orca and it's basically

8:31

designed to enhance the interpretability

8:32

and the explainability in complex

8:35

machine learning models now this

8:37

framework enables the actual method to

8:39

like refine its understanding by

8:42

incorporating new explanation traces

8:44

over time and this is by accumulating

8:47

and analyzing a diverse range of

8:49

explanation traces and with Orca you're

8:52

able to see that there's a progressive

8:54

adaptation and a work towards improving

8:58

its interpretation interpretation sorry

9:00

of complex systems and this is one of

9:03

the key contributions to actually

9:05

working towards the challenges of

9:07

existing models secondly there's a

9:10

complex explanation traces

9:12

of emphasizing that it works towards the

9:16

importance of complex explanation traces

9:18

and understanding the decision-making

9:20

process of different types of models and

9:24

these traces basically provide a

9:25

detailed record of the model's internal

9:28

operations which focuses on including

9:30

data flow of computation and decision

9:34

actual outcome sorry now by leveraging

9:38

these traces orcas captures the actual

9:40

relationship between the inputs as well

9:43

as the actual outputs that are used

9:45

within the model and what this actually

9:48

does is that enables a deeper

9:50

understanding of the model's Behavior

9:52

now a third contribution is the actual

9:55

net enhancements of its interpretability

9:58

and through the use of orca we're able

10:01

to see that there's a progressive

10:02

learning and a complex explanation

10:04

traces which are used to demonstrate the

10:07

distinguished improvements of the

10:09

interpretability of complex machine

10:11

learning models now the insights that

10:14

are gained from accumulated explanation

10:15

traces helps enable to identify

10:18

important features which helps

10:20

understand decision boundaries as well

10:22

as highlighting areas of uncertainty and

10:25

biases within the actual model now this

10:27

interpretability enhancement contributes

10:29

to the building trust and understanding

10:31

in decision making processes of these

10:34

models

10:35

lastly the last key contribution is the

10:39

comparative performance evaluation now

10:41

with Orca it provides a comprehensive

10:43

comparator comparative evaluation that

10:46

is against various Foundation models

10:48

which we showcased above in this table

10:50

and basically the valuations demonstrate

10:53

that orcup is able to outperform these

10:56

models in terms of accuracy performance

10:58

and interpretability on different

11:00

evaluation sets as well as benchmarks

11:03

now this comparative analysis highlights

11:06

Effectiveness and robustness of orca and

11:08

complex scenarios and by addressing

11:11

these challenges that we talked about

11:12

we're able to see that there is an

11:16

introduction of towards the progressive

11:18

learning framework which emphasizes a

11:20

solution for a complex explanation

11:22

traces and enhancing interpretability

11:24

and conducting through performance

11:26

evaluations

11:28

now in table one we're able to see that

11:31

it overviews the popular models that

11:34

have been in like actually used for

11:36

instruction tuned with open ai's large

11:38

foundational models now these lfms serve

11:42

as a basis for comparison in the context

11:44

of Aura now in Orca sorry in contrast to

11:48

these models Orca has been able to

11:50

leverage complex instruction and

11:52

explanation for Progressive learning and

11:55

about utilizing detailed and

11:56

comprehensive explanations Orca was able

11:59

to enhance its learning process and

12:01

improving its interpretability now this

12:04

approach set showcases that orcas are

12:07

part of different types of models that

12:09

rely solely on instruction tuning with

12:12

showcasings showcases the unique

12:15

contribution and effectiveness of orca

12:17

in the domain interpretability of

12:19

machine learning

12:21

in figure four we're able to see that it

12:24

illustrates the process of instruction

12:26

tuning with gbt4 and this approach where

12:30

the system takes the user instruction

12:32

for a specific task along with the input

12:34

and generates an appropriate response

12:37

now existing works such as alpaca by

12:40

kuna as well as their variants follow a

12:42

similar template for training smaller

12:44

models and they utilize the training

12:47

data in the format of user instruction

12:49

input as well as output to fine-tune

12:52

these models now instruction tuning

12:54

involves training the models to

12:56

understand and follow user instruction

12:58

in order to generate accurate and

13:01

contextual appropriate responses Now by

13:03

incorporating both the user instruction

13:05

and the input these models tend to aim

13:08

to capture the desired Behavior or

13:10

responses from the system now we can see

13:13

over here that alpaca of ikuna as well

13:16

as similar approaches that have been

13:18

utilized to train smaller models

13:20

leveraged instruction tuning techniques

13:23

to improve the performance of

13:25

effectiveness of these smaller models

13:27

Now by training on specific user

13:30

instructions as well as Associated

13:31

inputs these models can learn to

13:34

generate responses that align with

13:36

intended tasks or approaches

13:39

now in context or Orca figure 4 is able

13:43

to highlight the difference in

13:44

approaches now while instruction tuning

13:47

with gpt4 or GPT 3.5 relies on user

13:50

instruction as well as user inputs Orca

13:53

takes a distinct path by leveraging

13:55

complex instructions and explanations

13:57

for Progressive learning and this is by

14:00

utilizing detailed explanations in

14:02

addition to instruction and basically

14:05

with Orca it enhances its learning

14:08

process and offers an improved

14:10

interpretability which sets it apart

14:12

from models that solely rely on

14:14

instruction tuning techniques

14:18

in figure 5 it depicts the process of

14:21

explanation tuning with gpt4 now this

14:24

approach involves not only user

14:26

instructions and inputs but also system

14:28

instructions which guides the system to

14:31

generate well-reasoned and coherent

14:33

responses now the system instructions

14:35

are sampled from a diverse set that

14:38

includes a Chain of Thought reasoning

14:39

steps as well as explaining like I'm

14:43

five now being helpful which shows other

14:46

relevant structures of this model Now by

14:49

incorporating these rich and

14:50

well-structured instructions small

14:53

smaller models that can be tuned to

14:55

mimic and thinking processes of gpt4 can

14:58

be paired with Orca now explanation

15:01

tuning with gbt4 aims to enhance the

15:03

quality as well as the coherence of

15:05

responses generated by smaller models

15:08

about providing an explicit system

15:10

instruction these models are guided to

15:13

produce while recent as well as

15:15

contextual appropriate outputs these

15:18

system have been able to Showcase that

15:20

there's instruction contributions to

15:22

capture the thinking process as well as

15:25

the reasoning abilities of larger GPT

15:27

form models by utilizing orca

15:30

lastly I want to emphasize a little bit

15:32

more on the explanation tuning and we'll

15:35

end off this video by showcasing a

15:37

couple examples of what you can do in

15:39

terms of orca in comparison to different

15:42

types of models such as GPT 3.5 turbo as

15:45

well as gpt4 now we talked about

15:48

training before but training plays a

15:51

vital role in the development and the

15:53

deployment of orca and this section

15:55

basically showcases an importance that

15:58

all reviews the training process

16:00

including tokenization and sequencing as

16:03

well as loss computation now code

16:05

tokenization and the training of orca as

16:08

well as the Llama by bit pair encoding

16:11

tokenizer is employed to process the

16:14

input examples and this is the notable

16:16

feature of llama's tokenizer which is an

16:18

ability to split all numbers into

16:21

individual inputs as well as digits

16:23

which ensures this numerical information

16:26

is appropriately representative within

16:29

its data now sequencing is used after

16:32

tokenization which basically focuses on

16:35

the inputs examples that are organized

16:37

into sequences and basically they're

16:40

suitable for training the actual model

16:42

and these sequences ensure that the

16:44

information is properly structured and

16:47

can be effectively processed by the

16:49

model during the training phase and

16:51

lastly there is the focus of loss

16:54

computation and this is during the

16:56

training process where the model

16:57

performance is evaluated through a

16:59

computation of loss and the loss is

17:02

basically a function that quantifies the

17:05

disparent sorry the discrepancy between

17:08

the predicted output of the model as

17:11

well as the desired outputs

17:14

now the experiments conducted by orca to

17:18

evaluate its performance for conducted

17:20

in-depth with a focus that provides

17:22

detailed insights into its capabilities

17:25

the evaluation encompassed a variety of

17:28

tasks and benchmarks to assess orca's

17:30

language understanding reasoning

17:33

abilities as well as its overall

17:34

performance now the experiments involve

17:37

the rigorous evaluations as well as

17:39

protocols that considered different

17:41

abilities including writing

17:43

comprehension analytical mathematical as

17:47

well as logical reasoning now various

17:49

data sets and benchmarks are actually

17:51

utilized to provide a comprehensive

17:53

evaluation of orca's ability to Showcase

17:57

its performance across different domains

17:59

now for writing abilities we're able to

18:02

see later on in the actual research

18:04

paper that it showcases its evaluation

18:07

where orca's responses were assessed by

18:10

coherence contextual relevance as well

18:13

as overall quality and this is something

18:16

that was compared with different types

18:18

of large models such as GPT 3.4 as well

18:22

as or sorry not 3.4 3.5 and gpt4 and the

18:27

model's Proficiency in generating

18:28

well-structured and informative

18:30

responses that were thoroughly exempt

18:33

with this comparison now in terms of its

18:36

comprehension the orca's understanding

18:39

of complex textual information as well

18:41

as its ability to accurately interrupt

18:44

and respond to comprehensive based

18:46

questions were assessed and this

18:48

evaluation was aimed to measure the

18:51

model's comprehension on terms of its

18:54

skills and its capabilities to derive

18:56

relevant information from given tasks

18:59

now if you want to get more information

19:00

on these types of experiments if you

19:03

scroll down a little bit to the end

19:05

you're able to get a better

19:06

understanding of these different types

19:08

of responses that were compared with

19:11

different larger models such as vicona

19:13

gpt4 and Etc and we can see down over

19:17

here there's temporarily a section that

19:21

focuses on the types of comprehensive as

19:24

well as the performances across

19:27

different domains such as math and as

19:30

well as the different things that we

19:31

talked about before and with that

19:33

thought this concludes today's video on

19:35

this new Amazing Project of orca I hope

19:38

you got some sort of value out of this

19:40

research paper this is something that

19:42

I'll be showcasing later on as they

19:44

continuously work towards improving as

19:47

well as incorporating different

19:48

Innovative tactics to improve This

19:50

research paper and with that thought

19:52

guys thank you so much for watching hope

19:55

you got some sort of value out of this

19:56

if you guys haven't followed the Twitter

19:58

page please do so if you haven't turn on

20:00

notification Bell subscribe like this

20:03

video If you guys haven't seen any of my

20:04

previous videos I highly recommend that

20:06

you do so and with that thought thank

20:08

you so much for watching I'll see you

20:10

guys next time peace out fellas

What is Artificial Intelligence and how does it work? What are the 3 types of AI? The 3 types of AI are: General AI: AI that can perform all of the intellectual tasks a human can. Currently, no form of AI can think abstractly or develop creative ideas in the same ways as humans.  Narrow AI: Narrow AI commonly includes visual recognition and natural language processing (NLP) technologies. It is a powerful tool for completing routine jobs based on common knowledge, such as playing music on demand via a voice-enabled device.  Broad AI: Broad AI typically relies on exclusive data sets associated with the business in question. It is generally considered the most useful AI category for a business. Business leaders will integrate a broad AI solution with a specific business process where enterprise-specific knowledge is required.  How can artificial intelligence be used in business? AI is providing new ways for humans to engage with machines, transitioning personnel from pure digital experiences to human-like natural interactions. This is called cognitive engagement.  AI is augmenting and improving how humans absorb and process information, often in real-time. This is called cognitive insights and knowledge management. Beyond process automation, AI is facilitating knowledge-intensive business decisions, mimicking complex human intelligence. This is called cognitive automation.  What are the different artificial intelligence technologies in business? Machine learning, deep learning, robotics, computer vision, cognitive computing, artificial general intelligence, natural language processing, and knowledge reasoning are some of the most common business applications of AI.  What is the difference between artificial intelligence and machine learning and deep learning? Artificial intelligence (AI) applies advanced analysis and logic-based techniques, including machine learning, to interpret events, support and automate decisions, and take actions.  Machine learning is an application of artificial intelligence (AI) that provides systems the ability to automatically learn and improve from experience without being explicitly programmed.  Deep learning is a subset of machine learning in artificial intelligence (AI) that has networks capable of learning unsupervised from data that is unstructured or unlabeled.  What are the current and future capabilities of artificial intelligence? Current capabilities of AI include examples such as personal assistants (Siri, Alexa, Google Home), smart cars (Tesla), behavioral adaptation to improve the emotional intelligence of customer support representatives, using machine learning and predictive algorithms to improve the customer’s experience, transactional AI like that of Amazon, personalized content recommendations (Netflix), voice control, and learning thermostats.  Future capabilities of AI might probably include fully autonomous cars, precision farming, future air traffic controllers, future classrooms with ambient informatics, urban systems, smart cities and so on.  To know more about the scope of artificial intelligence in your business, please connect with our expert.

מהי בינה מלאכותית וכיצד היא פועלת? מהם 3 סוגי הבינה המלאכותית? שלושת סוגי הבינה המלאכותית הם: בינה מלאכותית כללית: בינה מלאכותית שיכולה לבצע את כל המשימות האינטלקטואליות שאדם יכול. נכון לעכשיו, שום צורה של AI לא יכולה לחשוב בצורה מופשטת או לפתח רעיונות יצירתיים באותן דרכים כמו בני אדם. בינה מלאכותית צרה: בינה מלאכותית צרה כוללת בדרך כלל טכנולוגיות זיהוי חזותי ועיבוד שפה טבעית (NLP). זהו כלי רב עוצמה להשלמת עבודות שגרתיות המבוססות על ידע נפוץ, כגון השמעת מוזיקה לפי דרישה באמצעות מכשיר התומך בקול. בינה מלאכותית רחבה: בינה מלאכותית רחבה מסתמכת בדרך כלל על מערכי נתונים בלעדיים הקשורים לעסק המדובר. זה נחשב בדרך כלל לקטגוריית הבינה המלאכותית השימושית ביותר עבור עסק. מנהיגים עסקיים ישלבו פתרון AI רחב עם תהליך עסקי ספציפי שבו נדרש ידע ספציפי לארגון. כיצד ניתן להשתמש בבינה מלאכותית בעסק? AI מספקת דרכים חדשות לבני אדם לעסוק במכונות, ומעבירה את הצוות מחוויות דיגיטליות טהורות לאינטראקציות טבעיות דמויות אדם. זה נקרא מעורבות קוגניטיבית. בינה מלאכותית מגדילה ומשפרת את האופן שבו בני אדם קולטים ומעבדים מידע, לעתים קרובות בזמן אמת. זה נקרא תובנות קוגניטיביות וניהול ידע. מעבר לאוטומציה של תהליכים, AI מאפשר החלטות עסקיות עתירות ידע, תוך חיקוי אינטליגנציה אנושית מורכבת. זה נקרא אוטומציה קוגניטיבית. מהן טכנולוגיות הבינה המלאכותית השונות בעסק? למידת מכונה, למידה עמוקה, רובוטיקה, ראייה ממוחשבת, מחשוב קוגניטיבי, בינה כללית מלאכותית, עיבוד שפה טבעית וחשיבת ידע הם חלק מהיישומים העסקיים הנפוצים ביותר של AI. מה ההבדל בין בינה מלאכותית ולמידת מכונה ולמידה עמוקה? בינה מלאכותית (AI) מיישמת ניתוח מתקדמות וטכניקות מבוססות לוגיקה, כולל למידת מכונה, כדי לפרש אירועים, לתמוך ולהפוך החלטות לאוטומטיות ולנקוט פעולות. למידת מכונה היא יישום של בינה מלאכותית (AI) המספק למערכות את היכולת ללמוד ולהשתפר מניסיון באופן אוטומטי מבלי להיות מתוכנתים במפורש. למידה עמוקה היא תת-קבוצה של למידת מכונה בבינה מלאכותית (AI) שיש לה רשתות המסוגלות ללמוד ללא פיקוח מנתונים שאינם מובנים או ללא תווית. מהן היכולות הנוכחיות והעתידיות של בינה מלאכותית? היכולות הנוכחיות של AI כוללות דוגמאות כמו עוזרים אישיים (Siri, Alexa, Google Home), מכוניות חכמות (Tesla), התאמה התנהגותית לשיפור האינטליגנציה הרגשית של נציגי תמיכת לקוחות, שימוש בלמידת מכונה ואלגוריתמים חזויים כדי לשפר את חווית הלקוח, עסקאות בינה מלאכותית כמו זו של אמזון, המלצות תוכן מותאמות אישית (Netflix), שליטה קולית ותרמוסטטים ללמידה. יכולות עתידיות של AI עשויות לכלול כנראה מכוניות אוטונומיות מלאות, חקלאות מדויקת, בקרי תעבורה אוויריים עתידיים, כיתות עתידיות עם אינפורמטיקה סביבתית, מערכות עירוניות, ערים חכמות וכן הלאה. כדי לדעת יותר על היקף הבינה המלאכותית בעסק שלך, אנא צור קשר עם המומחה שלנו.

Google Colab: Google Colab היא פלטפורמה מקוונת המאפשרת למשתמשים לשתף ולהריץ סקריפטים של Python בענן Graphics Processing Unit(GPU): GPU, או יחידת עיבוד גרפית, הוא סוג מיוחד של שבב מחשב שנועד להתמודד עם המורכבות חישובים הדרושים להצגת תמונות ווידאו במחשב או במכשיר אחר. זה כמו המוח של המערכת הגרפית של המחשב שלך, והוא ממש טוב לעשות הרבה מתמטיקה ממש מהר. GPUs משמשים סוגים רבים ושונים של מכשירים, כולל מחשבים, טלפונים וקונסולות משחקים. הם שימושיים במיוחד למשימות הדורשות כוח עיבוד רב, כמו משחקי וידאו, עיבוד גרפיקה תלת-ממדית או הפעלת אלגוריתמים של למידת מכונה. מודל שפה גדול (LLM): סוג של מודל למידת מכונה שאומן על כמות גדולה מאוד של נתוני טקסט ומסוגל ליצור טקסט בעל צליל טבעי. Machine Learning (ML): שיטה ללמד מחשבים ללמוד מנתונים, מבלי להיות מתוכנתים במפורש. עיבוד שפה טבעית (NLP): תת-תחום של AI המתמקד בהוראת מכונות להבין, לעבד וליצור שפה אנושית רשתות עצביות: סוג של אלגוריתם למידת מכונה המבוססת על המבנה והתפקוד של המוח. שדות קרינה עצביים (NeRF): שדות קרינה עצביים הם סוג של מודל למידה עמוקה שיכול לשמש למגוון משימות, כולל יצירת תמונה, זיהוי אובייקטים ופילוח. NeRFs שואבים השראה מהרעיון של שימוש ברשת עצבית למודל של זוהר תמונה, שהוא מדד לכמות האור שנפלט או מוחזר על ידי אובייקט. OpenAI: OpenAI הוא מכון מחקר המתמקד בפיתוח וקידום טכנולוגיות בינה מלאכותית שהן בטוחות, שקופות ומועילות לחברה. Overfitting: בעיה נפוצה בלמידת מכונה, שבה המודל מתפקד היטב בנתוני האימון אך גרועים בחדשים, בלתי נראים. נתונים. זה מתרחש כאשר המודל מורכב מדי ולמד יותר מדי פרטים מנתוני האימון, כך שהוא לא מכליל היטב. הנחיה: הנחיה היא פיסת טקסט המשמשת לתכנון מודל שפה גדול ולהנחות את הדור שלו Python: Python היא שפת תכנות פופולרית ברמה גבוהה הידועה בפשטות, בקריאות ובגמישות שלה (כלי AI רבים משתמשים בה) למידת חיזוק: סוג של למידת מכונה שבה המודל לומד על ידי ניסוי וטעייה, מקבל תגמולים או עונשים על מעשיו ומתאים את התנהגותו בהתאם. מחשוב מרחבי: מחשוב מרחבי הוא השימוש בטכנולוגיה כדי להוסיף מידע וחוויות דיגיטליות לעולם הפיזי. זה יכול לכלול דברים כמו מציאות רבודה, שבה מידע דיגיטלי מתווסף למה שאתה רואה בעולם האמיתי, או מציאות מדומה, שבה אתה יכול לשקוע במלואו בסביבה דיגיטלית. יש לו שימושים רבים ושונים, כמו בחינוך, בידור ועיצוב, והוא יכול לשנות את האופן שבו אנו מתקשרים עם העולם ואחד עם השני. דיפוזיה יציבה: דיפוזיה יציבה מייצרת תמונות אמנותיות מורכבות המבוססות על הנחיות טקסט. זהו מודל AI של סינתזת תמונות בקוד פתוח הזמין לכולם. ניתן להתקין את ה-Stable Diffusion באופן מקומי באמצעות קוד שנמצא ב-GitHub או שישנם מספר ממשקי משתמש מקוונים הממנפים גם מודלים של Stable Diffusion. למידה מפוקחת: סוג של למידת מכונה שבה נתוני האימון מסומנים והמודל מאומן לבצע תחזיות על סמך היחסים בין נתוני הקלט והתוויות המתאימות. למידה ללא פיקוח: סוג של למידת מכונה שבה נתוני האימון אינם מסומנים, והמודל מאומן למצוא דפוסים ויחסים בנתונים בעצמו. Webhook: Webhook הוא דרך של תוכנת מחשב אחת לשלוח הודעה או נתונים לתוכנית אחרת דרך האינטרנט בזמן אמת. זה עובד על ידי שליחת ההודעה או הנתונים לכתובת URL ספציפית, השייכת לתוכנית האחרת. Webhooks משמשים לעתים קרובות כדי להפוך תהליכים לאוטומטיים ולהקל על תוכניות שונות לתקשר ולעבוד יחד. הם כלי שימושי למפתחים שרוצים לבנות יישומים מותאמים אישית או ליצור אינטגרציות בין מערכות תוכנה שונות.