AI for Embedded Systems | Embedded systems podcast, in Pyjama

Summary

In this engaging discussion, a group of five individuals explore the practical applications of AI in embedded systems. They delve into the current capabilities of AI for tasks like reading and interacting with data sheets, with mixed results. The conversation covers the challenges of relying on AI for coding assistance, the limitations of AI in understanding specific documentation, and the potential for AI to generate code and unit tests. The group also touches on the broader implications of AI in the software development process, highlighting both its benefits and the need for cautious adoption.

Takeaways

• 😀 The group discusses the use of AI in embedded systems and its current applications, focusing on large language models.
• 🔍 Wasim shares his experience using AI to interpret data sheets, noting the model's mixed success in providing relevant information.
• 📚 The conversation highlights the limitations of AI when dealing with poorly documented or proprietary hardware data sheets.
• 🤖 A member of the group explores using local AI models like llama 3 for tasks to maintain data privacy, especially for company-specific hardware.
• 🛠️ The group acknowledges AI's utility in writing blockware code such as HTML and CSS, but its less effective performance with more complex or custom software engineering code.
• 🔧 Some participants find AI-generated code suggestions distracting and sometimes inaccurate, leading to a preference for disabling certain AI features.
• 🔄 The discussion points out AI's tendency to 'hallucinate' or generate incorrect information, necessitating verification of its outputs.
• 🔑 The importance of understanding AI's limitations is emphasized, such as its inability to understand the context as deeply as a human expert.
• 🔑️‍🔒 Privacy and security are considered when deciding to use local AI models to avoid uploading sensitive data to the cloud.
• 📈 The group sees potential in AI for reducing research time and providing meaningful responses for common queries found on the internet.
• 🛑 The script concludes with a cliffhanger about whether AI will replace embedded engineers, suggesting that it's currently far from happening.

Q & A

What is the main topic of discussion in the video?

The main topic of discussion is the use of AI in embedded systems, focusing on how AI, particularly large language models, can be applied in this field.

How is Wasim using AI in his current work?

Wasim is exploring the use of AI to read and chat with datasheets, although he has faced challenges with the accuracy of the responses.

What is the general consensus about the reliability of AI-generated responses for technical documentation?

The consensus is that AI-generated responses can be hit or miss, often providing incorrect or incomplete information, which can be unreliable for technical documentation.

What alternative method is being explored for using AI with local PDFs?

An alternative method involves using a local LLM model like LLaMA 3 and converting PDFs to text and embeddings for querying, thus avoiding cloud-based solutions for proprietary data.

What are some challenges mentioned regarding the use of AI for writing code?

Challenges include AI generating incorrect code, creating distractions with irrelevant autocomplete suggestions, and sometimes hallucinating incorrect solutions.

What are the advantages of using AI for scripting languages mentioned in the discussion?

AI is found to be helpful in writing scripts like Python and JavaScript, as well as generating boilerplate code for HTML and CSS.

How do the speakers use AI for generating unit tests?

They use AI to generate basic unit tests by inferring from the code, which can help in testing all combinations of input data types and expected outputs.

What is one of the significant limitations of AI in coding, according to the discussion?

A significant limitation is AI's inability to handle custom or complex codebases effectively, often generating more noise than useful logic.

What are the speakers' thoughts on the future improvement of AI in coding?

They believe that as AI gets used more often and receives more feedback, its accuracy and usefulness in coding will improve over time.

What is a common problem with AI-generated technical solutions as highlighted in the video?

A common problem is AI's tendency to hallucinate solutions that seem plausible but are actually incorrect, leading to confusion and mistrust in its responses.

Why do some speakers prefer to run AI models locally rather than using cloud-based solutions?

They prefer local models to ensure the privacy and security of proprietary data, which might be at risk if uploaded to cloud-based AI services.

What is the perceived gap between AI's current capabilities and the potential to replace embedded engineers?

The perceived gap is significant, as AI currently lacks the ability to fully understand and implement complex hardware and software integration, which is critical in embedded engineering.

Outlines

00:00

🤖 AI in Embedded Systems and Documentation Challenges

The group discusses the application of AI in embedded systems, starting with general AI usage and then focusing on its relevance in embedded systems. They share personal experiences with AI tools like chat PDF for reading datasheets, highlighting the varying degrees of success and the challenges of finding precise information. The conversation also touches on the reliability of documentation and the potential of AI to assist in programming tasks.

05:00

🔍 Exploring Local AI Solutions for Proprietary Hardware

Members of the group explore the idea of using AI locally to interact with PDF documents, particularly for proprietary hardware where cloud-based solutions might not be ideal. They discuss the use of models like llama 3 for local language model (LLM) implementations and the process of converting PDFs into text and embeddings for AI training. The summary also includes the challenges of getting AI to understand and interact with specific documents locally.

10:01

🛠 AI's Role in Code Writing and the Limitations Encountered

The discussion delves into the use of AI for writing blockware code, such as HTML and CSS, and the issues faced when attempting to write more complex, custom database scripts. It is noted that AI can be distracting and sometimes generates incorrect code, leading to a lack of trust in its output. The group also shares anecdotes about AI-generated code that required correction and the subsequent challenges in getting accurate responses.

15:04

🔄 AI's Predictive Text Capabilities and the Risk of Hallucination

The conversation examines the predictive text capabilities of AI, noting its tendency to 'hallucinate' or generate incorrect information. The group discusses the analogy of AI trying to 'wing it' without a clear understanding of the task, leading to inaccuracies. They also touch on the potential for AI to backtrack and correct its predictions based on probability and user feedback.

20:05

📈 AI's Evolution and Its Impact on Software Development

The group reflects on the rapid development of AI and its impact on software creation, comparing it to the speed at which the COVID-19 vaccines were developed. They discuss the importance of user feedback in improving AI and the potential for AI to become more accurate and creative over time. The conversation also includes the use of AI for summarizing documents and conversations, as well as its limitations in certain areas like code writing.

25:06

🍽️ Wrapping Up the Discussion and Looking Forward to Future AI Topics

The group concludes the discussion with a humorous note about AI replacing embedded engineers, acknowledging that while AI has come far, it is not yet capable of completely取代 human engineers. They plan to continue the conversation in a follow-up session, where they will explore the potential of AI to replace certain roles and the ethical considerations of AI development.

👋 Final Thoughts and Goodbyes

The final paragraph captures the group's light-hearted sign-off, with a member humorously referencing the 'monkeys typing' scenario to illustrate the potential for AI to eventually generate meaningful output. The group agrees to reconvene for further discussions on AI's role and capabilities.

Keywords

💡AI in embedded systems

AI in embedded systems refers to the integration of artificial intelligence capabilities within devices that perform specific functions, such as sensors or home automation systems. In the video, the theme revolves around the application of AI in these systems, discussing how it can be utilized and its relevance in the context of embedded systems. For example, the participants discuss using AI to interact with data sheets and the potential of AI in improving the functionality of embedded devices.

💡Large language models (LLMs)

Large language models are AI systems that have been trained on vast amounts of text data, enabling them to generate human-like text based on the input they receive. In the script, LLMs are mentioned in the context of their ability to process and interact with PDF documents, such as data sheets, which is a significant aspect of the discussion on how AI can be applied in various scenarios, including embedded systems.

💡Chat PDF

Chat PDF refers to an AI tool that can read and interact with PDF documents, allowing users to ask questions about the content. In the video, one participant mentions using a chat PDF tool to extract information from data sheets, highlighting both the usefulness and limitations of the AI's ability to understand and provide relevant information from these documents.

💡Data sheets

Data sheets are documents that provide detailed information about a product's specifications, typically used in engineering and technical fields. The script discusses the use of AI to read and interpret data sheets, which is crucial for understanding how AI can assist in tasks such as hardware design and embedded system development.

💡Proprietary hardware

Proprietary hardware refers to hardware that is unique to a company or individual and not available for public use. In the context of the video, the discussion about using AI locally for proprietary hardware data sheets emphasizes the importance of keeping sensitive information secure and within the company's control.

💡Local AI models

Local AI models are AI systems that run on a user's own device or server, as opposed to cloud-based AI services. The script mentions the desire to use local AI models for handling proprietary data, which is important for maintaining the confidentiality of sensitive information and avoiding reliance on external cloud services.

💡Code generation

Code generation is the process by which AI systems can create code snippets or entire programs based on user input or requirements. In the video, participants discuss the use of AI for generating code, particularly in the context of scripting and blockware code like HTML and CSS, but also highlight the challenges and inaccuracies that can arise in more complex or custom code scenarios.

💡Code bloat

Code bloat refers to the unnecessary increase in the size of a software program's code, often due to the inclusion of redundant or overly complex code. The script mentions AI's tendency to generate bloatware code, which may not be efficient or optimal, and the need for engineers to carefully review and refine the AI-generated code.

💡Unit tests

Unit tests are a type of software testing methodology where individual components or units of a software are tested to determine if they are fit for use. The script discusses how AI can be used to generate unit tests, which can be beneficial in ensuring the quality and correctness of the code, although it also notes the limitations and the need for human oversight.

💡Documentation gap

The documentation gap refers to the discrepancy between the information provided in design documents and what is actually needed by engineers to implement a system. In the video, the participants joke about the possibility of AI replacing embedded engineers, but acknowledge that the documentation gap is a significant reason why human engineers are still required in the development process.

💡AI-generated errors

AI-generated errors refer to mistakes made by AI systems when they attempt to perform tasks such as code generation or data interpretation. The script provides examples of AI providing incorrect information or suggestions, such as incorrect linker flags, which highlight the need for verification and correction by human engineers.

Highlights

The group discusses the use of AI in embedded systems and large language models.

Wasim explores using AI to read and chat with data sheets, with mixed results.

The group debates the reliability of AI-generated information compared to human-written documentation.

AI's limitations in understanding and providing accurate information from data sheets are discussed.

Exploration of using local AI models like llama 3 for proprietary hardware documentation.

The process of converting PDFs to text and training AI models with the text is outlined.

Google IO's announcement of AI models that understand and respond to PDF documents is mentioned.

AI's utility in writing blockware code like HTML and CSS is highlighted, with limitations in SE code.

The group shares experiences with AI-generated code inaccuracies and the need for manual correction.

AI's struggle with custom database constructs and the resulting 'noise' in code predictions.

The analogy of AI as a human trying to 'wing it' without a clear direction is used to describe its limitations.

The importance of feedback loops for AI improvement and the comparison to software development cycles.

AI's potential to replace embedded engineers is humorously dismissed by the group.

The group acknowledges the rapid pace of AI development, drawing parallels to the COVID-19 vaccine.

The practical applications of AI in summarizing documents and conversations are praised.

The group discusses AI's role in creating unit tests and its potential impact on software development.

Licensing and attribution issues related to AI-generated code are raised as concerns.

The group concludes with a cliffhanger on whether AI will replace embedded engineers, to be continued in the next session.

Transcripts

but all right hey everyone we back again

the five us uh you know

finally had like intersection of free

time and we have decided this time

around to discuss on the use of AI in

embedded systems yes well I think we'll

just focus well or at least start from

how we are using AI in general the large

language models and then maybe go from

there and talk about you know how

relevant it is you know can we use it in

embeded systems to what degree and so on

and so

forth so yeah whoever feels like jumping

in go ahead jump

in yeah I think Wasim had one

interesting yeah so currently I'm

[Music]

exploring was same into the under the

bus good yeah

nice so current I'm exploring like how I

can use AI to read or chat with the data

sheets so I tried but uh the AI model

which I was using it was chat PDF so

some of the answers was very relevant

and some of the answers it was not able

to find it out like exact register which

register I should update

to and so on and so also is this a tool

that's online where you have to kind of

push your PDF yeah it is a chat PDF I

see okay so what's this I don't think

that is ai's problem uh if I give that

data sheet to the person who wrote that

data sheet he won't be able to find

relevant information in

either you know my my okay this is

interesting because my nephew is trying

to program the stm32 and for some reason

he had to program the DC or something

yeah and there there was

like okay I'm paraphrasing him but the

he reported that there were few bits

that were undocumented and he found it

on Reddit or somewhere you know finally

cracked the problem that way uh but yeah

I think the documentation is also

reliable to like a good degree but you

know not 100% yeah yeah that said you

know along the same lines of what Wasim

was mentioning I was also on the site

trying to explore uh you know AIS that

can or rather the llm models that can be

used to chat with the PDFs but I wanted

to do so like locally so I was try

trying o lama lama 3 and any reason for

the local way well I suppose I was

trying it out for our audience majorly

uh in the sense that you know if you're

working on a proprietary piece of

hardware and the data sheet for that is

like local to your

company then it wouldn't be like a good

idea to push it onto yeah some cloud

based AI so the old Lama at least I have

had like good Su well good success I

I've had some success running it locally

at least the llm models run and it

responds back to General queries answers

like I don't rely on the answers usually

the PDF part I have not cracked it uh

there are a lot of uh tutorials that I'm

watching that go around kind of you know

they call it rag

some generating so what happens is they

there is like a bunch of uh python

libraries that you need to kind of you

know use to First convert the PDF

document into just texts and then

convert that text into like vectors or

embeddings and then feed it

as a training data to the llms and then

after all of that you know go ahead we

can go ahead and ask it questions so

I've just seen few T tutorials didn't

get to the point where you know I was

feeding it the local PDF but yeah that

seems something useful and relevant for

us

yeah I wonder how so most of these

models have now start are become model

right so we can attach the complete PDF

document can we can we upload the PDF

document into Bing or CH

gity I don't know but in in you know

Google IO they mention that you can uh

you know provide a PDF document and it

will pass the document and then it will

you know act as it knows about that PDF

and then you can ask any questions and

it will you know respond to your queries

I think there is might be because so the

most of these chat most of these models

are constrained by the context window

right so it's a 2 Milli supports 1.5

million context window so it can pass

big

documents right

nice but then it would be consuming this

PDF out of the drive right the Google

Drive locally maybe locally ultimately

that D uploaded to Cloud somehow yeah I

think

okay well I if you ask me I use AI

mostly to write my blockware

code

like what I have find is it is really

good when you're are writing uh scripts

like python or any kind of JavaScript

Etc it's really helpful to write

blockware codes like HTML and

CSS but what I have found is if you try

writing SE code it can predict to some

extent but if you are working on a

custom database like very which where

it's not generic construct it's very it

sometimes create more noise than actual

logic yeah

so yes so what I so most of these Co

code assists what they do is as you type

like Snippets like code Snippets or auto

complete will do the it will create some

uh code something in light gray giving

you a prediction you know maybe this is

what you are trying to type the

autocomplete feature in the IDS but the

thing is those are limited to some

extent like some words or some or one or

two lines but AI what I have seen is

mostly with co-pilot and others like it

tries to be smart and create the

complete logic out of it and it become

and to me sometimes that logic is not

right and I get distracted by that

because I want to read that what it

generated and that shifts my focus from

what I was thinking to what it then I

found then I find out it's useless I

type I get context switch back into what

I was typing I type that and it

generates new a new autocomplete

suggestion big which again distracts me

and and ultimately what I do is so I

have to disable the auto complete the

Cod and I use it in the console mode

like when I want it to work I go to

console and type it can you do this for

in the you know this this just reminded

me the whole point of it creating noise

so I also enabled this U again llama 3

based you know there's some extension in

vs code that can you know consume an LL

uh llm model and then try and predict

the code and my God it's just a pain

yeah whatever rajit mentioned is like

100% true you know I'm trying to type

something and it predicts like three

lines worth of garbage um and

you know pretty much the same situation

which is I think it's good for writing

scripts um to some degree it does like a

good job Python scripts or shell scripts

C code really

bad you know the worst scripts wherever

you want to write the functional logic

of the script right I have seen that I

I'm better off writing it myself what I

can do is this so for example even in my

C code what I use it for is let's say I

want a link implementation I just tell

it Implement a link list for me in which

the data structure should have XYZ

elements mhm and then it generates the

blockware code of what the node

structure should look like insert remove

insert delete except print and then I

write the function logic what I whatever

I want to do with that list so for these

purposes AI is really

good yeah I think that's fair by the way

one other thing I want to call out is um

there were cases in which I asked it to

generate code and it kind of generated

some incorrect code and then I call out

saying hey you know these XYZ lines are

wrong and then it says oh yeah yeah yeah

you know sorry I made a mistake and goes

ahead writes some new code which also

doesn't work like 60% of the time it

doesn't work and then I ask it to again

correct it and it uses the same library

or the same you know incorrect

statements from the first response like

for example if you some okay I'm

forgetting the respon um kind of example

that I asked

it uh I had yeah okay this is the one so

I was working on this course like videos

for the Linker script course and then at

some point I asked it hey you know how

can I ask the Linker to strictly follow

my Linker script because apparently

linkers what they do is um while

processing the Linker script they add

stuff that is convenient so for example

let

say um I wanted to demonstrate that I'm

only picking some section from a given

object file and what happens is that

object file has some content which is

kind of in available in another object

file like a global variable or something

what I wanted to demonstrate was that if

I specifically mention that only include

one object file it should only include

you know contents of that one object

file the Linker is very convenient if

you give it the inputs and the other

object file has the content it will just

add that in so I wanted to kind of you

know uh try and find out if there is a

flag compiler flag or not compiler but

Linker flag that specifically restricts

Linker to only process the Linker script

I provided so yeah the you know lri

model says oh there's like this minus s

option that you can use I try it doesn't

work then I tell it that hey you know

this option doesn't work is like oh

sorry my bad my bad try a RX then I try

ARX and you know R is like an unknown

option or something like that happens

and then I'm like man this is also wrong

sounds like tar Comm something like that

yes I think r r and a might be relevant

X also might be relevant but they don't

do what you know the model told me yeah

then I tell it that hey you know this is

also incorrect it's not working and then

it goes ahead and uses the minus s

option again and so this has happened

with me like this is one instance but

this has happened with me over and over

again enough times that I don't trust

any output it gives me I cross check it

on Google again or huc is big problem it

just confuses it just hallucinate like

hey it if you know let's say compiler

has an option then Linker must also have

a similar option just try out but you

know in that sense it sounds as though

it's becoming more and more human like

because it's like hey you know I'll just

wing it let's see how far this

goes oh this didn't work no problem try

RX

confidence

exactly it's good that we have a textual

proof that what it said otherwise

likey man when

did that's you know this reminds me

early days when this chat gbt Etc were

there and you know I was trying out so

at that time I was trying to work with

the trace 32 and something else uh so I

and very custom uh so that tool has a

very custom debug uh script uh syntax

and I'm like but it's quite common the

manuals are out there so I'm like you

know let's put AI to try it's uh rather

than me going and learning it I want

very small thing to do let me ask it and

so I asked uh I think I asked multiple

more model the same question and uh I

asked them can you generate a trace 32

script for me that can do XYZ

functionality and it generated the code

and I was like really amazed I'm like

okay it generated the code it is really

helpful I don't have to go through the

complete manual to understand it I run

that and the TR to complains that this

is not even the

syntax then I went back and read the

manual and that was not the

syntax

ah I think one thing that's starting to

surf is now is that the okay large

language models are more or less just

trying to predict the next word right

they're just trying to predict more and

more of the

statement and

uh yeah I mean if if you think of

someone who is trying to just predict

statements and you know that those

statements happen to be something that

look like

code uh doesn't mean that you know that

machine might be actually understanding

what it has written like so at least to

me that is super clear

now so one analogy comes to mind like

like the office scene where Michael spot

says right I just start a statement

knowing where without knowing where am I

going and just you know wing it on at

run time maybe yeah is also kind of

doing this who knows who knows or at

least well

fundamentally it just trying to predict

the next word the next word so not

everything it says is accurate

at least that's my learning and again

I'm only exp I'm not really expert on

this but is there any way it can

backtrack saying hey you know this

series of sentence doesn't lead me

anywhere let me backtrack to the like

start word and start again the do you

know if that is something how would you

decide when to termine it and question

it like I would imagine like the next

words probability are so low like you

know the best choic is even the

worst compared to the previous choices

like had you gone past you know some

words the probability list becomes so

narrow that you will you feel like hey

let me go back and try do you know if

that oh I I suppose that is where those

settings come in be more creative be

more

accurate uh I think tempure temperature

yeah temperature settings and I suppose

there is also one to help the AI kind of

backtrack a little bit which is like

regenerate the response it is in ch Chad

i' seen seen those options I think it's

it's there in every

mod Gemini also like they produce three

responses like I see okay okay

interesting okay but uh I I do see a

thing that things can improve and I

think as as it gets used more and more

often it will improve

again what I think it's one and a half

year and that's too small for a software

life

cycle

yeah I I I don't think any of the big

softwares were amazing within one and a

half year I suppose I suppose the last

time something got built with this rapid

Pace was the covid

vaccines and we don't know where they

are like going or what what kind of side

effects they will have later in life or

even

now but okay not a

medic he's saying he doesn't what did

you say man you have never seen a

software which is quite perfect in one

year of its initiation one and a half

year yes one and a half year I will show

you my calculator which I built in

college perfect got it right first right

my hello I'm sure if I give it to some

product manager he will he'll find 500s

in it

and then we'll ask you to implement

standups and multiple cycles and that

will take another one track progress

Sprints all of this provide a heat map

of how the progress is going Sprint

planning all of that so the thing is um

when you use it and nothing nothing

against it it's in general the more and

more users use it the more and more

feedback comes in because everyone

because like like AI we also hallucinate

you know if YZ feature was there in this

it would be really amazing and that's

what the creative part of humans is

right yeah this calculator should Al

also show temperature you know

exactly all of that good stuff

yes yeah I I think like uh like right

now if you use a for generic you know

questions like which are very popular in

the content is present on the net then

it's like it will give you know precise

answer that okay this is how uh you

would want to go ahead but then like if

you are asking for some specific area

then it might give you some results

which are which you don't know okay what

is correct and then also it hallucinates

so you cannot really see okay you know

whatever you have received may not be

correct right so you'll have to verify

that but I think personally for me uh I

don't use that much but wherever I use

like on the Google it generate that you

know summary s g right M and for generic

queries like you know if you are asking

okay how do I develop a you know

particular driver or a firmware or a

particular you know maybe algorithm or

software it will give you uh you know

some some SG uh content and that is you

know well formatt that you you know line

by line and then okay figure out okay

this is the content and then it will

give you some signat of the code right

so that you can figure out okay what is

this code and then again some content

and then some lines of code right so

that gives you you know how you can use

it how you can learn it to uh you know

for for your benefits right yeah I think

I would agree with that you know it does

narrow the search space a lot in terms

of like doing the research it can give a

lot of meaningful respons it feeds you

the right keywords where you can go on

about looking yeah that that I think is

a major upshot of you you know the

llms uh wherein they kind of reduce the

research time yeah that's I I think um

so some of the use cases like people

have been highlighting are really

amazing except from the Cod space if we

take it out from the Cod space like

summarizing documents summarizing

conversations those are really amazing

like so there were demos at the io

Google IO where they showcased its

integration with Gmail and workspace

where you can just rather than going to

each group and reading all the messages

you can add a chat bot into it and just

ask did anybody come to a conclusion on

whether the event should happen or not

it can and the answer will always be no

no one to

con I think as a productivity it might

and as I said right even in code assist

what I have found is it's really good at

creating bloatware code another use case

is creating unit tests for your code h

it can do really great job at that yeah

yeah I think that functional end like if

uh like the one where the where they say

swe right where the product manager put

a requirement and generates the code for

you and everybody's like you don't need

Engineers product managers can generate

yeah I think we are way far off from

that far away from

that man that's what Engineers would say

even if we are not like I mean just

yeah don't worry about it it's like far

away yes so so for unit test right you

cannot provide your actual code but you

have to specify okay what you are

looking for

right I I think you can you can actually

you can so what happens is how most of

these companies are offering like if you

look at co-pilot Etc your code base it

does not train data on your Cod or train

on your code base it infers on your code

base and even at they are not using it

for the generic co-pilot

training so never leaves your leaves the

server correct so uh you know the what

rajit is pointing to is given a code

that you have written in your ID the AI

will try and predict what kind of test

to wrap yeah it will do inferencing not

the training part okay and the kind of

inference it does then like the basic

ones is okay it figures out what are

like the data types of the inputs then

you know okay feed all combination and

check for all the written values and

done okay and I think the other problem

with the code assist kind of code assist

and I think that may resolve over time

is

licensing oh nice yeah so it's trained

on the data of the web and it can

generate code bases Etc and tell you and

you are not even aware where it actually

picked that came from

yeah attribution would become a problem

yeah definitely

Perfect by the way I have to call it out

right now and I think the internet will

also get to know but it's dinner time on

my end and my mom is calling me so can I

can I drop off I think we can uh you

know record Another followup where we

will have a question where okay will AI

replace the eded of I yeah we have few

questions which we would like to answer

these are kind of questions that comes

regularly on the LinkedIn and this is

like a cliff Cliffhanger but in

podcast okay so next time around

we so next time around we chat on okay

the probability of like AI replacing

embed engineer and and it is Zer Z

zero just kidding just no one know I

don't know no comments cool okay so

let's let comments next time maybe not

today not not now yes I I only believe

that they can just just I would say the

only way for AI to replace embed

Engineers if it starts building its own

Hardware oh because today the

documentation Gap that comes from

designer to embed engineer is the reason

why embed Engineers are needed in a lot

of don't give AI ideas

man I I I don't think I am giving ideas

people are already starting to think in

that direction yeah okay maybe Sam Al

that's why Sam Alon needs 7 trillion to

design AI chip maybe it's asking Chad j

to design that

chip could be could okay cool we discuss

next you know sorry I'll just make final

comment it looks to me like that you

know like that situation of if you let

the monkey put enough you know alphabets

in a

row

will long time yeah it will generate

that cool anyways maybe you know I'll go

have dinner and let's meet up next I see

a flying chle on your

way you should

now

not sorry well okay

I didn't get one mommy one

mom cut cut

cut both of them are super amazing human

beings I love both of them yes why did

you why did you win at the I didn't I

did I any anyhoo let's let's drop off

for now and let's catch up next time

take carebye bye