Andrew Ng - Why Data Engineering is Critical to Data-Centric AI

Summary

In this podcast, Andrew and Joe discuss the paradigm shift from model-centric to data-centric AI. They explore how data engineering is pivotal to successful AI implementation, especially with the rise of generative AI and large language models. Andrew highlights the importance of curating high-quality data for training and fine-tuning models. The conversation also touches on emerging AI applications in education and the future of work, emphasizing the potential for generative AI to transform these sectors.

Takeaways

• 📈 Data Centric AI is gaining momentum as a shift from the traditional model-centric approach, emphasizing the importance of data quality and engineering.
• 🔄 The process of training AI models involves a significant amount of data handling, from selection to fine-tuning, which is often more labor-intensive than model engineering itself.
• 🌐 Data issues are prevalent in generative AI, affecting training, fine-tuning, and practical deployment, highlighting the need for robust data strategies.
• 🔄 An example of data curation in AI is Meta's use of an earlier model, llama 3, to generate coding puzzles that trained the subsequent model, llama 3.1.
• 📚 Licensing and procurement of data are significant challenges, especially when aligning data with human values and labeling schemas.
• 🔧 Data engineering is mission-critical for AI, with decisions around data storage, cloud services, and database schemes being vital for performance.
• 🏢 Companies are urged to invest in data infrastructure with specific use cases in mind to avoid over-optimization and to build a strong foundation for AI applications.
• 🤖 Agentic workflows, where AI models engage in a process of thinking and refining their outputs, are emerging as a key strategy to enhance AI output quality.
• 📊 The intelligence of AI models is largely derived from the data they've been trained on, suggesting that data curation is paramount for achieving intelligent behavior.
• 🎓 Generative AI has the potential to transform education, making coding companions and customized courses more accessible and effective.

Q & A

What is the main topic of discussion in the transcript?

The main topic of discussion is Data Centric AI and its significance in contrast to the traditional model-centric approach in AI development.

What is the significance of data engineering in Data Centric AI?

Data engineering is critical to Data Centric AI as it involves structuring the data effectively to build successful AI systems, including error analysis, data curation, and making trade-offs between data quality and quantity.

How does Andrew view the shift from model-centric to data-centric AI?

Andrew sees the shift as a practical evolution where focusing on data engineering and curation is more fruitful than solely engineering the mathematical models or AI algorithms.

What is the role of data in training large language models?

Data plays a significant role in training large language models, with a large fraction of effort dedicated to acquiring the right data to feed into these models for effective training and fine-tuning.

What is an agentic workflow in the context of AI?

An agentic workflow in AI involves a series of steps where an AI model generates content, reflects on it, possibly does additional research, and iteratively improves the output before finalizing it.

How does Andrew perceive the current momentum in the Data Centric AI community?

Andrew finds the momentum in the Data Centric AI community quite exciting, noting the increased focus on data issues across various stages of AI model development.

What is the importance of data infrastructure in building AI applications?

Data infrastructure is mission critical in building AI applications as it provides the foundation for data storage, management, and accessibility, which are essential for developing and deploying AI models effectively.

What challenges do companies face regarding data in AI, according to Andrew?

Companies face challenges such as deciding how to store data, choosing cloud services, database schemes, and making trade-offs between cost and performance. Additionally, they struggle with architecting data for specific AI purposes.

How does Andrew suggest companies should approach improving their data infrastructure?

Andrew advises companies to invest in data for specific purposes, using AI wins as a way to drive data architecture improvements, rather than attempting to fix all data issues at once.

What is Andrew's perspective on the future of AI and its impact on education?

Andrew anticipates a transformation in education with generative AI playing a significant role, potentially acting as a coding companion or customizing courses, but he emphasizes the need for people to understand coding concepts to effectively use these tools.

What are Andrew's thoughts on the current state and future of Transformer architectures in AI?

While Andrew acknowledges the dominance of Transformer architectures, he also expresses interest in alternative models like SSRM and diffusion models, suggesting that the field should continue exploring new architectures.

Outlines

00:00

🤖 Introduction to Data Centric AI

The paragraph introduces a conversation between Andrew and Joe about Data Centric AI. Andrew explains the shift from model-centric AI, where progress was made by inventing new models and algorithms, to data-centric AI, where the focus is on curating and engineering data to improve AI performance. He discusses the importance of data in training large language models and how data issues persist through various stages of AI development, including fine-tuning and deployment. Andrew also highlights the Meta Llama 3.1 paper, where an earlier model was used to generate coding puzzles that trained the next version, showcasing the potential of using AI to create training data.

05:00

🔍 The Role of Data Engineering in AI

In this section, Andrew and Joe discuss the critical role of data engineering in AI. Andrew emphasizes the importance of data infrastructure and how it affects the success of AI systems. He mentions the challenges companies face in architecting data and the need for a purpose-driven approach to data improvement. The conversation also touches on the talent shortage in data engineering and how companies should approach investing in data for AI. Andrew suggests a cyclical approach where initial AI successes fund further data infrastructure improvements.

10:02

🧠 Exploring Agentic Workflows and Data Centric AI

The discussion continues with agentic workflows and their impact on AI. Andrew explains how agentic workflows, where AI models are prompted to think and refine their outputs iteratively, can lead to higher quality outputs compared to direct generation. He also talks about the use of these workflows in various applications and the importance of data curation to avoid model collapse when training AI on machine-generated data. The conversation highlights the need for a thoughtful approach to data generation and usage in AI training.

15:02

📚 Generative AI in Education

Andrew shares his thoughts on the potential of generative AI in education. He envisions a transformation in education where generative AI could serve as a coding companion, making coding easier and more accessible. He suggests that learning to code with the assistance of AI could align educational practices with the future of work, where coding companions are commonplace. Andrew also addresses the challenge of teaching students to use AI tools effectively while understanding the fundamentals of coding to avoid common errors.

20:04

🚀 The Future of AI Applications

In the final paragraph, Andrew expresses his excitement about the future of AI applications. He believes that while foundational work in AI, such as training models and improving data engineering, is essential, the real value will come from practical applications of AI. Andrew notes that despite the high costs associated with training foundation models, the economics at the application layer are favorable for innovation. He anticipates a surge in AI applications that will drive the field forward.

Keywords

💡Data Centric AI

Data Centric AI is an approach that emphasizes the importance of high-quality data over the model itself in building AI systems. It suggests that refining the data can often lead to better AI outcomes than solely focusing on the mathematical models or algorithms. In the transcript, Andrew discusses how the AI community is shifting from a model-centric world to one that values data curation and engineering, highlighting the momentum of the Data Centric AI community.

💡Model Centric

Model Centric refers to the traditional approach in AI where the focus was on developing and improving the algorithms or models to process data. The transcript mentions a historical perspective where AI progress was driven by people downloading datasets and then creating new models to perform better on those datasets.

💡Data Engineering

Data Engineering involves the practices and processes of managing and preparing data for use in AI systems. It is highlighted in the transcript as critical to Data Centric AI because it deals with the infrastructure, storage, and management of data, which directly impacts the effectiveness of AI applications.

💡Foundation Models

Foundation Models, also known as large pre-trained models, are AI models like large language models that have been trained on a vast amount of data. The transcript discusses the effort that goes into training these models and how data issues are a significant part of the process, from training to fine-tuning.

💡Generative AI

Generative AI refers to AI systems that can create new content, such as text, images, or music, that resembles the training data they were fed. The script mentions how Data Centric AI principles relate to the training of large language models, which are a key component of Generative AI.

💡Agentic Workflow

Agentic Workflow is a process where AI models are used iteratively to refine and improve outputs, similar to how a human might revise their work. The transcript gives an example where an earlier version of a model is used to generate coding puzzles, which are then used to train a newer version of the model.

💡Data Infrastructure

Data Infrastructure encompasses the systems, tools, and processes used to manage data within an organization. The transcript emphasizes the importance of getting data infrastructure right to support AI applications and how it can drive improvements in data architecture.

💡Model Fine-tuning

Model Fine-tuning is a process where a general AI model is adjusted or customized for a specific task or dataset. The transcript discusses how fine-tuning large language models involves a lot of data-oriented thinking and decision-making.

💡Data Curation

Data Curation is the selection and preparation of data for use in AI systems. The transcript mentions the ongoing challenge of curating high-quality data from the internet and how it's an important aspect of training better AI models.

💡Transformer Networks

Transformer Networks are a type of deep learning architecture that has become the standard for many AI models, especially in natural language processing. The transcript discusses how the intelligence of AI models comes not just from the architecture of Transformer Networks, but also from the data they've been trained on.

💡Educational Transformation

Educational Transformation refers to the changes and innovations happening in the field of education, often driven by technology. The transcript mentions the potential for Generative AI to transform education, suggesting that learning to code with AI companions could become the norm.

Highlights

Discussion on the shift from model-centric to data-centric AI.

Importance of data engineering in the practical application of AI.

The role of data in training large language models.

Data curation challenges in training AI models with machine-generated content.

Innovative use of an earlier version of a model to generate training data for its successor.

The significance of data in driving the intelligence of AI models.

The necessity of error analysis in data-centric AI systems.

The trade-offs between quantity and quality of data in AI applications.

The impact of generative AI on the future of work.

The role of data infrastructure in supporting AI applications.

The potential of agentic workflows in improving AI output quality.

The use of AI in education and the transformation it might bring.

The importance of understanding coding concepts when using AI coding companions.

The economic implications of AI applications and the potential for capital efficiency.

The future of AI and the focus on developing useful applications.

The potential of diffusion models for text generation as an alternative to Transformers.

Transcripts

hi

Andrew hey good to see you Joe good to

see you too how's

things uh the uh exciting times

say awesome well yeah uh thanks for uh

joining the show today it's good to see

you um yeah so we're here to talk about

a topic I think near and dear to you uh

which is data Centric AI uh but also how

data engineering um is critical to data

Centric AI so um I guess to back up uh

walk us through the beginnings of your

of your um thoughts around data Centric

AI because I I think when before that

there was a more of a model Centric

world we were in yeah so I think for

many decades um AI progress was driven

it feels like primarily by people say

downloading data sets off the internet

and then spending a lot of time trying

to invent new map or invent new models

to make it do better on that data and

that's fine nothing's wrong with that

because of that recipe AI made a lot of

progress but I think many practitioners

of AI including you and me and many

others have known that if we're trying

to build something you know practical

ship it um sometimes uring the data is

much more fruitful than trying to engine

in the math or the model and so I wanted

to uh creating and trying to popularize

this term data Centric AI to coales a

lot of the already ongoing work um on

entering the data rather than the model

and it's been quite uh exciting actually

to see how much momentum to dat the

centri AI Community

has I think we we have to ask because I

mean this is the main thing in the air

right now generative AI does data

Centric AI have something to do say

about how we train large language models

for example what are your thoughts on

that I think very much so in fact um uh

everything from training the foundation

model to you know post training maybe

fine-tuning to even some of the uh

deployment usage seems to keep on

running into data issues I know that in

the popular press people talk a lot

about scaling laws and building you know

bigger Transformer networks or whatever

to train even more data on and that is a

key part of it but when I talk to my

friends they're involved in you know the

actual dayto day of how do you get these

models to work um a large fraction of

the efforts I'm tempted to say more than

50% but L very large fraction of their

fers is actually thinking through how to

get the right data to feed into these

Foundation models um and then of course

you know even after someone else has

trained a large language model large

Foundation Model A lot of data workers

go into fine-tuning it and then also in

Practical deployment and usage um you

know if you're doing a few shot learning

lot lot of data oriented thinking there

as well so not everything is data

Centric AI but even in gen and

Foundation models a much larger fraction

of it is then I think people wiely

appreciate I actually talk about this

for a long time but you guys are oh no

please do keep going yeah we're here to

listen oh so maybe one one fun thing

when uh in the Llama 3.1 paper uh The

Meta release I think one of the coolest

things you a lot of cool things in that

paper but one of the coolest things was

meta used an earlier version of the

model used llama 3 um and then an

agentic workflow to basically use llama

3 to generate coding puzzles that were

then used as training data to train

llama 3.1 and I think this has always

been one of the puzzles of how you get

synthetic data work to train Foundation

models and um using an agentic workflow

we use the early model but let it think

for a long time iterate over something

over and over to come over good result

and you train the Next Generation model

to come over the equally good answer

very quickly rather than need to think

of over and over I thought that was um

one really nice recipe um uh you know

for for creating data to train

Foundation models and really when I

think when I talk my friends training

you know some of the very large

Foundation models um a lot of the head

space is boy can I sign the right

licensing deals with the Publishers to

get the data and of all these datas

which one do I invest dollars in to buy

um or for the preference tuning be rhf

or you know DPO to to line it with human

values what's the labeling schema how do

I get that data and then it turns out

that while there's certain Innovations

on trading Transformer networks and all

that it feels like uh there's there's at

least as much maybe even more hot to

compare that there Sly a lot of thinking

dayto day on um how to get the data to

chain these

models when you came up with the um I

think the original uh article on um data

Centric AI um and you're thinking around

that that was in the time before chat

GPT and I and I think that um how how do

you differentiate between uh or is there

a differentiation between data Centric

AI um and maybe more classical um I I

kind of let deep Lear into that to like

classical like three gen Ai and then um

gen AI is there is there a difference in

how you approach data Centric um AI or

is it all sort of the um same thing at

the end of the day I don't know maybe

are related to set of techniques on the

Continuum um feel like you know data CI

for vision is different than for text

it's different for audio it's different

when is a input modality like structure

data that humans kind process that well

ourselves so I think it is very I think

of is different for um different types

of data and types of modality but then

there are underlying principles um

really how do you systat engineer the

data to build a successful AI system so

things like error analysis to figure out

where are the gaps um in order to try to

get more data as well as what the

techniques for datation to get high

quality data and then the trade-offs

between you know small amounts of high

quality data versus large amounts of

slightly lower quality data I see I see

these themes that seem to be pervasive

um Al the way from you know training

computer vision models to the way we are

trying to figure out which small handful

of examples to put into a large language

model prompt because we're doing F short

learning um yeah I I think and then of

course um underlying that I see with you

know supervised learning and generative

AI I see a lot of businesses um actively

thinking about how to get the data

infrastructure sorted out and gen

certainly given Tailwinds to a lot of

companies and boards and cosos wanting

to get that data infrastructure right

and I think that's been a that's been a

positive change actually just increased

urgency when people say we got got to

sort out our data so that certainly has

uh create a lot more urgency you know to

to to the kind of stuff that you guys do

I I I imagine it would been good for

sales of your book as

well very much so yeah we we came off of

one hype cycle and now we we're kind of

riding another

yeah but it's it's it's interesting

though because I feel like the

um yeah I feel like there's a focus on

the fundamentals and making sure you can

invest in the the foundation to enable

um you know analytics and and certainly

machine learning and AI uh that was um

but I feel like it's a hard realization

for companies to get to because it's

like I think initially it's a uh

especially in the 2010s it's like well

let's just jump straight into doing

machine Learning Without uh much data at

all um and I think all of us kind of saw

how that worked out and so um yeah so

with data engineering I guess is ends up

being I guess fairly important to to all

this so understanding it a bit

um I guess what are your thoughts on uh

the the role of data engineering um with

uh with AI these days Andrew I think as

Mission critical I am seeing uh

significant you know uh Talent shortages

uh as in I I I you know travel around

the US travel around to different

countries is now than usual to visit

sometimes very large very profitable

companies um not not tech companies

sometimes sometimes tech companies

sometimes not tech companies and um

there are really smart people doing some

business but kind of struggling to Think

Through how the architect the data and I

think it is difficult you know I think

the number of decisions we have to make

um in terms of how to store the data

which cloud services to use you know

what what's the right database scheme

what are the trade-offs between cost and

performance um and then also one of the

other challenges is most companies it

turns out don't want to spend you know

whatever a million dollars to just to

improve the data because you you want to

improve the data for a purpose and so

hoping businesses get in that cycle of

when you can hopefully start to deliver

wins on the AI machine learning

generative AI site but concurrently even

as you're building valuable applications

um use that as a way to Drive how you

improve your data architecture so that

your foundations to keep building on top

of become even better so so I have say

there are there are some cosos that have

said great let me invest a lot of money

to fix all my data and then it'll be

beautiful and then I'll have wonderful

Ai and I think you would both advice

people like please don't do that right

you do need to invest in the data but

the data is usually built out for a

specific purpose because otherwise there

too many things we can optimize you make

it faster speedier more distributive

more robust what whatever it's two IND

decisions but if you have one or a

handful of use cases that you're driving

toward then that helps the data team um

make the right priorization decisions uh

in order to improve the data

infrastructure and then that creates an

amazing foundation for lots of people to

build lots of exciting applications on

top of but I know I I feel like this is

advice I give to kind of companies quite

often I I I'm I'm guessing you give very

similar advice to many businesses all

the time we do yeah and it's yeah it is

interesting I suppose but you know I

think every company we talk to has to

have an AI story right now so um whether

you like it or not you're going to have

to figure out a way as a company to to

um you know start working with AI on

that note too I would like your your um

thoughts on what you're seeing or

hearing when you talk to your friends

about uh small language models and also

AI agents um what are you seeing out

there so I think uh agentic workflows is

is one of the most exciting um

directions for AI uh so you know I think

the way a lot of people use the large

language model is we prompt it and then

we expect it to write an essay for us on

whatever we ask and that's a bit like

going to a person and saying hey buddy

please write an essay for me by typing

the essay from the first word to the

last word all in one go without ever

using back space and you know maybe you

can write like that but most of us don't

do our best writing that way in contrast

to agentic workflow might ask the large

language model to um brainstorm and

outline and then ask if uh you need to

do any online research if so go download

a few web pages and put them into the

large language model context then write

the first draft they read the first

draft in critique it and and and so on

um and so uh we're seeing with many

agentic workflows the quality of output

is much higher than you know anyone than

than possible with just having with

direct Generation Um and I'm seeing this

useful for for many applications so

actually you know de AI interally as

some agentic workflows AI fund which

also lead um has many projects in kind

of healthcare you know legal compliance

processing various types of complex

documents where we couldn't do the job

without an agentic workflow and then the

most interesting thing I think um uh was

it uh GP 401 uh uh preview uh just

released recently and then also

anthropic also been doing something

related for months uh which is a

fine-tune the large language model um to

generate you know thinking tokens um

along the way and I think this

incorporating an agentic workflow where

the Lun langage model is um pre-trained

or sorry fine tuned to do sort of Chain

of Thought reasoning so it spends more

time thinking uh before it outputs you

know the final answer I think

incorporating this directly into the LA

langage model is exciting Direction

actually I think for a few months uh uh

I I I I I think it's been you know

jailbroken for months right so that's

what we know anthropic uses a tag um XML

tag and thinking I think but it's you

know various people have jailbroken it

to make it basically review the tag and

show his internal thinking dialogue so

that this so so I I think this is public

I don't think I'm reviewing new already

the internet so I thought I thought

actually really clever but I think the

open eyes taken it to a whole other

level with this new release model and

then I think it was just one or two

weeks ago there was a um overhyped

reflection 70b model with the initial

claims turned out not to be quite

accurate but that was also exploring a

similar technique and even though the

hype and the inaccuracy of the initial

results was unfortunate you know the the

underly technique it seemed like an

interesting very interesting Direction

takes well so I think this is in the air

um agentic work those that people can

implement but in also Al uh that is

possible to fine-tune a large language

model to basically do Chain of Thought

reasoning internally and maybe use some

tags to have some thinking you know that

may or may not need to be reviewed to

the end user so I I think there's a

fascinating direction oh and then of

course uh a lot of the work to do this

is you know come of the data set right

to show it how to Think Through what's

the right thinking process different

types of tasks so but so so exciting

times that's interesting I guess what

are you gonna ask Matt oh I I was going

to ask and this is kind of related to

what you're talking about I think one of

the problems that's in the air right now

is so theoretically if you take the

output of a model and then retrain on

that output you degrade the weights

basically right it's like recording a

signal that has extra noise in it and

then there also been research papers

that have shown this experimentally um

what what are your thoughts on how we

solve this problem especially as more

and more of the content on the web that

we're using to train is machine

generated and deep learning

generated yeah good question so I think

I think the the data curation of uh

selecting high quality data off the

internet I think that's an ongoing thing

um and I think so it turns out if you

use a large language model to generate

text and you train a different model on

that um that is a good idea if you're

applying model distillation if you have

a large model generating very thoughtful

text you want to train a small model to

mimic the thoughtfulness of a big model

you know that's that works that's model

distillation but I think as you're

saying Matt what doesn't work is if you

train a model use that to generate data

and then use that data to try to train

an even better model in fact a few

researchers have shown that if you do

this process enough times generate data

train a new model have the new model

generate data use that to train the next

model then you actually end up with

model collapse right where where the

model starts generating very

uninteresting things but where this

technique does Really Work Well is If

instead of using direct generation

instead of copy pasting one model's

output into the training data of the

next model if you instead use this type

of agentic workflow where the first

model you might have it write an essay

reflect on it do some web search you

know then critique it and improve it so

it does a lot of work to come up with a

pretty good essay and then you try to

get the next generation of model to

generate that essay directly with much

less work than the first model had then

that does seem a whole maybe analogies

of human thinking are always dangerous

so I was nervous about making that but

you know I remember when um when I was a

kid you know practicing for math

competitions or whatever right I would

spend a long time trying to solve some

math problem but having solved that

myself I go oh next time maybe that's a

shortcut I could use to solve the next

problem so training on your own thinking

is okay if you learn to do quickly what

took you yourself a long time to do and

and so it's been so this does seem to

work for L languish models as well

inter so it's almost like a

self-governing approach that you're

looking for not just training on data

but actually critiquing and then like

you're saying an agentic almost

thoughtful process you go through yeah

yeah yeah so and I think um yeah and and

and maybe just I I I think uh lot of the

large companies training you large

Foundation models um kept lot the

details of what they do right some of

proprietary but I definitely you know uh

get a strong sense talking to multiple

people from multiple companies a lot of

the head space is not just in tuning the

foundation model and making sure the

gpus are reliable whatever there is a

lot of that too but a lot of head space

isn't freeing out the

data I think I think you know a AI right

AI models is uh the model Plus data and

in fact if you know if um actually

common experience for a lot of people if

you go through the math of what a

Transformer neuron network does you know

go through the attention mechanism right

blah blah blah um I've had a really

interesting experience a lot of people

learning that from the first time they

go what I don't get it how could you

know like a few lines of math

demonstrate this weird intelligent

behavior of a large language model and I

think the answer is um the magic is is

not only from the Transformer neuron

Network which is very clever you know a

lot of the intelligence of the L Anish

model comes not from the neuron Network

architecture but it comes from the data

and this is why when people wonder I

want understand lar langage models let

me study the Transformer neuron Network

very common reaction is okay I finally

worked through all this math but I don't

get it it doesn't make sense why this

MTH would be so intelligent and I think

the the Gap is lot that code

intelligence comes from these models

having sucked in massive Text data sets

you know generated by mostly humans we

hope uh and that's what's that data is

what's creating a lot of the

intelligence or appearance of

intelligence in these

models I guess you

um uh you have a lot of friends in the

space are you seeing um maybe an

evolution of something outside of the

Transformer architecture right now or

are we uh or are we stuck with this for

a while you know it's a great question

um I I I I think uh transform

architecture seems to have strong

Tailwinds uh the what ssrm stasis models

have been kind of around for a little

bit they have not really taken off yet

uh uh but you know still enough

researchers working on it I think it's

fascinating we keep an eye on scaled to

very long input context very interesting

ways uh and then uh the other one that I

don't never will take off is the

diffusion models for text generation uh

really

just very recently I see my paper on

this fascinating uh today I think the

dominant model for image generation or

diffusion models where you generate a

blurry image and then repeatedly kind of

quote remove noise to sharpen up the

image and so um uh I think Stephano man

and some folks came a way to generate a

quote blurry piece of text and then

slowly sharpen it up and and when

trained to be gpt2 size it seems to

outperform gpt2 but you know JW's out on

on what will happen as as this scills

out so I I don't know if it'll work but

this I think if we're stuck with

Transformers and nothing else for a long

time I think we'll be fine but you know

with ssrm and diffusion models and

others trying other things I think you

know collectively our community has a

few shots at coming up with something

even better no pun

intended um kind of Switching gears a

bit so education's also a big passion of

yours um you started companies around it

before

um and where do you think uh generative

AI fits into uh education these days say

I want to learn a topic um you know I'm

teaching a topic where do where does it

fit in yeah so I think I feel like there

is a coming transformation of Education

maybe that I don't feel like I know

exactly what it will be yet um you know

there stuff that um a few companies have

done you know I think uh K Academy built

kigo which seems to work well Cera has

CER coach which actually works really

well uh people haven't tried it I've

used a bunch it's actually you know

surprisingly good um but I think that uh

that's just one product idea um uh corer

uses gen for course Builders so quite a

lot of companies are using cuse Builder

to customize causes for specific

Enterprises needs so you know there's a

bunch of ideas like this um like a aita

and so on I think that they could be

bigger transformation in education

coming and I don't feel like I know

exactly what it is um but maybe uh when

I when I chat of uh um you know

University leaders one other thing I

often talk end up talking about is the

future of work uh because one challenges

maybe take coding as as an

example I think we should just all learn

to code with generative AI as a coding

companion um I know that some schools

are still debating what are not to ban

you know chat gbt for the programing

causes uh but honestly I think it's

clear the future of software engineering

will be coding alongside the coding

companion and we can decide you know is

it GitHub co-pilot or cursor or copy

pasting directly from gb4 or CLA or

Gemini which I still do a lot of

actually um but you never have to learn

to code alone and so I think aligning

the way we teach programmers um with the

future with not where the field has been

but where the feud is going I think we

need to do that and that's just computer

science I think in education you know

thinking through what will future

chemical engineer do what a future

doctor do I think that's actually a big

challenge of academic institutions but

on coding specifically I think every I I

would like I would like to see pretty

much everyone learn to code uh because

with a coding companion with Jenny I

hope coding is easier than ever before

um and the value of someone being able

to write a little bit code is higher

than ever before so I'm seeing you know

soft engineers get meaningful

productivity boosts we just do a lot

more right when you use geni but then

also among among my teams I'm seeing

kind of marketers and investors and kind

of people who job Ro is not software

engineer um write just a little bit of

codes uh to download web pages

synthesize it get insights and I find

that people that know just a little bit

of code um can do a lot more than often

do a lot more so that's why we we we

release this free sequence of causes uh

ai ai python for beginners to to help

people you know uh learn coding for the

first time so so one of the complaints

I'm seeing from open source maintainers

is that you have people who will use

like chat gbt to generate code and then

they try to run it and it doesn't work

correctly and it turns out there's a

really basic error like a variable is

named inconsistently right maybe it's

even capitalization or something and so

the variables don't match so it doesn't

work and they they know no coding and so

they don't know how to debug those

errors errors so how do we like teach

students to use the tool but also you

know have enough insight into how coding

actually works so they're actually

learning the the minutia that they need

to know to write good code yeah I know I

know social media tends explode with the

you know code and then why I build this

thing that's very cool that they did

that uh and I think at least for now

those are the exceptions and I'm seeing

that but hopefully there'll be more and

more of these exceptions but I think

I find people get a lot more traction

with low code than no code and if you

know just a little bit about what does

the word even exception mean you know

and and there these Concepts um I think

I I think for quite a long time someone

that knows a little bit of programming

Concepts uh will be to do a lot more

than someone that you know doesn't know

coding at all and just prompting um the

the the boundary is Shifting because of

Improvement in technology I think

anthropics um artifacts was really

clever right helping take Bas take

things to deployment and I know kind of

replate you know also making it easy

reducing friction deployment I uh uh and

I think um uh I personally build a lot

of um streamlet apps because it turns

out you know gbd4 is really good at

writing streamlet code so I don't worry

about the syntax just thrill stuff up in

minutes on on like a cloud you know or

whatever um so I think I think yeah I I

but but by find that people that

understand the bit of the coding

Concepts um you just get much further

much quicker and it's less like less

likely to hit a dead end I

think we have about three minutes left

um what are you most excited about over

the next

year

um excit about many many things but they

made me pick one it would be um

applications uh I think that uh there's

a lot of foundational work to be done um

including training Foundation models

better technology tons of work to be

done in data engineering but where I

think many people you know will get the

most value out of it will be the

applications um and we do need to work

on the foundations the shortage of

people that understand how to build the

data the foundation models the

infrastructure got to keep on working on

that and then I think ultimately you

know our field would be judged by our

success um at delivering useful

applications so I spend a lot of my time

focusing on applications uh but it

actually work on applications that then

you know causes me to sometimes have a

strong view maybe right maybe wrong that

boy we really need to get this data

infrastructure right or boy we really

need to get this orchestration later

right but but I I I think that we're

actually starting to see a rising flood

of applications oh and one one one

interesting thing about applications as

well I know people read in the news

about these you know billions of dollars

more than single digit billions of

dollars spent on gpus to train

Foundation models and people think doing

AI is really expensive right but it

turns out that because someone else has

spent you know these tens of billions of

dollars um is now very Capital efficient

to start to work on some data

infrastructure to start to work on some

application so I think at the

application layer the economics look

very favorable to to people you know

building and want to deploy stuff

it's

awesome Andrew it's uh great to talk

with you as always um thanks for thanks

for joining the podcast for taking the

time great yeah thank you great time you

guys thank you Joe thank you man yeah

cool right take care