#18: Keeping Humans "Near the Loop" with Jerrold Jackson

Welcome back to another episode of the Junction.

We are excited to have Jerrold Jackson

here with us today.

We got introduced to Jerrold through our founder.

He networked us up because we were talking

about these things called AI and ML.

And then we met Jerrold and realized

we were about ten years behind.

So, Jerrold, tell our listeners who you are, what

you do, what your background is, and then we'll

kind of get into some more of the nitty

gritty what you're doing with AI and ML today. Yeah.

So I'm Jerrold Jackson.

I am, broadly speaking, a

technology and data entrepreneur.

I've got a really mixed background.

A few careers ago, I was a psychotherapist at

an outpatient psychiatry clinic in New York City.

And then fast forward, a PhD business school and

about 15 years of technology experience later, I find

myself now leading a fitness and wellness startup based

in Houston, Texas, but also advising pretty widely across

the health and wellness space.

Slightly.

I'm curious, in the background, right, going from

kind of medicine then to business now to

technology, was there always this underlying theme of,

like, I mean, it sounds like in each

of those you just want to help people.

I don't know if you maybe felt that way,

but is there an underlying theme on that front?

There is, yeah.

The two themes are I want to have an

impact and help people, but I also want to

use data and technology to do that.

So what I failed to mention is even

going back to my undergraduate days, I was

actually it support at my undergraduate college.

So I was a nerd.

I mean, my father and I were building the

big computers with the big boxes and the towers.

I was building those as a kid,

was doing all kinds of things.

I convinced my parents to buy

multiple AOL dial up lines. Remember that?

Like, way, way back. Throwing it way back. Yeah.

So I was in all kinds of really cool stuff back then.

Was just like kind of early,

early stage Internet and web.

Built my first website when I

was ten, that kind of kid.

But then fast forward, even into college, like I

said, I was doing kind of tech support things,

was learning about early, early stage kind of statistics

and advanced statistics, and then eventually machine learning.

Even when I was doing psychotherapy work,

my clinic ran lots of clinical trials.

And so I became interested in not

only implementing the evidence based intervention, but

also running some predictive analytics on why. Right.

Why is something happening and what might happen if.

So, my phd was from Mount Sinai School of Medicine.

Of course, the advanced statistics work I did,

there was a great parlay into more of

the machine learning then, you know, moving into

industry, and even through business school, started to

do some machine learning based work.

Initially, the industry was actually for a hedge fund.

So I started kind of in that world,

which is highly regulated, or at least the

work we were doing was more regulated.

It was more financial services on behalf of the

hedge fund, and then, of course, into the wild,

wild west of health and wellness and worked in

healthcare payments for a little bit.

Had my first startup company in the health

and tech world about 13 years ago with

a colleague of mine at Mount Sinai.

Exited that company a while back.

But yeah, I've always been interested in kind of woven

impact and data through all that I do, man.

Tell me, obviously you've got this diverse background.

I think we're going to talk about some of the

studies that you've done or the work that you're publishing.

But I'm just curious, what's your take on the

last few weeks with all of the, maybe I'll

call it not like government politics, right?

But Sam Altman's out.

Anthropic does a big push right after that happens.

I don't know, what's the current sentiment, at least

from you, like in your mind, for December?

Yeah, it's a great question.

I look at all these big themes, right?

Thematically, blockchain was huge when it first

launched, and it tried to position itself

as, like, how is it practical?

How does someone actually use this technology to

make money, to make something more secure?

All those kind of things.

I put generative AI in a different bucket, but a

similar workflow, which is that people are still trying to

figure out an enterprise, how to use this to make

money or to make someone's life easier.

A lot of folks, there's lots of sentiment

around, am I going to lose my job?

But really it's like, what's the practical application of

this and the Sam Alton kind of movement?

First, he's fired things back and

all that kind of stuff.

That's just, to me, it's details, right?

It's kind of just this new technology

trying to settle itself in a marketplace.

I have to remind people that large language models

may be new, this transformer network that GPTs are

built on may be on the newer side of

things, but language models are not new, right?

Natural language processing is not new.

So as we think about, AI is now

synonymous with chat, GPT, or cloud bianthropic.

These newer large language models.

AI is synonymous with large language model, but

language models are not new at all. Right?

There's a whole field of natural language processing

that's been around for a long time.

What's new is cloud computing resources.

And I've got a sneaky suspicion,

chase that down the road.

We'll probably actually see a bit of a

reversion to the mean, where it's like, oh

my gosh, there's too much access.

It's too available, it's too

scary, it's too understood.

How do I go back to my private cloud? And then.

I'm not saying that we're going to be back

to on prem support, but it's almost like you

see these cycles where people get really excited.

They adopt these early adopters, something big happens.

People say, oh, crap.

Or maybe something a bit more explicit than that, and

then it reverts back to something even more secure.

You talk about people wanting to go back.

I don't know how far back people want to go, but

I have been known to tell you, Chase, that I would

go back to the days of a landline sometimes because I'm

so overwhelmed with all of the availability that we have.

It's a blessing that we've got all

these incredible technologies that make it.

People can work from wherever. Right.

But then I just think about, man,

wouldn't it be nice to just.

I just wasn't at the house.

Sorry I missed your call.

Sorry, I didn't get your text.

I think it would be great.

I think it would be great.

And I think what comes to mind in terms of old

technology that's currently having its heyday is the QR code.

Right.

The pandemic brought the QR code to us.

That's not new. Right.

But they're finding new applications, which

I think is really funny.

But it's a fairly secure way to transmit

information to understand who someone is, et cetera.

So I agree with you.

I think that chase, to your original question, I think

the last month has brought us maybe only the beginning

of what we'll see probably through the next, I'd say,

a handful of quarters at this point.

Some new release people are chasing Agi. Right.

Like some new application of what we're

trying to do will come out.

The one theme I'll say that I think is really interesting

is both in the work I currently do and in some

work I'll be writing about soon and publishing soon.

I think that doing this responsibly

is going to be the key.

So I think partially how this plane

lands at Enterprise and at scale is

doing all this generative AI work responsibly.

Yeah, you bring up a great point.

Mel and I were talking about

predictions for the next year.

And one of those, at least in my mind,

is people are going to be utilizing the technology

in a way where they're not really going to

talk about the underlying structure of what they're doing.

They're just going to come to

market with a new product.

And I think we maybe see a little bit about

that with how the large language models are trained.

Right.

Where are they getting the data set from? Right.

And it seems to be like, well, they just

scraped the entire web, and maybe there's some moral

implications behind that, but they did it.

And here we are.

And going back to your idea of the reversion of

the mean right now that their models and the weights

are out there for you to download, we have some

colleagues internally that are running these models on their servers

at home, and we've got all the vpns set up

right, and we can just play around and we're not

racking up giant electricity bills.

So I can see it going a ton of ways.

Actually, we were just, like, running cost

analysis to figure out relatively inexpensive gpus

that most gamers use a whole lot

less electricity if it's already trained. Right.

You're just kind of utilizing existing power, and maybe

you're ramping up the electricity bill a little bit,

but we're talking like maybe a couple of month.

I'm curious if, as we kind of pivot to

how your team and the folks that you are

working with and the different initiatives that you're working

on, are you bringing, where do you sit on

that spectrum of responsible use, and how are you

incorporating that into the work you're doing currently?

I think about this a couple of different ways.

As a technologist, I like solutions that

are less monolithic and more modular.

Those aren't complete opposites.

But what I mean by that is a monolith,

broadly speaking, is something you build for one purpose.

More of a modular architecture is an architecture are, that

literally is built as a sum of multiple parts.

And you can swap out this part

and optimize it whenever you want to.

You can swap out this part and optimize it.

Oh, this part over here failed. That's okay.

The whole system didn't fail.

You can swap out that one part.

So the way I think about this

to date has been very modular.

And the LLM component of any solution, broadly

speaking, any automated solution has to be modular.

So being LLM agnostic first and foremost to me

has been critical in launching products at scale, particularly

products that are b to b, because different businesses

may have a different perspective on what large language

model they want to use.

They may have a preference,

for example, cloud by anthropic.

By the way, I don't get paid

by any of these folks, right?

But cloud by anthropic is one example of

one that has been pretty forthright about.

They actually published white papers about

how they're doing, what they're doing.

And they've been very transparent and they kind of

have gone to market as this responsible, trustworthy LLm.

Others have not been quite as transparent

about what they're training on, how they're

training, their model, et cetera.

All of these can be considered a foundation model.

So as you all were kind of alluding

to, in my case in particular, kind of.

So far, I've launched these different solutions in an LLM

agnostic way, which means I can swap out llms as

I want to in a very modular sort of architecture.

And then I can very responsibly, based

on the use case, deploy this technology.

So, for example, a marketing copy use

case, people refer to llms as hallucinating.

It may actually be more of a feature versus a bug.

I want extreme creativity for marketing. Right.

Give me the top 5000 examples of

how to take these three concepts.

If I'm prompt engineering, take these three concepts or

this bucket of words and give me 5000 different

versions of sentences or whatever the use case is.

If I'm recommending some very specific actions

for someone to take, however, I don't

want that level of creativity.

You can actually ramp up and ramp down.

They're called hyperparameters.

How creative you want these llms to be?

Another way to make these things responsible is

to pair them with a recommendation engine or

some more structured way that you actually want

to deliver actions and sort know next steps.

So this is one critique I have about

some of the ogs in the recommendation space.

So take, for instance, Netflix.

Netflix is a prolific company with a

prolific recommendation engine along the lines of

Amazon prime or other companies.

These big, big, massive companies

worth billions of dollars.

These recommendation engines, which are not gen

AI, they primarily recommend that you just

consume more, buy more, watch more.

I've never had Netflix say, hey, hold

on, brother, you've watched too much today.

Or, hey, Amazon prime, you've bought too

many peepee pads for your dogs.

They're driving consumption, consumption, consumption.

Whereas I think a more responsible way to leverage, let's say,

gen AI is to, instead of just have it drive a

conversation kind of into a hole and potentially into a direction

that you don't really want is to pair it with a

very intelligent way of recommending something to someone in a way

that is a bit more guided and might even kind of

say, hey, maybe calm down today, maybe don't do so much

today, don't do the most today. Right.

So there are some responsible ways to leverage Gen AI,

and I think it really depends on the use case.

I like that thought a lot.

I even see some of this.

I don't know how much you've played around with Chat

GPT, but they have different versions of the same model.

And from version to version, maybe it gets

better or maybe it gets worse, right?

And I think a lot of in the

background they're trying to manage their electricity bill

just to be real broad, but also provide

more functionality, but at a lesser cost.

So they're constantly adjusting their weights or whatever,

but the end user doesn't know the difference.

Oh, you're so right.

Yesterday we were talking about the trend

around mental health and people using these

models as sort of therapists.

And one of the benefits that's

being touted is like the accessibility.

It's there all the time, you can always talk to it.

But hadn't really considered the possibility of like,

maybe there is such thing as too much

consumption and something that can help kind of.

We were talking about, well, with a therapist or

a counselor, you're kind of locked into your 30

minutes or your hour for that week.

But if it's Chat GPT or Claude, they're there for you.

So I'm sure there's pros and cons

on both sides, but I think that

idea around consumption is really interesting.

Well, you have this high level of like, well,

now you're talking to the therapist GPT or you're

talking to the coach GPT and different personalities almost.

But under that you could be flipping out models based

on how much knowledge we need this personality to have.

And you can make it a feature, right?

Like, well, here's the expert coach, right, that

knows everything about everything, but maybe you're paying

half as much and now you get, I

don't know, the high school coach.

I can see use cases across the board

and I too wouldn't call them bugs, right?

They're features.

It's providing a level of functionality to somebody

while at the same time flipping out.

Like the modular idea that you had where you can

flip out a model but not interrupt that user experience.

Just like you were talking about, Gerald.

Yeah, I think Chase, to your point earlier

too, about GPT 3.5 versus GPT four and

different versions and they'll have it.

I know OpenAI is working on a

version five right now as well.

If we're talking about OpenAI

specifically, what's tough about that?

If you're just purely going through OpenAI to

build an entire product, and once again, I

don't get money, nor do I don't get

anything pros or cons from speaking about it.

But just me observing as a consumer and someone

that builds products using these things, it is kind

of terrifying because literally they could have an announcement

tomorrow that says that they've come out with GPT

five and they're downgrading 3.5.

You don't know what that means.

Is the knowledge base changing?

Is the word vectorization strategy

on the back end changing?

If and when you're working with a product or a

solution that's not as transparent, that's what you get.

On the other hand, there are the major cloud providers,

Microsoft Azure of course, as being the main one.

They're serving up these models, really all the major

ones, cloud, GPT, et cetera, they're setting them up

as managed services, meaning you pay those companies for

a stable version of those models, that is not

going to change as OpenAI kind of wishes.

So in doing so, what you get is if you go

through Azure and pay a little bit more money, you get

a more stable product that you can rely on.

Well, I don't know how close you stayed to the dev

day, the OpenAI dev day I put my application in, but

I'm apparently not, I was waiting on my press pass.

Yeah, we're not well known enough

to get access to that.

But within that one day, right, they blew all of

the chat bots that have popped up out of the

water and just, I mean, maybe didn't completely put them

out of business, but I know I saw several Reddit

posts that were like, I just lost my entire business

that I've been working on for the past six months.

And it's like, yeah, well, yeah, when the technology is

shifting that much, you have to be prepared for that.

But I like your thought on the azure piece, right?

It provides a level of stability for

the SMB space like the enterprise.

Hey, I want to build a product off of this, but

if you want bleeding edge, you got to be prepared to

go off the edge almost, because tomorrow it could change.

GPT five might do everything that

we're already talking about and more. Yeah.

And I gave the blockchain example.

We saw the same thing in crypto. Right?

So if you were invested in bitcoin

or ethereum, you're great big, stable.

If you have one of the meme coins, you might

be able to afford a new shirt on one day,

and you might lose two shirts the next day.

It's all over the place.

It's exciting, but I think it also kind of facilitates

for me, this is what I'm writing about these days.

I'll say as well, in general, and this is

kind of a general place to find these things.

Ww dot neartheloop.com.

Near the loop, as in humanintheloop.

So L-O-O-P.

So, neartheloop.com is where I'll publish

these musings, kind of over time.

But my thought is the following.

Long before chat, GPT, and even OpenAI kind

of changed the game, really, this year.

Last couple of years, there's been a lot of

thought around AI systems without a human in the

loop, and I think there are a number of

industries where that can certainly work.

Once again, I mentioned the idea of marketing copy.

I'm not a marketer, and nothing against marketers,

but there's a certain world where you can

generate a bunch of copy really fast.

Do I want my attorneys to be billing

me a billable hour, but using GPT to

spin up legal documents, probably not, right?

So I want a human in that loop for sure.

Do I want my local doctor's office using

some GPT based or enhanced radiology technology?

No, I probably don't. Right?

I want an actual human.

It's okay if a machine takes things a

certain percentage of the way with certainty, but

I still love a human in the loop.

Well, the concept I have is, what if a human

was near the loop, maybe not in the loop the

whole time, and maybe not fully autonomous, but what if

a human was near the loop to where, especially in

some regulated industries or industries where the stakes are pretty

high, you have a built in mechanism to where you

have a human who can still make sure know the

actual machine or the bot is still on the rails.

Because I think it's funny, last time I was

in San Francisco, I tried the autonomous driving cars.

I applied.

I was on a list for Waymo for a couple years, got

off the list recently, tried a waymo for the first time.

It was awesome.

But anyone that's been to San Francisco knows that you oftentimes

see a human in the driver's seat of a waymo.

You're like, wait a minute, I

thought it was self driving.

There's still occasionally a human, and

there's articles published about this both

academically and non academically.

There are still occasionally humans

in that loop, right?

There are humans in the driver's seat who are

assisting that vehicle's data collection with all of its

thousands of cameras, because we still can't quite get

away from a human giving feedback.

Yeah, I always do a gut check with chase.

Every so often, I'll play this game with him.

Like, let's go back and forth and list the different

things that can't be automated or taken over by AI.

And I still have, like, there's those,

like, fly fishing or being some kind

of in person yoga instructor or something.

Like, the things, to your point,

Gerald, that you were talking about.

Like, we still desire to have a human near the

loop, even if we know additional insights or data.

Because I think even when I brought up

the fly fishing example, you're like, well, what

if I had the ability to.

It could correct my movements and stuff. Great.

But you're the one still doing it.

You're still a human near the loop or in.

I don't know, maybe I'm not using the right.

No, that's example for that.

But it's just a different use case.

I don't foresee anybody absolutely loving the day

where they don't have to interact with anybody

whatsoever, but there's also nobody to interact with.

Man on the moon, all by themselves, totally disconnected,

and all they've got is their large language models.

I don't somebody being like, give that to me any day.

I'll do it right now.

On average.

What about when the robot is talking to another robot?

The agent idea, the auto GPT, some of

these GitHub repos, right, that are trying to

get the large language models to talk to

each other before they ask for a recommendation.

Do you see that really exploding and

really being, like, a really big deal?

And now you've got, like, 20 different

agents, right, working on a project?

Or is it really just in your mind going to be

like, yeah, there's one agent, one super strong agent, and does

most of the work, and you have the human next to

that agent, and they work side by side.

I definitely see a world where industry, right?

We're going to try it.

We're going to try this agent to

agent set of interactions at scale.

I know it's being tested, and I'm certain that there

are industries where this will make a lot of sense.

I see that entire sort of collection of

efforts hitting a ceiling at some point where

there's some pretty measurable negative impact.

And I mean, not just people losing their jobs.

But once again, long before generative

AI, there have been concepts of

reinforcement learning, for example, right?

It's a basic behavioral kind of concept rooted in

all sorts of behavioral theories where with the right

reinforcements in place, effectively, this is not an official

definition for my behaviorist out there, so don't pin

me to the stake on this one.

But if you have a series of consequences built

in, both negative and positive, you can teach a

machine, or in this case a bot or an

automated system to almost learn over time.

There's some famous examples of people attempting this

work and then kind of on the side,

hush hush, secretly teaching it rules to expedite

the process of learning through reinforcement.

Because true reinforcement learning can take, like in a

fully automated way, can take a really long time.

It's not very efficient.

But I do think that there will be

industries where we will attempt to have agent

to agent and maybe even multiple agents interacting.

But I think it will hit a ceiling at some

point and people will realize that for interactions that have

actual consequence, like real day to day consequence, we will

need to have a human near the loop.

That makes a lot of sense too, right?

Fully automated, too much risk, right.

It could just explode and your whole business

model goes under, out of the loop.

Or maybe like, no, AI is maybe not necessarily

the best either because now your competitors are going

to find a way to use it.

But in the loop where the AI is doing enough

work to expedite, what you're doing actually scales, right?

Because now you can have these agents or

whatever they end up being, help you do

your work or elevate these things and makes

you faster or better than your competition.

And in the business world, not that customers aren't

our primary focus, but if you're not in business

because your competitor puts you out of business, you

really can't help a whole lot of people, right?

So it's in most businesses interests, right, to pick up

this stuff, start to learn it, start to figure it

out, at least be aware of what's going on.

But Gerald, I've got to ask.

I know most of us, definitely, you

deal in the objective world, like statistical.

Give me the numbers.

Let me look at the data.

But you mentioned something that

I think, well, the rumor.

So here we're going subjective for a second, right?

But the rumor was that Sam was out because

they developed a model that can learn, right?

And maybe the rumor was it's not super

crazy, but it demonstrated the ability to learn

without rules or sidebars or things like that.

So I'm curious, are the rumors just totally not

a thing because it's not even possible yet?

Or is there a chance that they came out

with something, or they've built something that can learn?

And maybe it's rudimentary, but if it can

learn at all, and it can learn at

any rate whatsoever, that's game changing.

What are your thoughts?

Yeah, so once again, I think that

the core of this is reinforcement learning.

There's even reinforcement learning

with human feedback.

These are all things that for a while now, current

and former employees of really kind of big tech, right?

So meta and Alphabet have published

on these for a long time.

There's really a few players, the two of them

and maybe a couple of others, that actually have

the data sets big enough to train these massive,

massive foundation models on billions of parameters.

The human feedback part typically comes into play

when you have some expert level feedback, right?

But the use cases get really nuanced.

You can't collect human feedback on

every topic known to man, right?

That's just something that's not even feasible.

So I definitely think the technology is

there to give this general purpose foundation

model that can learn autonomously.

I definitely think the technology is there.

The data sets are there.

My concern still is in the nuances

and specifically the nuanced use cases.

So once again, do I want

an autonomous system doing my taxes? No.

Do I want autonomous system even making my coffee?

Heck no.

I think there are still some specific use cases

where, depending on, at the end of the day,

things that we as humans really, really care about.

And there's a long list of

them, and they're all individualized.

Those are the things where I think that, as

a professor of mine said many years ago, models

are beautiful, but they're not real people.

So we're still in that realm where we've got

models that can do really amazing things, but at

the end of the day, they aren't real people.

That sounds a lot of like what we were talking

about the other day, is that original content is king

from a market, like in the marketing space.

I don't know if you'll ever be

able to really, truly replace the humanity

in humans with something like Agi, right?

Because maybe they won't ever get there.

I won't say never, but I think there's

a long stretch of time you can get

pretty far now with deep fakes, right?

I've seen some pretty convincing deep fakes.

And if you use that as a medium to deliver

a message in a very visual way, combined with, let's

say, a script that's either generated dynamically or previously.

You can do a lot to communicate that.

But I think we also run up against regulation, right?

So our current administration has put out probably the

biggest, most opinionated piece on how some of this

might begin to shape up in terms of regulation,

and who knows where that's going to go.

I'm not going to touch that topic with

a ten foot pole, but I think that

regulation will likely come into play here.

And there seem to be a few lines in

that sort of initial regulation regulatory piece that could

potentially put some pressure on deep fakes, for the

primary reason, of course, preventing bad actors from convincing

people that something is real when it's actually just

generated by some automated system.

You're reminding me of something that I was thinking about the

other day, how I think of the Cold War, right?

And it was a race to arms, right, the nuclear arsenal.

And not to be drastic, but we're definitely on

a race with other countries to get this technology

harnessed to figure out how to use it well.

And whoever is really first to beat the

others may never get out of first place.

So there's a whole geopolitical thing going on there that we

probably don't have a ton of time to talk about.

But I'm curious.

You obviously know what you're talking about.

You've got a ton of experience in this space.

If I'm a listener wanting to find more out about

you and what you've done, do I go to LinkedIn?

Are we on TikTok? Where should I go?

Find more about what you've done in your

research and at work and things like that.

Yeah, I'm not nearly cool enough to be

on TikTok, so I'll just start that.

Definitely easy to find on LinkedIn

or just googling my name.

So Gerald M.

Jackson or Gerald Jackson.

J-E-R-R-O-L-D. Jackson?

Geraldjackson.com leads to my LinkedIn, so that certainly

is handy, as I said, as well.

Certainly heading into Q one of 2024 and

ongoing, I'll be publishing a lot of my

thoughts here on the near the loop concept. So.

Www.neartheloop.com.

I think there are pretty easy applications in

the world that I currently play in the

sort of health, wellness and fitness world.

And I think certainly beyond, as you think about things

that people care about, financial services, how we travel from

point a to point b safely, et cetera.

Well, we look forward to continuing to

follow what you do on this topic.

Gerald, I know this has been

a really fun conversation for us.

We geeked out on our intro conversation prior to that.

So thank you for spending the time with us.

And we look forward to, like I

said, seeing where all this stuff goes.

We'll probably have to check in with

you at some point next year. Totally.

We'll have you back.

We'll have you back and we'll figure out how much

of our predictions were right and how much they were

so far off because we weren't thinking that far.

Gerald, thanks for joining us.

Everything that we've talked about today, like Mel said, is

top of mind for a lot of what we've got

going on in house that we're doing for our clients.

If you're listening today's episode, check out Gerald on

LinkedIn and give us your thoughts and feedback.

We'd love to hear what you guys have to think,

but until we see you again, keep it automated.