Sam Altman and the team behind OpenAI have been kind of quiet for a while.

A calm before the storm, if you will.

But in the last few days, we've been getting some glimpses into what potentially is coming next.

Let's take a look at what they've been cooking up.

Interestingly enough, this seems to be a nod back to their original goal, their original stated mission of being the AI base layer.

Here I think is a great illustration of what they meant by that.

Imagine this sort of big block here.

That's layer one.

That's the base layer AI.

That's the GPT.

That's the Sora.

That's the DALL·E.

And on top of that, we build the various companies, various apps, various software that takes advantage of this base layer of this AI to build specific functionality.

For example, for science, for chat applications, for medical AI, etc.

When ChatGPT exploded onto the scene, OpenAI did say that this was kind of the plan for the future.

They were going to be the base layer and people could build on top of it.

Tons of people and apps did build on top of it.

But then OpenAI was like, I am death and just killed them all by building their own sort of versions of it, their own applications.

That's probably not 100% fair because a lot of these startups, they were trying to build technology to fill in the gaps that OpenAI was missing.

For example, things like whisper, things that were similar to Code Interpreter or advanced data analytics that it's called.

In other words, a lot of these startups, they were building kind of these obvious apps that OpenAI itself probably had in the works and they kind of got rolled over by OpenAI, by the base layer, if you will.

But it seems that now the game has changed.

OpenAI seemingly, now that it has all the pieces in place, it has the Sora, the whisper, the voice engine, it has advanced data analytics, it has DALL·E with inpainting and editing and all that stuff.

Like it's built all of the different pieces of the base layer, their full functionality.

You can talk back and forth to ChatGPT.

It's able to understand your speech.

It's able to talk back to you.

The base layer, a lot of it is complete.

All the companies that try to plug missing pieces of the base layer are dead.

Seemingly OpenAI is saying, okay guys, we've got the foundation, we've got the base layer, build stuff on top of it.

That's my opinion, but let me show you a few key pieces of news that came out just in the last few days that I think support this viewpoint.

As you look at this, ask yourself this question, is OpenAI rolling out the red carpet for the layer two startups to build on top of it, on top of OpenAI's base layer.

Let's get started.

This drop today, introducing improvements to the fine tuning API and expanding our custom models program.

Developers get more control over fine tuning and new ways to build custom models with OpenAI.

It begins, there are a variety of techniques that developers can use to increase model performance in an effort to reduce latency, improve accuracy, reduce costs.

Whether it's extending model knowledge with RAG, retrieval, augmented generation.

Basically having that model as example would be to search the internet for information like Perplexity, it returns to you the various links it finds.

And then kind of summarize them.

Or if you've ever done the chat with a PDF thing, right, the PDF is your sort of database, extra data, extra info that you have in there.

The model, instead of kind of making up an answer, it retrieves information from that PDF and then answers with that information in mind various custom models, fine tuning, custom trained models with new domain specific knowledge, and they're launching new features to give developers more control over fine tuning.

And they're introducing more ways to work with our team of AI experts and researchers to build custom models.

And they're not kidding.

As you'll see in a second, they've partnered with quite a number of startups of different companies that are outside of OpenAI to bring some pretty interesting features.

We'll cover that right after this.

They start by talking about some fine tuning of API features, or rather fine tuning of these models using an API.

For those that are not familiar, fine tuning is basically if you think of the GPT-4 model as this kind of big blob that can do anything, but maybe it doesn't do everything well, right?

What we can do is let's say we wanted to do exclusively coding or customer service with it.

We can fine tune this model to do customer service.

I'm going to say support because that's easier to write.

We can take this big huge model and fine tune it to be great at support customer service to answer questions that's specific to our company, our business, training how to answer questions correctly in the style that we want.

Of course, there's going to be tons of other areas where it's going to get worse because this support model is going to be almost a brand new model, right?

It's a fine tuned GPT-4.

It's a custom model for support.

And chances are it might get a lot worse in various other things, like it might not be that good at coding anymore.

It might get bad at poetry, but it gets good at support.

Again, our sort of like little demonstration here, this little representation here, right?

The base layer, right?

That's let's say GPT-4, right?

We take that, we pull it out and we fine tune it to be a science GPT-4 or a medical GPT-4.

All the stuff that GPT-4 already has comes with it, right?

It understands the language.

It can reason, right?

It has certain knowledge.

It has certain skills.

We're just taking it and shaping it into this specific task.

And the other really big thing about fine tuning is you can achieve higher quality results while reducing cost and latency.

Since we're pulling out just what we need for our customer support agent, for example, there's tons of things that are still in this big GPT model that we don't need, that we kind of leave behind.

This model could be cheaper to run, much faster, etc.

As an example, they give indeed a global job matching and hiring platform, wanted to send personalized recommendations to job seekers, highlighting relevant jobs based on their skills, experience, etc.

And they were able to fine tune GPT-3.5 Turbo to generate higher quality and more accurate recommendations.

They were able to cut their token costs by 80%. So it was able to improve cost and latency, how fast it was doing at cutting the number of tokens, like the worst that were run through the model by 80%. This allowed them to scale from less than one million recommendations to job seekers per month to roughly 20 million.

By the way, as a quick aside, so on this channel, we cover a lot of research from companies like Apple, Microsoft, Meta, OpenAI, Google, DeepMind, etc.

And I'm beginning to notice a pattern here that usually we see the research papers come out let's say three to six months before sort of the things that are talked about in those papers start hitting the market.

The Orca 2 paper from Microsoft Research kind of showed that teaching small language models had a reason that this was extremely effective.

Orca 2 models match or surpass all other models, including models five to ten times larger.

Smaller fine tune models built for specific tasks are incredibly effective, incredibly fast, and incredibly cheap.

We've covered Apple's research into vision models, you could say.

We covered this a few days ago.

They had this model that was 80 million parameters that was in some cases beating GPT-4, getting close in other cases, but 80 million parameters is a tiny, tiny model.

GPT-4 is gargantuan, clocking in at we think 1.7 trillion or you know, that's what people are guessing.

It's probably somewhere in the low trillions.

Here's a model that is, if I did my math correctly, this is like a thousand of a percent the size of GPT-4 that is getting close to it for performance for certain specific tasks.

Not for everything obviously, but for a specific task, this microscopic model is as good as GPT-4.

If I had to guess, I would guess that we're going to be seeing a lot more things like this, where small fast models, GPT-3.5 turbo and the like, get fine tuned to do pretty incredible things in specific domains for pennies, for very cheap and they'll do it very, very fast.

And the people that will be building this for businesses, those people will be printing money.

That's my guess.

And OpenAI is building services to make that easier.

They're going to be working with some companies on their own.

It's going to be an OpenAI plus, this is a company partnership, but I know a percentage of you listening right now are builders in the space, are developers.

I mean, these B2B solutions, if you're looking to make money, seem like a very, very lucrative thing to approach.

OpenAI is launching new features to give developers even more control over their fine tuning jobs.

By the way, I've said this before, but just let me reiterate because I don't think I've said it recently, for the next, let's say five years to a decade, one place where people are going to make a lot of money and one place where we're going to see a lot of progress is this idea of shoving LLMs into everything.

In other words, for any digital task running, there's probably some solution where you take a custom, fine-tuned Large Language Model, or maybe some custom small model, something like Orca 2, which Microsoft, by the way, was kind enough to open source, so you're able to see exactly how they built it.

And then apply to something like this.

Think about how much, indeed, a global job matching firm.

How much were they spending on these customized personalizations?

How important it is to their business model?

It probably has a very high importance and probably had a high cost, right?

If you're able to build this custom model that cuts down how much it costs them to do and how fast it's done by 80%, how many millions are you saving them per year?

Certainly, you can probably charge a pretty penny for doing something like that.

And I would guess that the talent pool for people that know how to do this isn't massive.

Like, it's not, like, I doubt that we have an overwhelming amount of people that are highly skilled in something like this.

And in just a second, I'll show you examples of how this is applied.

But here's the new features that OpenAI is rolling out.

It's adding a sort of model checkpoints.

During each training epoch to reduce the need for various retraining, etc.

Kind of like a version control, side-by-side Playground UI for comparing model quality and performance.

Apparently, somebody made two lems fight each other in Street Fighter 2, so they trained them to output what it should do, fireball, kick, move closer, etc., to see which LLM is the best Street Fighter 2 player.

That's kind of what I'm imagining here.

I doubt it's that cool.

And it also supports integration with third-party platforms, starting with weights and biases.

That's this platform where you're able to track experiments, evaluate model performance, etc., where you have the model registry, the added OpenAI.

OpenAI uses weights and bias models to track all their model versions across 2,000-plus projects, millions of experiments, and hundreds of team members.

You have visibility into model development process with just a few lines of code.

You know, you're able to answer questions like what exact version of the dataset was this model trained on?

I'm going to skip some of these parts.

I will leave a link.

If you're really interested in this, if you're in this space, I highly encourage to read this, obviously.

But for the rest of us, I think it might be more interesting to quickly go over the big points.

And then let's dive into exactly what this looks like.

What are the actual final outputs of this OpenAI as a base layer thing that's happening?

First of all, they are talking about something called assistant fine-tuning.

This is where OpenAI helps train models for a specific domain in partnership with a dedicated group of OpenAI researchers.

Since then, we've met with dozens of customers to assess their custom model needs and evolve their program to further maximize performance.

And they've published this.

We're going to take a look at that next.

But before we look into those specific examples, here's kind of the end of this blog post where OpenAI is saying, well, here's what's next for model customizations.

They're saying that we believe that in the future, the vast majority of organizations will develop customized models that are personalized to their industry, business or use case.

I think this is very important, especially if you're if you're in this for the money.

Not everybody listening to this channel is approaching this from kind of a business centric perspective.

Some of us are just interested in the research and kind of personal use of AI, how can we take advantage of it for work, for our careers.

I understand that this might not apply to everybody, but I still think this is very important to understand because this gives you a glimpse into what's going to happen over the next decade.

Here's 0% and here's 100%. This is a line, these are all the tasks that can be either automated or significantly improved or maybe made cheaper by applying AI.

From taking either a biological neural net, like a human being, an employee, or even taking code, like a lot of stuff runs on code, on something that a smart software developer sat down and typed out on their keyboard. And now that code runs to complete a certain task, like in accounting or stocks or, I mean, nowadays pretty much everything, right?

Not that long ago, close to zero of those business operations ran on neural nets.

Neural nets, that's the AI that we're talking about.

That's ChatGPT, that's Sora, that's all the AI music, all the AI images that you're seeing.

A lot of Google stuff is machine learning how they run their ad serving platform, content recommendations, stuff like that.

But you go back a few years and maybe the number of those things were like 1%, right?

That could be automated neural nets or improved in neural nets in some way.

That was like 1%. Since ChatGPT came out and all that stuff started coming out in the last few years, maybe we're at like, I don't know, 1.5%, 2%, I don't know.

This might be even like a horrible overestimation. Now that I think about it, it's probably some fraction of 1%. Over the next decade plus, all of those processes will be replaced by AI, by neural nets.

Some of them will be a one model to rule them all.

For example, some people will just use ChatGPT for whatever task they need, right?

They'll just use the base layer, kind of like that non-custom, non-fine-tuned version of GPT-4 to ask questions.

One interesting thing about Perplexity, for example, that's having incredible success and progress, it's not really doing that much on top of the base layer.

It's taking all the language models that already exist.

You have your GPT-4 Turbo, Claude3 Sonnet, Claude3 Opus, Mistral Large.

They have a few that you can play around in the Playground with.

Then you have their own fast ones that they've kind of fine-tuned.

But for the pro users, they're just using the base layer, whether that's Claude3 or GPT-4, with heaps of software on top of it.

They're taking GPT-4, they're building a little Perplexity, little search engine on top of it, and they're adding heaps of software to search the internet, pull back answers.

Literally in this image, let's say this is GPT-4, this is Perplexity, and the heaps of software is the features of that Perplexity search engine.

And they're valued at billions of dollars.

Bezos is investing in it, tons of other people are investing in it.

And rumor is Apple is in talks of potentially buying them.

That's one example where literally billions of dollars are created just with this.

But these Large Language Models or other things like vision models and speech models and image generation models, they're coming for everything.

They're going to be in your thermostat.

They're going to be in the food photography, right?

Some sort of an enhancement.

You know, this is a Magnific AI or Magnifique AI, right?

That upscaler.

I've featured it on a channel.

Some people use it to upscale their food photography, right?

Look at that thing.

You can see every little grain of pesto, every pine nut.

I'm guessing this is cracked black pepper.

I mean, that looks delicious.

I mean, take a look at this burger.

Like, they're going to use this in product photography.

The left is sort of the un-upscaled.

The right is the upscaled.

I mean, most product images will use this in one way or another.

We've covered Google Shopping AI with their new thing they're rolling out where you're basically able to use AI to create product visuals, to put them into different environments, to create virtual sort of those models, right?

That kind of showcase your product.

It's able to translate everything into different languages.

You can basically have an international store selling to the entire world in their own native language.

You don't have to do product photography, right?

You can pull out your phone.

If you have a prototype, take a picture of it and the AI will create 3D images and variations and put them in the hands of models and just do all of that for you.

As you'll see in a second, OpenAI is partnering up with people in the legal field to create custom trained models for them, another company that does health insurance, another company that does education, JetBrains for coding, email assistance, Salesforce, health and weight loss, productivity, clinical trials, creativity for farmers to increase their income, personalized fitness and health coaching, content creation.

Like, I can sit here all day and just name use cases, but I think this kind of chart, I hope, represents it a little bit better, right?

We're at close to zero right now of penetration of this technology in the market that's going to change everything.

It's going to reduce costs, improve automations, create new use cases that we haven't even thought of, right?

Everybody's going to want it.

Everybody's going to need it.

And over the next however many years, five, 10, 20, it's going to get rolled out from almost zero to eventually approaching 100.

And for every model that gets rolled out, people pay money and this money, it won't be an issue.

It won't be hard to sell these services because they're going to be saving money.

They're going to be making things faster.

It's hot.

It's sexy.

Everybody wants it from your local mom and pop store.

That's local.

That's brick and mortar.

That's selling, I don't know, quilts.

I've been doing e-commerce for the last 10 years.

I'm telling you, there's like off the top of my head, I can think of a number of places where if I could get a custom trained fine-tuned model to just do that task for me, it would probably be able to do it faster, better, cheaper than I could from inventory management to customer service, to a lot of aspects of marketing.

What I'm saying is this may be a good opportunity for those of you who like money.

Opening it continues saying with a variety of techniques available to build a custom model, organizations of all sizes, and this is important.

This is kind of like the big thing too of all sizes because with enterprise level software, like software that you sell to hospitals and government organizations, it was hard to build.

You had to have a massive team to even approach projects like that.

There's really no such thing as like artisan software, how you can go to a farmer's market and buy some artisan loaves of bread.

That's really delicious.

With software, you wanted a certain amount of scale because you had to build that software.

The bigger of a customer size that it served, the better the economics of that software would be.

It's a little bit different here.

We covered this blog post from Semi Analysis a while back.

Semi Analysis is a website that talks about all the news and events and does analysis of semiconductor and AI industries.

Sort of the hardware, the chips behind AI, all that stuff, NVIDIA and TSMC and all of that.

And they dropped this little lie that I thought was fascinating.

They're saying we're training a C and N. So CNN is a convolutional neural network to accelerate the frequent satellite imaging of data centers to expand our tracking to every data center across every country.

This is kind of like a publisher.

They have a website, a newsletter, they sell their reports, and they have this really specific use case.

They have access to satellite images.

I'm sure a lot of that stuff is public.

They track 1,100 data centers and their deployments with publicly available information, including but not limited to property records, power usage.

This is Freedom of Information Act requests and satellite images.

They have these very specific use cases where it would help them to have a neural net, a CNN in this case.

It's not a Large Language Model.

It's not a transformer-based architecture, a convolutional neural network in this case.

It kind of like labels and is trained on images.

But they're building their own custom AI model to help them analyze images of these data centers, looking at the satellite images and Figure out what's happening there.

They might do the same thing for these other things, for property records, power usage, right?

All this stuff.

How many other people in the world might be interested in doing something like this to analyze satellite images for this specific AI accelerator data centers?

Probably not that many.

Maybe a handful.

The point being that there's going to be a great demand for these custom small AI models for very specific use cases for organizations of all sizes that can develop personalized models to realize more meaningful, specific impact from their AI implementations.

The key is to clearly scope the use case, design and implement evaluation systems, choose the right techniques, and be prepared to iterate over time for the model to reach optimal performance.

Again, the people doing this will be making lots of money.

With OpenAI, most organizations can see meaningful results quickly with the self-serve fine-tuning API for any organization that needs to more deeply fine-tune their models or imbue new domain-specific knowledge into the model.

Our custom model programs can help.

One, this is where they have their own in-house AI people.

Two is where OpenAI helps them develop it, probably for bigger companies with specific use cases that can afford to pay for that.

But I think there's another obvious sort of category, and that's people that are outside of a corporation, right?

Like a third party that just goes to businesses and says, hey, I can create this for you and charge them a certain amount of money.

A consultant or whatever, AI consultant, automation consultant, whatever you want to call that.

And again, we're at close to 0% now and it's heading to 100.

The early movers will probably have a massive advantage.

But you might be saying, okay, but like, what are some of the use cases that make sense?

Is all this just hype?

Is all this just like marketing hype?

People are getting overexcited for nothing.

By the way, Demi-Sasabi said an interesting thing, right?

Here's AI Breakfast saying, in the words of Demi-Sasabi, the AI landscape is overhyped.

Yes, it is overhyped, but underestimated.

It's funny how both can be true, but it makes sense.

There's a lot of hype, but people aren't fully grasping the massiveness of everything that's happening here.

Because what we're talking about is just a tiny sliver of it, like these fine-tuned models for specific business cases.

That's just a tiny, tiny part of it.

But let's see what these use cases are.

These are customer stories from OpenAI.

This is kind of what they're talking about where they work together with companies to create custom fine-tuned models for their specific solutions.

Here's one that kind of jumped out at me.

This is Super Human.

That's the company name.

And they're introducing a new era of email with OpenAI.

They're building a suite of next generation AI email products that are saving users time, driving value, and increasing engagement.

We've all struggled with the overflowing inbox, right?

There's too many things, too many appointments, too many meeting minutes to keep up on.

Super Human is reimagining how that could be using AI to help you write emails, to rewrite certain ideas in your voice.

Or for example, just say it out loud, kind of dictate it and have the AI kind of fill the gaps, write it in your voice, make sure everything's correct.

And then send it.

I'll summarize a quick summary of each email that's arriving.

Instant reply allows you to reply from certain contextual options, similar to how a text message, you have those little options to reply.

It's important to understand that yes, so GPT, ChatGPT can do a lot of this, but this we're talking about custom solutions.

That's somebody sitting there and specifically trying to make this as good as possible for the specific use cases.

But okay, so maybe this isn't as sexy as I made it sound out to be, but here's what caught my attention.

They're saying here, okay, so what's next?

Super Humanist Company imagines a world where GPT-based agents will soon be able to filter, triage, and respond to email automatically, scheduling meetings and appointments, and taking basic actions online and in the real world.

Again, I mean, they're talking about building an email agent or a scheduling agent.

Kind of like an executive assistant for you that has access to your emails, to your calendar, and you can kind of like manage and stay on top of it based on what you tell it to do.

The idea of opening up an email inbox and then reading the first email and then the second email, the third email, or only doing triage kind of thinking, okay, like what can I get away with not answering today?

Let me see what's the most important things.

Like the idea of you doing that yourself with your brain, your eyes, and you clicking the mouse or whatever.

Like in five years, that's going to be laughable.

That's going to be on par with manually adjusting the heat in your house every 10 minutes instead of using a thermostat.

How many people have that now?

Close to zero, right?

Maybe some fraction of 1% of the people in the world are using that in their inbox, right?

It's not 1%. It's less for sure.

If you look at the entire world, even the United States, it's less.

In five years, it'll be 100% or close, right?

Somewhere getting closer to 100%. Next, we have Harvey.

It's a custom trained model for legal professionals.

Of course, lawyers and technology, the combination of the two is hilarious, right?

The I am not a cat, your honor video is perhaps the funniest video I've ever seen in my life.

I think the first time I heard myself laughing, I'm not even exaggerating.

Then of course, you have the lawyers that use Chagmit to, I don't know what they were trying to do, but it basically made up a bunch of court cases to cite their arguments, and it got them into a lot of trouble.

At the time, a lot of the clueless reporters that cover AI wrote articles saying their conclusion was AI bad.

The reality was they shouldn't have used GPT-4 kind of that base model, right?

They needed something that was custom trained, that had some architecture, that had some rag, that retrieval augmented generation, right?

As you can see here, this is the model that was trained.

It's custom trained.

And as you can see here, it's giving you little citations of the case law.

For everything that it writes, it supports it by giving you a little, okay, so I'm referencing this specific case law, right?

Stone versus writer or whatever, right?

To support its claims.

It's not guessing, it's not hallucinating, it's not taking the best shot at it.

It's using actual databases to support those claims.

It's custom trained not to take a stab at it or whatever.

It's custom trained to answer just as accurately as possible.

That company that's doing it, they're called Harvey, and they've worked with 10 of the largest law firms to test this model.

And they were surprised by how strong the reaction was from those law firms.

97% of the time, the lawyers prefer the output from the case law model from this custom trained model, because it was longer, more complete answer.

It went into the nuance of what the question was asking and covered more relevant case law.

Hallucination reduction was one of Harvey's motivations for building a custom model.

Again, the problem wasn't quote unquote AI.

The problem was it just didn't have the correct architecture needed to provide the correct answers and the people using it didn't have the correct training to understand what it was good at, where it could fail.

Again, that's another part of this whole thing is training the staff, the people using it to kind of understand its limitations, etc.

Having ChadGPT represent you in a legal case is shockingly not a good idea.

Where are they seeing this whole thing going next?

Well, they're saying this, don't build for the current capabilities of these models today.

Because remember, when you do that, what happens?

OpenAI comes in, it's like, I am deaf.

And it just like reaps everybody, right?

Build for where the models are going to be.

Tackle more complex versions of problems so that when better versions of the model come out, they aren't solved as a side effect.

And what's Harvey working on next?

Let's see.

One of their main focuses is agents.

Again, surprise, surprise.

AI autonomous agents are once again, the next frontier.

Here's yet another company that's like, well, that's the very next step.

It's in this case, how to combine multiple model calls together in a single working output.

This would simplify the user experience and reduce the amount of prompt engineering and typing users needed to do.

Another one is Oscar, bringing AI to health insurance.

The OpenAI models together with Oscar help ensure HIPAA compliance, doubling productivity, automated documentation, and claims processing.

Medical care workers cut their time spent documenting various medical care conversations, stuff like that, by 40%. So that's your doctor, that's your nurse, almost half of the time, instead of thinking like what your needs are, they're sitting there filling out this paperwork, draining their mental energy, their physical energy.

Think about how many millions of dollars we spend having them do that instead of something like a Large Language Models taking care of it for pennies.

This reduces burnout and allows nurses and clinicians to focus on higher order tasks.

How many lawyers and doctors and nurses use various tools like this to help them in their work right now?

Probably close to zero, right?

We're definitely not near 1%, right?

It's some tiny fraction of 1%. How many of them will be using it in five to 10 years?

Probably a lot closer to 100%, right?

Next, they're also talking about claims, improving accuracy, automating 40,000 tickets by the end of the year.

All right, so this is what a claim looks like, right?

You click this little magic button, help me with this inquiry, and it fills it out, I assume.

Next, they're saying creating an AI flywheel.

A flywheel in business usually refers to kind of like a virtual cycle where something almost kind of like builds on itself to keep improving that business or that habit or whatever.

The more users a social network has, the more value there is for existing users and the more users it will attract.

This is kind of like a direct network effect.

They're saying we don't want to just nibble around the edges of administrative use case simplification, right?

Because this is what we're talking about, right?

It's just simplifying administrative use cases, right?

Important thing, very valuable thing, will save a lot of money, will save the mental energies of doctors and nurses that need to focus on their patients.

It's important, it's critical, but they're saying that's just the beginning.

They're looking to bring down the cost of seeing physicians and being in the hospital by a factor of 10 in the next three to five years.

Our customer was Zelma, who's using GPT-4 to make education data accessible.

Our customer success story is Healthify, an app that helps with weight loss.

This is for a specific population.

It looks like they're really focusing on India, it seems, and specifically classifying traditional Indian foods.

Focusing on certain custom populations can be a really good idea.

Here they're saying SNAP achieved around 80% accuracy for single Indian foods, right?

Here's where it gets a little bit interesting if you're looking at it from a business perspective.

Like how important is it for businesses to have something like this?

Well, increased engagement.

Because of the fine-tuned models, the AI, the engagement increases.

People are more interested, people tend to use it more.

Users track 50% more often with these new models.

Users engage more for nutrition and fitness coaching.

Clients engage with AI-supported coaches 18% more.

And, and this is kind of interesting, with users permission, the agents will even be able to order food or book gym classes.

A brand new industry that's going to emerge that we haven't yet seen, but a few people are beginning to mention it in the future when we have agents that control our lives.

That sounds harsh.

That help us run our lives.

They're kind of like our assistants, but they do connect us to information, to services, to everything.

They're going to recommend certain products, right?

They're going to order food for us, book gym classes, etc.

A lot of marketing dollars are spent on marketing to the person, right?

Billboards, ads, it's focused on you, right?

It's focused on trying to get you, the human, to buy something.

Slowly over time, we're going to see more and more money going into marketing to AIs, to these agents.

What is that going to mean?

Well, I mean, the simplest way is spending ad dollars to get the company behind these agents to recommend your products over the competitors.

That's the most obvious one, but there might be things like with search engine optimization, right?

It's how you optimize your webpage, the links that are pointing to your webpage, that determines how high you appear in Google Search.

Could there be an SEO, but for autonomous AI agents and AO, if you will, autonomous AI agents optimization, it'll be interesting to see how that unfolds.

But you know, companies will be spending money trying to Figure out how to get autonomous agents to recommend them over their competition.

Like that's going to happen.

100% chance of that.

And gen AI isn't just for us soft first-worlders only.

Digital Green uses OpenAI to increase farmer income in countries including India and Kenya.

It's important for these farmers to be able to teach each other best practice for growing new crops, share various local weather conditions.

And there's a lot of problems with that, right?

Not just connecting them, but also having a lot of different languages that are being spoken, right?

They need to connect with their local suppliers and provide market and pricing information.

It's like Eve online, but for farmers, farmer to farmer training videos, increase farmer income by an average of 24%. This was interesting.

We, I mentioned this in yesterday's video.

There's some things that governments and various well-funded government like entities can do that may not be necessarily something that large corporations are willing to do where they can use their resources to build something that's useful for the whole country, right?

For example, a database for all the countries and that cultural AI to be trained on, right?

Microsoft probably won't do that.

Google won't do that, but maybe something like DARPA might consider it.

I don't know if that's exactly in their wheelhouse or not, but certainly it's something that would help everybody in that country that's in AI to be able to do something like this here.

India's Ministry of Agriculture validates all documents in the knowledge base to ensure accuracy and reliability.

And then with GPT-4, they're using RAG, Retrieval Augmented Generation, to pull the needed stuff from there, various crop research, fact sheets, etc.

And the cost of this farmer chat, these extension services, they went from $35 per farmer to 35 cents per farmer.

Again, massive, massive leap.

And again, right now this technology is available to close to 0% of the world's population of farmers with those needs, right?

And over the next decade, hopefully it will start approaching 100%. And there's like 20 of these little customer profiles.

Each one has its own mind-blowing application.

Like there's not one use case here that isn't kind of revolutionary in its own way, where it provides more features, more use cases, it drops the price of doing something.

It frees up the humans, the experts, to focus on what they're supposed to be doing instead of being bogged down in the details.

Where this medical group believed they would need specialized medical models to get good results.

And they were shocked to find that GPT-4 outperformed a team of highly trained human experts.

They said it, not I. They were shocked.

I'm sensing that the title of this video is going to have the word shocking in it.

What do you think?

Am I correct?

Is my intuition accurate?

I'll leave this page down below.

This is the OpenAI.com.

This is not their blog, which I think a lot of people are familiar with.

This is/customer -stories.

And I think it just got populated with all of this stuff.

There's a whole bunch of them here.

And most of them aren't like fluffy.

They're not just whatever nonsense.

They're impactful.

And they showcase how powerful these custom trained models can be.

I'll leave it off there.

I think I've made my point.

Back when software was rolling out, that created a lot of wealth, a lot of change in the world.

Software was eating the world, they said.

This is the next step where AI is doing that thing.

We probably shouldn't say AI is eating the world.

We should say AI is helping the world.

I think that's the A16Z motto.

AI is going to save the world.

And I agree for a lot of people that are scared of AI, I don't know, becoming the Terminator and turning us into paper clips or whatever that scenario is.

I hope they take a look at this as the ability to make our lives easier to eliminate busy work, to help people understand various medical procedures better, to deal with email, to understand legal cases better, to make education data more accessible, so more people understand what decisions are being made in our school systems.

Help people in farming communities that are struggling to a little bit better.

This is where AI is right now.

This is what it can do right now.

And over the next five years, ten years, it's going to be rolling out from zero to 100% of whatever use cases we can apply it to.

The world will change.

Some people will make a lot of money.

And it's going to be a heck of a wild ride.

I don't know about you, but I'm pretty excited about what's coming.

I'm excited about the empowerment that AI is going to bring to education, to business, to creating music, to creating movies with things like SunoAI and Sora.

It's going to empower people to do more.

And we'll have some bumps in the road, I'm sure.

But wherever you are in life, I would say, get excited.

If you're in a position to become this AI automation consultant and help people incorporate neural networks into various parts of their business to create these fine-tuned custom models for them, I think that's going to be a millionaire making industry, probably for some people.

If you're running a business or want to think about where things like that could help you in the business, where they could help you do more or do more with less.

And if you're wondering why I keep showing you this wolf from Puss in Boots, I honestly don't know.

It's just a great character.

He just cracks me up.

This is the guy that plays them, by the way, if you weren't aware.

Anyways, my name is Wes Roth and thank you for watching.