In this episode, we take a deep look at the two weeks that changed the world. From GPT-4 to Google Bard, Midjourney v5 and even talk of AGI from Microsoft, it’s all right here.

Podcast: https://youtu.be/dp_A32XkXYs

Intro Track: https://youtu.be/IECITKmNnzM

Track during Midjourney section: https://youtu.be/N8HYLv4WZWY 

Outro Track: https://youtu.be/8nTMej5WIOw

Runway-1 video credit at 10:45 : https://twitter.com/nickfloats

» ColdFusion Discord:  https://discord.gg/coldfusion
» Twitter | @ColdFusion_TV
» Instagram | coldfusiontv
» Facebook | https://www.facebook.com/ColdFusioncollective


Producer: Dagogo Altraide

Hi, welcome to another episode of ColdFusion.

The growth in both the prevalence and capacity of artificial intelligence systems seems to be exponential.

With the release of ChatGPT, suddenly everything from education to copy editors was disrupted.

Then came Microsoft's announcement, The technology would be integrated into their search engine.

For those following along, these events sparked a battle in the tech industry, so entertaining that the only thing that 1 could do was sit back and enjoy with some popcorn.

Google and Microsoft would go head to head in what's being dubbed the artificial intelligence wars.

In the latest developments, OpenAI reveals the next generation of their language model, GPT-4.

It's smarter than ever before, can recognize images, and has left users dumbfounded.

Once again, the internet was set alight with incredible use cases.

Many were left wondering, what industries are going to remain untouched at this rate?

But within a matter of days, the floodgates were blown right open.

Google's barred AI finally drops to consumers.

ChatGPT starts taking plug-in extensions for third-party software and this includes access to the internet.

Microsoft Bing receives AI generation image support.

Microsoft's Office suite receives AI capabilities as well, and everyone from Adobe to NVIDIA made their AI announcements.

Microsoft researchers admit that GPT-4 is showing, quote, sparks of Artificial General Intelligence or AGI.

A feat that wasn't supposed to happen for the next 20 years.

Also in this episode, we'll feast our eyes on some of the latest images coming from AI.

So definitely be sure to stick around for that.

And just a disclaimer, with the rapid speed of AI progress demonstrated in the rest of this episode, some viewers may find it concerning.

But my aim here isn't to scare anyone, but to give you the courtesy of showing you what's around the corner.

There's even 3 prompting tips in here to help you get ahead.

And with that, hold onto your seatbelts because it's going to be a wild ride.

We're starting to get the first studies done on this new early stage of generative AI.

An MIT paper found that professionals using systems like ChatGPT are now able to complete tasks in half the time while increasing the quality of the output by more than 20%.

In the ColdFusion episode covering ChatGPT's launch, I mentioned that this would be the equivalent of having an expert assistant in any field with you at all times.

We all know it's not perfect, but it's still helping people even in its current form.

New applications of the technology were being discovered every day.

Even its creators had no idea what it could do or even the precise means of how it does it.

So what happens when the world hasn't even digested the implications of ChatGPT, and yet a new and improved version is unleashed.

Just recently, OpenAI once again expanded the capabilities of its now famous text generation AI.

In the latest version, GPT 4.0, not only has the size of both its input and output been expanded from 3, 000 words to 25, 000 words, but the model has now become multimodal.

In other words, it not only understands text, but images now too.

It's very interesting and particularly intriguing because above just understanding the image, it can also interpret and explain what's depicted within the image.

For example, GPT-4 can now look at the groceries in your fridge and even generate a list of dishes that can be made with them.

It can understand and solve an image-based problem, and it can identify humour in multiple photos, even explaining why it's funny.

And another mind-blowing example is creating a website from scratch while only being shown a shoddy drawing on a piece of paper.

Here's an example of GPT-3 versus GPT-4 on the same logic problem.

User, in a room there are 100 murderers, you kill 1 of them, how many murderers are left?

If there were originally 100 murderers in the room and you killed 1 of them, there would still be 99 murderers left in the room.

Assuming the murders are people who have committed murder and you kill 1 of them, there would still be 500 people who have committed murder in the room, including yourself, as you would now also be considered a murderer.

On first glance, I can even see a lot of humans tripping up on this 1, so I find this particularly impressive.

On top of all of that, its deep learning capabilities have also received a boost.

It's improved its overall reasoning abilities.

As a result, its performance across multiple reasoning benchmarks have increased significantly, reaching a human-like level of common sense.

For example, on the bar exam, the GPT series went from the bottom 10% to the top 10% performance level.

Its reasoning capabilities surprised even the OpenAI team.

During tests, they discovered that it was able to solve hindsight neglect, a decision-making problem where it had struggled greatly in the past.

In this figure, you can see GPT 3.5 scored almost 0 and GPT 4, a perfect score of 100.

Improved reasoning also boosted a sore spot for previous models, the factualness of the model.

According to the company GPT-4 is 40% more likely than its predecessor to produce a factual response and 82% less likely to respond with quote, disallowed content.

Yet of course, despite these improvements, the research notes that this model is still biased and can output some questionable information.

Now the interesting thing is, if you're curious about how the team at OpenAI achieved these results, there's little to no information on the subject.

In their release paper, the creators state that this is mainly due to 2 reasons, competition and security.

It seems that Microsoft is putting a lot of pressure on the back end of this technology to remain undisclosed.

This is a huge departure from OpenAI's original ethos and basically the opposite of the norm for the current AI industry up until this point.

It seems now like Microsoft is in full battle mode and the AI wars seem to be in full swing.

There may be some potential risks ahead, but more on that later.

About 1 week after GPT-4's public release, a team of Microsoft scientists published and released a paper highlighting that the AI system is showing, quote, sparks of human level intelligence, or AGI.

Quote, we demonstrate that, beyond its mastery of language, GPT-4 can solve novel and difficult tasks that span mathematics, coding, vision, medicine, law, psychology and more without needing any special prompting.

Moreover, in all of these tasks, GPT-4's performance is strikingly close to human level performance and often vastly surpasses prior models such as CHAT-GPT.

We believe that GPT-4's intelligence signals a true paradigm shift in the field of computer science and beyond.

Given the breadth and depth of GPT-4's capabilities, we believe that it could easily be viewed as an early, yet still incomplete, version of an artificial general intelligence system.

So, it's not full AGI just yet, but it looks like we'll be here soon enough if progress continues at the same rate.

So in all of this, you might be confused.

You might be thinking, I've used chat-gpt or gpt4, but it doesn't seem that great.

Well it turns out that many people are using it wrong.

Of course, there's plenty of prompt engineers out there to give you tips, but here are some of the best ones that I've heard.

The trick is to make it do, and not think.

A good template to follow is as such, give a specific objective for the output, a vague input will give you a vague answer.

And lastly, give a specific list of things that it should avoid.

You can't give a worker vague instructions and expect an amazing result versus being specific in what you want.

Here are some examples of some people who have figured it out and got GPT-4 to do some amazing things.

In just over a week since its release, people around the world have started to discover the whole world of new possibilities that this new AI model has to offer.

Due to its boost in deep learning, it got significantly better at coding.

1 user was able to create 5 microservices for a new product in less than 3 hours with the help of the new ChatGPT.

For context, the same service usually takes a human developer around 2 weeks and it would cost in the range of $5, 000.

Another user was able to create a Google Chrome extension in a matter of minutes from the ground up.

What about an iPhone app built from scratch using just GPT-4 and it's already on the App Store?

Its capabilities to create video games has also improved dramatically as well.

It can now create games all the way from classics like Snake and Tetris to more advanced 3D prototypes like the original Doom.

With its extended input and output, it can now process a lot more information.

It was able to find multiple vulnerabilities on a live Ethereum contract.

The extended output also facilitates creating deeper content.

For example, the CEO of LinkedIn was able to come up with the first AI-created book.

Staying on the creative side, by combining GPT-4 with other AI tools such as Mid Journey, Runway Gen 1 and Boomi AI.

And here's another example of a short clip completely made by AI.

By giving it a current drug it can find compounds with similar properties, modify them, and make sure that they're not patented.

And it can even help you purchase these compounds from a supplier.

1 user asked the system, I have $100, How can I make as much money as possible?

GPT-4 then acted basically as a company founder.

And at the time of writing, it's been 4 days since the experiment began.

You've probably been hearing a lot about ChatGPT.

It's this powerful artificial intelligence bot that can write papers for you and maybe even pass a bar exam.

But what about having it just make you some money?

This week 1 person gave it a simple task explaining it on Twitter.

He said, I gave chat GPT for a budget of $100 and I told it to make as much money as possible.

I'm acting as a human liaison buying anything it says to.

So 24 hours later, the stakes were raised significantly.

I've given Hustle GPT a formal challenge to get to $100, 000 cash on hand as quickly as possible.

Just sitting, I thought it would be a fun idea.

I went on Twitter and it really kind of blew up very quickly.

And so now we've got a community on Discord of 200 people all trying this hustle GPT challenge and about 30 people starting their own projects with it.

So I'll tell you exactly what's going on is it suggested this name, Green Gadget Guru.

I asked chat GPT, come up with a business idea, come up with a name, I'll go out and buy the domain, I'll get everything set up.

I'm just the human liaison like you said.

On another note, not related to GPT-4 specifically, but still extremely interesting nonetheless, a Chinese company took this a step further.

It's in charge of analytics, leadership decisions, risk management, and workplace efficiency.

Since its introduction, the company's stock has constantly outperformed the overall market for 7 months now.

Companies are also moving rapidly to create new services based on the capabilities of GPT-4.

Creating personalized bedtime stories for your kids to fall asleep to.

AI-powered matchmaking to make your dating life easier.

And in the case that you want to file a quick lawsuit, there's also a solution for that.

Some of these services are complete game changers for people.

It's an app that helps the visually impaired get some of their day-to-day functions back.

Duolingo has created a new feature in its subscription program that allows you to access a private AI tutor.

Khan Academy also introduced AI-based tutors for both students and teachers.

It seems that the knowledge gap can start to be closed with the introduction of these new tools.

It's not only much cheaper, but it offers high levels of personalized education, a problem that has forever plagued the field.

So, with human level reasoning, how far are we from starting to see this type of AI acting on its own?

According to its creators, there is evidence that this new model was developing independent long-term plans while accumulating power and resources to carry them out.

During the testing project by OpenAI, something very peculiar happened.

While testing for potentially risky behaviour, GPT-4 lied to a gig worker on TaskRabbit, telling them that it was visually impaired, and it hired this human worker to get around capture restrictions.

Moreover, GPT-4 can even just recognise the capture images itself when asked.

Despite these worrying signs, the model was not able to better itself on its own even if it had the research.

Well, in an open AI study, here are the most at-risk jobs and sadly for many people it's not looking great.

According to them, almost no 1 will escape.

80% of US jobs will be impacted by 10% or more, and the figures should be similar in the rest of the developed world.

The worst affected are those in high finance, insurance, data processing, publishing, scientists, and most interestingly, monetary authorities like central banks.

Now me personally, I don't think that last 1 is such a bad thing.

The least affected were agriculture, wood manufacturing, food, mining and construction.

The physical jobs will be the last to go.

But the sad thing is, a lot of the world still doesn't even know this is happening, so they're going to be completely blindsided.

And this brings us to an interesting discussion.

While more competition tends to be good for the consumer, in a rare case, this time around, it may not be so.

As companies move faster and faster to integrate AI into their solutions, the guardrails of these technologies are starting to be removed.

GPT-4 was not supposed to be released this soon.

In the official document, the researchers state that they do not endorse GPT-4's release or Microsoft's deployment plans.

It seems like the launch may have been accelerated by corporate pressure from Microsoft.

A leaked audio recording from Microsoft's VP, John Montgomery, seems to indicate the same.

He stated to his team that, quote, the pressure from CTO Kevin Scott and CEO Satya Nadella is very, very high to take the most recent open AI models and the ones that come after them and move them into the consumer's hands at a very high speed.

Concerns continued to grow as there was a shocking announcement made.

Microsoft had fired their entire AI ethics team that the company just removed the brakes.

As the pace becomes more rapid, more and more companies will lose the incentive to withhold these new models and launch them as they are as soon as possible.

The consequences of such actions could be drastic, and this is mainly because the next version may behave in ways that we simply cannot predict.

At the same time, other big tech companies announced their own intentions for creating large language models.

Among these companies are Apple, Amazon and Baidu.

And while these models are likely to compete for size and quality, a major breakthrough in costs came from Stanford.

Its own language model, Alpaca, was trained using other existing language models.

As a result, while not competitive with the latest GPT-4, the cost reduction was insane.

To train a model like GPT costs somewhere in the ballpark of $4 million.

To put this in perspective, this type of cost reduction was projected to only happen in the year 2030.

What was supposed to take 8 years of work only took 5 weeks.

Please just stop and think about that for a second.

This opens up the gates for more entities to create their own models for better or for worse.

And to add to that, While the majority of the cost comes from running the actual model, due to Alpaca's small size, it can also run on a local computer, reducing its operational cost.

This type of accessibility can be an absolute game-changer for many companies.

It's pretty much locked in that in a few years' time, we'll all be able to run localised AI chatbots on our smartphones.

Next, we'll move on to the latest in the corporate AI war.

But before we get to that, let's take a look at something else going on in the world of AI.

The AI image generator, Midjourney, has just ticked into its fifth iteration.

Some of the images that have come out of it are so lifelike that it's insane.

We'll sleep for a while, and I will dream We'll sleep for a while, and I will dream Sleep for a while I will return I will return I will return

And remember, this is the worst that it would ever be.

So now, all of this brings us to the latest developments in the AI war.

A lot has happened in a short period of time, and the speed of updates only seems to increase.

During the same week, both Google and Microsoft announced their own work-related AI integration applications.

After sounding the alarm earlier this year, Google tried to have the first move in this space and as such on March 14th, they released new AI-powered features to Google Workspace.

However, this victory was short-lived as only a few days later, Microsoft revealed its own slew of AI-powered updates to their office applications.

And with that, it seems that both companies are looking to steal market share from 1 another in the field of work-related applications.

The search engine battle is also in full swing.

It's clear that Microsoft is going all in.

While GPT-4 is only available to subscribers, Anyone can access it using Bing.

Microsoft was using it for Bing the whole time.

While we don't have the numbers to see if the strategy is working, Microsoft is certainly spending a lot of resources to disrupt Google's business.

While the main news is for sure the launch of GPT-4, its integration of third-party extensions has caused a stir.

Now it can have access to the internet, book travel, buy groceries, all with direct support from companies that specialize in these fields.

Many fear that this is all moving way too fast.

There's already been a data leak of users' chat history to other users.

If GPT-4 has only been out for about 2 weeks, how do we know what happens when it can read and write to the internet in real time?

I guess we're all going to find out together.

But aside from that, there's been some other important recent events in the AI industry.

Starting off with Google, it released an API for its own language model, Palm.

This is important news, since the development of AI apps by users will be a fundamental component for the competitiveness of Google in the future.

This was followed by the release of their BARD AI to a select group of users.

Unfortunately for Google, the reaction was more or less consistent across users.

BARD is significantly worse than even ChatGPT.

But I think Google could leapfrog, they have the resources to do it.

So we're all going to just have to wait and see.

In other news, Adobe released their AI art generator, Firefly.

NVIDIA announced Foundations, a system that will allow any company to create generative AI given enough data.

Jensen Huang, CEO of NVIDIA, also stated that, quote, every single pixel will be generated soon, not rendered, generated.

This flurry of a breakneck speed innovation in the field of AI has caused none other than Bill Gates to speak up.

Love him or hate him, Bill Gates knows a lot about 1 thing, the world of computing.

On March 21st, 2023, he wrote a personal note on his website, Gates Notes, and declared that we have entered the age of AI.

The penny dropped for him when he saw GPT achieve the equivalent of an A or A plus grade on a college level biology test and that was last year.

Since then things have progressed a whole lot.

I think it's safe to say that we're in a new era.

As I stated before, we're moving from the information age into the knowledge age.

I might keep providing updates like this every now and again if you guys are willing to hear me talk about them.

See all ColdFusion transcripts on Youtube

AI is Evolving Faster Than You Think [GPT-4 and beyond]