The ‘AI Company’ Backed in Unison by Tech Giants That Astonished the World

There is a technology that had been developing beneath the surface and has now begun to break above the waterline.

At least the terms themselves are extremely familiar to us: artificial intelligence (AI) and machine learning (ML).

We use AI to build chatbots, detect plagiarism, enhance security, and unlock devices through facial recognition.

And with ML, we get content recommendations, identify consumer purchasing patterns, and detect fraudulent payments.

The applications of this technology are so broad that, even without a detailed explanation, we can easily infer just how enormous the added value it can generate is.

From the very common approach of using AI and ML to build products people care about and bundling ads with them, to leveraging these technologies to deliver better outcomes, we can get a glimpse of how powerful they truly are.

Reference:

DALL·E 2 - OpenAI (text-to-image AI)

Portrait AI (photo-to-artwork AI)

ChatGPT - OpenAI (AI chatbot)

One example is Amazon’s acquisition of robot vacuum maker iRobot in August.

Why did Amazon acquire a robot vacuum manufacturer? To build robot vacuums and sell them through Amazon?

It was because Amazon needed it for its AI assistant, ALEXA.

ALEXA is an AI smart speaker first released in November 2014. Through a wide range of voice commands, ALEXA converses with users and provides information such as weather and traffic updates.

For now, many say this division is a drag on Amazon’s earnings, but the combination of ALEXA and robot vacuums is clearly something worth watching.

When you think about combining Alexa with a robot vacuum, this is probably the kind of image that comes to mind.

But among all the functions of a robot vacuum, what Alexa truly needs is not the cleaning itself, but the ability to recognize the layout of the room and the objects in it.

With this capability, Alexa can figure out an extraordinary range of things: the characteristics of the people living in that home, what kind of daily routines they need, what their hobbies are, and much more.

If dog toys are scattered all over the living room, chances are that household has a pet dog, and Alexa can update this information and start serving that person ads for dog products.

There are obvious concerns about privacy, but the ability to deliver ads that are genuinely relevant and thereby raise conversion rates is undeniably valuable.

Another example is art competitions that use artificial intelligence.

Jason Allen took first place in the digital art category at the Colorado State Fair fine arts competition by using AI.

It would certainly have been problematic if he had simply asked the AI to paint a picture and then gone off to watch Netflix while the AI did all the work.

In response to the controversy over his use of AI,

he said, “I spent a long time researching a special set of prompts to create this piece. Using them, I generated close to 100 images, then spent weeks on fine-tuning and curation. After that, I upscaled them with Gigapixel AI and printed the best three works on canvas.”

He showed that he did not merely rely on AI, but actively used it as a tool, and the organizers sided with him, noting that he had disclosed his use of AI before entering the competition.

The point is not that AI “replaced” something specific, but that he used it to create something better than what existed before.

The world is advancing at a breathtaking pace, and companies across industries are adopting increasingly sophisticated technologies.

But we should recognize that this astonishing technology, which has only just begun to surface, is still in its infancy.

Greg Brockman, OpenAI Co-founder & CTO

There is a product that the CTO of OpenAI, the company behind the image AI and chatbot that recently stunned social media, has praised highly: Ray.

This can be interpreted as meaning that without Ray, it would have taken much longer to reach today’s level of technological sophistication.

Ray is a framework created by Anyscale for building machine learning infrastructure,

and major companies such as IBM, Meta, Uber, and Riot Games are currently using Ray.

So what exactly makes this framework so powerful?

Anyscale, which recently completed the Ray 2.0 upgrade and raised an additional 99 million dollars in Series C funding, describes Ray’s strengths as follows.

Distributed computing means making multiple computers work together to solve a single problem.

One example is storing the countless photos and videos uploaded to Facebook across many different computers.

It offers benefits such as scalability, which lets you add another computer as the workload increases; availability, which keeps the system running even if one computer fails; and efficiency, which optimizes resource usage.

However, because distributed computing spans hardware, middleware, and software and requires multiple systems to operate together, it has the drawback of making system development difficult.

Ray addresses this by providing an API that makes the process flexible and simple.

It supports mapping one value to another and can simplify a variety of parallel patterns.

A workload refers to a collection of resources and code.

AI and ML involve an enormous number of workloads, and managing them becomes increasingly difficult in proportion to that number.

This huge volume of workloads can also put pressure on data centers, for example through power and cooling issues.

So there are times when you need to scale workloads down, and conversely, times when higher specs are required and you need to scale them up.

However, during the process of adjusting workload scale, cache misses can frequently occur,

When a cache miss occurs, the server has to send a request to the database for the missing data, which can create additional load.

Ray allows you to adjust the scale of model training and can automatically scale up or scale down based on resource requirements.

To complete an AI system, you need many components beyond just training.

You need everything from basic engineering to hyperparameter tuning and data processing, and each of these used to require a different API, which was a major inconvenience.

In particular, in the data processing stage, conventional data tools typically target structured data, whereas AI relies heavily on unstructured data such as images and video, which has been a significant pain point.

On top of that, the exponential growth in workloads is also driving demand for more advanced hardware.

If building a single application requires A, B, and C, you need a separate system for each of A, B, and C. And if you later find that you also need D, you have to add system D on top of that. Just hearing this is enough to know how difficult it would be to manage.

Ray has built a platform that, through library integration, eliminates the need to constantly switch between different frameworks,

It offers an integrated toolkit and allows you to specify a variety of hardware at instantiation, creating an environment that is highly favorable from a developer’s perspective.

According to Globe Newswire, over the past few years the computational demand for ML has increased by a factor of 10 to 35 every 18 months.

Many AI projects are clearly solving real problems, but this surge in computational demand and the complexity of engineering often collide head-on.

Roughly a quarter of companies running AI projects are seeing project failure rates approaching 50%, and 78% of AI and ML projects stall before they ever reach deployment.

In addition, 81% of teams reported that training AI models turned out to be more difficult than they had expected.

In fact, many AI projects at large corporations have either caused serious problems or ended up being shut down.

Amazon’s AI recruiting system was scrapped in 2017 because it failed to learn criteria that could remove gender bias in technical roles.

Amazon Rekognition, the company’s demographic analysis program, struggled to identify the gender of people with darker skin tones and even made errors such as matching photos of members of Congress with images of criminals.

Google’s AI for diagnosing diabetic retinopathy had difficulty making accurate diagnoses when images were incomplete.

IBM’s Watson, a natural-language question-answering system, at times recommended dangerous treatments to certain cancer patients.

Microsoft’s chatbot Tay was designed to casually banter with people, but after being trained on biased data it quickly began posting inflammatory and abusive content on Twitter.

Even when each project requires years of development and costs that scale with that timeline, failure cases remain common, and an even larger number of projects are simply stuck in limbo.

With the advent of Ray, however, engineering has become far more intuitive, and there is now little need to rewrite code when moving to other applications.

It has become easier, faster, and more convenient to process unstructured data.

Users can now focus solely on training machines without worrying about scalability.

AI projects can still fail. But as trial and error is repeated, the failure rate will drop significantly.

Ray can dramatically reduce the cost and time companies spend going through trial and error. With those savings, companies can take on more challenges, and the market will continue to grow.

The pace of technological progress is accelerating by the day.

If you compare the time gaps between major advances—from paper and ink to landline phones, from landlines to mobile phones, and from mobile phones to smartphones—you can infer that the speed of technological progress is not an upward-sloping straight line, but an upward-sloping curve.

One of the factors that adds another notch of acceleration is, of course, improving work efficiency. If credit, or today’s corporate structures, did not exist, technological progress might well have followed a straight line.

And Ray is one of those factors that improve work efficiency. In other words, it is a factor that adds acceleration to progress.

Nick Washburn, Managing Director at Intel Capital

This may read like an IR deck for Anyscale, but the point I want to make boils down to just one thing.

Now that something truly useful has emerged, the pace of progress in AI and ML is clearly something worth looking forward to.

Questions like when the market will grow to a certain size or exactly which company to invest in may not be particularly meaningful at this stage,

but there will be a clear difference in investment outcomes between those who have a rough sense of the landscape and stay engaged, and those who do not. Especially when it comes to technological progress.

Three-line summary:

1. Ray by Anyscale is a framework used to build machine learning infrastructure, dramatically reducing the complexity of traditional engineering workflows.

2. This allows teams to significantly cut costs from trial-and-error and greatly accelerate development speed.

3. As Ray continues to evolve, the growth of the AI and ML markets is expected to accelerate exponentially.