"The Internet is going to die, and only we can save it"

Original source: Silicon Star People

Image source: Generated by Unbounded AI‌

**How long has it been since you heard a startup company say that it is going to save the Internet? **

Today’s entrepreneurs seem to be either silently making money, or discussing using AI to save (or destroy) all mankind. In the Internet we live in every day, it seems that no one cares about his life and death.

And to Chri, it looked like it was going to die.

When I heard him say this to me in the bright sunshine of San Jose, it seemed a bit sensational. But when you stay with this German with lion-like hair for a long time, his logical and German-style English will easily make you think that this is indeed a big problem.

Chri Besenbruch is the co-founder and CEO of Deep Render. Simply put, what his company is doing is "using AI to compress the size of video files to very small sizes."

He claims the algorithm is up to five times "better" than the industry standard codec HVEC, though he didn't specify by what metric, and can run in real time on modern chipsets from Qualcomm, Apple and Nvidia. A "codec" refers to software and sometimes hardware that can encode and decode data, usually for audio or video data.

Some people who are familiar with Silicon Valley may be confused when they hear this, thinking that I am retelling the script of "Silicon Valley". In this American drama, which is described as a perfect display of the entrepreneurial ecosystem of Silicon Valley, all the farce is because the protagonist has developed an amazing compression algorithm.

"Yes, many people have told me this. We do look a lot like pied pipers." This person with a complex background who grew up in Germany, studied mathematics as an undergraduate, went to Switzerland to study art school, and then went to the UK to study computer science, and " The protagonist Richard in "Silicon Valley" also has a strong nerd atmosphere, but unlike the ridiculous compression algorithm breakthrough process in the TV series, and the lack of business common sense in the TV series, Richard is often hesitant. His entrepreneurship is as rigorous as solving mathematical problems, and he is very serious about the company. The business model is also clearly thought out.

In his opinion, the Internet is heading for destruction, and that's the biggest problem. He wanted to solve this biggest problem. After an analysis, they believed that the root cause was that the old technology did not adapt to the new world:

*After video became popular and became the main form of communication on the Internet, we have actually entered a new world. A world where the scale of data is getting more and more terrifying, and there is no going back. *

*Old technologies are no longer suitable for this new world. Continuing to make minor innovations and improvements along the same old path will only lead to a dead end, and ultimately the collapse of the Internet. *

"We're going to save the Internet by inventing new compression algorithms. And what makes this possible is AI."

Along these lines, Chris founded Deep Render in 2018 with his fellow graduate students in the UK—he was not one of those startups chasing the ChatGPT wave.

"Basically, we have little to do with LLM. It may be related to the diffusion model, but it is more about potential possibilities in the future." He told me. "We started from first principles and wanted to understand what AI is suitable for and the essence of compression. Ultimately, we believed that a method that relies entirely on AI must be a new technical route."

However, LLM's "restlessness" actually helped its company. "We were actually gambling at the time, because to achieve our goals, in addition to my own algorithms, we also needed significant progress in software and hardware. We made a prediction at the time, assuming that these basic technologies could all make leaps and bounds. And today all of this has happened Now." He said to me, his tone full of excitement.

"Maybe we're just lucky."

**Looks very humble, but what he did not say is that if his predictions come true, his company may be one of the most important Internet technology companies in mankind in the next few decades. **Their technology will be provided to major companies in a licensed form. While solving the Internet bandwidth problem, they will also become the type of company that collects utility bills and taxes.

In fact, everything has started to change.

Since this year, this company that has been ignored has entered the vision of Silicon Valley and well-known investment institutions around the world. After completing a financing in March, Chris began to shift his focus to Silicon Valley, where there is money and people who are interested in them. Big customers.

According to his statement, their current small team is already overwhelmed when facing interest and demand from "big companies that cannot be named but they are very important."

And he revealed to me that if everything goes well now, its technology will be used in a very well-known application service in the second quarter of 2024.

"Tens of millions of users will have access to it."

The Chinese market, which has huge demand for video and data compression, has also begun to appear in various forms in front of this CEO who has not been to China.

"We have many investor friends who have strong ties to China and are helping some Chinese companies contact us. We are already in contact with several of China's largest Internet companies," he said. These demands came so suddenly and violently that he spent half a day asking me basic questions about "what is it like to work with Chinese Internet companies?"

"It seems that if you want to save the Internet, you can't do it without the Chinese market." I told him half-jokingly after telling this technology entrepreneur who spent most of his life in Europe stories about the East as much as possible.

"You're right." He replied with a smile. But before thinking about the Chinese market, he first had to conquer Silicon Valley.

At the end of the American drama "Silicon Valley", Richard's company has been valued at US$8 billion after six years of ups and downs. However, just a few days before the product was officially released, they discovered that the compression algorithm was different from the one they developed. The AI that improved the network improved each other and turned into a "monster" that they couldn't understand but could break through all systems. After weighing it, they finally decided to "save the world" - by using a gorgeous fool to make everyone stop. I have thoughts on this technical route.

This is not a good ending for a startup. When I raised this topic to Chris, who will be in his sixth year of starting a business next year, he thought for a while, and then just smiled and said:

"Hopefully our ending will be better."

The following is a transcript of the conversation

Q: Briefly introduce yourself and your company.

A: Of course. I'm Chris Besenbruch, CEO and co-founder of Deep Render. What Deep Render is doing is basically compressing the size of video files and so on to a very, very small size. Why we need to do this is because the data on the Internet is growing exponentially, and 90% of the data was generated in the past two years. They need to travel through fiber optic networks around the world, which is a very expensive infrastructure at this scale. The world's entire fiber optic infrastructure costs $5 trillion. That's not comfortable because if data doubles every two years, what does that mean for infrastructure, it needs to double as well. But this is a trillion-level doubling. This is impossible. This is the reason why I do this.

The idea is that if we can't make the pipeline bigger and faster, we'll make the data in the pipeline smaller. So it pointed to compression. This is our foundation and vision.

Q: So you first deduced one of the core problems faced by the Internet today, figured out its solution, and then used it to determine your entrepreneurial direction.

A: Yes, that’s it. Because without a vastly better compression method, the entire Internet will soon collapse. And I love the Internet and I want to keep it going.

Q: So what is the difference between your technology, because compression is not a new thing, it has been around for a long time, and people have even gotten used to it and forgotten it exists. Maybe you could describe your technology in terms that a fool can understand.

A: Hahaha, I'll try my best. We need to look at the entire industry from the past 60 years. When traditional compression technology came out, it was a huge breakthrough. Everything was based on DCT* (DCT stands for Discrete Cosine Transform, which divides the image into components composed of different frequencies. Small blocks. During the quantization process, high-frequency components are discarded, and the remaining low-frequency components are saved and used for subsequent image reconstruction. Editor's note)*, this method that makes today's video compression technology possible was invented in the 1970s and 1980s or so, and it dominated the industry thereafter. **But since then, this technology has only improved a little bit every 10 years, and it has always been the same technical idea. We cannot expect a technology to achieve a leap in effect just by iterating the same idea hundreds of times. **It is great technology, but its innovation cycle has ended or is ending.

So we need something new. And this “new” is AI. Because AI can handle images and videos very well, this is not an unthinkable route. So AI technology and compression technology began to be combined. There were two waves after that. The first one happened in 2017, when super-resolution appeared and Magic Pony invented it. Their idea was to keep the traditional compression method in the middle of the pipeline and add AI at the front and rear.

**Deep Render does not believe this is the correct approach. Because what remains in it is the traditional compression technology that we believe has completely collapsed. It doesn't change much from what it came before, and it's still hardly useful. **

We are the second wave, completely abandoning traditional compression technology and just using neural networks. This also means re-inventing compression. Re-invented compression around machine learning technology has the highest effect. The neural network gets the data and compresses the file. Send it to the Internet, and the network receives the compressed file and returns the video to you. This is an AI-only solution. We're leading the charge, and we're actually the only company doing it.

Q: So basically you are using AI to replace the part that traditional technology is responsible for, so why can AI do it better?

A: Two reasons. Video or video compression is essentially a trade-off between file size and image quality. You either have very large files and great looking video, or very small files with crappy definition. So good compression is about good trade-offs. In terms of file size, the key is redundancy. If you can predict what the next pixel will be, you don't need to send redundant data, which saves file size. This is where AI can make better predictions. If the prediction is good, redundancy can be removed and less data can be received. And AI is fundamentally stronger because it reacts to data.

The second is about video quality. Any compression algorithm will introduce errors in the video output, this is the definition of lossy compression, accepting some errors for the sake of smaller file sizes. But what really matters to humans is the distribution of these errors. We are more sensitive to some and may be fine with others. **So our AI imitates the human visual system so that these errors are hidden from humans. **

A great example is that people hate when lines become blurred. People feel unhappy if a sharp line becomes blurred because this is biologically determined. For example, when a tiger starts moving from behind a rock, we need to be able to see it, and we will stare at it. The accuracy of the color is not guaranteed, and no one really notices if there is a loss. So you can give away some color accuracy without blurring the edges. People prefer this video quality.

Our idea is always to start from first principles, what is reasonable, what are the mathematical principles behind it, and which tools can be used to solve real problems. AI happens to be the most suitable tool for this.

Q: It’s very interesting, so this is the most basic idea, and the next step is execution. As a five-year-old company, do today’s AI changes have any impact on you?

A: At the software level, we have actually always been rooted in the research circle, and our software products come from the research circle. Today's changes in AI have little impact on us. I have been involved in these research circles since 2015, when the AI revolution began.

On the contrary, the hardware level is actually fresher. We have great software-level compression technology but in the past it only worked in the cloud, so it was a good research topic, not a product. **Only when it enters every terminal so that millions of people can use it on their own devices, it becomes a product. **

Therefore, we actually made a bet on future technological changes. In addition to betting that AI would be effective, we also bet that the hardware we need would appear, specific AI acceleration hardware, such as NPU, and various hardware chips from Apple, Qualcomm, and Google. will be made. **We made a bet in 2018, and today we have it in 2023. **

**Either we had a good view or we were just lucky. **

Q: And these major hardware manufacturers are also interested in you.

A: **Yes, because they have these hardware, they started looking for killer apps. We went to them and said, hey, everyone watches the video, this is the killer app you want. This time is also very suitable. **

Q: In addition to algorithms, when we chatted before, you mentioned that data is also a threshold, but obviously large companies have more data, will this be a problem?

A: Very interesting question. (Contemplation) I think with AI compression, it will ultimately be found that the algorithm is more important. Because I don't think the field as a whole is mature enough to require data quality decisions.

Q: So that’s the second phase question.

A: Yes, it is still in the first stage. We cannot say that we have a perfect algorithm. Our algorithm is making great progress every month and changing very, very fast. When the algorithm is strong, it is the data. But we are also accumulating, because data has diminishing returns. Now we have one to two million video sequence data, which we obtain from open source and trusted channels or purchase from video websites. At this point, they are quite open, but everyone Seeing that ChatGPT is changing the degree of openness, we are lucky to have purchased this data earlier. This is also an advantage, but it is not a truly decisive advantage at this stage.

Q: Do LLM and diffusion models bring anything new to your algorithm?

A: There are relatively few. There are some ideas that can be used for reference. The problem is that these models seem to be able to enter our process and bring us improvements. For example, Stable Diffusion may help me improve compression capabilities, but there are still trade-offs - for example SD is processed once every 10 seconds, but we need to process it 300 times in 10 seconds. How to choose between real-time and effects is a problem. They have the potential to make an impact in the future.

Q: If we look at it from a competitive perspective, where do your current challenges come from?

A: The algorithm, and then the people behind it. We study many basic algorithms ourselves because there are not many papers to read. The most critical challenge is to form a good team,** because there are many people in the market studying AI, not so many people studying information theory, and there are almost zero people studying both at the same time. So we need to team up and train them, which takes a long time** because they don't have much to do with each other. So you need to train a group of people first, and then let them lead the people below. This is a pyramid structure.

Q: Have you already built this pyramid?

A: We have more than 30 AI research superstars. But we spent a long, long time on this.

Q: So the only 30 people in the world are here with you.

A: I would say that. There are also some labs looking at these topics, and they have good people, but by far Deep Render is the largest organization in this field in terms of scale.

Q: I saw you mentioned using the recent new financing to expand your business and presence in Silicon Valley.

A: This is a bit strange, we are a British company, but the companies that use the most Internet resources are in China and the United States. Moreover, Europe is indeed not a friendly place for technological development. We have a lot of demand from the United States, so it is natural to enter the United States, and the same is true for China.

Q: Have any Chinese customers contacted you?

A: I can’t say the name, I signed a long, long agreement so I have to be careful haha. But when I contacted these companies, I found that China indeed has higher bandwidth requirements than the United States.

Q: You are already in contact with customers. If we judge by the PMF that people like to discuss, what stage are you at now?

A: This is always hard to define. I would consider us to have reached PMF because we have more demand than we can handle. We are really rejecting many large companies because our own bandwidth resources are not enough. Many companies have booked us, and we can only handle the needs of 4 to 5 large companies at the same time. Once the technology is more mature, commercialization will be easier. We're past the proof-of-concept stage. If all goes well, you can see our technology being applied to a very well-known Internet service in the second quarter of 2024, and tens of millions of users will use it by then.

Q: Faced with crazy demand, limited manpower, and your own technical status, it is your responsibility as a CEO to make balanced choices.

Q: Haha, this is the challenge of my job. I'm going to try to make a decision. Moving too fast always consumes resources, and recruitment will bring challenges. Our employees generally need 4 months of training before they can be productive. So you are right, this is an optimization problem. Of course, I may be willing to pay for someone who uses AI to develop a tool.

Q: Haha, maybe LLM can participate here.

A: Yes hahaha.

Q: Everything seems to be going well. What is the business model you have designed for this business?

A: We are currently in the form of B2B. We provide product authorization to customers. Customers can save a lot of money, and part of it can be given to us. This is a business model based on licensing. If you look at the data here, it's actually crazy. By 2030, the cost of transmitting content worldwide will reach $125 billion. **If you want to destroy Netflix, use 4K and watch Netflix 24 hours a day for a month. **

Q: It's best not to do this.

A: Hahaha, but if you can reduce the file size by, say, 90%, then based on $125 billion, companies can save a lot of money.

Q: This is your chance to make money.

A: This is a business where everyone benefits. I won, the company won, the users won, and the big Internet companies won. No one loses anything.

Q: What about saving the Internet? Have you ever imagined what the future Internet your company will help build will look like?

Q: Of course, I dream about it every day hahaha. **Our vision is to turn bandwidth into an unlimited commodity, so that everyone no longer has to worry about network speed. For everyone, they can enjoy extremely high-quality videos at home, and for companies, they can get very cheap or even free Internet resources. Information could flow, even data that would be considered heavy today could flow freely. This is the future the Internet was supposed to have. **

Q: I’m sure many people have asked you about this last question. Talking about it made me feel even more that the story in the American TV series “Silicon Valley” seems to be the same as yours.

A: Right, right. This is pretty much my favorite show. But what’s actually interesting is that my co-founders and I didn’t know about Deep Render until a year and a half after we founded it, because HBO wasn’t that popular in the UK. But what’s interesting is that our business model and the stages we went through are almost the same as the story in this play. Especially since we saw it later, we both said, wait a minute, this is just like us.

Q: You also know the ending of this drama.

A:... Hahaha, I hope our ending is better.