The artificial intelligence industry is developing rapidly, including in terms of adoption of the technology among ordinary users. Tools like ChatGPT, Stable Diffusion, and ElevenLabs have enabled millions of people around the world to interact with AI.
Does ChatGPT detect intelligence? Will technology put people out of work? Is it ethical to use AI in war? About this and not only ForkLog spoke with the founder of Pheon, a digital human cloning startup, and in the past, the owner of the outsourcing company Hey Machine Learning, Yura Fitzgerald from Kharkiv.
About ChatGPT
ChatGPT. Literally everyone is talking about him. What do you think about this technology?
I think it's a great technology. She did not appear yesterday, it was a long time coming. The evolution took five years [since the appearance of the first version of GPT]. And now we are at the point where there is ChatGPT, GPT 3.5, and soon the fourth version will be released.
Google is also doing some experiments with its language model. They most likely use TheMDA. One of the successful experiments is the application of a language model in the planning function.
That is, the language model is given a task, for example, "I need to bring a bottle of beer." Then the language model generates an action algorithm: "go to the refrigerator - raise your hand - open the door - take the bottle - close the door - turn around - bring the bottle."
Next, this algorithm is parsed and executed. The results were good.
Can this be called a manifestation of intelligence?
Language models, in particular GPT, are already a good manifestation of intelligence. Five years ago, when AI performed highly specialized tasks, I said: "people will understand that artificial intelligence has already arrived, when algorithms will perform a wider range of tasks, if not better than humans, then at least on par."
ChatGPT and GPT in particular are a huge step in this direction. In fact, this is one model that solves many tasks well, even those that were not intended.
This is such a multitasking thing that will develop towards multimodality, that is, it will combine different algorithms into single systems. More precisely, it is already happening. Have you seen Nothing Forever on Twitch? Where a picture generator and a text model are combined, which continuously create a script and draw images.
If these models have been around for a long time, what is the secret of ChatGPT's success?
In my opinion, a very convenient interface for interaction is a good step. It's like with [protocol] HTTP. It is convenient to watch and debug, and only then was it awarded the Internet, which is familiar to all of us.
OpenAI is notable for the fact that they are, in fact, monopolies. As technology pioneers, they have an excellent team and virtually unlimited resources from Microsoft [thanks deal for $10 billion].
I'm sure it's not all about money. A lot is decided by the Azure service and their servers, to which OpenAI has unlimited access.
Nowadays, it is especially valuable, because there is a common lack of computing power. Amazon and Google don't have enough of them. Even we, as a small startup - we don't need many servers - regularly face problems. They said, here is our money, but they cannot take it because they do not have available resources.
And now it is very difficult for an average startup to compete in the fundamental direction of dialogue models. To train a model from scratch is expensive, very expensive, and such resources are not available at the starting line.
Therefore, ChatGPT is a very strong monopoly.
About synthetic people
If the conversation has already started about your startup, Pheon, tell us more about it .
This is a digital-cloning startup. Technology of cloning people, creating digital copies. Basically, a generated video in which a person looks and sounds exactly like they do in life and says roughly the same thing as the original.
Let's say a clone of Elon Musk. To the question "where do you work" he will answer: "I am the CEO of Tesla Motors, SpaceX, Neuralink, Twitter" and what else he has there.
How did you get this idea?
It all started with a search. At this stage, we went through all possible options for AI products with new and promising technologies. Many options were collected, from which the five best were chosen and presented to profile investors.
The idea with digital people attracted the most interest, so we decided to focus on it.
In addition, it has been talked about for a long time, TV series like "Black Mirror" are being filmed. A customer came to us [at Hey Machine Learning] who wanted to do something similar - to "reanimate" his late grandfather. We explored the possibilities and then things were bad.
Currently, the issue of technological risk does not arise. They already exist in one form or another.
Are synthetic people a promising niche?
It's like GPS, when it stopped being a purely military system and "went to the people." On its basis, services such as Uber, Glovo, Google Maps appeared, and the drone industry developed.
So it is with digital people - a fundamental technology on top of which many different applications can be built. Celebrities can be digitized and linked to educational courses and language learning. For example, learn Spanish with Beyoncé.
It can be a consulting story. Many legal cases, such as starting a company under the laws of the state of Delaware, filing tax returns and preparing reports, are subject to formalization. A digital lawyer can easily cope with such a volume of work that a person cannot handle.
Another example is a motivator coach who helps achieve a goal such as regularly visiting the gym. He will be able to remind you about the need to go to training, monitor the performance of exercises for different parts of the body, argue about something.
And there are many applications that we are not even aware of. This industry is just starting to emerge. We are currently looking for a large market for this story.
How does the digital cloning process work? Let's say I'm a celebrity and I want to create a copy of myself. What do I need to do for this?
We already have a self-onboarding solution where you can create a clone. Now it is in a simple version, where you describe a short biography of a person, important facts about him, character. And you download a video, taken at least from the selfie camera of a smartphone, where he says something.
This data is used by neural networks to generate personalized video responses.
Sounds simple somehow. I remember the case when the Slovak basketball player Luka Doncic was digitized. He was photographed for a long time in the studio from different angles, voice samples were recorded, etc. Does your approach suffer greatly in terms of output quality?
In the beginning, we also had high content requirements. For this, it was necessary to rent a studio, which is not cheap in America. Pay for the work of the cameraman, producer, shoot content for several hours, take care of the perfect light, the position of the head in the frame.
Subsequently, the requirements for content have decreased significantly. To the selfie video for five seconds.
Do you have protection against unfair use? To avoid creating clones of stars and using them to spread toxic content?
Of course. Our neural networks filter content. There is a model that is trained on such datasets to minimize the amount of obscene, rude or toxic content. This is about text requests.
In terms of video, all this can be solved by watermarks, disclaimers in the application itself.
But so far, the generation technology has a number of limitations. Sometimes artifacts can slip through some frames, the resolution of the picture is also limited. That is, by such markers, you can determine whether the content is real.
But it is a matter of time when the technology will be different from the video recorded on the camera in 99% of the cases.
Have you caught attempts to generate something unacceptable? Have you noticed errors in the program itself?
It is not uncommon for a person to create a doppelgänger, but instead of his selfie, he uploads a video with some ducklings. Or records YouTube along with the interface.
Although we have simplified the entry threshold, for a large number of users, it is not an easy process to capture quality content. For a number of technical and psychological reasons.
If someone copies an image of, say, Kim Kardashian without permission. Who is responsible for this?
If you make your own program and generate content, then you own the rights to use the image.
We had a situation with the AppStore when we put together an app for one celebrity. Apple rejected the application and requested documents confirming the rights to use the image.
We sent them the relevant papers and, as a result, the appendix was accepted for publication.
on UGC sites, users are responsible for content. The platform should only moderate. In case of disputed situations, it is necessary to understand whether rights have been violated or not.
About war
The main part of your team was concentrated in Kharkiv. How did the start of a large-scale invasion affect work?
This is a rhetorical question for everyone who was in Ukraine at the beginning of the war. Of course, it affected us negatively. Processes were disrupted, security issues came to the fore. Kharkiv had to be evacuated.
Some people left. And I am a great opponent of remote work: I believe that the team should work together, because the speed of communication and the communication itself decide a lot.
A lot of cool ideas appear in random dialogues. Yes, it's banal to explain something, show, talk about working things - it's faster to do it in a face-to-face format.
Did you manage to save the composition of the team?
We have one person who went to fight. The rest of the teams remained.
Almost a year later, did you manage to return to the previous pace of work?
Yes, the performance returned to the pre-war level. The first few months were difficult.
Speaking of war, how ethical do you think it is to use AI on the battlefield?
Absolutely suppose, why not? Why should natural intelligence be used ethically, but artificial intelligence not? Their difference is only that the natural one was born, and the artificial one was collected.
And if robots can fight each other, people will stop suffering. But such a utopia is unrealistic.
About general Artificial Intelligence
Now AI has become a mass phenomenon, although until recently it was more interesting to geeks and the target community. What has changed in recent years?
5 years ago, I gave a presentation on AI at the Kharkiv National University of Radio Electronics. However, it has not lost its relevance since then. Some new developments have appeared, the same Diffusion or ChatGPT.
The predecessor of this was iron, the availability of computing power. The community is growing organically, more specialists, "stars" of the industry are appearing. Accordingly, this community is doing more research, more good new tools.
There is more data, it has become easier to store and cheaper to process. That is, the prerequisite is the economy.
In your opinion, there was no turning point, and everything developed in its own way?
And what is a turning point?
Something happened that divided it into "before" and "after".
And what is "before" and what is "after"?
For example, when DALL-E came out and it turned out that images can be generated by text request.
DALL-E is far from the first, there were many other solutions. They were worse in quality, generating more "LSD" pictures.
Of course DALL-E, GPT are milestones. To some extent, these are all turning points. But for me it is one natural continuous evolution.
Five years ago, we discussed chatbots and said that this technology was already fading into the background. Could you have guessed then that in 2023 a chatbot would be so popular and in demand?
At that time, I did not think that a chatbot is a convenient interface for artificial intelligence.
But even now there is little difference between a person communicating with another person or a bot. Even a very smart robot.
There is already a bigger barrier in psychology. Friendship is not just correspondence. This is a long process of building relationships, having shared moments, memories, hobbies.
Communication in the format of correspondence is one of the components of friendship. And chatbots do not replace it.
But even in their current form, they can create a certain affection. This is especially noticeable among lonely people who are looking for support.
But all this will evolve, will become overgrown with psychological factors. In this way, robots will be perceived more alive.
And if not as communication, but service. If a robot served you in a restaurant, would you feel comfortable?
Of course, there is a need for human communication, but at the same time there are no complaints about bots. I recently went to a cafe where cars are prepared. There is only one person working there who installs capsules with pasta and sauces for these robots. They mix it all, heat it up, cook it, and you watch the process and after 15 minutes you have a ready order.
The food tastes no different from the chef's dishes. It's certainly not Michelin, rather closer to homemade pasta. But this is ordinary, edible food.
Fine cuisine can also come to this in the process of natural evolution.
Yes, it's nice when a waiter comes and takes care of the guest's comfort. Machines cannot replace them yet, because there are no such technologies. If a robot comes instead of a person, that's only great.
Which II sectors do you consider to be the most promising?
But in general, II is a very promising industry. As Andrew Yin said, AI is the new electricity.
What will develop? From what is currently in trend, actually, language models. They will become the foundation for II. If we talk about the vector of development - multimodality.
New interfaces will be added on top of the models, in addition to the text ones. It can be decision-making systems for robots, video script generators, military technology.
How much will automation affect the labor market? Will people be out of work?
People will not be left without work. And work can be invented from any activity. You can retrain for another profession.
Some areas will begin to transform. From the obvious - copywriting.
Although algorithms can create large volumes of images, they will not replace designers. They are transforming the craft.
With the same GPT - the request must be correctly formed. So such work may appear - prompt engineering. A specialist who will form the right task for AI.
At the moment, man has a great advantage. You can ask him when something went wrong. You can't ask a chatbot. This is another reason why people will not be out of work anytime soon.
I recently came across a picture on the Internet where a cleaner is cleaning the floors in a shop with robotic vacuum cleaners. I always remember her when they say that people will be out of work.
What about common AI, how soon will it come? And do we need it at all?
He has already appeared. The same GPT.
One can speculate on the topic of "what is common AI" because there is no consensus. In my understanding, this is one system, one brain, an architecture that can solve a wide range of tasks.
ChatGPT is like that. She solves a wide range of tasks that she has not even studied. And this ability will become stronger and stronger.
In theory, ChatGPT would be able to pass the Turing test, and an ordinary person would not guess who they are talking to?
Even here, people who communicate with a clone ask: Are you a living person? Let's call and talk." And they drop the phone number in the chat.
People have a grain of doubt. Therefore, the Turing test has been passed at this stage.
Five years ago, AGI was much dumber. Even now, he is far from human. But some time will pass and AI will catch up with people. This is great, it will advance development.
Currently, researchers and mathematicians are very limited in their cognitive abilities. We have a barrier: the size of the brain, the number of neurons. And we cannot overcome it.
And the advanced intelligence will have an advantage, it will be able to find some deeper patterns that we do not even suspect. To invent new meanings, inaccessible to the human mind.
AGI will be able to create some new devices, generate new concepts, and everyone will be fine.
Unless, of course, robots destroy us all. But the good news is that it is unlikely to happen in our lifetime.
To always be aware of the most important things, read us at Telegram
