General AI Has Arrived: Interview with Pheon CEO Jura Fitzgerald

Artificial intelligence is rapidly evolving, including in the adoption of technology among everyday users. Tools like ChatGPT, Stable Diffusion and ElevenLabs have allowed millions around the world to interact with AI.

Does ChatGPT display intelligence? Will the technology leave people jobless? Is it ethical to use AI in war? ForkLog spoke with Jura Fitzgerald, founder of the AI startup Pheon, which focuses on digital cloning of humans, and formerly owner of the outsourcing company Hey Machine Learning.

About ChatGPT

Bogdan: ChatGPT. Literally everyone is talking about it. What do you think about the technology?

I think it’s a wonderful technology. It didn’t appear yesterday; the development has been under way for years. The evolution took about 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 arrive.

Google is also conducting experiments with its language model. They are probably using LaMDA. One of the successful experiments is applying the language model in a planning function.

That is, you input a task to the language model, for example, “I need to bring a bottle of beer.” Then the language model generates an action algorithm: “drive to the fridge — raise your arm — open the door — take the bottle — close the door — turn around — deliver the bottle”.

Then this algorithm is parsed and executed. The results were good.

Bogdan: can this be called intelligence?

Language models, in particular GPT, are already a good manifestation of intelligence. Five years ago, when AI performed narrowly specialized tasks, I used to say: “people will understand that artificial intelligence has arrived when algorithms can perform a broader range of tasks if not better than humans, then at least on par.”

ChatGPT and GPT in particular — a huge step in this direction. Essentially, it’s one model that solves many tasks well, even those that were not anticipated.

It’s such a multitasking tool that will develop toward multimodality, i.e., it will combine various algorithms into unified systems. More precisely, this is already happening. Maybe you’ve seen the Nothing Forever series on Twitch? Where they combined an image generator and a text model that continuously create a script and render images.

Bogdan: if these models have existed for a long time, what’s the secret of ChatGPT’s success?

In my view, the key move is having a very convenient interface for interaction. It’s like the HTTP protocol. It’s easy to view and debug, and only then did the familiar Internet grow on top of it.

OpenAI is notable in that they are effectively monopolists. As pioneers of the technology, they have an excellent team and almost unlimited resources from Microsoft [thanks to the $10 billion deal].

I’m sure not all of the money is in cash. A lot depends on Azure services and their servers, to which OpenAI has unlimited access.

Right now this is especially valuable because there simply aren’t enough compute resources. Amazon and Google don’t have them in sufficient quantity. Even we as a small startup — we don’t need many servers — regularly face problems. Here is our money, but they cannot take it because no resources are available.

And it is very hard for a typical startup to compete in the fundamental area of dialog models. Training a model from scratch is expensive, very expensive, and such resources are not available to ordinary startups.

Therefore ChatGPT is a very strong monopoly.

On Synthetic Humans

Bogdan: now that we’re talking about your startup, Pheon, tell us more about it.

It is a digital-cloning startup. The technology of cloning humans, creating their digital copies. Essentially, a generated video in which the person looks and sounds the same as in life and says roughly the same thing as the original.

For example, a clone of Elon Musk. When asked “where do you work?” he would answer: “I’m CEO of Tesla Motors, SpaceX, Neuralink, Twitter,” and what else he has.

Bogdan: how did you come up with this idea?

It started with exploration. We evaluated many possible AI products with new and promising technologies. We gathered many options, chose the five best, and presented them to profile investors.

The idea of digital humans drew the most interest, so we decided to focus on it.

Also, this has been talked about for a long time; shows like Black Mirror have depicted it. A client at Hey Machine Learning came to us who wanted something similar—“bring my dead grandfather to life.” We explored possibilities and back then things were not good.

Today, the question of technological risk does not stand. In some form, the necessary developments already exist.

Marina: synthetic humans — a promising niche?

It’s like GPS, which stopped being purely military and “went to the people.” Based on it services such as Uber, Glovo, Google Maps emerged, and the drone industry developed.

So with digital humans — a fundamental technology on which many applications can be built. You could digitize celebrities and tie them to educational courses or language learning. For example, learn Spanish with Beyoncé.

It could be a consulting story. Many legal tasks—such as opening a company under Delaware law, filing tax returns, and drafting reports—are amenable to formalization. With such volume of work, a digital lawyer could easily handle it.

Another example — a coach-motivator helping to achieve goals like going to the gym regularly. It could remind you to train, monitor exercises for different body parts, argue about things.

And there are many applications we haven’t yet imagined. This industry is just starting to emerge. We are currently seeking a large market for this concept.

Bogdan: how does the process of digital cloning work? Suppose I’m a celebrity and I want to create my copy. What do I need to do?

We already have a self-onboarding solution where you can create a clone. Now it’s the simple version, where you describe a short biography of the person, important facts about them, character. And you upload a video, shot even with a selfie camera, in which they say something.

These data are used by neural networks to generate personalized video responses.

Bogdan: it sounds simple. I remember a case when we digitized the Slovak basketball player Luka Dončić. He was photographed in a studio from different angles, voice samples were recorded, and so on. Does your approach suffer in terms of output quality?

In the beginning we also had high content requirements. To do this you would need to rent a studio, which in America is not cheap. Pay for the operator, producer, several hours of shooting content, ensuring perfect lighting and head position in frame.

Over time the content requirements dropped. To a five-second selfie video.

Marina: do you have protection against misuse? So that clones of stars are not created and toxic content is not spread using them?

Of course. Our neural networks filter content. There is a model trained on such datasets to minimize the amount of obscene, crude, or toxic content. This is for text queries.

In terms of video this can be addressed with watermarks, disclaimers in the app itself.

But for now the generation technology has a number of limitations. Sometimes artifacts may slip into some frames, image resolution is also limited. In other words, by such markers you can determine whether the content is genuine.

But that is a matter of time, when the technology will be indistinguishable from video recorded on a camera in 99% of cases.

Bogdan: have you recorded attempts to generate something inappropriate? Or noticed errors in the app itself?

It is not uncommon for someone to come to create a double, but instead of their own selfie they upload a video with ducklings or something. Or records YouTube along with the interface.

Although we have simplified the entry threshold, for a large number of users capturing quality content is not an easy process, for a variety of technical and psychological reasons.

Bogdan: if someone copies a person’s image, say Kim Kardashian, without permission. Who bears responsibility for this?

If you build your own app and generate content, you should own the rights to use the image.

We had a situation with the App Store when we built an app for a single celebrity. Apple rejected the submission and asked for documents proving rights to use the image.

We sent the appropriate documents and ultimately the app was allowed for publication.

On UGC platforms, responsibility for the content lies with the users. The platform should merely moderate. In case of disputes, one must determine whether rights were violated or not.

On War

Bogdan: the core of your team was based in Kharkiv. How did the start of a full-scale invasion affect the work?

This is a rhetorical question for everyone who has been in Ukraine since the start of the war. Of course, it affected us negatively. Processes were disrupted; safety concerns came to the foreground. We had to evacuate from Kharkiv.

A portion of the team dispersed. And I am a strong opponent of remote work: I believe the team should work together, because speed of communication and direct interaction matter a lot.

Many great ideas surface in casual conversations. And explaining, showing, and discussing work is faster in a face-to-face setting.

Bogdan: were you able to preserve the team?

One person went to fight. The rest of the team stayed.

Marina: after nearly a year were you able to return to the prewar pace?

Yes, performance has returned to prewar levels. It was tough for the first couple of months.

Marina: speaking of the war, how ethical do you think it is to use AI on the battlefield?

Absolutely permissible, why not? Why is natural intelligence OK to use and artificial not? The difference is that natural intelligence was born, while artificial was assembled.

And if robots can fight each other, humans will stop suffering. But that’s a utopia, not very realistic.

On General AI

Bogdan: AI has become a mass phenomenon, though until recently it was more of interest to geeks and a niche audience. What has changed in the past years?

About five years ago I gave a presentation about AI at Kharkiv National University of Radio Electronics. It remains relevant. New developments emerged, such as diffusion models or ChatGPT.

The hardware that preceded all this was the availability of computing power. The community is growing organically, more specialists appear, the “stars” of the field. Accordingly, this community drives more research and better new tools.

There is more data, it has become easier to store and cheaper to process. In other words, the premise is economics.

Marina: do you think there wasn’t a tipping point, and everything just evolved?

What is a tipping point?

Marina: for example, when DALL-E came out and it turned out that images could be generated from a text query.

DALL-E was not the first; there were many other solutions. They were poorer in quality and produced more “LSD-like” images.

Of course, DALL-E and GPT are milestones. In a sense, they’re turning points. But for me this is one natural, continuous evolution.

Marina: five years ago we discussed chatbots and said this technology would fade into the background. Could you have imagined that by 2023 chatbots would be so popular and in demand?

I did not think then that a chatbot would be a successful interface to artificial intelligence.

Yet there is still a difference between interacting with another human and with a bot. Even a very smart bot.

Here the barrier is more psychological. Friendship is not just texting. It is a long process of building relationships, shared moments, memories, and interests.

Text-based communication is only one component of friendship. And chatbots do not replace it.

But even in their current form they can create an attachment. This is especially noticeable among lonely people seeking support.

But all this will evolve, accruing psychological factors. In this sense, bots will be perceived as more humanlike.

Marina: and if not for conversation, but for service. If a robot served you in a restaurant, would you feel comfortable?

There is a need for human interaction, but there is no objection to bots. I recently visited a cafe where machines do the cooking. There is only one person who installs capsules with pasta and sauces into these robots. They mix, heat, and prepare the food, and you observe the process; in about 15 minutes you have a ready meal.

The food tastes no different from dishes prepared by a chef. It isn’t Michelin-grade — more like homemade pasta. But it’s ordinary, edible food.

Fine dining could also arrive at this through natural evolution.

Yes, it’s pleasant when a waiter takes care of the guest’s comfort. Machines cannot yet replace them because such technologies don’t exist. If a robot were to replace a human, that would be fine.

Bogdan: which AI sectors do you consider the most promising?

Yes, AI is a very promising field overall. As Andrew Ng has said, artificial intelligence is the new electricity.

What will develop? Of what is currently trending, namely language models. They will become the foundation for AI. In terms of the development vector — multimodality.

On top of the models, new interfaces will be added beyond text. These could be decision-making systems for robots, script generators for video, military technologies.

Bogdan: how strongly will automation affect the job market? Will people be left without work?

People will not be left idle. Work can be created from any activity. One can retrain into another profession.

Some areas will transform. An obvious example is copywriting.

Even though algorithms can generate large volumes of images, they will not replace designers. They will transform the craft.

With the same GPT — you must phrase the prompt correctly. So there may emerge a job — prompt engineering. A specialist who will formulate the right task for AI.

Right now humans have a big edge. You can ask a human when something goes wrong. You can’t ask a chat bot. This is another reason why humans won’t be out of work soon.

I recently came across an image online showing a store with robot vacuum cleaners where a cleaning lady is mopping floors. I always think of her when people say humans will be left without work.

Bogdan: what about Artificial General Intelligence, how fast will it arrive? And do we even need it?

It has already appeared. The same GPT — this is AGI.

On the topic of what general AI is there is room for speculation, because there is no single consensus. In my understanding — it is one system, one brain, an architecture capable of solving a wide range of tasks.

ChatGPT is such a system. It solves a broad range of tasks, including those it wasn’t trained for. And this ability will only grow stronger.

Bogdan: in theory could ChatGPT pass the Turing test, such that a real person would not guess who they’re talking to?

Even among our own users, those who chat with the clone ask: “Are you a real person? Let’s talk on the phone.” And they drop a phone number into the chat.

There are seeds of doubt in people. So the Turing test at this stage is passed.

Five years ago AGI was much dumber. Even now it is far from human. But over time AI will approach human levels. That’s great; it will push development forward.

Today researchers and mathematicians remain limited in cognitive capacity. We face a barrier: brain size, number of neurons. And we cannot overcome it.

But a more advanced intelligence will have an edge; it will find deeper patterns we don’t even suspect. It will devise new meanings beyond human understanding.

AGI will be able to create new devices, generate new concepts — and it will be good for everyone.

If robots, of course, do not wipe us all out. But the good news is that this is unlikely to happen within our lifetimes.

Interviews conducted by Bogdan Kaminsky and Marina Glayboroda