It broke the internet overnight. A new artificial intelligence tool is going viral. It's called ChatGPT. It can answer follow-up questions; it can admit its own mistakes.

'write a poem about unicorns.' And it might spit out something that sounds just like it was written by a person. That really struck a chord with a lot of people.

It also struck a chord with investors. Tech companies and their investors have sunk billions of dollars into building

AI systems with a bet that these tools will keep getting more sophisticated and one day prove wildly profitable. There's this concept in AI that if you just keep feeding an AI model more and more data,

it will start to teach itself and become smarter and smarter. But progress seems to be getting a little trickier for some of these big companies. As expenses skyrocket, an industry known for moving fast and breaking things could be slowing down.

I think the progress is going to get harder. When I look at 2025, the low-hanging fruit is gone. They're trying to figure out how to make these models—that are getting increasingly expensive, eating up increasing amounts of computing power—worth that trade-off,

right? If you're not getting this incredible boost in performance, what do you do? AI really started in the 1950s, with Alan Turing being one of the early academics. You might be familiar with the Turing test?

Periods of AI innovation, and then what you might call an AI winter. But fast forward to today. In the past couple of years, this intense pickup in AI came from the breakthroughs we saw with OpenAI and ChatGPT.

Search, as we know it, is going to change, and the web as we know it is going to change. I was very skeptical; I didn't expect ChatGPT to get so good. When ChatGPT came out, it looked like, wow,

overnight we had this huge advancement—it sort of came out of nowhere for most people. I think it really kickstarted a lot of interest in innovating in AI, and also investing in AI, and companies trying out generative AI.

ChatGPT and AI systems like it work by using massive software engines called Large Language Models, or LLMs. A Large Language Model is an AI system that's trained on lots of data sourced from the internet, typically, and it can respond to written prompts

with amounts of text that sound really like they were written by a human. To do this, these models use an algorithm to process the parameters of any given prompt, and this is how that unicorn poem comes into being. And people keep expecting these companies to keep rolling out better and better

models that are more and more capable. What if I were to say that you are related to the announcement, or that you are the announcement? Me? The announcement is about me? Well, color me intrigued. For a while, these models have been getting better quickly.

But right now, at least three of the top companies—OpenAI, Anthropic, and Google—are having issues training their models to the level they would like. The easy gains are gone. We will keep scaling them, but it won't be at the

same rate of pure scale that we've seen in recent years. So we're working on new things, too. Good, curated data sets that are created by humans are increasingly scarce. I mean, the internet at this point has been largely scraped by these companies. So if

you want to make models better than what they are now, you need even more data. That's harder and harder to find. Some companies are going as far as paying people with advanced degrees to help train their models so that they can get this expertise to get better data.

So how do you continue to teach an AI? Where do you get that data from? Especially once this AI is getting so smart that it needs expert-level data—it needs the kind of data that a PhD student or a Nobel Prize winner would be

able to give you. How do you get all of that? How do you feed all of that when it's not as simple as just scraping the web anymore? Some projects are experimenting with what's called synthetic data, which includes training AI with content that is itself AI-generated.

Synthetic data basically means you take the output of an AI model and actually use that to start training more AI models. But, of course, that has its challenges, and that's still a technique that's being tested. And we don't know how much companies can really rely on

that or they will need to continue to also actually get human-created, higher-quality data. While we have seen meaningful new products and advances this year, tech companies are struggling to clear the high bar for improvements that would justify the tremendous amount of money they're spending. Because let's be real,

AI is not cheap. The CEO of Anthropic, which is a leading AI lab and probably a number two competitor to OpenAI in the startup world, has said it costs a hundred million dollars to train a new AI model, and in the coming years, that could increase to a hundred billion dollars.

As we're spending more and more money, the engineering complexity of getting it right is increasing, and that's why companies need to be larger, need to have more talent. It's hard to say if anyone's making money off of AI right now. We do know that

OpenAI has quite a number of paying business customers, but we don't know how that breaks down in terms of the exact number of companies that are paying for it, or what kind of benefits these companies are actually seeing. ChatGPT is

one of the most quickly growing consumer software products of all time, but when that will start to match the costs is in question. But it's not all about the money, is it? Let's keep in mind, OpenAI, for example,

was established as a non-profit research company focused on advancing AI for the benefit of humanity. Of course, it plans on shifting to a for-profit model, but the timeline on that is unclear. And still, despite the wintry forecasts, the money keeps pouring in.

They have plenty of money at the moment, but it's not totally clear if companies are going to be using it long-term, or what kind of return they're getting on their investment. These companies are having to do bigger and bigger fundraising rounds, billions and billions of

dollars, and at a certain point it's unclear where that money's actually going to come from if it's not coming from customers. Investors may be doubling down because of the promise of even more mind-boggling advancements just around the corner.

OpenAI came out with a new reasoning-based model recently. And the idea there is that if you give these AI models more time to sit and think about a problem, they can reason their way to a more accurate or

more intelligent answer. Another breakthrough that may be happening is also around things like agents, which are this idea of an AI that can just talk to you like a chatbot can, but actually complete tasks for you; like, let's say, book travel or actually integrate code into an

application. Perhaps the most exciting thing for AI enthusiasts and the most horrifying for its detractors is the prospect of artificial general intelligence, or AGI. AGI would be a system that can reason and think like a human, applying its

synthetic brain across disciplines. Not just reacting to prompts with poems or six-fingered hands, but accomplishing complex tasks independently of us humans and maybe even surpassing us. Could we one day be working for them? No one exactly agrees on when we're going to reach AGI or what exactly it will

look like. A lot of people are predicting that it can come very quickly. You see other people saying it could take decades, it could take a century. It could even never happen, and some of these setbacks that we're seeing give people reason

to maybe reassess or say, "Hey, this path to AGI may not be as simple as some people wanted."

How AI Got a Reality Check

Bloomberg Originals offers bold takes for curious minds on today’s biggest topics. Hosted by experts covering stories you haven’t seen and viewpoints you haven’t heard, you’ll discover cinematic, data-led shows that investigate the intersection of business and culture. Exploring every angle of climate change, technology, finance, sports and beyond, Bloomberg Originals is business as you’ve never seen it.


Bloomberg Originals

ChatGPT의 입소문 성공과 투자자의 관심

후속 질문에 답하고 실수를 인정하는 ChatGPT의 기능 덕분에 입소문이 났습니다.

ChatGPT의 인간적인 시 쓰기는 많은 이들에게 깊은 인상을 남겼습니다.

투자자들은 미래의 수익성에 베팅하며 AI에 수십억 달러를 쏟아 부었습니다.

데이터와 컴퓨팅 성능이 증가하면 스스로 학습하는 AI가 될 것이라는 믿음은 도전받고 있습니다.

수익성이 낮은 과일이 사라지고 비용이 증가함에 따라 진전이 더뎌지고 있습니다.

AI의 역사에는 혁신과 정체의 시기가 있습니다.

최근 OpenAI와 ChatGPT의 발전은 관심과 투자를 촉진했습니다.

ChatGPT의 빠른 발전은 많은 사람들을 놀라게 했습니다.

ChatGPT는 AI와 제너레이티브 AI에 대한 관심과 투자를 촉진했습니다.

대규모 언어 모델(LLM)의 작동 방식과 그 한계

LLM은 방대한 인터넷 데이터를 학습하여 사람과 유사한 텍스트로 응답합니다.

AI 모델의 지속적인 개선에 대한 높은 기대치가 도전받고 있습니다.

최고의 AI 기업들은 모델을 원하는 수준으로 훈련시키는 데 어려움을 겪고 있으며, 손쉬운 이득은 이제 끝났습니다.

고품질 데이터는 점점 더 희소해지고 있으며, 이를 확보하기 위해 더욱 정교한 방법이 요구되고 있습니다.

기업들은 교육용 데이터 품질을 개선하기 위해 전문가에게 비용을 지불하고 있습니다.

전문가 수준의 데이터에 대한 필요성으로 인해 AI 학습의 난이도가 높아지고 있습니다.

AI가 생성한 콘텐츠를 학습에 사용하는 합성 데이터도 연구되고 있습니다.

합성 데이터에 대한 의존도와 사람이 만든 데이터의 필요성은 불확실합니다.

기업들은 AI 개발의 높은 비용을 충분한 개선으로 정당화하기 위해 고군분투하고 있습니다.

새로운 AI 모델을 학습시키는 데 드는 비용은 점점 더 커지고 있습니다.

AI의 수익성은 불확실하며, 비즈니스 고객 혜택에 대한 정보가 제한적입니다.

ChatGPT의 빠른 성장은 높은 개발 비용과 맞물려야 했습니다.

비영리 단체로 시작한 OpenAI의 시작과 영리 모델에 대한 계획은 그 복잡성을 강조합니다.

어려운 상황에도 불구하고 AI 개발에 상당한 자금이 계속 유입되고 있습니다.

대규모 모금 라운드는 자금의 장기적인 지속 가능성에 대한 의문을 제기합니다.

새로운 추론 기반 모델을 통해 보다 신중하고 정확한 대응이 가능합니다.

AI 에이전트는 챗봇처럼 상호작용하며 작업을 완료할 수 있습니다.

인공 지능(AGI)은 향후 발전 가능성이 있는 분야로, 그 시기에 대한 예측은 다양합니다.

최근의 좌절로 인해 AGI에 대한 경로를 재평가해야 할 수도 있습니다.