“The model that’s being introduced today, while it looks great, isn’t actually the best model available,” said Tom Goldstein, a computer science professor at the University of Maryland. “As a result, the models you see have a lot of weaknesses.” Weaknesses that might be avoided if you don’t mind the cost, such as the tendency to spew out biased results or outright falsehoods. can be mentioned.
Tech giants betting their future on AI rarely discuss the cost of their technology. OpenAI (ChatGPT developer), Microsoft and Google all declined to comment. But experts say it’s the most obvious obstacle to Big Tech’s vision of generative AI permeating every industry, reducing headcount and increasing efficiency.
Because AI requires intensive computing, OpenAI withheld its powerful new language model, GPT-4, from the free version of ChatGPT and still runs a weak GPT-3.5 model. ChatGPT’s underlying dataset was last updated in September 2021 and is not useful for researching or discussing recent events. Also, even someone paying $20 a month for GPT-4 can only send 25 messages every 3 hours because it’s so expensive to run. (Response is also quite slow.)
These costs may also be one reason why Google has yet to incorporate AI chatbots into its flagship search engine, which handles billions of queries daily. When Google released his Bard chatbot in March, the company chose not to use the largest language model. Dylan Patel, principal analyst at semiconductor research firm Semianalysis, estimated that a single ChatGPT chat could cost up to 1,000 times more than a simple Google search.
In a recent report on artificial intelligence, the Biden administration pointed out that the computational cost of generative AI is a national concern. The White House said the technology is expected to “dramatically increase computational demands and associated environmental impact” and that there is an “urgent need” to design more sustainable systems.
More than any other form of machine learning, generative AI requires a dizzying amount of computing power and specialized computer chips known as GPUs that only the wealthiest companies can afford. The escalating battle for access to these chips has elevated major providers into tech giants in their own right, helping them hold the keys to what has become the technology industry’s most valuable asset.
Silicon Valley came to dominate the internet economy by providing the world with services like online search, email and social media for free, initially losing money but eventually personalized advertising. generated enormous profits. And ads will likely appear in AI chatbots. But analysts say advertising alone probably won’t be enough to make cutting-edge AI tools profitable anytime soon.
On the other hand, companies that offer AI models for consumers must balance their desire to gain market share with the economic losses they are accumulating.
And the quest for more reliable AI will primarily benefit the chipmakers and cloud computing giants who already dominate much of the digital space, along with those who own the hardware needed to run their models. Probability is high.
It’s no coincidence that the companies building the major AI language models are either the largest cloud computing providers like Google and Microsoft, or have close partnerships like OpenAI does with Microsoft. is not.companies to buy Clem DeLang, CEO of open-source AI company Hugging Face, says these companies’ AI tools are tied to subsidized services that cost far more than they’re currently paying. He said he didn’t realize that.
OpenAI CEO Sam Altman indirectly acknowledged the issue during a Senate hearing last month, while Senator John Ossoff (D-Georgia) called ChatGPT addictive in the way OpenAI harms children. He warned that Congress “will look very hard” if it tries to be one. ” was written. Altman said Ossoff shouldn’t worry, saying, “We try to design systems that don’t maximize engagement. In fact, GPUs are so scarce that people who use our products The less, the better.”
Developing and training AI language models is expensive and requires huge amounts of data and software to identify patterns in language. AI companies typically hire star researchers whose salaries rival those of professional athletes. This presents an initial barrier for companies wanting to build their own models, but some well-funded startups have been successful, such as Anthropic AI, founded by OpenAI graduates with funding from Google.
Each query to chatbots such as ChatGPT, Microsoft’s Bing, and Anthropic’s Claude is then routed to data centers, where supercomputers speed up models and perform many fast calculations at the same time. First, interpret the user’s prompt, then anticipate the user’s prompt. The most plausible response is he one “token” at a time, a sequence of four characters.
This kind of computing power requires a GPU (graphics processing unit). Initially made for video games, GPUs turned out to be the only chips capable of handling heavy computer tasks such as large language models. Only his Nvidia company currently sells the best, and it costs tens of thousands of dollars.Recent Nvidia Ratings Expected sales jumped to $1 trillion. The value of TSMC, the Taiwan-based company that makes many of these chips, has skyrocketed as well.
“At the moment, GPUs are a lot harder to get than drugs,” said Elon Musk, who recently purchased about 10,000 GPUs for his AI startup, at the Wall Street Journal summit on May 23. Told.
These computing requirements also help explain why OpenAI is no longer the non-profit organization it was founded for.
Launched in 2015 with a stated mission to develop AI “in a manner that is most likely to benefit humanity as a whole, unconstrained by the need to generate economic returns”, it will continue until 2019. switched to a commercial model to attract investors to That includes Microsoft, which invested $1 billion to become OpenAI’s exclusive computing provider. (Microsoft then poured another $10 billion to integrate OpenAI’s technology with Bing, Windows, and other products.)
Exactly how much it costs to run a chatbot like ChatGPT is a moving target as companies strive to make their chatbots more efficient.
Shortly after launch in December, Altman estimated ChatGPT’s cost to be “probably in the single-digit cents per chat.” That may not sound like a big deal until he scales the number of daily users to more than 10 million, according to analyst estimates. SemiAnalysis he said in February, based on the processing he needed to run GPT-3.5, which is his default model at the time, ChatGPT’s computing costs alone he said to OpenAI that he We calculated that it cost about $700,000.
Multiply these computing costs by the 100 million people you use each day. Microsoft’s Bing search engine Or, with more than a billion people reportedly using Google, we begin to see why tech giants are reluctant to release their best AI models to the public.
“This is not a sustainable equation for generative AI, the economy, the democratization of the environment and its widespread availability,” said the founder of d-Matrix, a startup working to develop chips for more efficient AI. CEO Sid Sheth said.
In February’s Bard announcement, Google said it would initially run on a “lightweight” version of the company’s LaMDA language model, because it “requires significantly less computing power, allowing more We will be able to respond to users.” In other words, even a wealthy company like Google wasn’t prepared to cover the cost of putting its most powerful AI technology into a free chatbot.
Cost cutting has taken a heavy toll. Bird stumbled on a basic fact in its launch demo, ripping $100 billion from the value of Google stock. Bing, on the other hand, didn’t get off to a good start early on, leading Microsoft to scale back both its personality and the number of questions users can ask in a given conversation.
Such errors, also known as “hallucinations,” are a major concern with AI language models as both individuals and businesses increasingly rely on them. Experts say these are a function of the basic design of the model, built to produce probable sequences of words rather than truthful statements.
Another Google chatbot called Sparrow was designed by the company’s subsidiary DeepMind to search the internet and cite sources with the goal of reducing falsehood. But Google hasn’t released it yet.
Meanwhile, each of the big players is vying for ways to make AI language models cheaper.
Query execution costs on OpenAI’s new lightweight GPT-3.5 Turbo model are less than 10 times less than the top-of-the-line GPT-4. Google, like startups like d-Matrix, manufactures its own AI chips that it claims are more efficient than Nvidia’s. And many startups are building on open source language models such as Meta’s LLaMA, so you don’t have to pay OpenAI or Google to use it. However, these models are still not very performant and may lack guardrails. to prevent abuse.
Goldstein, of Maryland, said the push for smaller, cheaper models marks a sudden reversal in the industry.
“We spent the last four years just building the biggest model we could,” he said. But the goal then was to publish a research paper, not to release an AI chatbot to the public. “In the last few months there has been a complete turnaround in the community and suddenly everyone is trying to build the smallest possible model to control costs.”
For consumers That could mean the days of free access to powerful general-purpose AI models are over.
Microsoft is already experimenting with embedding ads into AI-powered Bing results. At the Senate hearing, OpenAI’s Altman didn’t rule out doing the same, but said he prefers a paid subscription model.
Both companies have said they are confident the economic situation will eventually clear up. “There’s so much value here, I can’t believe you don’t know how to ring the cash register,” Altman told technology blog Stratechery in February.
But critics point out that generative AI also comes with a cost to society.
“All of these treatments have an impact on greenhouse gas emissions,” says Vaskar Chakravorty, dean of global business at Tufts University’s Fletcher School. Computing requires energy that can be used for other purposes, such as other less prevalent computing tasks than AI language models. This “could even delay the development and application of AI to other more meaningful uses such as healthcare, drug discovery and cancer detection,” Chakravorti said.
Based on estimates of ChatGPT’s usage and computing needs, data scientist Kasper Groes Albin Ludvigsen estimated that ChatGPT used the equivalent of 175,000 people in January. This corresponds to a medium-sized city.
For now, Goldstein said tech giants are willing to lose money to gain market share with their AI chatbots. But what if you can’t make a profit? “Eventually we reach the end of the hype curve, at which point investors only look at earnings.”
Still, Goldstein predicted that many people and companies would encounter generative AI. A tool that is hard to resist, Even with all its shortcomings. “Even though it’s expensive, it’s still a lot cheaper than human labor,” he said.
Nitasha Tiku contributed to this report.
