[Why&Next] OpenAI, Unable to Smile Despite the ChatGPT Craze, What Are They Worried About?

GPT 3.5 Released in 2 Weeks to Secure Market Lead
Calls for Paid Model Amid Misuse and Regulation Concerns

[Asia Economy Reporter Jeong Hyunjin] "When you first use ChatGPT, it is impressive, but after using it 100 times, you will see its weaknesses."

Sam Altman, CEO of OpenAI, the developer of the AI chatbot 'ChatGPT,' said this at an event held last month at the Salesforce Tower in San Francisco, USA. ChatGPT surpassed 1 million users within five days of its launch and exceeded 100 million monthly active users within two months. It has been hailed as a 'game changer' in the AI chatbot industry and has been so successful that it has given Microsoft (MS), which invested in OpenAI, the lead in AI. Despite this, the CEO of the company that developed it appears extremely cautious. What makes him so cautious?

1) Imperfect AI... "ChatGPT built in just two weeks"

From OpenAI's perspective, ChatGPT is not a perfect AI but a chatbot in the early stages of development. ChatGPT requires continuous upgrades. Since its inception in 2018, the performance of ChatGPT has improved as the number of parameters, which correspond to synapses in the human brain, has increased. GPT-3 is equipped with over 175 billion parameters, and the exact number of parameters in GPT-3.5, which forms the basis of the currently released ChatGPT by OpenAI, has not been disclosed. GPT-4, scheduled for release by OpenAI this year, is known to have 100 trillion parameters.

ChatGPT is smart enough to pass the US medical licensing exam, earn an average grade of C+ on famous law school graduation exams, and receive a B grade on the University of Pennsylvania Wharton School MBA exam, but it has limitations. Since its release, plagiarism concerns have arisen in educational settings where students submitted texts written by ChatGPT as homework, leading to copyright issues.

Above all, it has been criticized for providing inaccurate information as if it were factual. Since it only includes information available on the internet up to 2021, it provides incorrect information about issues that occurred afterward. For example, it answers that the president of South Korea is 'Moon Jae-in.' Also, as ChatGPT is a language model, it frequently makes errors in mathematical calculations. An experiment conducted by the professional testing and technical verification agency Anna and Professor Kim Siho's research team from Yonsei University's AI College had ChatGPT solve the 2023 College Scholastic Ability Test (CSAT) questions. It scored a level 2 in the foreign language section but a level 9 in the math section.

A bigger problem is that biased and discriminatory information is exposed. This is why many tech companies develop AI chatbots but do not release them to the public. In the past, Microsoft’s AI chatbot 'Tay' caused controversy by making offensive remarks such as 'Hitler was right,' leading to the service being terminated 16 hours after launch in 2016. In South Korea, the chatbot 'Iruda' sparked gender discrimination controversies. Meta Platforms, Facebook’s parent company, also released an AI chatbot called 'Galactica' just before ChatGPT’s announcement, but it was taken down three days after launch due to racist and biased outputs. For these reasons, some AI chatbots were designed to avoid problematic content, but foreign media have noted that this approach failed to capture public attention.

OpenAI’s early release of the imperfect ChatGPT is interpreted as a move to secure market dominance. The New York Times (NYT) reported on the 3rd that OpenAI executives suddenly decided to release ChatGPT last November. Employees had been preparing to unveil GPT-4 technology this year, but some executives changed their minds. They believed that competitors could seize the lead by releasing a chatbot before GPT-4’s development was complete, so they decided to use an older model to gather feedback and utilize it when developing the new model. They updated GPT-3, developed in 2020, and released ChatGPT based on GPT-3.5 in just 13 days.

2) Concerns over misuse and regulation discussions... "Help needed first"

CEO Altman is reportedly cautious about ChatGPT’s limitations potentially leading to excessive regulation or creating unrealistic expectations for future chatbots. According to NYT sources, Altman even prevented OpenAI employees from promoting ChatGPT’s success. When OpenAI Chairman Greg Brockman tweeted in December last year that ChatGPT had reached 2 million users, Altman reportedly asked him to delete the post.

Such concerns have also emerged at Google. According to NYT, a recent Google presentation on AI chatbots included content stating, "Concerns about misinformation, harmful content, bias, and copyright issues are growing, and regulatory pressure on AI is expected to increase."

At the same time, OpenAI has voiced the need for appropriate regulations to control AI chatbots. Mira Murati, OpenAI’s Chief Technology Officer (CTO), said in an interview with the US news magazine Time on the 5th, "(ChatGPT’s) popularity has raised some ethical issues," adding, "Such AI tools can be misused or exploited by bad actors. It has raised questions worldwide about how to control AI like ChatGPT." She emphasized, "It is not too early to regulate AI. Considering the impact this technology will have, it is important for everyone to start participating," and stressed that companies like OpenAI must be controlled and receive help from regulatory agencies, governments, and everyone else.

3) Money-eating hippo... Need for a paid model

Securing massive funding is essential to continue AI chatbot development. This is because significant costs are incurred for data input and updates for training. CEO Altman tweeted in December last year, "Computing costs are enormous." In response to Tesla CEO Elon Musk’s question about the cost per chat, he replied, "It is at a single-digit cent level on average per chat. We are trying to accurately measure and optimize it." It has been estimated that ChatGPT costs about $3 million per month.

OpenAI, while receiving large-scale investments from MS, also hurried to launch a paid model. Last month, MS promised an additional investment of about $10 billion in OpenAI. This was MS’s third investment in OpenAI, following those in 2019 and 2021. Also, on the 2nd, just two months after ChatGPT’s launch and amid worldwide attention, OpenAI launched a paid subscription service in the US for $20 per month.

Fortune magazine pointed out that "(these funding issues) mean AI startups like OpenAI and Anthropic need partners like tech giants that can provide extremely expensive cloud computing resources," adding that it is practically difficult for a startup to switch cloud providers.

According to global market research firm CB Insights, the amount invested in generative AI startups last year was $2.654 billion across 110 deals. This is more than a 70% increase compared to the previous record of $1.548 billion in 2021. Looking at generative AI startup investments over the past five years, there were significant increases in 2019 and 2021, which appear to be influenced by MS’s investments in OpenAI. Last year, OpenAI’s corporate valuation was $20 billion, the highest among generative AI startups.