According to news on December 25, when OpenAI launches its chatbot ChatGPT in late 2022, there are no other similar services on the market. OpenAI has successfully attracted a lot of attention with its superior technology and billions of dollars in funding. As the first artificial intelligence company to launch breakthrough products, OpenAI quickly gained widespread popularity.
However, by the end of 2023, many well-funded technology companies have launched artificial intelligence services comparable to ChatGPT. In September this year, Mistral, a French startup valued at approximately US$2 billion, released a powerful open source large language model (LLM). LLM is the basic technology supporting artificial intelligence chatbots. Just four months later, Elon Musk also launched Grok, a practical chatbot. Earlier this month, Google released the initial version of its long-awaited Gemini model.
Although these tools have not yet surpassed OpenAI's GPT-4 model in performance, in many cases their performance is very close to ChatGPT, and they pose a certain threat to ChatGPT by competing for users and commercial customers. Moreover, users can easily switch between different services without any restrictions.
While ChatGPT excels in terms of versatility, many users seem to prefer using more specialized chatbots. For example, estimates from data firm SimilarWeb show that CharacterAI, a service that allows users to interact with AI versions of video game characters and other characters, has taken hold.
"The situation now is similar to the early days of mobile phones, where everyone is trying different approaches," said Aravind Srinivas, CEO and co-founder of PerplexityAI, which operates research chatbots.
PerplexityAI aims to differentiate itself from other chatbots through more accurate and timely responses. While other leading chatbots, including ChatGPT, have struggled to obtain accurate and reliable real-time information, PerplexityAI is not developing its large model entirely from scratch, but rather building on existing models, including open source models from Mistral and Meta.
Open source models perform inconsistently across a wide range of topics, from calculus to poetry, and are not always as consistent as GPT-4. But average users may not notice these differences as much as researchers do, and many may not care about calculus at all.
"Users don't care whether your answer comes from GPT 3.5, Anthropic's Claude or GPT-4," Srinivas noted. His product also allows users to use ChatGPT's API. "They're looking at whether the answers are accurate, whether the product is fast enough and whether it has the right features. That's where they'll complain."
OpenAI's advantage has always been based on the power provided by its large language models. These models are able to ingest massive amounts of online data to better respond to a variety of requests and use cases. However, a growing number of AI startups are looking to build models with comparable performance but smaller scale.
"While the training process of GPT-4 is amazing, I don't think there is anything magical about it," said Qiu Kanjun, CEO and co-founder of artificial intelligence startup Imbue. "Despite its large scale, we see a lot of evidence that we can achieve similar performance with smaller, more efficient models." Imbue is backed by Nvidia and is developing its own basic artificial intelligence model.
Qiu Kanjun also mentioned that her company developed a tool internally that can fix errors in Python (a popular programming language) "like a toy." Because this tool is task-specific, the company doesn't need to build models as powerful as GPT-4 to support it.
It’s worth noting that ChatGPT is no longer the only artificial intelligence chatbot out there amid OpenAI’s internal turmoil last month. Many rivals including Cohere, Anthropic and Google are seeing a rise in customer interest after briefly ousting its CEO and suddenly putting its future into question.
Andrew Ng, co-founder of the education platform Coursera and co-founder of Google Brain, said: "GPT-4 is still the best large language model currently, there is no doubt about it. But I think competition is definitely intensifying, and open source and other competitors have a good chance to catch up with GPT-4. Of course, OpenAI will not be satisfied with the status quo, they are working on GPT-5."
Andrew Ng is also the CEO of LandingAI and a professor at Stanford University. In recent years, he has spent much of his time educating the public about artificial intelligence and resisting AI doomsday theories. Last month, Coursera released an online course on generative artificial intelligence. Ng said the course broke the record for the fastest-growing course on the site, with 44,000 people signing up in the first week.
Ng stressed in two interviews in the past month: "Despite these concerns, I am enthusiastic about the prospects of the industry. I think it is time for everyone to start understanding these technologies."
The following is a summary of the media interview with Ng Enda:
Q: What are your special thoughts on Google Gemini? Does it meet your expectations?
Andrew Ng: I think Google engineers have made good progress, although they are still a long way from catching up with GPT-4. Secondly, I think Google made a marketing mistake in the way they released their videos. I find this unfortunate, because if we ignored the marketing missteps, most of us would be cheering for Google engineers to build an impressively large model.
Q: How do you think OpenAI’s internal turmoil has impacted the industry as a whole?
Ng Enda: It’s too early to draw a conclusion at this time. What I feel is that OpenAI's missteps, including the firing of Sam Altman and subsequent team reorganization, have really shaken some people's confidence. I've heard that many teams are making sure they have alternative OpenAI services. However, there are some positive aspects. For example, they did add Larry Summers and Bret Taylor to the board. OpenAI is actively engaging with governments, which is beneficial to a certain extent. Summers is a hugely influential figure, and OpenAI will need to hire an army of lobbyists to achieve the impact he can have with just a few phone calls.
Q: You once said that some technology companies are using the fear that "artificial intelligence will lead to human extinction" to push for regulation. Can you explain why and give an example?
Ng Enda: I think all parties have motives to create panic and hype. Some tech companies may be reluctant to compete with the open source model. Others are motivated by concerns, even if they are not very realistic, which creates favorable conditions for lobbyists and regulators. In the United States and Europe, this widespread panic about artificial intelligence, whether about extinction or other causes, is creating opportunities for lobbyists and regulators to write biased rules.
Q: I understand that tech companies may create fear and hype to push for regulation because of competing with open source models. But why are your academic colleagues, including AI luminaries like Geoffrey Hinton or Yoshua Bengio, worried about the existential threats posed by AI? They don't appear to be directly connected to the hype.
Andrew Ng: I had a conversation with Hinton just a few weeks ago. While I agree that his concerns are genuine, I think his perspective is misplaced. I'm not singling out Hinton or Bengio specifically. It has nothing to do with them as individuals. But I think that for a lot of people, there's also an incentive to panic.
Fear grabs people's attention. I think if it weren't for the role some people play in inciting fear, they might not be in the conversation. And, frankly, I know some people who use exaggerated fears to attract attention and then receive substantial speaking fees. It turns out there's actually little incentive not to play up fear. We see on social media that fear and anger are amplified, while a generally positive view of the future is less interesting to the media.
Q: Altman was once your student. Can you share with us some of his experiences when he was a student at Stanford?
Ng Enda: Ultraman is very smart, but he is very stubborn and always does things his own way. He joined my lab at Stanford to use reinforcement learning to control robots. This is already well known, but he actually worked with classmates in my lab to build a snake robot and have it climb up steps.