OpenAI is far from the only game in town. Some of the earliest work on large-language models came from Alphabet Inc.’s Google, for example. — Photo by Jonathan Kemper on Unsplash
OpenAI has stunned the world of techies and investors with its viral artificial intelligence products and its jaw-dropping US$10bil (RM42.5bil) in backing from Microsoft Corp. Now, a growing number of large and small companies are racing to try to overtake the startup in the suddenly hot world of AI services.
"There’s obviously a whole crew of startups that are trying to chase after them – or leapfrog them,” said Guido Appenzeller, a former Intel Corp. AI executive and an Andreessen Horowitz adviser.
AI is a rare bright spot in a contracting, job-cutting tech industry. Generative AI companies – so named for their ability to generate new content from digital troves of text, photos and art – are attracting vast sums of venture capital dollars. In 2022, they raised about US$920mil (RM3.9bil) in the US, according to PitchBook data, up 35% from the year before.
Less than three months into 2023, multiple generative AI companies have raised or are in talks to raise upwards of US$700mil (RM2.9bil) cumulatively, according to reports of funding rounds – not including OpenAI’s Microsoft backing. A running list maintained by the Homebrew AI Club, a group intended as a meeting place for AI workers, counts more than 150 startups in the sector.
"If there is a single shining star in the sea of gloom, it is generative AI,” said Venky Ganesan, a partner at Menlo Ventures. "That’s why it’s also hyped up the way it is. You cannot go to a coffee shop in Palo Alto, or the Village Pub in Woodside, without overhearing three different conversations about generative AI.”
Within the tech industry, OpenAI is famous for its large-language models, massive AI systems that process text from the whole of the internet and use it to generate language. The startup’s sprawling, general models are intended to serve as the foundation for many uses rather than focusing on a single set of narrower applications. Its technology is called GPT, which stands for Generative Pre-trained Transformer. And it’s expected to get radically better with the release of GPT-4, which could debut in the coming months.
OpenAI has drawn wide admiration for the products it has built with GPT so far. Dall-E, a text-to-image AI program, lets users create an image of virtually any scenario in any art style, based on just a few words of prompting. In November, it released the chatbot ChatGPT for wide testing. Soon, the program was spitting out elegant poetry and passed sections of the bar exam. There’s also Codex, a product for computer programmers that uses GPT to suggest the next few lines of code while the programmer types – Microsoft uses it as the basis for its GitHub Copilot programming tool.
But OpenAI is far from the only game in town. Some of the earliest work on large-language models came from Alphabet Inc.’s Google, for example. Google, as well as a host of startups, are currently working on other similar projects.
"OpenAI has had the lead with a number of large-language models but there’s a lot more coming online this year,” said Niko Bonatsos, managing director at General Catalyst. "There’s going to be more competition.”
Here are some of the most significant companies in the running to be the next big AI success story:
Stability AI: When OpenAI released Dall-E last year, Stability AI wasn’t far behind. The startup quickly released its own AI image generator called Stable Diffusion, which became Dall-E’s main competition.
While the two companies’ products are similar, a key difference is that Stability is open-source, which means companies can analyze, tweak and build on its models. At OpenAI, the dataset and other technology that make up Dall-E is proprietary and confidential, although companies can integrate Dall-E into their own products as well.
Stability is planning more products for later this year. The startup plans to release a ChatGPT rival, said Chief Technology Officer Tom Mason. It’s also aiming to debut systems for AI-generated video and is focusing heavily on serving companies in the film industry. "We’re working on video models this year, which is my passion,” Mason said.
While the company’s products are open source, it’s planning to make money from offerings such as helping customers through the process of curating and preparing their data to be used with Stability AI’s systems. "You need engineers who know what they’re doing to handhold you through that process,” Mason said. "Our commerce strategy is to help big companies.”
In October, Stability raised US$101mil (RM430mil) in a seed round led by Coatue Management and Lightspeed Venture Partners at a US$1bil (RM4.2bil) valuation. It’s also working with Amazon.com Inc. to help with the massive cloud infrastructure costs required to power its systems.
But even as investors swoon, the company is being assailed by legal challenges. In January, Getty Images Inc. sued Stability AI in a London court alleging the artificial intelligence software illegally copied and processed vast amounts of copyright-protected photos.
Anthropic: Founded in 2021 by former OpenAI leaders, including siblings Daniela and Dario Amodei, Anthropic in January released a limited test of a new chatbot to rival to ChatGPT. Its name is Claude.
Claude puts a particular emphasis on ethics. Co-founder Daniela was OpenAI’s vice president of safety. And Dario worked at OpenAI variously overseeing safety at the company and as the vice president of research, leading work on both GPT-2 and GPT-3. "We first built Claude as a test bed for AI safety, seeking to develop insights into how to make AI systems that are helpful, honest, and harmless,” Dario said.
While Claude is less capable than ChatGPT at coding, a spokesman for Anthropic said that it’s harder to get Claude to say something offensive. The company has gone through extensive testing in which humans attempt to make the program bend the rules. Scale AI, a startup that helps companies build AI applications, was given access to Claude to test it and concluded, "Claude is not only more inclined to refuse inappropriate requests, but is also more fun than ChatGPT.”
Anthropic’s backers include Facebook co-founder Dustin Moskovitz, former Google Chief Executive Officer Eric Schmidt and now-disgraced FTX co-founder Sam Bankman-Fried. Last week Google invested almost US$400mil (RM1.7bil) in Anthropic and inked a deal in which the startup will use Google’s cloud.
AI21 Labs: Israeli startup AI21 Labs has developed a GPT-3 rival called Jurassic, as well as tools that use AI to help customers write. "Our focus has been to change how we read and write,” said co-founder Yoav Shoham, a former director of the AI lab at Stanford University. The company’s first large-language model was about the same size as GPT-3, even slightly bigger, but more recently AI21 has put out a much smaller version. The performance has been impressive, Shoham said. About 25,000 developers have signed up to use Jurassic and in November the company made it available through Amazon’s cloud AI service.
AI21 raised US$64mil (RM272.5mil) in July, a deal that valued the company at US$664mil (RM2.8bil), according to reports. The startup is likely to raise more soon, Shoham said.
Character.AI: Want to talk to Joe Biden? How about God? Character.AI’s technology allows users to create chatbots that simulate both, along with other celebrities. The company was founded in 2021 by Noam Shazeer, a former Google Brain researcher and one of the inventors of the transformer – a key component of new language models. It launched its beta product less than a year later.
"Our goal is to put this in users’ hands,” Shazeer said. "We’ve done this so far – train the site and launch Character.AI, where users can instantly define their own use cases.”
Right now, the company is seeking to raise a massive US$250mil (RM1bil) funding round. So far, its investors include Gmail creator Paul Buchheit and former GitHub CEO Nat Friedman.
Cohere: Aidan Gomez, co-founder of Cohere Inc., describes his company as similar to OpenAI in that it’s also developing large-language models that can carry out conversations. But the audience for Cohere is not consumers. "What we’re really focused on is bringing this technology to enterprises, developers and startup founders,” Gomez said. That means a heightened focus on strong data privacy protections, which are often demanded by corporate customers.
Gomez used to work at Google Brain, and his startup has a deal to run its systems on Google’s cloud. The Wall Street Journal reported last year that Google was also considering investing US$200mil (RM851mil) in the startup.
In a twist that sets Cohere apart from other AI competitors, Gomez said, "We definitely do have revenue.”
Google: In some ways, it’s surprising that Google isn’t already the dominant name in the conversation about artificial intelligence. The company was a pioneer in the field of large-language models with BERT (Bidirectional Encoder Representations from Transformers), a system used to fuel the company’s market-dominating search engine.
But lately Google’s vaunted AI research operation seems mired in dilemmas over whether to release its work and how to innovate without imperiling the company’s core search engine and ad business. In December, Google employees asked CEO Sundar Pichai and AI research chief Jeff Dean about competition from ChatGPT. According to CNBC, the executives responded that while startups can release new tools to the public fast, Google faces vast reputational risk from any mistakes or errors.
Pichai and the company’s management have since mobilized teams of researchers to respond to ChatGPT, declaring the situation a "code red” threat. The company is also testing rivals to ChatGPT internally, according to reports last week. One is called Apprentice Bard and is based on LaMDA, Google’s Language Model for Dialogue Applications system. Additionally, at Google’s DeepMind AI lab in London, researchers have published work on a conversational AI agent called Sparrow.
Last week on a conference call about Alphabet’s earnings, Pichai said Google will make artificial intelligence-based large-language models like LaMDA available "in the coming weeks and months.” He added that users will soon be able to use language models "as a companion to search.”
"We have long been focused on developing and deploying AI to improve people’s lives,” said Google spokesperson Lily Lin, noting that the company also aims to consider its societal impacts. "We continue to test our AI technology internally to make sure it’s helpful and safe, and we look forward to sharing more experiences externally soon."
Amazon Web Services: Amazon’s cloud unit is using partnerships with companies like Stability and AI21 to supplement its in-house AI expertise, said Bratin Saha, vice president of machine learning and AI services, in an interview. The company also has a service called CodeWhisperer, which suggests code to computer programmers as they type, competing with OpenAI’s Codex and Microsoft’s GitHub Copilot, which is built on Codex.
"A lot of our roadmap is driven by what customers tell us, and this is such a vast space that we think our partners play a big role,” Saha said. "There’s a lot of innovation yet to be done here and we will be partnering with a lot of companies to enable that innovation for our customers.”
Baidu: The Chinese search giant is planning to roll out an artificial intelligence chatbot service similar to ChatGPT, a person familiar with the matter told Bloomberg last week. It could debut in March, initially being embedded into Baidu Inc.’s main search services. The tool, whose name hasn’t been decided, will allow users to get conversation-style search results. Baidu has spent billions of dollars researching AI. Its Ernie system – a large-scale language model that’s been trained on data over several years - will be the foundation of the project, the person said. – Bloomberg
