- “Moat” is beloved term by tech to describe how a company can stop rivals taking their business.
- In the fast-changing world of AI, OpenAI appeared to have created a powerful moat through ChatGPT.
- But Meta rolling out its own powerful language model, Llama 2, with few restrictions puts ChatGPT at risk.
It seems fair to say that OpenAI has been king of the hill in generative AI ever since ChatGPT’s launch. That might not be the case for much longer.
OpenAI’s chatbot has upended teaching, writing, technology, and many other sectors since its arrival in November 2022. It put the world’s biggest tech companies like Meta and Google on the back foot. Across the US, blue-chip CEOs spent the last six months running to roll out ChatGPT plugins.
It looked like OpenAI’s competitive moat was infallible — until now.
Not the only game in town
The term “moat” refers to a company’s competitive advantage and was popularized by Warren Buffett. One of the most famous moats in tech was, for example, Microsoft bundling its own software with Windows, meaning generations of kids grew up using products like Word and Internet Explorer — until Google rolled out free, rival products like Google Docs.
In much the same fashion, OpenAI started off strong out of the gate. The free version of ChatGPT was the fastest-downloaded internet app of all time when it arrived in November, with 100 million users in just two months. That millions of people had their first exposure to advanced AI powered by large language models through ChatGPT was a major competitive advantage, and it was a genius move by OpenAI to launch a free version of the chatbot.
The competition is now following suit.
This week, Meta announced the launch of Llama 2, a huge new AI model through a partnership with Microsoft. The companies described this as “open source,” and say it’s freely available to developers for both research and commercial use.
“We believe an open approach is the right one for the development of today’s AI models, especially those in the generative space where the technology is rapidly advancing,” a blogpost announcing Llama 2 said.
Now, the open-source community disputes that Llama 2 is open source at all, as Insider’s Kali Hays and Rosalie Chan reveal. Companies with over 700 million monthly active users will need to get special permission from Meta to use Llama 2, per the terms and conditions for the AI model’s use. That essentially blocks Meta’s biggest competitors like Google from using the model but puts it at odds with open-source definitions set out by the nonprofit Open Source Initiative.
And unlike OpenAI’s ChatGPT, Llama 2 doesn’t come with an easy-to-use consumer interface. It’s targeted at technical and corporate users who will use the model to power their own apps and services.
That said, Llama 2 is now a powerful tool to be used by companies across the board who are desperate to reap the benefits of generative AI. Even if it isn’t strictly open, it is broadly free, whereas later OpenAI models are not. Early impressions suggest Llama 2 stacks up pretty evenly against ChatGPT.
Nathan Lambert, a research scientist at AI company Hugging Face, noted in a Substack post on Tuesday that Llama 2 involves “substantial technical expansions” in terms of data quality and training techniques, and its “fine-tuned chat models seem on the same level as ChatGPT.”
“It is a huge leap forward for open-source, and a huge blow to the closed-source providers, as using this model will offer way more customizability and way lower cost for most companies,” Lambert wrote.
A question of quality
The second issue for OpenAI is quality.
There is a growing sense throughout the AI community that the responses from GPT-4, the advanced, multi-modal AI model underlying the premium version of ChatGPT, have gotten a little dumber.
Insider reported this month that some of this deterioration may have resulted from a radical redesign of the model that seems to have sped up the response time of the chatbot while decreasing the quality of the responses.
A paper published by researchers at Stanford University and UC Berkeley on Tuesday added to the thesis that the model is getting sloppy in its ability to solve math problems, generate code, and respond meaningfully to sensitive questions.
It’s an issue OpenAI will want to be on top of as competitors emerge – especially as the initial novelty of ChatGPT wears off. This month, web analytics company Similarweb estimated worldwide desktop and mobile traffic to the ChatGPT website dropped 9.7% from May to June.
As the first popular entrant into the generative AI space, OpenAI’s chatbot had the advantage of capturing the imaginations of millions before Google, Meta, and the rest.
But open models were always the threat.
In May, the Semianalysis blog published a memo purportedly from an anonymous Google engineer warning that open-source models would eventually overtake anything closed models could do. The paper was titled: “We Have No Moat, And Neither Does OpenAI.”
OpenAI may need to come to terms with that before Llama 3 comes around.