Last month, JPMorgan Chase let us in on the development of a ChatGPT-like model, called IndexGPT, which gives AI-powered investment advisory to select customers. Similarly, a few months ago Goldman Sachs announced that it is working on an AI-powered in-house chatbot to assist developers in writing codes. Amazon, too, is working on developing its own language model that aims to outperform OpenAI’s GPT models, and so is Apple.
Why is everyone making their own LLMs?
Interestingly, a majority of these companies are the ones that have previously banned the use of ChatGPT for internal purposes. Besides these companies, Samsung, Bank of America, and Wells Fargo are also some of the players who have restricted their employees from using the publicly trained chatbot over the fear of leaking sensitive/confidential information.
One of the biggest reasons for a company to build its own LLM is security. Companies are sceptical about accessing proprietary LLMs like GPT as the companies’ data might be accessible to someone else. For this, the simplest approach is to take an open source LLM and fine tune it.
Building your own LLM is a lot like building your own chips. While most can use the existing models, and specialise them slightly, very few would want to build something for very specific purposes. The rest of them would likely use the same LLMs, which is more practical.
For instance, a company might choose to integrate OpenAI language models into their automated customer service chatbot to make it more human-like. This approach often proves superior to developing their own large language models, which is costly, time-consuming, and may not yield comparable results.
On the other hand, even though OpenAI’s APIs grant access to cutting-edge models like GPT-3, they come with significant drawbacks. The cost can escalate quickly for high-traffic applications due to usage fees. Additionally, your application’s functionality is bound by OpenAI’s terms of use, potentially limiting your startup’s flexibility and adaptability.
Moreover, with GDPR in Europe, companies are restricted from using APIs through OpenAI or Google. These LLMs are incompatible with the GDPR policy as the regulation calls them a breach of privacy. This is another reason to use models rather than calling for an external API.
Simply because they can
Traditionally, it has been believed that only a few research companies with substantial resources, such as OpenAI, Google DeepMind, Anthropic, Cohere, to name a few, possess the necessary capital, research expertise, and computational power to create, train, and protect large and sophisticated AI models. Consequently, these companies were expected to be the ones pioneering the development of groundbreaking, high-performance models for the foreseeable future as they have valid reasons to consider building their own specialised models as they can afford to.
But, what about those that can not afford it? That is where open-source models
like LLaMA, Vicuna, Falcon, etc are helping small and medium businesses and companies to leverage and develop their own LLMs for specialised as well as general use cases, or for research purposes. Again, it depends on factors such as the company’s maturity, workforce, financial resources, etc.
Researchers point out the rapid pace at which new models are being developed and shared on open-source platforms like Hugging Face and GitHub. These models are often more compact, faster, customisable, require less development time, and exhibit comparable or even superior capabilities compared to the massive models created by well-funded players like Google and OpenAI.
Open Source to the rescue
While companies have been relying on APIs from OpenAI to utilise them in their own work, recently, a document purportedly written by a Google researcher surfaced on a Discord channel which revealed that some individuals within these prominent research companies perceive open source as a significant threat.
The open-source AI community is indeed experiencing a surge in activity. Hugging Face, a major repository of open-source machine learning models, data sets, and tools, reports that over 15,000 companies are utilising their platform, sharing more than half a million models, data sets, and demonstrations.
Since the release of ChatGPT in November, developers have contributed over 100,000 public models for different tasks such as next-word prediction, mask filling, token classification, sequence classification, and so on. This trend aligns with the growing interest of large corporations across various industries in leveraging AI text and image generation models to transform core business functions, such as content creation, marketing, and customer service, as is the case with JPMorgan Chase and Goldman Sachs.
When it comes to deploying generative AI, many companies like Salesforce, and Snapchat, opt to access models developed and hosted by Google or OpenAI through their APIs. That is because they do not have to comply with GDPR and do not care about the leakage of sensitive information.
It all boils down to a simple question for companies — do we want to build something from scratch or trust large models with our privacy and security? The former might be difficult for a lot of them so they use APIs by OpenAI or Google as they serve as a foundation for their future use cases while the latter is just a compromise.
The post Build LLMs From Scratch or Surrender: Cos Caught in the Privacy-Precision Paradox appeared first on Analytics India Magazine.
