AI, especially generative AI solutions, is being introduced into all kinds of products and services that we use every day. Entities and organizations that hold the largest market share in the models used as the basis for these everyday apps (known as foundation models) are holding the keys to the kingdom in the spirit of the fast-paced gold rush that is generative AI. You will hold it.
Foundation model that will become a new AI platform
The underlying model is not just an AI model built to handle a single task, as was the case with previous models. Because the underlying model is trained on a very large corpus of data, it is broadly applicable to a very wide range of tasks and performs them at a level better than most single-objective models. For example, the fundamental large-scale language model (LLM), which is very popular today, is also capable of sentiment analysis, chatbots and conversational systems, document Q&A, document analysis, and OCR of images. The ability to do all of this in one big model is incredibly compelling and has pushed generative AI to the forefront of every AI conversation today.
The most popular base model is quickly becoming famous. OpenAI's GPT models, especially the ChatGPT application based on them, Google's Gemini model, Anthropic Claude, and many others. Many of these models, including those mentioned above, are unique in that the underlying training data, trained model parameters, model tuning settings and hyperparameters, and other aspects are kept tightly guarded and not made public. It belongs to.
Incorporating someone else's AI models into your system can be problematic, especially if you have little control or visibility over those models. However, despite increasingly walled gardens of proprietary large-scale language models, organizations are still not fully aware of the potential drawbacks of using non-open machine learning models. Is not … After all, most organizations and individuals' experience with AI is still very new and green.
In fact, when organizations say they are using AI or building AI systems, they are almost always making API calls to versions of their own models hosted in the cloud. While organizations may be running models locally or on their own infrastructure, the lightning-fast pace of AI iterations is forcing companies to take the opportunistic step of using only what is currently available. It has become. The need for local control or customization of the model is pushed to later iterations or not at all.
Governments are paying close attention to their own foundation models and their investments
Incumbent technology vendors like Microsoft, Amazon, Google, Meta, Oracle, and IBM have rightly recognized that AI-based models are the platform of the future and that they need to have a hand in it if they want to have a stake in the future. Masu. Each of the major players in AI models is in their own pie. This has driven incredible investments in companies like OpenAI, Anthropic, and Mistral, which have raised tens of billions of dollars in partnership investments, far exceeding what the venture capital community is capable of investing.
These vendors are competing for AI dominance, as detailed in the AI Today podcast on this topic. This is not surprising as AI is the biggest market opportunity of the past decade. There are really only two ways he can dominate these markets with new technology. It's either build or buy. However, the industry progresses too quickly to forge a path to superiority. Therefore, these vendors must be acquired or strategically partnered to gain an advantage. (Disclosure: I am the managing partner and co-host of the AI Today podcast)
This extensive partnership and investment activity has attracted the attention of federal regulators. The Federal Trade Commission (FTC) recently launched an investigation into these generative AI investments and partnerships. They will scrutinize the competitive dynamics of the rapidly emerging AI market to ensure innovation and fairness, and better understand market trends and practices to identify potentially anti-competitive behavior. I'm interested in checking. However, there is no need to wait for unplanned regulatory action to determine an approach to ensuring greater freedom of choice and control in the use and development of AI models.
What are the alternatives to open AI and other emerging AI foundation models?
The challenge with proprietary models is that as organizations become dependent on proprietary models for their AI capabilities, they have limited flexibility and freedom in how to deploy, configure, and tune models for specific purposes. This is because without the underlying data, model parameters, and additional details, those models cannot actually be reconstructed and can only be used as consumers.
The way the underlying model behaves and its performance capabilities may change, and you may run into issues where the model refuses to respond to prompts or input as a form of content moderation, limiting the types of data that can be brought in and taken out. there is. To them. If it's not your platform, you can't control it. Organizations are starting to realize the problems with this approach.
Organizations have more and more choices regarding the underlying models they use for their applications. The fast pace of AI innovation means that models are being developed both in open source models that share all aspects such as data, training information, and model parameters, as well as other proprietary models or fine-tuned versions of models already in use. It has been.
New models emerge because proprietary models are in many ways the AI version of vendor lock-in. Elon Musk specifically sued Open AI for violating his original promise to develop AI technology “openly” and just recently opened his Grok model as a competitor to proprietary models on the market. Released as source.
What is the difference between open source AI and proprietary AI?
This led to the growth and emergence of open source AI-based models. Open source models are free to download, use, and embed. You can also see the model's code, details, weights and parameters, and the final trained model and make your own changes. You can also check all aspects of this. Just as much of the technology base upon which AI systems run is open source, so too are efforts to make the models produced by those systems open and available. There's a lot to be said about transparency, especially in the context of building trusted AI solutions.
Many of these open source models perform as well as less open alternatives on the market. Many open source models perform as well, or in some cases better, than proprietary commercial models. Open source solutions give you freedom in how you deploy and operate your models, and the ability to fine-tune and retrain your proprietary data deeper and more precisely without sharing it with third-party vendors. You can Open source large language models are transparent and can be more easily customized to specific needs than closed versions. Open source also has an active community and active community support that fosters innovation.
Emerging popular open source models include LlaMa/Llama 2, Bloom, MosaicML MPT-7B, Falcon, GPT-NeoX and GPT-J, PaLM2, Dolly, Vicuna, OPT-175 B, and many more. Contains models. This is an area of continuous innovation and development, and one that could definitely receive a lot more attention.
Open source, large-scale language models that you can own and run mean you can control the data you share and maintain the privacy of your prompts and large-scale language model responses. This has gotten a lot of attention recently, with organizations and government agencies announcing that they are no longer allowing their employees to use hosted LLM models, citing data privacy and security requirements.
Using an open source model can also reduce costs and vendor dependence. Open source large-scale language models can lower costs in the long run, especially if you embed them in many systems, as there are no licensing costs to use the model. However, like all open source technologies, the model can be expensive to implement and run.