OpenAI's strongest "enterprise version" bombing field, will the B-side large model market "winner takes all"?

Double the speed, secure encryption, call GPT-4 without cap, this morning, OpenAI's officially announced enterprise-specific version of ChatGPT is called the "strongest ChatGPT version"!

Compared with the regular version of ChatGPT, the enterprise version has more powerful performance, including: providing unlimited GPT-4 access, 2x faster speed, supporting unlimited advanced data analysis, supporting 32k tokens context window for 4x input and files, providing shareable chat templates and free API interfaces, ensuring enterprise data privacy and security, etc.

In addition to the current version for large enterprises, OpenAI will soon launch a self-service ChatGPT Business product for all types of small teams, serving organizations of all sizes and types. That is to say, starting today, OpenAI has sounded the clarion call to attack the B-end market in an all-out way.

Naturally, this brings up a question: will the B-side AI big model market for enterprises be "winner-take-all"?

Recently, Guru Chahal, a partner at Lightspeed Venture Partners, analyzed the issue in depth.

The author believes that the most likely path on the B side is for enterprises to use large models in the exploration stage, and gradually shift to smaller, specialized (adjustment + refinement) models in the production stage as the understanding of large models deepens in actual use.

Chahal also mentioned the factors that enterprises need to consider when choosing a model, as well as the opportunities for AI infrastructure, including evaluation frameworks, model operation and maintenance, enhancement systems, operational tools, data utilization, and more.

The article is very dry, I believe it will be of great benefit to friends who want to understand the B-side AI market, AI infrastructure, future opportunities, etc.

The following is the full text, everyone enjoy~ ✌️

Directory:

● Large model ecosystem classification

● Match use cases to models

● Where are the future opportunities?

Over the past decade or so, as part of the Lightspeed team, I have experienced amazing innovation in artificial intelligence and machine learning, thanks in large part to our deep collaboration with exceptional entrepreneurs.

Now, we're working with their companies, the platforms they've built, and the customers they serve to get a more systematic picture of how enterprises think about generative AI.

Specifically, we delved into the big model ecosystem, trying to explore questions such as "Will the most powerful model take all the winner?" Will enterprises rely on OpenAI's APIs when using them, or will they choose more diverse real-world use cases?" "This type of problem.

The answers to these questions will determine the growth direction of this future big model ecosystem, as well as the flow of computing power, talent and capital.

Large model ecosystem classification

Based on our research, we believe that the field of artificial intelligence is experiencing a "Cambrian-style" model explosion. In the future, developers and enterprises will choose the most suitable model according to actual needs, although the use in the exploration phase may be more concentrated.

The most likely path on the B side is for enterprises to use large models in the exploration stage, and gradually shift to smaller, specialized (adjustment + refining) models in the production stage as the understanding of large models deepens in actual use.

The diagram below shows our view of the evolution of the ecosystem of the base model.

We believe that the field of AI models can be divided into three main but somewhat cross-cutting categories:

Category 1: Giant brain models

These are the best models and pioneers in the field of models. They produced amazing presentations that caught our attention. When developers try to explore the limits of AI's potential for their applications, these models are often the default starting point.

These models are expensive to train and complex to maintain and scale. But the same model can tackle the Law School Admission Test (LSAT), the Medical School Admission Test (MCAT), write high school essays, and interact with you like a chatbot friend. Currently, developers are experimenting on these models and evaluating AI usage in enterprise applications.

It is important to note that these models are expensive to use, have high inference latency, and can be overly complex in well-defined constrained use cases.

At the same time, these models are generic models that may not be accurate enough for specialized tasks (see, for example, comprehensive studies by Cornell University et al.).

Moreover, they are also black boxes that can pose privacy and security challenges for businesses, which are exploring how to exploit these models without exposing data.

OpenAI, Anthropic, and Cohere all fall into this category.

Category 2: Challenger model

These models are equally highly capable, second only to leading models. Llama 2 and Falcon are the best in this category. They are usually as good as the "N-1" or "N-2" models in the Category 1 model.

According to some benchmarks, the Llama 2 is even comparable to GPT-3.5-turbo. By tuning on enterprise data, these models can be comparable in their capabilities to specific tasks to those in Category 1.

Many of these models are open source (or very close). Once released, they tend to be quickly improved and optimized by the open source community.

Category 3: Long-tail models

These are "expert" models. They are built for specific goals, such as classifying files, identifying specific attributes in images or videos, identifying patterns in business data, and more. These models are flexible, inexpensive to train and use, and can be run in the data center or at the edge.

Just browse Hugging Face to get a glimpse of the breadth of this ecosystem, which will continue to expand in the future as it serves a variety of use cases!

Match use cases to models

Although it's still early days, we're already seeing some leading development teams and enterprises start thinking about this ecosystem in this granular way. They are eager to match use cases to the most appropriate models, and may even use multiple models for more complex use cases.

When choosing which/which models to use, the following factors are usually considered:

a. Data privacy and compliance requirements, which affect whether the model runs in the enterprise infrastructure or whether data can be sent to an externally hosted inference endpoint.

b.Whether the ability to fine-tune the model is critical to this use case or is strongly desirable.

c. The desired level of inference "performance" (latency, accuracy, cost, etc.).

The actual list is often longer than the above, reflecting the diverse use cases developers want to address with AI.

Where are the opportunities

This emerging ecosystem has had several important impacts:

(1) Evaluation framework: Companies will need tools and expertise to evaluate which model fits which use case.

Developers need to decide how best to evaluate whether a particular model is suitable for the "required work". The evaluation needs to consider several factors, including not only model performance, but also cost, level of control that can be exercised, and so on.

(2) Run and maintain models: Platforms are expected to emerge to help enterprises train, fine-tune and run models, especially the third type of long-tail models.

These platforms used to be often referred to as ML Ops platforms, and we expect this definition to expand to include generative AI. Platforms such as Databricks, Weights and Biases, Tecton, and others are rapidly moving in this direction.

(3) Enhanced systems: Models, especially managed LLM (Retrieval Augmentation Model), need to deliver superior results through enhanced generation.

This involves making secondary decisions, including:

o Data and metadata ingestion: How to connect structured and unstructured enterprise data sources, and then ingest data and metadata about access policies, and more.

o Generate and store embeddings: Select the model used to generate embeddings for your data. Then, how do you store these embeddings: which vector database to choose based on the desired performance, scale, and functionality?

Here, there is an opportunity to build RAG (Retrieval Enhanced Generation) platforms for the enterprise to simplify the complexities that come with selecting and combining these platforms:

(1) O&M tools: Enterprise IT departments need to establish regulatory measures for engineering teams to manage costs, etc.

As with all the work done for software development today, they need to expand these tasks to include the use of artificial intelligence. Areas of interest for IT include:

o Observability: How does the model perform in production? Has their performance improved/deteriorated over time? Are there usage patterns that might influence model selection in future application versions?

o Security: How to ensure the security of AI native applications. Are these applications vulnerable to new ways of attacking and requiring new platforms?

Compliance: We anticipate that the use of AI native applications and LLM will need to comply with the framework that the relevant authorities have begun to develop. This is in addition to existing compliance regimes such as privacy, security, consumer protection, fairness, etc. Businesses will need platforms to help them stay compliant, conduct audits, generate proof of compliance, and more.

(2) Data: The rapid adoption of platforms that help businesses understand their data assets and how to extract maximum value from those assets through the use of novel AI models is expected to emerge.

As one of the largest software companies on the planet once said to us, "Our data is our trench, our core IP, our competitive advantage." ”

Monetizing this data by leveraging artificial intelligence in a way that drives "differentiation" without weakening defenses will be key. Platforms like Snorkel play a key role in this.

We believe that now is the perfect time to build an AI infrastructure platform.

While the adoption of AI will continue to transform entire industries, enabling every enterprise to adopt this powerful technology will require supporting infrastructure, middleware, security, observability, and operational platforms.

Source: Hard AI

This article is from Wall Street News, welcome to download the APP to see more