Over the past few weeks, I’ve spent some time understanding the promise of generative AI infrastructure. In this article, I aim to provide a clear overview of key components, emerging trends, and highlight early industry players driving innovation. I will explain basic models, computation, frameworks, computations, orchestration and vector databases, fine-tuning, labels, synthetic data, AI observability and model security.
My goal is to understand and explain these concepts in a simple and straightforward way. Additionally, I hope to be able to use this knowledge to invest in future growth.
At the end of this post, I'll connect all of these concepts by explaining how both companies leverage the infrastructure stack in a consistent way.
Large languages and underlying models
Let's start with a large language model. In short, LLM is a computer program trained using large amounts of text and code, including books, articles, websites, and code snippets. The ultimate goal of LLM is to truly understand the meaning of words and phrases and become good at generating new sentences. It is combined with deep learning to achieve this.
Base models are another name for these LLMs and play a vital role as they provide the basis for a wide range of applications. In this study, as the name itself suggests, we will focus most of our efforts on this fundamental aspect.
These models leverage huge data sets to learn a variety of tasks. While they may occasionally make mistakes or display biases, their abilities and effectiveness are constantly improving.
To bring this concept to life, let's consider a practical example. Imagine you are a writer looking for new ideas for a story. By feeding a few words into the model, it can generate a list of potential concepts. I use this to receive suggestions for titles for this article. Likewise, scientists facing a problem can harness the power of the underlying model to discover the information they need from large amounts of data by typing in a few words.
Fundamental models triggered a major shift in the development of artificial intelligence systems. They power chatbots and other artificial intelligence interfaces, and their progress is largely due to self-supervised and semi-supervised learning. Now, what exactly do these terms mean?
In self-supervised learning, models learn from unlabeled data by deciphering word meanings based on frequency and context. Semi-supervised learning, on the other hand, involves training a model using a combination of labeled and unlabeled data. Labeled data refers to instances of data that have specific information assigned to them, such as a dataset with labeled images of bicycles and cars. The model can then use labeled images to differentiate between the two and further refine its understanding of unlabeled images. I’ll dive into the concept of fine-tuning soon.
Now, when it comes to building applications on top of the underlying model, a key consideration arises: Should developers choose an open-source model or a closed model?
The underlying code and architecture of open source AI models are publicly accessible and free for anyone to use, modify, and distribute. This openness fosters a collaborative environment where developers and researchers can contribute to model improvements, adapt it to new use cases, or integrate it into their own projects.
Closed-source AI models, on the other hand, keep their code and architecture private, limiting free access to the public. The use, modification, and distribution of these models are often tightly controlled by the company that developed it. This approach is designed to protect intellectual property, maintain quality control and ensure responsible use. Although external developers and researchers cannot directly contribute to model improvements or adjustments, they can often interact with the model through predefined interfaces or APIs provided by the entity that owns the model.
Choosing between open and closed models can present challenges. Choosing an open source solution means taking on the responsibility of managing infrastructure needs, such as processing power, data storage, and network security, which are typically provided by closed model providers.
In writing this article, I wanted to understand the unique advantages and selling points of these models. Most importantly, I seek insights from builders in the field.
While the perspectives I've encountered may vary, a few key themes emerged when choosing a base model: the precision required by the application, the readiness of the developer team to handle their own infrastructure, and if not enough exploration is done, tend to stick to familiar content. Not done.
First, accuracy is crucial. Depending on what the model needs to accomplish, the tolerance for error may vary. For example, a sales chatbot can handle occasional errors, making it suitable for building on top of existing base models. However, consider the case of self-driving cars, where mistakes could have catastrophic consequences.
Secondly, cloud hosting plays an important role. For agile startups aiming to maintain lean operations, dealing with computing power, data storage, and technical complexity can distract from their core goals. That’s why many startups choose to build on top of off-the-shelf closed-source platforms like Chat-GPT. On the other hand, larger companies with in-house expertise in infrastructure management may prefer the open source route to retain control over all aspects and gain a deeper understanding of the system's outcomes.
Finally, business goals come into play. Different companies have different agendas, which can impact the decision-making process. For example, Zoom has invested in and leveraged Anthropic, a model tailored for enterprise use cases and security. While Anthropic may not have a better system than OpenAI, Zoom may want to avoid the risk of its data being used by OpenAI/Microsoft, which competes with Teams. These strategic considerations play an important role in determining the partner companies choose to build their systems.
The promise of large language models (LLMs) continues to expand. There are some leading models here, such as OpenAI’s GPT4 and DALL-E, Cohere, Anthropic’s Claude, Meta AI’s LLaMA, StabilityAI, MosaicML, and Inflection AI.
OpenAI is a cornerstone of the artificial intelligence industry, known for its advancements in GPT4 and DALL-E. ChatGPT is a closed-source model with an impressive conversational AI interface that enables bots to have complex conversations with people, while DALL-E can generate unique images based on text descriptions.
MosaicML is an open source AI startup that develops a platform for training large language models and deploying generative AI tools. Recently acquired by Databricks, MosaicML's unique open source approach will continue to help organizations create their own language models.
Meta AI’s contribution to the AI field LLaMA is an open source model. By encouraging other researchers to use LLaMA, Facebook aims to stimulate the development of new applications and improve the accuracy of language models.
Known for systems such as Dance Diffusion and Stable Diffusion, StabilityAI is a leader in open source music and image generation systems. Their goal is to inspire global creativity. The company also owns MedARC, a foundational model for medical AI contributions.
Anthropic, a closed-source company co-founded by OpenAI veterans, created Claude, a secure and powerful language model. Claude stands out as a new model for processing data, setting an early benchmark for responsible artificial intelligence.
Inflection is a well-funded AI-based model startup with a bold vision to build “personal AI” for everyone, and its powerful language model recently powered the Pi conversational agent. The company is backed by Microsoft, Reid Hoffman, Bill Gates, Eric Schmidt and Nvidia.
Finally, Canadian startup Cohere offers a reliable and scalable large-scale language model designed for enterprise use. Their models meet the specific requirements of enterprises, ensuring reliability and scalability.
Semiconductors, chips, cloud hosting, inference, deployment
Generative AI models rely on powerful computing resources to train and generate output.
While I started with the basic model, GPUs and TPUs (specialized chips) and cloud hosting do form the basis of the generative AI infrastructure stack.
Computing, the ability to process data (and perform calculations), plays a crucial role in artificial intelligence systems. GPU, CPU and TPU are different types of computing. The most important part of the generative AI stack is the GPU, which was originally designed for graphics tasks but excels at computationally intensive operations such as training networks for generative AI. GPUs are optimized for parallel computing processing, which means breaking up large tasks into smaller tasks that can be handled by multiple processors simultaneously. AI/ML tasks are highly parallelizable workloads, so GPUs make sense.
Generative AI requires massive computing resources and large data sets, which are processed and stored in high-performance data centers. Cloud platforms such as AWS, Microsoft Azure, and Google Cloud provide scalable resources and GPUs for training and deploying generative AI models.
GPU leader Nvidia's market capitalization recently topped $1 trillion, and new entrants like d-Matrix are entering the space, launching high-performance chips for generative AI to aid inference, that is, using trained generative models The process of making predictions on new data. d-Matrix is building a new inference chip that uses digital in-memory computing (DIMC) technology to significantly reduce per-token latency compared to current compute accelerators. d-Matrix believes that solving the problem of memory computing integration is the key to improving AI computing efficiency, thereby handling the explosive growth of inference applications in an efficient and cost-effective manner.
Lambda Labs helps enterprises deploy artificial intelligence models on demand. Lambda provides workstations, servers, laptops and cloud services for power engineers. Recently, Lambda launched GPU Cloud, a GPU cloud service dedicated to deep learning.
CoreWeave is a professional cloud service provider focused on large-scale, highly parallelized workloads. The company has received funding from Nvidia and GitHub founders. Its customers include generative AI companies such as Stability AI, and supports open source AI and machine learning projects.
Additionally, there are dedicated companies dedicated to supporting generative AI. HuggingFace is essentially LLM's GitHub. It provides comprehensive AI computing resources through a collaboration platform called Hub, promoting the sharing and deployment of models on major cloud platforms.
Interestingly, cloud providers are aligning themselves with key foundational model players; Microsoft has invested resources and significant funds in OpenAI, Google has invested in Anthropic and supplemented its Google Brain initiative, and Amazon has formed an alliance with HuggingFace. The conclusion is that AWS's previous dominance based on credibility and innovation is no longer the default option for companies that might want to use one of the specific underlying models.
Orchestration layer/application framework
The next level in the stack is the application framework, which facilitates seamless integration of AI models with disparate data sources, allowing developers to quickly launch applications.
The key takeaway from application frameworks is that they speed up the prototyping and use of generative AI models.
The most famous company here is LangChain, which started out as an open source project and grew into a real startup. They introduced an open source framework specifically designed to simplify application development using LLM. The core concepts of the framework revolve around the concept of “chaining” various components together to create chatbots, generated question answers (GQA), and summaries.
I caught up with founder and CEO Harrison Chase. He said: "Langlian provides two major added values. The first is a collection of abstractions, each abstraction represents a different module required to build a complex LLM application. These modules provide all the integration/implementation within that module Standard interfaces so providers can be easily switched with a single line of code. This helps teams quickly experiment with different model providers (OpenAI vs. Anthropic), vector libraries (Pinecone vs. Chroma), embedded models (OpenAI vs. Cohere), etc. Second The big added value is in the chain - a common way to perform more complex sequences of LLM calls to enable RAGs, digests, etc."
Another player is Fixie AI, founded by former engineering leaders from Apple and Google. Fixie AI is designed to connect text generation models like OpenAI's ChatGPT to enterprise-grade data, systems and workflows. For example, companies can leverage Fixie AI to incorporate language model capabilities into customer support workflows, where agents can process customer tickets, automatically retrieve relevant purchase information, issue refunds as needed, and generate draft responses to tickets.
vector database
The next level up the stack is a vector database, which is a special type of database that stores data in a way that helps find similar data. It does this by representing each piece of data as a list of numbers, called a vector.
These numbers in the vector correspond to features or attributes of the data. For example, if we are working with images, the numbers in the vector might represent the color, shape, and brightness of the image. In vector databases, an important term to master is embedding. Embeddings are a form of data representation that encapsulates semantic information that is critical for AI to understand and maintain long-term memory, which is critical for performing complex tasks. Embeddings are a form of data representation that encapsulates semantic information that is critical for AI to understand and maintain long-term memory, which is critical for performing complex tasks.
This is a concrete example. A picture of a bicycle can be effectively converted into a series of numerical values, including characteristics such as size, wheel color, frame color, and handlebar color. These digital representations facilitate seamless storage and analysis, providing advantages over mere images. The conclusion is that vector databases have the ability to process and store data in a way that is easy for machines to understand.
These databases can be conceptualized as tables with infinite columns.
In my previous experience building conversational AI, I mostly worked with relational databases that store data in tables. However, vector databases are good at representing the semantics of data, supporting tasks such as similarity search, recommendation, and classification.
Several companies have developed vector databases and embeddings.
Pinecone is the creator of the category. They have a distributed vector database designed for large-scale machine learning applications. In addition to the generative AI company, which has customers like Shopify, Gong, Zapier and Hubspot, it offers enterprise-grade solutions with SOC 2 Type II certification and GDPR readiness. GDPR compliance is important because if a developer has to delete a record, it's not that hard to do in the database, but it's much harder to remove bad data from the model because of the way the model is structured. Pinecones also help remember chat experiences.
Another vector database worth noting is Chroma, a new open source solution focused on high-performance similarity search. Chroma enables developers to add state and memory to their AI-enabled applications. Many developers have expressed a desire for AI tools like "ChatGPT but for their data," and Chroma serves as a bridge by enabling embedding-based document retrieval. Since its launch, Chroma has received over 35,000 Python downloads. Additionally, its open source alignment with the goal of making artificial intelligence safer and more consistent.
Weaviate is an open source vector database ideal for companies looking for flexibility. It is compatible with other model centers such as OpenAI or HuggingFace.
fine-tuning
The next layer of the infrastructure stack is fine-tuning. In the field of generative AI, fine-tuning involves further training a model for a specific task or data set. This process enhances the performance of the model and tunes it to meet the unique requirements of that task or data set. Just like how versatile athletes specialize in specific sports to excel in them; broad-based AI can also focus its knowledge on specific tasks through fine-tuning.
Developers build new applications on top of existing models. While language models trained on massive datasets can generate grammatically correct and fluent text, they may lack accuracy in certain fields such as medicine or law. Fine-tuning the model on domain-specific datasets enables it to internalize the unique characteristics of those domains, thereby enhancing its ability to generate relevant text.
This is consistent with the previous point about being a base model for other services and product platforms. The ability to fine-tune these models is a key factor in their adaptability. Fine-tuning an existing model can simplify the process and be cost-effective, rather than starting from scratch (which requires a lot of computing power and a lot of data), especially if you already have a large specific data set.
One well-known company in this field is Weights and Bias.
Label
Accurate data labeling is critical to the success of generating artificial intelligence models.
Data can take many forms, including images, text, or audio. Tags serve as descriptions of data. For example, an image of a bicycle could be tagged "bicycle" or "bicycle." One of the more tedious aspects of machine learning is providing a set of labels to teach the machine learning model what it needs to know.
Data labeling plays an important role in machine learning as algorithms learn from data. The accuracy of labels directly affects the learning ability of the algorithm. Every AI startup or corporate R&D lab faces the challenge of annotating training data to teach algorithms what to recognize. Whether it's a doctor assessing the size of a cancer on a scan or a driver marking a street sign in a self-driving car video, labeling is a necessary step.
Inaccurate data can lead to inaccurate model results.
Data labeling remains a significant challenge and obstacle to the advancement of machine learning and artificial intelligence in many industries. Allocating time for this is costly, labor-intensive, and challenging for subject experts, leading some to turn to crowdsourcing platforms with minimal restrictions on privacy and expertise. It is often considered a "cleaning" job, although the data ultimately controls the behavior and quality of the model. In a world where most model architectures are open source, private, domain-relevant data is one of the most powerful ways to build an AI moat.
Snorkel AI is a company that speeds up the labeling process. The company's technology began as a research initiative at the Stanford Artificial Intelligence Laboratory to overcome the labeling bottleneck of artificial intelligence. Snorkel’s platform helps subject matter experts label data programmatically (via a technique called “weak supervision”) rather than manually (one-by-one), putting humans in the loop while significantly increasing labeling efficiency. This can shorten the process from months to hours or days, depending on the complexity of the data, and make the model easier to maintain in the long run because as the data drifts, new error patterns are discovered, or the business can Easily revisit and update training labels. Goals change.
Alex Ratner, co-founder and CEO of Snorkel AI, said: “Behind every model-centric operation like pre-training and fine-tuning are more important data-centric operations that create the data that the model actually learns from. .” “Our goal is to make data-centric AI development feel less like a manual, ad-hoc effort and more like software development so that every organization can develop and maintain models that work with their enterprise’s specific data and use cases. ” Snorkel’s data-centric platform also helps systematically identify model errors so labeling efforts can be focused on the most impactful pieces of data. Today, Fortune 500 companies use it in data-intensive industries such as finance, e-commerce, insurance, telecommunications, and pharmaceuticals.
Labelbox is a leading artificial intelligence labeling company. I spoke with CEO Manu Sharma. Labelbox helps companies like OpenAI, Walmart, Stryker, and Google label data and manage processes. "Labelbox makes basic models useful in an enterprise environment". Developers use Labelbox's model-assisted labeling to quickly transform model predictions into new automatically labeled training data for generating AI use cases.
Other companies develop interfaces and labor specifically for performing manual annotation. One of them is scale, focusing on government agencies and businesses. The company provides a visual data labeling platform that combines software and human expertise to label image, text, speech and video data for companies developing machine learning algorithms. Scale employs tens of thousands of contractors to label data. They initially provided labeled data to autonomous vehicle companies and have expanded their customer base into government, e-commerce, enterprise automation, and robotics. Customers include Airbnb, OpenAI, DoorDash and Pinterest.
Comprehensive data
Synthetic data, also known as artificially created data that mimics real data, offers a variety of benefits and applications in the fields of machine learning and artificial intelligence (AI). So why should you consider using synthetic data?
A major use case for synthetic data occurs when real data is unavailable or cannot be leveraged. By generating artificial datasets with the same characteristics as real data, you can develop and test AI models without compromising privacy or encountering data limitations.
There are many advantages to using synthetic data.
Synthetic data protects privacy because it lacks personally identifiable information (PII) and HIPAA risks. Ensure compliance with data regulations such as GDPR while utilizing data effectively. It enables scalable machine learning and artificial intelligence applications by generating data for training and deployment. Synthetic data enhances diversity, minimizes bias by representing diverse populations and scenarios, and promotes fairness and inclusion in AI models. “Conditional data generation” techniques and synthetic data can also solve the “cold start” problem for startups that don’t have enough data to test and train models. Companies will need to synthesize proprietary datasets and then enhance them using conditional data generation techniques to fill in edge cases that they can't collect in the wild; this is sometimes called the "last mile" of model training.
When it comes to synthetic data solutions, there are several companies that offer solid options. Gretel.ai, Tonic.ai, and Mostly.ai are noteworthy examples in this space.
Gretel.ai allows engineers to generate artificial datasets based on real datasets. Gretel combines generative models, privacy-enhancing technologies, and data metrics and reporting to enable enterprise developers and engineers to create accurate and secure domain-specific synthetic data on demand. All three founders have a background in cybersecurity and have served in various roles in the U.S. intelligence community, and their chief technology officer is an enlisted officer in the Air Force.
Tonic.ai, for example, promotes its data as “real fake data,” emphasizing the need for synthetic data to respect and protect the privacy of real data. Their solutions are suitable for software testing, machine learning model training, data analysis, and sales presentations.
Model Supervision/AI Observability
The next level of the stack is AI observability, which involves monitoring, understanding, and interpreting the behavior of AI models. In short, it ensures that AI models function properly and make fair, harmless decisions.
Model supervision is a subset of AI observability that focuses specifically on ensuring that AI models are fit for purpose. It involves verifying that the model is not making potentially harmful or unethical decisions.
Data drift is another important concept to consider. It refers to changes in data distribution over time, which can cause AI models to become less accurate. If these changes favor certain groups, the model may become more biased and lead to unfair decisions. As the data distribution changes, the accuracy of the model decreases, potentially leading to erroneous predictions and decisions. AI observability platforms provide solutions to these challenges.
To shed some light on the need for observability in AI, I reached out to Krishna Gade and Amit Paka, CEO and COO of Fiddler.ai. Gade previously served as head of engineering for Facebook News Feed, where he saw firsthand the challenges companies face in understanding their own machine learning models.
“As these systems become more mature and complex, it becomes extremely difficult to understand how they work. Questions like ‘Why am I seeing this story in my feed? Questions like, "Why did this news story go viral? Is this news true or false? It's hard to answer." Gade and his team at Fiddler developed a platform to address these questions, increase transparency of Facebook's models, and Solving the “AI black box” problem. Now, Krishna and Amit Paka have launched the Fiddler platform, which helps companies like Thumbtack and even In-Q-Tel (the CIA's venture fund) provide model interpretability, modern monitoring and bias detection, giving enterprises a centralized way to manage these. information and building the next generation of artificial intelligence. Amit shared with me: “AI observability has become very important for safe and responsible AI deployment. It has now become a must-have for every company launching an AI product. We believe that without AI observability sex, we wouldn’t have enterprises adopting AI, and AI observability is forming a critical third layer in the AI stack.”
Arize and WhyLabs are other companies that have created powerful observability solutions for LLM in production. These platforms solve the problem of adding guardrails to ensure appropriate prompts and responses to LLM applications in real time. These tools can identify and mitigate malicious prompts, sensitive data, toxic reactions, problematic themes, hallucinations, and jailbreak attempts in any LLM model.
Aporia is another company emphasizing the importance of AI observability platforms, recognizing that trust can be lost in seconds and take months to regain. Aporia specializes in customer lifetime value/dynamic pricing and is currently delving into generative AI using its LLM observability capabilities.
model safety
The top of the stack is model safety. A significant risk with generative AI is biased output. AI models tend to adopt and propagate biases present in the training data. For example, an AI resume screening tool favored candidates named “Jared” with high school lacrosse experience, revealing bias in the data set. Amazon faced a similar challenge, with their AI resume screening tool exhibiting an inherent bias against male candidates due to the training data consisting primarily of male employees.
Another concern is the malicious use of AI. Deepfakes, which involve spreading false information through credible but fabricated images, videos or texts, can become a problem. A recent incident involving an AI-generated image of an explosion at the Pentagon caused fear and confusion among the public. This highlights the potential for AI to be weaponized by misinformation, and the need for safeguards to prevent such misuse.
Additionally, as AI systems grow in complexity and autonomy, unintended consequences may arise. These systems may exhibit behaviors that are not anticipated by their developers, creating risks or leading to undesirable outcomes. For example, chatbots developed by Facebook began inventing their own language to communicate more effectively, an unintended consequence that underscores the need for strict monitoring and safety precautions.
To mitigate these risks, techniques such as bias detection and mitigation are critical. This involves identifying biases in model output and taking steps to minimize them, such as increasing training data diversity and applying fairness techniques. User feedback mechanisms (where users can flag problematic output) play a crucial role in improving AI models. Adversarial testing and validation challenge AI systems with difficult inputs to uncover weaknesses and blind spots.
Powerful intelligence helps enterprises stress-test their AI models to avoid failure. Robust Intelligence's primary product is the AI Firewall, which protects companies' AI models from bugs through continuous stress testing. Interestingly, this AI firewall itself is an AI model tasked with predicting whether data points will lead to false predictions.
Arthur AI debuted in 2019 with the primary goal of helping enterprises monitor their machine learning models by providing an LLM firewall similar to the Robust Intelligence solution. The solution monitors and enhances model accuracy and interpretability.
CredoAI guides businesses in understanding the ethical implications of artificial intelligence. Their focus is on AI governance, enabling enterprises to measure, monitor and manage risks arising from AI at scale.
Finally, Skyflow provides API-based services for secure storage of sensitive and personally identifiable information. Skyflow is focused on meeting the needs of various sectors such as fintech and healthcare, helping to securely store critical information such as credit card details.
How does it all come together?
To gain a deeper understanding of the leading companies using these tools, I spoke with Science CEO Will Manidis. io. ScienceIO is revolutionizing the healthcare industry by building state-of-the-art foundational models purpose-built for healthcare. Hundreds of the most important healthcare organizations use ScienceIO models at the core of their workflows, giving Will unique insights into how to deploy LLM in production. This is what he saw:
- Compute: ScienceIO relies on Lambda Labs to leverage local clusters for its computing needs. This ensures efficient and scalable processing power that is more cost-effective than hyperscale services such as AWS or GCP.
- Base Model: ScienceIO creates its own base model using its internal data. The core of their business is an API that facilitates the real-time transformation of unstructured medical data into structured data (named entity resolution and linking), which can then be used for search and analysis purposes. Many of their customers choose to link ScienceIO with more general models in their workflows to perform tasks such as information retrieval and synthesis.
- Vector: One of ScienceIO's core products is the embedding product, built specifically for high-quality embedding in the healthcare space. One of Will's core beliefs is that custom embeddings will become increasingly important, especially as a complement to generic models. ScienceIO uses Chroma extensively to store and query these vector embeddings.
- Orchestration: For application development, ScienceIO relies on LangChain. Internal model storage, versioning and access are powered by Huggingface.
- Fine-tuning: While ScienceIO's core base model was specifically trained from scratch on healthcare data, meaning they've never seen piles of junk social media data or anything like that, many customers are interested in giving it additional fine-tuning Example. ScienceIO has launched Learn & Annotate, their nudge and human-computer interaction solution to address these use cases.
I also spoke with Pedro Salles Leite, CEO of Innerplay, a company that uses artificial intelligence to help people and companies become more creative. Innerplay helps companies create videos in a faster way, including script creation.
Pedro has been researching and building artificial intelligence use cases for eight years. Regarding his infrastructure stack, he said his job is to make sure the product makes sense to users... rather than setting up the orchestration or underlying model - which just adds another level of complexity. Here is his stack:
- Base models: Innerplay uses 14 different base models to turn ideas into reality. They use a closed model mainly because "there are no GPUs until the product is market fit".
- Vector database: Innerplay uses a vector database to perform tasks such as processing PDF documents. They generate scripts from PDFs and require a vector database to do this.
- Fine-tuning: Innerplay is a big believer in fine-tuning. The company prepares data sets manually but plans to use artificial intelligence to prepare the data for future fine-tuning.
- Prototyping: They use it to evaluate output and compare models. Spellbook by Scale is often used to quickly test iterations in a machine learning process before moving into a Python/production environment.
- AI Observability: They are now starting to consider AI observability to improve their AI in a privacy-focused manner. As a content creation platform. "Innerplay needs to make sure people use it for good," Pedro said.
in conclusion
The exploration of generative AI infrastructure has only scratched the surface, and the rapid progress in technology development and investment in underlying infrastructure components is remarkable. Companies like MosaicML being acquired for staggering amounts of money, and the number of players in the space continuing to grow, demonstrates the tremendous value and interest in this space.
This is a complex and evolving scenario with multiple layers, from base model to fine-tuning, from semiconductor to cloud hosting, from application framework to model supervision. Each layer plays a crucial role in harnessing the power of generative AI and making it applicable across industries. In this study, many companies that started in one area expanded into other areas.