Why Databricks Bought Mosaic and The Rapidly Decreasing Costs of AI


Hatched by Glasp

Sep 17, 2023

6 min read


Why Databricks Bought Mosaic and The Rapidly Decreasing Costs of AI

In recent years, there has been a significant decrease in the costs of AI, particularly in the realm of training and fine-tuning models. This has been made possible by companies like MosaicML, whose vision is to make it cost-effective for companies to train and refine their own models. This aligns perfectly with Databricks' mission to help companies rapidly adopt machine learning and stay ahead of the competition.

The decrease in training costs has been remarkable, with a 10x reduction in less than a year. For example, it now costs $50k to train stable diffusion and $200k to train a high-quality LLM. In comparison, Stability, another company, reported that it cost them $600k to train stable diffusion. MosaicML has published several papers demonstrating their ability to train stable diffusion for just $50k. This decrease in training costs can be attributed to two main factors.

First, companies like MosaicML are making significant algorithmic improvements. These advancements allow for more efficient and cost-effective training of models. Second, the cost of GPUs has also decreased by 3x in approximately three years. For instance, in August 2019, the cost of a Nvidia T4 GPU for one hour was $0.95. Today, that same GPU is available for just $0.35 per hour. These reduced GPU costs contribute to the overall decrease in training expenses.

The implications of these decreasing costs are significant. Inference costs, for example, have dropped by 10x in just 16 months. This means that the cost of processing 1k tokens went from $0.006 for Curie generations to $0.0005 for Curie quality generations today. For most SaaS companies at scale, the target is to achieve 80% gross margins or keep COGS (cost of goods sold) at 20% of revenue. Typically, software companies spend around 50% of their COGS on cloud costs, which means that cloud costs should account for approximately 10% of revenue. However, AI features come in addition to regular cloud costs, making it crucial for companies to find cost-effective solutions.

The decrease in training costs is driven by both the reduced prices of GPUs and the efficiency of algorithms. This trend is likely to continue, leading to more model providers and increased competition at the model layer. As training costs become more affordable, companies may be more inclined to opt for open-source model providers rather than closed-source ones. This shift can further drive down costs and provide more options for companies looking to adopt machine learning.

People Who Have “Too Many Interests” Are More Likely To Be Successful According To Research

Contrary to the common belief that specialization is the key to success, research suggests that individuals with a wide range of interests, known as generalists or polymaths, are more likely to achieve success. Many of the most impactful individuals throughout history, such as Elon Musk, Steve Jobs, Richard Feynman, Ben Franklin, Thomas Edison, Leonardo Da Vinci, and Marie Curie, were polymaths who excelled in multiple domains.

If we examine the founders of the five largest companies in the world - Bill Gates, Steve Jobs, Warren Buffett, Larry Page, and Jeff Bezos - we find that they are all polymaths who have also followed the 5-hour rule. This rule emphasizes the importance of continuous learning and dedicating time to acquire new skills or knowledge. To achieve something extraordinary in life, individuals have two paths: becoming the best at one specific thing or becoming very good at two or more things.

The future belongs to integrators, as stated by educator Ernest Boyer. Modern polymaths are individuals who become competent in at least three diverse domains and integrate them into a top 1-percent skill set. They bring together the best of what humanity has discovered from various fields to enhance their effectiveness in their core field. By embracing a wide range of interests, polymaths can leverage their diverse skill sets to achieve exceptional results.

The advantages of being a polymath are numerous. Firstly, combining two or more skills, even at a competent level, can lead to the development of a world-class skill set. Secondly, creative breakthroughs often arise from making atypical combinations of skills. Studies have shown that the most successful projects cite a mix of conventional citations from their own field and citations from other fields. This interdisciplinary approach fosters innovation and unique problem-solving.

Moreover, it has become easier and faster than ever to acquire new skills. The abundance of free or affordable content from top experts in various mediums allows individuals to learn at their own pace. Standing on the shoulders of giants, as famously stated by Isaac Newton, polymaths can build upon the knowledge and expertise of those who came before them.

Being a polymath also positions individuals to pioneer new fields, industries, or skill sets. Combining existing knowledge with new ideas is a common way to drive innovation and create something truly groundbreaking. Additionally, the number of new academic fields and business industries is growing exponentially, providing ample opportunities for polymaths to make their mark.

Polymaths are well-equipped to solve complex problems due to their diverse skill sets and ability to connect seemingly unrelated disciplines. This multidimensional approach allows them to consider problems from various perspectives and develop innovative solutions. In a global economy where competition is fierce, being a polymath sets individuals apart and increases their chances of success.

To fully embrace the polymath mindset, it is essential to cultivate original thinking. Asking oneself what unique belief they hold that no one else agrees with can stimulate creativity and originality. Challenging the consensus view opens doors to new possibilities and opportunities. As billionaire investor Peter Thiel suggests, "You can't make money agreeing with the consensus view."

In conclusion, the decreasing costs of AI training and the advantages of being a polymath are two significant trends that have the potential to shape the future. Companies like Databricks and MosaicML are driving down training costs, making it more accessible for companies to adopt machine learning. On the other hand, individuals who embrace a wide range of interests and develop multidisciplinary skills are more likely to achieve success in today's fast-paced and interconnected world.

Three actionable pieces of advice can be derived from these trends:

  • 1. Embrace continuous learning and explore diverse interests: By cultivating a wide range of skills and knowledge, individuals can develop a unique skill set that sets them apart from others. This can lead to increased opportunities and success in a global economy.
  • 2. Stay updated on advancements in AI and machine learning: As the costs of training models decrease and new algorithmic improvements emerge, it is crucial to stay informed about the latest developments. This knowledge can help individuals and companies make informed decisions and leverage the benefits of AI.
  • 3. Foster interdisciplinary collaboration: By connecting with professionals from various fields, individuals and organizations can foster innovation and solve complex problems. Embracing diverse perspectives and integrating knowledge from different domains can lead to breakthroughs and competitive advantages.

By combining these actionable pieces of advice with the understanding of the trends discussed, individuals and companies can navigate the rapidly changing landscape of AI and maximize their potential for success.

Hatch New Ideas with Glasp AI 🐣

Glasp AI allows you to hatch new ideas based on your curated content. Let's curate and create with Glasp AI :)