20VC Are Foundation Models Becoming Commoditised Do OpenAI and Anthropic of the World Have a Sustaining Moat Why Smaller Models May Work Better Why Incumbents with Data Power Win the AI War with Christian Kleinerman, SVP Product @ Snowflake

In a deep dive on generative AI with Harry Stebbings on 20vc, Christian Kleinerman, SVP of Product at Snowflake, discusses the democratization of data access and the future of AI in the enterprise. Kleinerman, with a robust background at Google and Microsoft, emphasizes the importance of talent and scalability in startups and the simplicity of products. He predicts that generative AI will significantly impact creative industries first and foresees a shift in the value of UI with AI advancements. He also addresses the challenges of model adoption in enterprises, the importance of data strategy, and the potential of AI to boost productivity across sectors. The conversation touches on concerns about the commoditization of AI models, the need for clarity in product decisions, and the role of incumbents with substantial data in driving the AI industry forward.

• Small companies can create models comparable to those by OpenAI or Anthropic for certain use cases.
• The commoditization of data and models continues until the next major innovation.
• Model size impacts cost and latency, with smaller models potentially being more efficient.

"We've seen companies that with seven employees have creating models that are comparable for some use cases to what OpenAI or anthropic do."

This quote emphasizes that even small teams can compete with industry giants in creating effective AI models, highlighting the democratization of AI technology.

• Christian Kleinerman is the SVP of Product at Snowflake.
• His background includes senior positions at Google (YouTube) and Microsoft.

"For this deep dive on generative AI is Christian Kleinerman, SVP of product at Snowflake."

The quote introduces Christian Kleinerman as the guest, setting the stage for a discussion on his experience and insights into generative AI.

• Christian has extensive experience in data systems, including roles at Google and Microsoft.
• His understanding of data and appreciation for Snowflake's technology led him to his current position.

"I joined Microsoft, did a long stint in data all the time... And from then I went over to YouTube at Google where I was responsible for the infrastructure, including data systems."

This quote provides a brief overview of Christian's career path, emphasizing his deep involvement with data systems and infrastructure.

• Compromising on talent can negatively impact outcomes.
• Building a scalable business is challenging, as evidenced by Christian's startup experiences.

"I would say talent being the driver of truly great outcomes... And the other thing that has been very clear is building a scalable business is difficult."

Christian stresses the importance of hiring talented individuals and the challenges of creating a scalable business model, drawing from his startup experiences.

• Simplicity in products is key to gaining a following.
• Consumer products are influenced by factors beyond technical challenges, such as timing and consumer behavior trends.

"I think I got the ease of use and value of simplicity in products... Consumer products have many more elements beyond just technical difficulty."

These quotes highlight the importance of simplicity in product design and the complex nature of consumer product success.

• Making products as simple as possible is generally beneficial.
• Over-simplification can be an issue, but simplicity tends to enhance user experience.

"I would say yes, you'll say, all things being equal, there are points where you will oversimplify. But I do think that make things as simple as possible and no more."

Christian argues for simplicity in product design, acknowledging there is a balance to be struck to avoid over-simplification.

• Products that work as advertised and have low latency offer a "magical" user experience.
• Simplicity and reliability are crucial in product design.

"Of making something work as advertised make a very big difference. So simplicity is part of it. Even things like latency make a big difference."

Christian would advise his younger self to focus on creating products that deliver on their promises and are simple to use.

• There is hype and FOMO around generative AI, but underlying innovation is significant.
• Generative AI has the potential to transform human-computer interactions.

"I would say that for sure there is hype... But if you look past that, there is fundamental innovation there."

Christian acknowledges the hype surrounding generative AI but also recognizes the substantial innovation and potential it holds.

• Generative AI is seen as having an impact on the scale of the internet or mobile technology.
• It is expected to enhance various aspects of life and business.

"I think it's comparable. I think it's of the scale of the Internet. I think it's of the scale of mobile..."

This quote compares the expected impact of generative AI to that of other major technological revolutions like the internet and mobile.

• Generative AI provides a boost to creative industries.
• Adobe's integration of generative AI into their products is praised.
• Generative AI has potential applications across all industries.

"I would say that this is a real shot in the arm to the creative businesses... So creative industries are probably the sweetest spot."

Christian highlights the particular benefits of generative AI for creative industries, citing Adobe as an example of successful integration.

• There is a lack of enterprise education on how to implement AI.
• Implementation services could be a significant business opportunity in aiding enterprises to integrate AI.

"You know what I'm finding though, because... They've got no freaking clue how to do it, Christian."

This dialogue reflects the gap in knowledge within enterprises on implementing AI, suggesting a market for educational and implementation services.

• Data maturity directly correlates with the speed of AI adoption across various industries.
• Financial services are at the forefront due to their long-standing ability to organize and leverage data.
• Retail and Consumer Packaged Goods (CPG) companies are also proficient in using data.
• The public sector may be slower to adopt AI due to regulatory constraints and data savviness.

"I think it completely correlates with data maturity... From that perspective, I would say financial services are at the forefront... Retail and CPG companies have also been very wise at using data... Maybe I would point at public sector."

This quote emphasizes the relationship between an industry's data maturity and its readiness to adopt AI technologies, highlighting financial services and retail as leaders and the public sector as slower due to regulatory challenges.

• The alignment of incentives is crucial for AI adoption, especially in public sectors like healthcare.
• Current AI use cases are more about enhancing productivity rather than replacing jobs.
• The potential misalignment of incentives may become more prominent in the future as AI capabilities advance.

"I think of most of the use cases right now are around productivity boost or assistance copilots as opposed to replacement."

The quote discusses the current state of AI use cases in public services, which focus on aiding productivity rather than replacing human workers, suggesting that incentives for adoption should currently align with organizational goals.

• Generative AI (GenAI) has the potential to simplify data access and understanding for non-technical business users.
• Traditional Business Intelligence (BI) required some level of technical knowledge, which GenAI could circumvent.
• GenAI could allow natural language processing to facilitate data retrieval and insights, making data more accessible to everyone.

"I think Genai has the opportunity to turbocharge this type of translation where the language is natural language and the answers come in natural language."

This quote highlights the transformative potential of GenAI in bridging the gap between complex data and business users by facilitating interactions through natural language.

• There is a distinction between public data and private enterprise-owned data.
• The current trend is towards reconsidering data policies due to the use of enterprise data in training AI models.
• Changes in data policy may affect the economics of data usage and AI model training.

"The interesting trend to watch there is the notion of many companies realizing that their data is being used and monetized by these models."

The quote reflects on the emerging awareness among companies about the use of their data in AI models and the potential shifts in data policy that may result from this realization.

• Data is considered significantly more valuable than the models themselves in the context of AI.
• The differentiation in AI is shifting towards unique data sets rather than the technology used to build models.
• There is speculation on whether the public data available for model improvement may eventually be exhausted.

"The vast majority goes to data, 90 plus percent... that is becoming less a differentiating aspect. And what's becoming bigger is data."

This quote signifies the speaker's belief in the overwhelming value of data over AI models, suggesting that unique data is the key differentiator in the field.

• The next big innovation could involve advancements in computer vision and multimodal AI, integrating images, speech, and text.
• Both established AI companies and new startups are likely to pursue these innovations.
• There is a distinction between startups that merely wrap existing AI models and those that contribute to core technological advancements.

"Do you think what has happened for language models is coming for computer vision, for images democratization?"

The speaker anticipates that advancements similar to those in language models will occur in the field of computer vision, suggesting a potential area of significant innovation.

• Model size is important for generic consumer products that require extensive knowledge.
• In specialized enterprise use cases, smaller, fine-tuned models may be more efficient and cost-effective.
• There is ongoing research into model compression to improve latency and cost, regardless of advancements in compute efficiency.

"For specialized use cases, which is what I see more in the enterprise model matters less model size."

This quote conveys the idea that in specific enterprise applications, the size of the AI model is less critical than its ability to perform effectively for the intended purpose.

• The cost of training AI models is a significant barrier, with some models costing millions to train.
• There is a possibility for democratization in training costs, allowing startups to fine-tune and train models themselves.
• The future may see a balance between reducing model sizes and improving compute efficiency to manage training costs.

"How do we think about the cost of training changing over time?"

This question raises the issue of the high costs associated with training AI models and whether these costs will become more manageable for smaller entities in the future.

• The cost of training AI models is expected to decrease due to advancements in compute efficiency and the reinvention of core training processes.
• A common subset of data might be used for foundational models, with fine-tuning on top to reduce costs.
• Good results have been observed using the approach of leveraging a common subset and fine-tuning.

"All of this, the compute is trending down. But the other piece is there's a lot of reinvention of the core training." "Are there ways to take a common subset and then use fine tune on top of it and avoid the cost? I think we've seen reasonably good results under that path."

These quotes highlight the trend of decreasing costs in AI training and the potential strategy of using a common data subset for foundational models to further reduce expenses.

• The longevity of AI models is questioned with predictions that current models may not be used in a year.
• A distinction is made between versions of models and their core types, suggesting that foundational models may persist while versions evolve.
• Continuous innovation and refinement in AI models are expected.

"Depends on how you define a model." "But the reality is there will be new models and new refinements on an ongoing basis."

These quotes discuss the nuanced view of AI model longevity, emphasizing the difference between model types and versions, and the constant evolution in the field.

• Flexibility in transitioning between AI models is crucial for companies to maintain competitiveness.
• Building systems with optionality allows companies to adapt to new models and innovations.
• Companies should avoid being too tightly coupled to a single model to preserve future flexibility.

"100% agree. There is so much innovation in the landscape of models that anyone that builds too tightly coupled to a given model is sort of giving up optionality for the future."

This quote underscores the importance of flexibility and optionality in AI systems, given the rapid pace of innovation in the field.

• An abstraction layer for AI models is recommended to facilitate the transition between different models.
• This layer would handle the translation of specific application requests to the appropriate model, accounting for each model's unique characteristics.

"You should have a model abstraction layer that knows how to translate a specific request that your application needs to do to a model with its intrinsies or specific characteristics."

The quote suggests that a model abstraction layer is essential for managing the transition between different AI models and their unique responses to prompts.

• Correctness and dependability of AI-generated answers are primary concerns.
• Secondary issues include security, privacy of data, and rights to the answers provided by AI models.
• Complex legal questions arise regarding the ownership and rights to AI-generated outcomes, especially in financial contexts.

"There are a lot of issues. Probably the most obvious one is around the correctness and dependability of answers." "A Wall street company asked me if we feed a number of portfolio trading strategies into a model and it makes a recommendation, and the recommendation makes money, could anyone have claims on that answer?"

These quotes highlight the multifaceted challenges associated with AI model adoption, ranging from the accuracy of AI responses to legal implications of AI-generated results.

• Enterprises seek to gain AI benefits without compromising data security.
• Solutions include creating private and secure endpoints for AI models that are close to the data sources.
• Major cloud providers are developing services to facilitate secure access to AI models.

"Data evolve. There are platforms where it's easy to bring llms to the data, as opposed to send large data volumes to where the llms are." "The trend is create private and secure endpoints that can run close to your data and by implication, not only don't have to move and copy a lot of data, but more important, there are some assurances on what is done with your data."

These quotes discuss the evolution of data platforms and the trend towards secure, localized endpoints for AI models, addressing the concern of data privacy and security.

• Enterprises are concerned about potential lawsuits related to the use of AI in their products and services.
• Microsoft's commitment to support customers from a copyright perspective has been a significant assurance for enterprises.
• This commitment may influence other providers to offer similar assurances to promote enterprise adoption of AI.

"Because enterprises are worried about if they incorporate Gen AI into any of their products or services in a way that they truly depend on them. And at some point lawsuits start to fly everywhere they're exposed." "Microsoft statement is immaterial in alleviating those concerns that are very real."

The quotes reflect the concerns of enterprises regarding the legal implications of incorporating AI into their offerings and the positive impact of Microsoft's supportive stance on copyright issues.

• The conversation around AI use cases and their regulation is complex.
• Some argue existing laws cover potential misuse of AI, but new categories of work products need to be considered.
• Transparency in AI models is necessary for regulatory purposes, and companies are moving towards providing more information about their training data and processes.

"Much has been said. Oh, don't worry about it. Most of the bad things that you can do with Gen AI are already regulated and illegal, so nothing new." "IBM was announcing their own gen AI models and they were talking about being fully transparent on the data that went into the models."

These quotes discuss the regulatory landscape for AI, with a focus on the need for transparency and the complexity of applying existing laws to new AI-driven work products.

• The debate on whether startups or incumbents will accrue more value in AI is ongoing.
• Incumbents with access to large data sets are well-positioned to lead in AI outcomes.
• However, incumbents are also creating platforms for startups to innovate, complicating the value accrual landscape.

"I would bias towards incumbents that have the data." "But I do think that data is what powers outcomes."

The quotes suggest that while incumbents have an advantage due to their data access, the dynamic between incumbents and startups remains complex due to the platforms that enable startup innovation.

• The definition of "open" in AI systems is nuanced, with considerations beyond just the availability of model weights.
• Open models facilitate research and innovation but may come with use case restrictions.
• Commercial solutions are likely to be cloud-hosted services, with the quality of service being a significant factor regardless of the openness of the model weights.

"I think there is nuance on what open means." "I have no idea if it matters or not. At the end of the day, who has the best answers or the best service integrated for customers."

These quotes address the complexity of defining open AI systems and the potential impact of open models on innovation and commercial AI solutions.

• The impact of AI on jobs and economies is a concern, but the actual pace of adoption may be slower than expected.
• The productization of AI is more challenging than demos suggest, leading to a natural throttling of adoption speed.
• Incremental productivity gains are anticipated, with businesses needing to decide how to leverage these gains in terms of workforce deployment.

"All of this is harder than people realize. The demos are awesome, the productization takes longer time." "So I would say all of us should go forward as fast as we can because there's going to be natural difficulties that will just throttle us."

These quotes convey a realistic perspective on AI adoption, emphasizing the challenges of productization and the potential for gradual productivity improvements rather than immediate job displacement.

• AI is leading to a decreased emphasis on traditional user interfaces (UI).
• Personalization and customization are becoming more tailored to individual users.
• The gap between device and user is widening in ways not seen before.

"I think that with AI you see the reducing value of UI. You will see personalization and customization according to each user."

This quote suggests that as AI becomes more advanced, the traditional UI becomes less critical due to increased personalization capabilities.

• AI shifts the value of UI, but does not eliminate the need for rich interaction in many use cases.
• Personalization has been a long-term trend, and AI continues to support this journey.

"So I would say yes, some use cases, it shifts the value of UI, but many others has the opportunity to continue to enrich them."

Christian Kleinerman points out that while AI may reduce the need for UI in some scenarios, it can also enhance the user experience in others by providing richer interactions.

• Snowflake is perceived primarily as a data warehousing service.
• The company has expanded beyond its original data warehousing capabilities.
• There's a need to communicate Snowflake's AI and Gen AI capabilities effectively.

"We need to make sure that organizations across the world understand that they can do AI and Gen AI close to their data within Snowflake without having to copy the data to a different platform."

Christian Kleinerman expresses the challenge Snowflake faces in changing the public's perception of the company to include its AI capabilities.

• Early success in a particular area can create a lasting public perception.
• Snowflake is still often seen as a cloud-based data warehouse despite its growth.
• Companies frequently remain associated with their original use case for an extended period.

"For many years we said Snowflake is the data warehouse built for the cloud and that still keeps getting repeated over and over."

Christian Kleinerman explains that Snowflake's initial positioning as a cloud data warehouse has created a persistent perception that they are working to evolve.

• Christian Kleinerman has become more assertive in pushing his opinions over time.
• Initially, he focused on being a great manager by accommodating all opinions.
• A unified product vision sometimes requires a more top-down approach.

"I've been more willing to push opinions in a slightly more top down way as more time has gone by."

Christian Kleinerman discusses how his product leadership style has evolved to sometimes require a more directive approach for the sake of product consistency.

• The balance between debate and execution depends on the product's nature.
• Core technologies require careful design to avoid data corruption or errors.
• For less critical features, speed and iteration based on user feedback are appropriate.

"I think it depends on the nature of the technology or the nature of the product."

Christian Kleinerman provides insight into how the decision-making process varies between foundational technologies and user-facing features.

• People are the central factor in all aspects of business.
• Outcomes, relationships, and overall success are all dependent on people.

"I for sure know that it always comes down to people in terms."

Christian Kleinerman emphasizes the importance of people in determining the success of various business aspects.

• PMs generally need to be deeply technical to succeed.
• There may be exceptions for certain types of products.

"For the most part, no. There may be a few types of products that you might get by, but I like deep technical pms."

Christian Kleinerman argues that being technically proficient is typically necessary for product managers to be effective in their roles.

• New PMs should deeply learn and use the product they are managing.
• Understanding the technology is crucial for success.

"Learn the product that you're a PM of. Go be a user. Go as deep as you can know the technology."

Christian Kleinerman advises new product managers to immerse themselves in their product and understand the technology behind it.

• Innovation in AI is not confined to Silicon Valley.
• Talent is distributed globally, and significant innovation occurs elsewhere.

"There is amazing talent throughout the world."

Christian Kleinerman points out that innovation in AI is happening globally, not just in Silicon Valley.

• Focusing on ease of use has been a significant product decision for Snowflake.
• Simplifying products for users can lead to better outcomes than releasing products quickly.

"Focusing snowflake and ease of use."

Christian Kleinerman reflects on the impact of prioritizing ease of use in Snowflake's product development.

• The AI community needs to recognize Snowflake as a suitable platform for AI.
• Changing this perception is seen as a key goal.

"They need to know that snowflake is a great platform for AI."

Christian Kleinerman expresses the desire to shift the AI community's perception of Snowflake as a capable platform for AI applications.

• Initially, Christian Kleinerman doubted Satya Nadella's understanding of the enterprise business.
• Nadella's subsequent decisions demonstrated his deep understanding and leadership.

"In the beginning, I was super wrong."

Christian Kleinerman admits to initially misjudging Satya Nadella, who later proved to be a highly effective leader.

• Satya Nadella is a clear thinker who drives relentlessly towards desired outcomes.
• His clarity helps him avoid getting bogged down by obstacles.

"He is very clear what the needed outcome or desired outcome is."

Christian Kleinerman highlights Satya Nadella's ability to think clearly and focus on achieving specific goals.

• Frank Slootman is a clear thinker, which simplifies and accelerates decision-making.
• His clarity is a valuable trait in leadership.

"He has that commonality with Satya. He becomes a clarifying force."

Christian Kleinerman shares his admiration for Frank Slootman's clear thinking, which he finds to be a common trait with Satya Nadella.

• AI will significantly boost productivity across various domains.
• The impact on society and GDP is expected to be net positive.

"Productivity boost on pretty much everything we do. Everything is going to be simpler, easier, faster."

Christian Kleinerman predicts that AI will greatly enhance productivity, making tasks simpler and more efficient.