Sep 22, 2023 · 46m · news
Christian Kleinerman: Do OpenAI and Anthropic Have a Sustaining Moat? Who Wins the AI Wars? | E1063 · 20VC with Harry Stebbings
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this interview, Christian Kleinerman, SVP of Product at Snowflake, discusses the pragmatic path of generative AI adoption, emphasizing data maturity, product simplicity, security barriers, and why value in the AI era will ultimately accrue to proprietary data owners and incumbents.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Harry holds 30.1% of the talking time here. How this is scored →
Harry as informed peer 3.3 Guest teaching 3.2 Guest disagreement 0.8 Harry pushing back 2.2
speaking balance: gold is Harry, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 11:16 Rejection of healthcare AI job loss premise
Christian directly pushes back against Harry's thesis that government incentives block AI due to fears of firing nurses, stating that current AI models are co-pilots rather than replacements.
Hardest push from Harry ▶ 16:09 Challenging low model valuation allocationHarry forcefully pushes back on Christian's claim that models only account for 10% of total value, demanding to know why OpenAI, Anthropic, and Bard command multi-billion dollar valuations if that were true.
Biggest teaching moment ▶ 33:52 Explaining nuances and limitations of open weightsChristian educates Harry on the exact definition of open weights versus true open source, pointing out hidden use-case restrictions in models like Meta's LLaMA 2.
Harry holds his own ▶ 22:08 Citing Character.ai $2M training cost data pointHarry brings concrete domain knowledge to the discussion by citing Character.ai founder Noam Shazeer and exact training cost figures to drive the conversation on AI compute economics.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Harry as informed peer | Guest teaching | Guest disagreement | Harry pushing back | Why |
|---|---|---|---|---|---|---|
| Christian's Career Journey and Path to Snowflake | 1 | 2 | 0 | 0 | Harry welcomes Christian and asks about his career journey from early startups to Microsoft, YouTube, and Snowflake. Christian shares foundational lessons about talent and business scalability without any friction. | |
| Key Takeaways from Microsoft and Google | 2 | 2 | 0 | 1 | Harry asks Christian about his key takeaways from Microsoft and Google, probing gently on whether simplicity is always better in product design. Christian explains SQL Server's simplicity strategy against Oracle and IBM. | |
| Navigating the Generative AI Hype Cycle | 2 | 3 | 1 | 1 | Harry introduces generative AI hype as the main topic. Christian makes a quick joke before offering a balanced perspective comparing AI to the scale of mobile and the internet. | |
| The Evolution of the AI Technology Stack | 4 | 3 | 2 | 3 | Harry puts forward a strong VC thesis that implementation services will be huge because enterprises lack education on AI rollout. Christian partially pushes back by emphasizing that the technology stack itself is still rapidly evolving. | |
| The Alignment of AI Co-pilots with Job Incentives | 3 | 4 | 3 | 3 | Harry asks if misaligned political incentives hinder AI adoption in sectors like healthcare due to fear of job loss. Christian explicitly rejects the premise, clarifying that current AI tools are co-pilots rather than wholesale worker replacements. | |
| The Convergence of Models on Public Data | 3 | 4 | 1 | 2 | Harry questions whether data remains a moat given how freely accessible public data seems. Christian breaks down public versus private enterprise data and explains how web owners are actively locking down crawling permissions. | |
| Assigning Value: Data vs. Model | 5 | 4 | 3 | 6 | Christian assigns 90% of value creation to data, arguing models are becoming commoditized. Harry directly challenges this framing, pointing out the immense market valuations assigned to OpenAI, Anthropic, and Google Bard. | |
| Evaluating 'GPT Wrappers' vs. Real AI Startups | 3 | 4 | 1 | 2 | Harry asks if VCs are being short-sighted by dismissing startups as mere 'GPT wrappers'. Christian differentiates between shallow wrappers built in a weekend and deep domain-specific integrations. | |
| The Changing Dynamics of Model Training Costs | 5 | 3 | 0 | 2 | Harry demonstrates domain expertise by citing Noam Shazeer from Character.ai on $2M model training costs and Alex from Nabla on model flexibility. Christian concurs on the necessity of model abstraction layers. | |
| Enterprise Obstacles: Privacy, Safety, and IP Rights | 3 | 4 | 0 | 2 | Harry asks about obstacles to enterprise model adoption. Christian highlights output correctness, data privacy, and legal rights over generated outputs using a hedge fund portfolio strategy example. | |
| Alleviating Enterprise Copyright Concerns | 3 | 3 | 0 | 3 | Christian highlights Microsoft's commitment to defend enterprise customers against copyright lawsuits. Harry asks why this backstop is so material for enterprise adoption confidence. | |
| Hallucination Control via Attribution | 4 | 3 | 0 | 1 | Harry asks about model opacity and hallucination control. Christian details how acquisition Neva used citations and source attribution to keep LLMs accurate, while Harry notes he was an early investor in Neva. | |
| Open Weights vs. Commercial Cloud Hosted Models | 3 | 4 | 1 | 1 | Harry asks whether open or closed models will dominate. Christian educates on the nuances of 'open weights', noting licensing restrictions like Meta's LLaMA 2 terms. | |
| Snowflake's AI Strategy and Platform Evolution | 4 | 3 | 0 | 3 | Harry asks how Snowflake balances maintaining core enterprise revenue with investing in fast-moving AI shifts. Christian details their strategy of bringing compute and LLMs directly to customer data. | |
| Overcoming the Data Warehousing Perception | 4 | 3 | 1 | 5 | When Christian admits Snowflake's biggest obstacle is being perceived purely as a data warehouse, Harry pushes directly, asking why customers still hold that perception after six years. | |
| Balancing Quality and Speed of Execution | 5 | 3 | 0 | 2 | Harry quotes Spotify CPO Gustav Söderström on product iteration versus internal debate. Christian contrasts core database infrastructure engineering with rapid UI experimentation. | |
| Quick Fire: Lessons from Satya Nadella and Frank Slootman | 2 | 2 | 0 | 0 | During quick fire, Christian reveals he was 'super wrong' about Satya Nadella when Satya first took over enterprise at Microsoft. The segment concludes with light banter and a dinner bet. |
Statements from this episode (42)
Kleinerman: AI will bring incremental productivity boosts, not immediate mass layoffs
“I don't think that it's a mass firing happening next week. It's more incremental productivity boosts happening over the next six, 1224 months.”
Kleinerman: Never compromise on talent for good intentions
“I would say don't ever compromise on talent. Don't ever say, yeah, this person doesn't have a background, but maybe Has the right intention. Just take a bet. No, I think that nothing substitutes talent.”
Kleinerman: Extreme product simplicity drives adoption over feature completeness
“If you turn something that is a subset of the capability, but dramatically easier to use, That gets a following, and that was a very, very clear lesson learned, and I've seen it over and over, and by the way, the Snowflake story follows a big part of that jour…”
Kleinerman: Generative AI will simplify nearly every human-computer interaction
“I do think that there is the opportunity to change pretty much every single interaction between humans and computers into something that is friendlier and simpler. So yes, there's fog, and there's noise around it, but it will clear up, and at the end of the da…”
Kleinerman: AI is on the same scale as mobile and the internet
“I think it's comparable. I think it's on the scale of the internet. I think it's on the scale of mobile.”
Kleinerman: AI hallucinations are features, not bugs, in creative fields
“What in. Other industries we would call a bug or a hallucination. In the creative world, there are features. They're, so it's a goodness to come up with something that has not been done before, or it's a mix and match of existing things. So creative industries…”
Kleinerman: Every Snowflake customer is actively trying to adopt AI
“Every customer that I talk to these days, they're looking into doing something. They're all trying to figure out how to get started and how to get there”
Stebbings: The largest AI businesses will be implementation services companies
“And I think the biggest businesses in AI will be built in the implementation services businesses, helping large enterprises implement AI in a meaningful way into their enterprise over the next decade.”
Kleinerman: There is no AI strategy without a data strategy
“I think it completely correlates with data maturity. You may have heard some of us at Snowflake talk about There's no AI or gen AI strategy without a data strategy. It's not just a line. It's a truth that we strongly believe in.”
Kleinerman: Financial services lead enterprise AI adoption
“And from that perspective, I would say financial services are at the forefront. Most of the financials have figured out for a long time how to organize data and leverage data for competitive advantages.”
Kleinerman: AI job displacement backlash will become a major issue by 2025
“So at this point, I would say the incentive should be working maybe, I don't know, A year, a couple of years from now, what you're saying becomes true.”
Kleinerman: Generative AI will democratize enterprise data access via natural language
“I think Gen AI has the opportunity to Turbo charge this type of translation where the language is natural language, and the answers come in natural language, but along the way, there's traditional database lookups, traditional retrieval, and has the opportunit…”
Kleinerman: AI models trained on public data will converge on identical answers
“If things did not change, the language models, or the models in general, would all converge towards, they're all training on the same data, and at some point it's what mix of data you use, but you'll trend towards the same answer.”
Kleinerman: Web data policies for AI training will shift within 12 months
“I think all of this will shift. In the next six, 12 months, this is all starting already, because in the same way that search changed the rules of engagement with public data, Gen AI is doing the same thing, and the companies that are behind that data are doin…”
Kleinerman: AI data monetization will shift toward paid licensing terms
“I don't know if they're going to be able to keep all the existing revenue, but for sure they should be able to capture some amount of revenue. So I don't know exactly what would be the replacement ratios. But clearly their data is valuable. Right now it's not …”
Kleinerman: Seven-person startups can match OpenAI and Anthropic for specific uses
“We've seen companies that with Seven employees have creating models that are comparable for some use cases to what OpenAI or Anthropic do.”
Kleinerman: AI foundation models will commoditize until the next architectural innovation
“These are strong words, but I think they're getting commoditized until the next big innovation comes and you allocate some more value to the model.”
Kleinerman: Computer vision democratization will follow LLMs into seamless multimodal AI
“What has happened for language models is coming for computer vision, for images. The democratization, democratization, how do you simplify it? Then there's the intersection of those true multimodal languages, which there are many of them out there, but how do …”
Kleinerman: Shallow GPT-4 wrappers built in weeks are not real companies
“There are some very shallow wrappers on top of a GPT-IV. I don't place much value on them. I usually ask, hey, how long did it take you to build this? Oftentimes it's a week or two. I don't think there's a company there.”
Kleinerman: Deep GPT-4 wrappers with domain expertise build real value
“I do think that there are some very deep wrappers on top of GPT-IV that apply domain specific legal or other domain that I think you end up with a true way to bring GPT-IV to a given market or industry. I think those are value.”
Kleinerman: OpenAI and Anthropic are effectively research companies, not software platforms
“Probably the opening eyes and Anthropics are continuing their innovation, which those are effectively research companies.”
Kleinerman: Fine-tuned smaller models outperform large generic models on specialized tasks
“And now there's plenty of examples that have been run where a smaller model fine tune where a specific purpose or a specific data set produces results better than a generic model.”
Kleinerman: Tightly coupling software to one AI model sacrifices future optionality
“Anyone that builds too tightly coupled to a given model is sort of giving up optionality for the future.”
Kleinerman: AI software requires a model abstraction layer for architectural flexibility
“In the same way that you have a hardware abstraction layer or a cloud abstraction layer, 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 intricacies or specific char…”
Kleinerman: AI adoption faces privacy, security, and IP rights challenges
“There are a lot of issues. Probably the most obvious one is around the Correctness and dependability of answers. If you ask folks, one of the key concerns is like, well, these models make up stuff. So that, that is the obvious one. But then there are second or…”
Kleinerman: Enterprise platforms are bringing LLMs to data, not vice versa
“What you're seeing many of us, Snowflake for sure on the list, is evolve their platforms where it's easy to bring LLMs to the data as opposed to send Large data volumes to where the LLMs are.”
Kleinerman: Copyright fears stall enterprise AI; Microsoft's legal indemnity is material
“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, they're exposed, and because of those concerns, it has held back enterpris…”
Kleinerman: Enterprise AI will require explicit source attribution for accuracy
“I would say that that is the world that we're headed towards, at least in the enterprise, like forget the creative side of things, but at least in the world where you need correct answers, you need to be able to attribute where things came from.”
Kleinerman: Data-rich incumbents will win the majority of AI value
“I would bias towards incumbents that have the data. So I would say, think of all the data that a company like Google has. Or that's public data, or not semi-public, but a lot of user data, or think all the private enterprise data that a company like Snowflake …”
Stebbings: Incumbent co-pilot AI strategy will win over the next 2-3 years
“I also don't think you can underestimate, especially in the short term, the power of distribution, and I think, like, the co-pilot strategy is, like, a very incumbent focused strategy, which will win in the next two to three years, for sure.”
Kleinerman: Meta's Llama 2 license restricts competitor training and massive-scale deployment
“For example, LAMA II is very significant as a development in the industry. But they were very clear. Thou shall not use LAMA-II for training other models. Thou shall not use LAMA-II for, I think it's seven hundred million user use cases.”
Kleinerman: Commercial AI solutions will largely be managed cloud services
“I also think that commercial solutions will largely be hosted cloud services, in which case, in the same way that it is true with open source, I think it will be true about open weights.”
Kleinerman: Productizing AI is harder and takes longer than building demos
“All of this is harder than people realize. The demos are awesome. The productization takes longer time.”
Kleinerman: Generative AI Will Not Render Visual UIs Obsolete
“I would say that for certain use cases, that's entirely true. Like if I have a UI to configure a cluster, like I don't need to know all the options. I just can specify what I want. There are many use cases where You may want a richer way to interact with data …”
Kleinerman: Successful early positioning eventually hinders product expansion and repositioning
“When you're very successful with some positioning that comes and bites you later that you are too successful with that positioning.”
Kleinerman: Consistent product vision requires top-down leadership over consensus
“In some instances where you want a product to come out with a consistent view, as if it came from a single unified set of principles and individuals, sometimes you have to go and push for something like, Hey, this is what we're doing.”
Kleinerman: Core infrastructure needs meticulous design, while UIs require rapid iteration
“So the core subsystem that does say clustering of data on disk I think you want to design that thing really, really well. Measure a hundred times and cut once because nobody wants their data to get corrupted or their results to be wrong if that thing is not bu…”
Kleinerman: Product managers today cannot succeed without deep technical knowledge
“For the most part, no. There may be a few types of products that you might get by, but I like deep technical PMs.”
Kleinerman: AI founders do not need to be in Silicon Valley
“Absolutely not. Well, there's amazing talent throughout the world. And even though the Valley has something special from the community and the ability to bounce ideas of one another, It's very clear right now. There's a lot of innovation happening elsewhere.”
Kleinerman: Shipping faster is worse if it compromises product simplicity
“It is something a little bit counterintuitive that you may put out a product faster to the market. If you just say, I don't know if we should be used this way or that way. So you just surface choices to users. And sometimes it takes longer for us to take the a…”
Kleinerman: Satya Nadella proved early Microsoft doubters wrong
“I was super wrong. When Satya came into the enterprise business, this is before he was CEO, he came in in very short order and made a lot of really difficult decisions. How the org was structured how contractors were hired, how we thought about the cloud versu…”
Kleinerman: Frank Slootman's leadership clarity accelerates decision-making
“He's also such a clear thinker. He has that commonality with Satya that he becomes a clarifying force. Oftentimes if you're just picturing yourself telling Frank about a problem and a couple of options, Just in the formulation, it becomes very obvious that you…”