AI companies fall into incumbents, foundation-model providers, and application or tooling companies, with foundation models carrying the greatest risk of concentrated power.
2
OpenAI's enterprise privacy guarantees are difficult to audit because companies cannot easily trace how their data may influence future models or training sets.
3
Open-source AI can spread capability and contributions more widely, while excessive capital and closed platforms can slow progress or make the ecosystem more dependent on a few companies.
Summary
Kyle Harrison discusses how the ChatGPT moment made existing progress in transformer-based models visible to the wider market. He divides AI companies into incumbents with distribution, foundation-model providers, and companies building applications or tooling. His main concern is the concentration of capability, capital, GPUs, and influence among foundation-model companies. Harrison argues that OpenAI's privacy and security assurances do not fully answer whether enterprise data can influence future models or hidden test sets. He expects regulated companies and teams with strong internal expertise to prefer local or open models over the easiest hosted API. He compares Meta's PyTorch strategy with Google's TensorFlow and says Meta's lack of a cloud business gives it different incentives around openness. The conversation also covers defensibility, traditional machine learning, agents, and AI's long development curve. Harrison is skeptical of both overnight AI hype and predictions of an immediate AI winter.
AI companies occupy three very different positions
Kyle Harrison groups AI companies into established incumbents such as Microsoft, Canva, Adobe, and Notion; foundation-model providers that need training data, access, and large amounts of capital; and the much larger group building tooling, orchestration, fine-tuning, or use-case products. He says the middle group has the greatest potential to centralize power because many other companies depend on its models. Application companies may generate revenue quickly, but that does not mean they have a lasting advantage. OpenAI can copy products and already has the attention, capital, and distribution to make that threat serious.
A fast-growing AI product may have first-mover advantage rather than a moat
Harrison says that if OpenAI can reproduce a product within months, the product probably did not have a real moat. Copy.ai and Jasper are examples of companies that reached the market early by using GPT for copy generation, but he describes that position as first-mover advantage rather than defensibility. Companies need to be unusually good at something that large model providers are not trying to commoditize. Replit is his example. It is a developer ecosystem for writing code, hosting applications, and building products, with AI features such as Ghostwriter feeding an existing product flywheel. Nomic's Atlas gives it a separate strength in visualizing and understanding model data.
Closed foundation models make attribution and accountability harder
Harrison says OpenAI has increasingly limited visibility into GPT-4's data, methods, weights, and other components. Earlier AI systems were often treated as black boxes, then transformer research made model construction easier to study. He argues that OpenAI is moving toward a black box again because a system people understand less easily is easier to monetize. The company can describe security controls and data encryption, yet customers may still struggle to establish that their information could not influence future models or internal training sets. Harrison connects this problem to lawsuits involving Stability AI and GitHub Copilot, where questions about training data and attribution are already active.
Concentration of AI capability can increase risk in the name of safety
Citing Clem Delangue of Hugging Face, Harrison calls concentration of power the main danger in AI. Limiting access in the name of safety can leave companies, nonprofits, and governments dependent on one central failure point. He argues that powerful technical capabilities eventually spread to bad actors, so keeping everyone else capable of responding can provide more protection than putting capability in the hands of a few companies. Harrison also explains why incentives differ across large technology firms. Microsoft benefits from moving AI workloads onto Azure, while Meta has no cloud-computing business and can gain more by making models widely available for use in its products.
Meta's open-source position has a business explanation
The conversation compares Meta's PyTorch with Google's TensorFlow. Harrison says PyTorch quickly became dominant among engineers, and one engineer described it as having 'eaten TensorFlow's lunch.' He sees a precedent for Meta becoming a major open-source force, although people are surprised because they do not usually view Meta as a moral leader. Harrison separates the business incentives from the moral story. Meta benefits when AI capabilities spread because its core products can use those capabilities. He also points to Databricks as a company that has built a sustainable business while using open source meaningfully.
Regulated companies may reject the easiest hosted AI option
Harrison expects healthcare, financial services, and government organizations to examine hosted foundation models more carefully than a small team seeking the fastest route to a working application. ChatGPT Enterprise offers controls such as SOC 2 compliance and encryption, but those controls do not automatically make model influence or data attribution auditable. Companies may want to run systems locally and retain more control over their data. Harrison also expects adoption of internal AI capabilities to grow as technical talent becomes less scarce. He compares this with Uber's use and release of open-source tools, where companies with enough expertise can build on an open foundation instead of depending entirely on a vendor.
Companies should add AI to an existing strength instead of chasing the label
Harrison rejects a general rule that companies should replace traditional machine learning with large language models. He says product decisions based on hype, such as adding AI merely to a landing page, are dangerous. Ramp is his example of a company using machine learning and AI within its existing strength of managing a financial stack. That can include answering questions about finances and extracting more value from the underlying data. He says companies should decide how the technology supports their own capability. Calling a system 'LLM' rather than 'machine learning' does not make the product strategy better.
Agents need better interfaces and much higher reliability
Harrison sees useful possibilities for purpose-built agents, especially when a task is well-defined and repeatable. He says language can limit what people are able to express, while a packaged agent could perform a known task without making the user describe every step. The harder problem is production performance across several systems. An agent that works inside one application may be effective, but booking a trip requires it to move through preferences, hotels, airlines, prices, and multiple interfaces. Harrison says he has not seen a system handle that sequence well. The more interaction layers an agent crosses, the more opportunities it has to fail.
AI will progress along a curve rather than changing everything overnight
Harrison rejects both extreme AI hype and predictions of an immediate AI winter. He compares AI with e-commerce, which rose steadily, jumped during COVID, then settled above its earlier trend rather than staying at the COVID peak or returning to its old level. He thinks AI has a long way to go before it becomes a general-purpose production experience. The gaps include model performance, context, tooling, orchestration, user interfaces, and people's ability to understand useful applications. Excitement can speed progress by attracting talent and contributions, especially when systems are open. He also warns that too much capital can create a category-wide setback when investments fail, as happened with parts of climate technology and crypto.
"If OpenAI can get it as broadly distributed as possible as quickly as possible, then when you do have that outcry in a couple years, it's so much harder to get them out because they're everywhere."Kyle Harrison36:56
Who should watch
You are deciding whether to build on a hosted foundation model or keep more of your data and model stack under internal control.
Your AI application is growing quickly, but you are unsure whether its early revenue reflects a durable product advantage.
You are building agents and need a grounded view of why multi-step, cross-application workflows still fail in production.
