This interview argues that the real value in AI isn't in the underlying models but in the applications built on top of them. Investor Sarah Guo says the 'GPT wrapper' narrative is dangerously wrong. She highlights three key bets: Harvey, an AI for lawyers that's already generating tens of millions in revenue; HeyGen, a video-generation tool that recently became good enough to sell; and Mistral, a French AI company focused on efficiency over brute force. She also warns that challenging chip giant NVIDIA is much harder than it looks, due to the massive validation costs for large-scale training.
At a Glance Sarah Guo, Founder and CEO of Conviction, shared her experience of leaving Greylock to establish an early-stage AI venture capital firm on the Invest Like the Best podcast. Core view: AI is the most important technological advancement of our time, and Conviction focuses on serving AI sta
Sarah Guo, founder and CEO of Conviction and former partner at Greylock, shared her journey of leaving an established institution to launch an AI-focused venture capital firm on the Invest Like the Best podcast. The central theme: the real value and investment opportunities in the AI application layer. The most impactful takeaway from the episode: Sarah Guo believes "that last mile (the application layer) is, in my view, 99 miles out of 100" — the value of the AI application layer is far from being fully recognized by the market, and the GPT wrapper narrative is a dangerously misleading notion.
Sarah Guo argues that believing the popular narrative that "the AI application layer has no value" is the greatest danger for investors.
She points out that in early 2023, the market was dominated by the narrative that "everything is a GPT wrapper, and the application layer has no value," but this narrative serves specific interests. "When I say it's dangerous to believe popular narratives, first, that narrative serves certain people; second, it goes against first-principles thinking." Sarah Guo emphasizes that the foundational framework for evaluating AI application companies is consistent with traditional software—distribution capabilities, talent quality, and market size—but requires additional judgment: where the boundaries of foundation model capabilities lie, and which capabilities will be replicated by model providers.
Key judgment: Application companies should not compete with foundation model providers on their core roadmap. Sarah Guo proposes the principle of "avoiding the incumbent's advantage path": large model companies will not win in all markets. "Technology may be universal, but the distance from technology to end users is very long." Taking Harvey (legal AI) as an example, OpenAI or Anthropic are unlikely to build a pure legal model, as this is not on their core roadmap.
Data support: Harvey has already achieved tens of millions of dollars in revenue, with clients including top law firms and consulting firms such as Allen & Overy and PwC. The product's value is reflected in two dimensions: making lawyers who charge $2,000 per hour more efficient, and completing tasks that were previously impossible to scale (such as searching for specific clauses across 20,000 contracts).
Sarah Guo proposes "Minimum Viable Quality" as the core framework for evaluating AI products, fundamentally different from the traditional SaaS concept of "Minimum Viable Product Scope."
Traditional software requires defining a feature scope (e.g., CRM needs to assign sales representatives, manage quotas, and display BI), whereas the core question for AI products is "whether the model's output is good enough to sell." Sarah Guo uses HeyGen (an AI video generation platform) as an example: only in the past year did HeyGen cross the minimum viable quality threshold—users can now generate commercially usable personal presentation videos from just two minutes of webcam-quality footage. This breakthrough has driven the company's rapid growth.
Sarah Guo points out that crossing the "uncanny valley" depends not only on model improvements but also on system design: "How can the cost of managing errors be made extremely low? Increasingly, model systems generate a large number of candidates, then improve user experience through validation or ranking." She cites code generation as an example: Google's team generates 1 million candidates, then returns the top three optimal results, significantly enhancing user experience.
The uncanny valley yet to be crossed: Code generation—current benchmarks achieve only a 13–20% success rate. "If an intern only gets it right 20% of the time, they would be fired." Multi-step reasoning and hallucination management are the core challenges that next-generation models need to address.
Sarah Guo expects a "diverse bloom" in the foundation model market but believes the risk of building a general-purpose foundation model startup is extremely high.
She has invested in Mistral, citing its differentiation in prioritizing efficiency. "Blindly believing in popular AI narratives is very dangerous—there was once a narrative that efficiency didn't matter, but there has never been a moment in computing history where efficiency was unimportant." Sarah Guo points out that efficiency will become a key competitive dimension in both inference and training.
On market structure, Sarah Guo assesses: "I hope to see multiple options, open-source choices, sufficient competition, and robustness so that application-layer companies can build real economic value." She believes Meta (Zuck)'s commitment to open-source models makes the competitive landscape more interesting, but enterprise clients still have concerns about Facebook as a long-term partner.
Risk warning: "Most people should not train general-purpose foundation models—this is a fast way to lose money. When the cost of entry exceeds $500 million to conduct meaningful pre-training, it is a very easy way to burn cash." She argues that the market's demand for another undifferentiated LLM player is not unlimited.
Sarah Guo believes the difficulty of challenging NVIDIA is underestimated by the market, with the core obstacle being the validation problem of large-scale training.
She shared a case from a portfolio company: a lab training large-scale models using non-NVIDIA chips, despite having top-tier researchers, still faced enormous engineering and infrastructure management challenges. “If you don’t have people running training workloads at thousands-of-node scale on your chips, you don’t know if they actually work—such failures only surface at scale.”
Sarah Guo points out that Google’s chips work because Google runs its own training workloads on them. For other challengers, they either need a “patient zero” customer or must bear the massive cost of deploying 10,000 chips themselves.
On the cost curve, Sarah Guo is optimistic: “Innovation across the entire ecosystem—from memory bandwidth to chips to systems—will yield benefits over the next 5–10 years. There is no reason to believe this won’t happen.” She believes the current AI infrastructure is in a “crazy, immature” phase, analogous to the evolution from on-premise servers to virtualization to serverless computing, with AI currently only at step one or two.
Sarah Guo believes the greatest opportunities lie where technical capabilities and domain expertise rarely intersect.
She cites Harvey as an example: a lawyer working 90-hour weeks wonders, "Can ChatGPT do my job?" and happens to have a roommate who is an outstanding researcher—a combination that is extremely rare. Similar opportunities include:
On service-based business models: Sarah Guo says she would be excited about them, but only if founders "have the ambition for continuous improvement and see if they can make margins as good as software." She believes the most interesting AI application companies may be consumed as "service experiences," but in reality, they are replacing what were previously services.
Sarah Guo distinguishes between long-term risks and short-term misuse: "Biological risks, out-of-control models, Chinese weapons systems—we should figure out whether these risks are real, but I think they are a huge distraction from near-term misuse." She is more focused on fundamental abuses such as misinformation and fraud, which will be significantly amplified with the help of AI.
Regarding the outlook for the 2030s: Sarah Guo believes there is a "genuine probability of abundance," but society's capacity to absorb so much change is limited. "There will be winners and losers, and the dynamics of job displacement are real—but probably not as fast as people fear."
| Position | Guest Stance | Key Data |
|---|---|---|
| Harvey | Bullish (first investment) | Tens of millions in revenue, clients include Allen & Overy, PwC |
| HeyGen | Bullish (board member) | Rapid growth over the past year, crossed the minimum viable quality threshold |
| Mistral | Bullish (first-round investor) | Emphasizes efficiency differentiation, open-source strategy |
| Figma | Positive review (Greylock-era investment) | Invested 10 years ago, held long-term |
| Awake | Positive review (Greylock-era incubation) | Network analytics security company, sold to Arista |
| NVIDIA | Neutral (acknowledges deep moat) | 90% gross margin, CUDA ecosystem, large-scale training validation barrier |
| OpenAI | Neutral (partner) | Impressive revenue growth but extremely high OPEX |
| Anthropic | Neutral (partner) | Alongside OpenAI as a major foundation model provider |
| Neutral | Chips (TPU) are effective due to internal large-scale training workloads | |
| Meta | Neutral to slightly positive | Open-source model commitment makes the competitive landscape more interesting |
| Base10 | Bullish (board member) | Provides serverless infrastructure for GPU inference |
1. Sarah Guo: “The last mile, in my view, is 99 miles out of 100” — The value of the AI application layer is far from being fully recognized by the market, and the GPT wrapper narrative is a dangerously misleading one. Supporting argument: Application companies must handle a vast amount of work, including distribution, customer relationships, proprietary data, and change management.
2. Sarah Guo: “It is very dangerous to blindly believe the prevailing narrative in AI — that narrative serves certain people.” Supporting argument: The 2023 narrative that “everything is a GPT wrapper” caused investors to miss real application-layer opportunities such as Harvey and HeyGen.
3. Sarah Guo introduces the concept of “Minimum Viable Quality”: The core issue for AI products is not “whether the feature scope is broad enough,” but “whether the model output is good enough to sell.” Supporting argument: HeyGen only crossed this threshold in the past year, fueling the company’s rapid growth.
4. Sarah Guo proposes the principle of “Avoiding the Incumbent’s Path of Advantage”: Application companies should not compete on the core trajectory of foundation model providers. Supporting argument: It is predictable that OpenAI would move into search, but legal AI is not on its core roadmap.
5. Sarah Guo: “There has never been a moment in computing history where efficiency was unimportant.” Supporting argument: Mistral’s differentiation lies in efficiency, and efficiency will become a key competitive dimension in both training and inference.
6. Sarah Guo: “The difficulty of challenging NVIDIA is underestimated by the market — if you don’t have people running thousand-node-scale training workloads on your chips, you don’t know if they actually work.” Supporting argument: Google’s TPU is effective because Google runs its own training workloads on its own chips.
7. Sarah Guo: “Most people should not train general-purpose foundation models — that is a fast way to lose money.” Supporting argument: The cost of entry exceeds $500 million to conduct meaningful pre-training, and the market has limited demand for another undifferentiated LLM player.
8. Sarah Guo: “The biggest opportunities lie where technical capability and domain knowledge rarely intersect.” Supporting argument: Harvey’s founders, a combination of a lawyer and a researcher, are extremely rare; similar opportunities exist in areas such as enterprise software configuration maintenance and materials science.