Tubegraph

Product market fit in AI

6 min read

From: redpointai

Characterizing genuine product market fit in AI today is challenging, especially for early-stage investments where success is unclear [00:24:28]. The focus for investors is often on finding innovations that, if successful, will clearly be adopted [00:24:41].

Identifying Product Market Fit

The definition of product market fit in AI evolves, but historically, a bar of $100 million in revenue for a use case has been considered a strong indicator [00:26:09]. This addresses skeptics who view AI tools as mere “toys” lacking reliability [00:26:00].

Early examples of applications demonstrating product market fit include:

The list of “form factors” that demonstrate product market fit is expected to grow over time [00:26:35].

Recent Additions to PMF Category

  • Deep research tools: Anything that facilitates long-term, agentic reporting and automates more aspects of a job shows product market fit [00:26:42]. OpenAI’s Deep Research feature, despite initial pricing adjustments, generated significant revenue, indicating strong demand [00:27:09].
  • Customer support agents: Companies like Sierra are tackling this market, indicating a belief in its long-term viability and ability to provide a durable business [00:31:26].
  • Voice AI: These applications do not require 100% precision and recall to be valuable [00:33:17]. For instance, scheduling intake companies for home services, where 50% of calls are currently missed, can significantly increase revenue with an AI that is only 75% effective [00:33:30].
  • Specific niche applications: Focused solutions, such as AI for appointment scheduling for veterinarians, demonstrate that addressing a significant pain point can lead to willingness to pay [00:34:15].

Evolution of AI Applications

The first wave of AI applications primarily focused on cost cutting [00:32:02]. These often targeted areas where businesses were already comfortable outsourcing and accepting some performance reduction for cost savings [00:32:30]. While these areas might face fierce price competition, the next wave is expected to focus on growth and revenue generation [00:32:06]. This includes capabilities that directly increase top-line revenue, such as AI for go-to-market strategies, which may prove more defensible and command higher prices due to clear ROI [00:32:56].

Future Potential Applications

Areas with emerging product market fit include:

Even with current model capabilities, there are “trillions of dollars of application value to be unlocked” [00:36:56]. For example, amazing education apps could be built with the existing generation of models [00:37:03]. The challenge in areas like education may be more societal and human-related rather than technological [00:37:24].

Defensibility at the Application Layer

Defensibility at the app layer is often underestimated. Key factors include:

  • Network effects: Prioritizing the multiplayer experience over just the single-player experience can create a robust defense [00:39:16]. Chai Research, a character AI competitor, outlasted others by building a network of people submitting models, demonstrating the power of network effects over proprietary IP [00:40:18].
  • Brand recognition: A company can quickly become synonymous with an entire category within six to nine months, giving them a significant advantage in customer access [00:41:17]. This brand premium can even lead to higher average contract values (ACVs) [00:41:41].
  • User experience and design: The “thousand small things” a company does to create a delightful user experience contributes significantly to defensibility, akin to traditional SaaS applications [00:42:06].
  • Velocity: The speed at which a company can develop a broad product and adapt to new models (which can be an “existential event” every 3-6 months) is crucial [00:42:12].

Initially, many AI app developers mistakenly believed that defensibility would come from unique datasets or proprietary models [00:41:51]. However, this has proven to be a “total head fake,” as the ability to adapt and provide a superior user experience is more critical [00:41:59].

Infrastructure and Model Specialization

The application layer is often considered more interesting than infrastructure due to the ability to charge for utility rather than just cost-plus [00:43:33]. While there are interesting infrastructure plays, especially those related to model integration (e.g., code execution, memory, search, cybersecurity), companies focusing solely on serving models face a capital-intensive business model [00:44:51].

It is questioned whether new model companies are still viable, as the larger players like Google and OpenAI are moving more into the product space [00:18:07]. The trend has shown that general-purpose models tend to outperform hyper-specific models in quality [00:19:18]. For example, a specialized Bloomberg GPT model was ultimately surpassed by general OpenAI models [00:19:42]. However, the data pipelines and teams assembled for such specialized models remain valuable [00:20:12].

The shift towards enterprise use cases and model specialization is observed, particularly in areas like robotics, biology, and material science, where unique datasets are crucial [00:17:34]. The general LLM agent space remains highly competitive with tech giants [00:18:02].

Overhyped vs. Underhyped

Overhyped

  • Agent frameworks: The proliferation of agent frameworks is seen as overhyped, as the underlying workloads are in such flux that it’s difficult to build stable frameworks [00:10:43]. The current stage is compared to the “jQuery era” of single-file, big frameworks, while a “React” equivalent for AI is still needed [00:11:25].
  • New model training companies: There is skepticism about the need for many new companies focused solely on training models, especially given the dominance of large players aiming for AGI [00:16:16].

Underhyped

  • Memory / Stateful AI: The ability for AI agents to maintain memory beyond simple conversation history, such as storing knowledge graphs, is underhyped [00:14:03]. This is a hard problem but crucial for smarter, learning agents [00:14:42].
  • Apple’s Private Cloud Compute (PCC): This technology, which extends on-device security to the cloud, is considered under the radar but has significant potential for secure multi-tenant AI environments where GPUs are scarce [00:12:45]. This addresses the need for single-tenant guarantees in multi-tenant environments [00:13:55].
  • AI for DevOps (AISR): While currently limited by technology, the potential for AI to significantly improve mean time to resolution (MTR) in operations is an interesting opportunity [00:48:50].
  • Agent authentication: The emerging need for agents to authenticate themselves when accessing websites or services on behalf of a user is a crucial, yet under-discussed, problem that requires a new form of “SSO for agents” [00:54:41].

Unanswered Questions with Broad Implications

  • Scalability of RL to non-verifiable domains: The biggest unanswered question is whether Reinforcement Learning (RL) can be successfully applied to domains where outcomes are not easily verifiable, such as law (contracts) or marketing/sales (simulating conversations) [00:50:15]. If not, agents may remain confined to verifiable domains, while co-pilots will be needed in non-verifiable ones where human “taste makers” are still essential [00:50:46]. This could lead to a “weird future” where AI excels in coding and math but struggles with basic sales emails [00:51:00].
  • Hardware scaling and competition: How the industry will scale compute given the “rule of nines” (an order of magnitude increase in compute for each jump in reliability, e.g., 90% to 99%) [00:51:34]. The continued dominance of Nvidia and the potential for competitors like AWS, AMD, Microsoft, and Facebook to offer viable alternatives are key [00:52:00]. The general-purpose nature of GPUs has allowed Nvidia to remain dominant, but if transformer architecture remains stable, specialized ASICs could emerge [00:53:57].

Graph View

Backlinks