AI product market fit and emerging applications

From: redpointai

This article summarizes a crossover episode of Unsupervised Learning and Latent Space, an AI podcast and technical newsletter for AI engineers [00:00:08]. The discussion features insights from individuals deeply involved in the AI industry, covering what surprised them, what they are focusing on, and the current landscape of AI product market fit and emerging applications [00:00:27].

Surprises in the AI World

The past year has brought several surprises to the AI landscape [00:01:10]:

• Recent Model Releases: The rapid release of new models, particularly after talks suggesting that “scaling is dead,” was a significant surprise, demonstrating quick transitions in the field [00:01:25]. The timing, for instance, of new models emerging right as pre-training seemed to tap out, felt “suspiciously neat” [00:01:51].
• Open-Source Model Adoption: A surprising lack of widespread adoption of open-source models in enterprise settings was noted, with one estimate putting usage at around 5% and decreasing [00:03:05]. Enterprises are often in “use case discovery mode,” prioritizing powerful models over open-source options [00:03:14]. However, the speed at which open-source models, specifically DeepSeek, caught up to closed-source models was a major surprise [00:03:36].
• Low-Code Builders Missing AI Market: It was unexpected that low-code builders like Zapier, Airtable, Retool, and Notion did not capture the AI builder market [00:07:51]. Despite having the “DNA,” reach, and distribution, they integrated AI by improving existing features rather than creating entirely new software paradigms [00:09:09]. This opened the door for new founders to build from scratch without preconceptions [00:09:13].
• Apple Intelligence: The perceived missteps by Apple in its AI products, such as Apple Intelligence, were also noted as a surprise [00:10:16].

Overhyped vs. Underhyped AI Concepts

Overhyped

• Agents Frameworks: These are considered overhyped due to the rapid flux in workloads, making it difficult to establish stable frameworks [00:10:43]. The current state is compared to the “jQuery era” of single-file, big frameworks, rather than the more stable “React” era [00:11:25]. It’s suggested that focusing on protocols rather than frameworks might be more effective [00:11:55].
• New Model Training Companies: There is surprise at the continued emergence of new companies focused solely on training models, especially given the established dominance of major players [00:16:16]. The general trend indicates that general-purpose models are preferred, and hyperspecific models, while cheaper, do not necessarily offer higher quality [00:19:18].

Underhyped

• Memory/Stateful AI: The concept of stateful AI, beyond conversational memory, is seen as underhyped. This involves storing knowledge graphs of facts to exceed context length and enable smarter, on-the-job learning for agents [00:15:02].
• Apple’s Private Cloud Compute (PCC): Despite Apple’s struggles with Apple Intelligence, their Private Cloud Compute (PCC) initiative is considered significant. It aims to bring on-device security to the cloud, allowing for multi-tenant architectures while maintaining single-tenant guarantees, which is crucial given GPU availability constraints [00:12:45].

AI Product Market Fit Today

The discussion identifies several areas that have demonstrated genuine product market fit:

• Co-pilot: A clear example of an AI application that has achieved significant market adoption [00:26:17].
• Deep Research / AI-powered Search: This category, exemplified by products like Perplexity and OpenAI’s Deep Research feature, shows strong market fit. These tools offer long-term agentic reporting and are increasingly taking over knowledge work [00:26:42].
• Customer Support Agents: AI applications in customer support, like Sierra, are seen as having substantial product market fit. This area targets a significant cost center for businesses [00:30:40].
• Voice AI in Services: For businesses like home services, where 50% of calls are missed, even 75% effective AI can lead to significant revenue increases, showing clear value [00:33:30].

Emerging Applications and Future Focus

The conversation shifts to anticipating future areas of AI application and market potential:

• Shift from Cost-Cutting to Growth: The first wave of AI apps focused on cost-cutting (e.g., outsourcing to BPOs). The next wave is expected to focus on revenue generation, where businesses are willing to pay more for tools that increase topline [00:32:02].
• Specific Emerging Areas:
  • Screen Sharing: AI assisting users by observing their work [00:35:56].
  • Outbound Sales: AI assisting with proactive sales efforts [00:36:06].
  • Hiring/Recruiting: AI supporting the recruitment process [00:36:14].
  • Education: Personalized teaching applications with AI, despite societal challenges, hold significant potential in private education [00:36:18].
  • Finance: Numerous use cases for AI in finance [00:36:31].
  • Personal AI: While harder to monetize, personal AI assistants are an area of interest [00:36:37].

Defensibility at the Application Layer

When considering defensibility for AI applications, the emphasis is less on unique datasets or proprietary models (which were seen as a “head fake” [00:41:51]) and more on traditional software strengths:

• Network Effects: Prioritizing multiplayer experiences and building network effects is crucial for long-term robustness [00:40:00]. An example is Chai Research, a Character AI competitor, which built a marketplace of models and users, creating a network effect despite not owning proprietary IP [00:40:18].
• Brand and Product Surface Area: Establishing a strong brand identity within a category quickly can lead to significant market presence and command higher ACVs (Annual Contract Values) [00:41:17].
• User Experience and Velocity: Defensibility comes from “a thousand small things” that create a delightful user experience, coupled with the speed to develop a broad product and adapt to new model releases every 3-6 months [00:42:06]. This is reminiscent of traditional application SaaS companies [00:42:30].

Infrastructure and Model Labs

The “LLM OS” concept, which encompasses the infrastructure around models, is considered a valuable area [00:42:55]:

• Key Infrastructure Areas: Code execution, memory, and search are highlighted as interesting infrastructure plays [00:43:12].
• Cybersecurity: The application of AI to cybersecurity, particularly in areas where AI is used by attackers, is seen as crucial. This involves leveraging models to understand semantics (e.g., in binary inspection) rather than just syntax [00:43:42].
• Application Layer over Infrastructure: While infrastructure is important, the application layer is currently seen as “way more interesting” because it allows charging for utility rather than merely for cost, which is typical for infrastructure that often reduces to cost-plus pricing [00:47:10].

Concerns exist about certain infrastructure categories:

• Finetuning Companies: Struggling to see this as a standalone big business, suggesting it needs to be integrated into broader enterprise AI services [00:47:56].
• AI DevOps (AISR): While there’s potential, the technology isn’t fully there for autonomous operations [00:48:50]. It’s primarily seen as anomaly detection, not a new LLM use case [00:49:25].
• Voice Real-time Infra: While hot, its ultimate market size is questioned [00:48:40].

Model companies like OpenAI are increasingly moving into the product and developer tools space, turning previously standalone startup categories (like search, code execution) into features within their platforms or as APIs [00:45:21].

Biggest Unanswered Questions in AI

• RL in Non-Verifiable Domains: Can Reinforcement Learning (RL) be successfully applied to non-verifiable domains like law (contracts) or marketing/sales (simulating conversations)? If not, agents may be limited to verifiable domains, while non-verifiable ones will rely on co-pilots with human oversight [00:50:22]. This could lead to a “weird future” where AI excels at coding and math but struggles with basic sales emails [00:51:00].
• Hardware Scaling and Availability: How will the industry scale compute according to the “rule of nines” (order of magnitude increase in compute for each jump in reliability)? [00:51:34] Will Nvidia maintain dominance, or will competitors like AWS, AMD, Microsoft, and Facebook successfully challenge the ecosystem built around CUDA to increase GPU availability? [00:52:00]
• Agent Authentication: A critical emerging question is how to authenticate agents accessing websites on behalf of users. A new “SSO for agents” is needed, possibly involving biometrics or crypto solutions [00:54:41].