Why did data catalog companies struggle in the modern data stack?

Many data catalog companies struggled because essential cataloging features were integrated into existing tools like Snowflake and DBT, making standalone solutions less necessary for human users. The opportunity might lie in machine-focused metadata services.

What is considered 'crazy' in the current AI startup funding environment?

Crazy funding involves raising upwards of $100 million in a seed round with a long-term vision but no clear near-term roadmap, often prioritizing high valuations over concrete milestones.

Why are AI startups raising such large seed rounds without clear roadmaps?

High valuations and large raises make it easier to attract talent, as candidates are drawn to unicorn status. Startups may also offer enticing equity packages based on these inflated valuations.

What is the current sentiment around world models in AI research?

There's significant skepticism surrounding world models, partly due to unclear definitions and the immense success of LLMs, which some feel negates the immediate need for them. Generalizability across different use cases remains a challenge.

What makes memory management and continual learning important for AI applications?

These capabilities are crucial for personalization and user retention in AI applications. As AI models become more dynamic and adapt to user preferences and changing data, they can offer a more compelling and less static experience.

Are RL environments a fad, and what makes a good one?

Some view RL environments as a fad, especially those that are mere clones of existing apps. The best RL environment is considered the real world, or using logs and traces from actual applications like Cursor does, emphasizing real-world interaction over simulated ones.

What kind of startups excite investors like Sarah Catanzaro?

She is excited by startups building infrastructure, platforms, and tools, particularly those that solve hard technical research problems to unlock novel applications or dramatically improve existing user experiences, like in areas of continual learning or rule-following.

[State of AI Startups] Memory/Learning, RL Envs & DBT-Fivetran — Sarah Catanzaro, Amplify

Latent Space Podcast

Science & Technology4 min read29 min video

Dec 30, 2025|1,961 views|38|5

Save to Pod

Key Moments

TL;DR

AI startup funding is booming with large seed rounds, but focus is shifting to memory, personalization, and practical applications over pure hype.

Key Insights

The merger of DBT and Fivetran signifies an acceleration toward IPO rather than the end of the modern data stack.

Data cataloging struggled as a standalone category because existing tools offered sufficient features for human users.

The funding environment in 2025 saw massive seed rounds ($100M+) for AI startups, often with unclear near-term roadmaps.

Memory management and continual learning are crucial for AI application retention and personalization, addressing user churn.

RL environments are potentially a fad, with the real world and existing applications serving as more effective learning grounds.

The most exciting startups solve hard technical problems to deliver significantly improved user experiences or enable new capabilities.

THE EVOLUTION OF DATA STACKS AND AI INTEGRATION

The conversation begins by addressing the perceived 'end' of the modern data stack following the DBT and Fivetran merger. This is reframed not as a decline, but as a strategic move by two healthy, growing companies to accelerate their path to IPO, requiring larger revenue thresholds. Notably, frontier AI labs are leveraging these data tools, highlighting the symbiotic relationship between data management and AI development. The demand for data analytics tools remains strong, even if the roles of data scientists and analytics engineers have evolved.

CHALLENGES AND OPPORTUNITIES IN DATA CATALOGING

A key insight reveals that data cataloging, once predicted to be a significant category, struggled to gain traction as a standalone product. This was largely because existing platforms like Snowflake and DBT incorporated sufficient data cataloging features for human users. The speaker posits a potential missed opportunity: building data catalogs not for humans, but for machines, akin to metadata services for agents or microservices. The focus may have been misplaced on discoverability rather than the more critical aspect of governance.

INFRASTRUCTURE SCALABILITY AND DATA LOADING EFFICIENCY

The scaling of existing data infrastructure to meet AI use cases has been surprisingly elegant, though the sheer scale of AI companies is immense. Efficient data loading, especially to GPUs, is critical to avoid idle compute costs. Companies are developing specialized file formats, like 'Vortex' from Spiral, to optimize this process. While current transactional databases are often sufficient, the potential explosion in transactions from widespread agent use could necessitate future database paradigm shifts.

THE DYNAMIC AND ANXIOUS FUNDING LANDSCAPE

The 2025 funding environment is characterized by unusually large seed rounds, sometimes exceeding $100 million, often granted with long-term visions but without clearly detailed near-term roadmaps. This trend causes investor anxiety, as responsible fundraising necessitates defining milestones and required resources for the next 12-24 months. Founders are encouraged to articulate concrete plans, not just grand ambitions. The pressure to raise at high valuations also impacts hiring, as candidates are attracted to 'unicorn' status and substantial equity offers.

THE CRITICAL ROLE OF MEMORY AND CONTINUAL LEARNING

Memory management and continual learning are identified as major themes with significant market potential, particularly for AI applications. These capabilities are crucial for improving user retention and personalization, combating the high churn rates often seen in rapidly growing AI apps. Implementing effective memory goes beyond basic user recall; it involves learning new skills and adapting to the dynamic nature of the world. This presents complex systems and infrastructure challenges, involving stateful inference and dynamic weight updates.

WORLD MODELS AND THE REALITY OF RL ENVIRONMENTS

The concept of 'world models' is met with skepticism, partly due to a lack of clear definition and generalization issues across use cases. While applications in video editing, autonomous driving, and coding are emerging, their efficacy depends heavily on specific definitions. Similarly, Reinforcement Learning (RL) environments are viewed by some as a potential fad. The argument is made that the most effective RL environment is the real world, or simulations based on logs and traces from actual applications, rather than mere app clones.

IDENTIFYING THE MOST EXCITING STARTUP ARCHETYPES

The most compelling startup investments often lie in infrastructure, tools, and platforms that address evolving needs, such as those in continual learning. However, the most exciting companies often emerge when they tackle hard research and engineering problems as a prerequisite to delivering a significantly improved user experience or enabling entirely new capabilities. This principle applies from application-level advancements like Retrieval-Augmented Generation (RAG) to foundational challenges like rule-following for customer support.

Mentioned in This Episode

●Products

●Software & Apps

●Companies

Common Questions

Data infrastructure tools like DBT and Fivetran are increasingly crucial for AI, extending beyond traditional analytics to manage and prepare training datasets for LLMs and other AI models.

Topics

RL Environments Personalization Data Catalogs

Mentioned in this video

companyWarpStream

Mentioned as a potential user of Antithesis's technology.

companyAntithesis

A company that recently emerged from stealth, raising a $100 million seed round led by James Street for AI testing and deterministic simulation.

productVortex

A file format developed by Spiral, designed to make data loading efficient, particularly for GPUs.

companyCognition

An AI application company that users might switch to, highlighting the importance of user retention and personalization.

softwareAugment

An enterprise AI adoption tool that would benefit from personalization and model learning similar to consumer applications.

companySpiral

A portfolio company that has developed a file format called Vortex to make data loading efficient, specifically for GPUs.

companyJames Street

Led a $100 million seed round for AI testing company Antithesis.

companyAtlan

A company in the data catalog space that, along with others, has struggled as a standalone category.

companyThinking Machines

A company where DBT was already an important part of their stack for managing training datasets within weeks of its formation.

companyGirl Data World

A company mentioned in the context of the struggles of the data catalog category.

companyPeriodic

An example of a company that requires significant funding (over $100 million) due to the costs associated with building out a high-throughput wet lab for biology.

toolHex