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

Summary

• The dbt–Fivetran merger is an IPO-scale response, not evidence that the modern data stack is ending. Sarah Catanzaro calls that reading a “fundamentally wrong take”: both companies were beating revenue targets, while today’s IPO bar is above $600 million and the combined business could approach that level. Frontier labs already use both products, making AI an extension of the data market rather than its replacement. Demand for analytics engineers and data scientists did not explode into “armies,” but the tools remain prevalent.
• The 2025 funding market frequently produced $100 million seed rounds before founders could articulate a six-month roadmap. Sarah sometimes saw billion-dollar asks paired with a seven-day decision window; her preferred calculation is 12–24 months of milestones, headcount, compute and equipment, plus roughly a 20% buffer. She has made exceptions when she knows the team well, but warns prospective employees that “valuation until a company exits is an entirely made-up number.”
• World models have commercial promise, but the category still lacks a stable definition or proven generality. The host noted there are about three definitions; Sarah sees uses in video, autonomous driving and potentially coding, yet a model built for video-game generation might not transfer to factories or robotics. She chose “might” strategically: cross-domain generalization could be solved, but “we’re not there yet today.”
• Memory and continual learning could become retention infrastructure for AI applications whose initial magic no longer prevents churn. Sarah agrees personalization is a plausible 2026 theme across consumer and enterprise products. It must extend beyond stored facts into learning skills as users, repositories, languages and frameworks change. The hard implication is stateful weights in an inference stack designed to be stateless: “Human intelligence is incredibly dynamic, and yet artificial intelligence is just so static today.”
• Sarah thinks RL environments are “just a fad,” despite labs paying seven or eight figures for them, though she allows they may create short-term value. She notes that labs could build them in-house but do not, comparing this with their past spending on “piss-poor data annotation.” Her durable call is that “the best RL environment is the real world,” augmented by carefully designed tasks and rubrics rather than simple app clones.
• The startup archetype Sarah wants pairs an application with research that directly unlocks product quality. Harvey’s retrieval work, Sierra’s rule-following focus and Runway’s models fit the pattern. Sarah says she does not think Runway would have built models if it had not had to. The best companies begin with “I’m trying to do this thing,” then solve the gnarly technical constraint blocking it.

Deep dive

1. AI is extending the modern data stack, not killing it

• On dbt–Fivetran, Sarah rejects the “end of the modern data stack” thesis: both companies were growing healthily and beating revenue targets. The host called them presumptive category winners, and Sarah agreed. The merger accelerates liquidity in an IPO market demanding far more than $100 million of revenue. She believes the combined company could be close to $600 million, though she did not have the exact figure.

• AI labs strengthen the thesis. Many big frontier labs use both products; within weeks of Thinking Machines’ formation, dbt was already important for managing training datasets and understanding user interactions. Analyzing interactions with agents and LLMs can be more complicated than conventional analytics.

• Conventional analytics workloads were surprisingly predictable because much of the work came from deterministic systems such as BI dashboards. Dataset analysis, curation and preparation are more ad hoc and therefore less predictable. Sarah says that could affect work on learned indexes and learned optimizers, although it is not yet clear that it changes the overall approach to data infrastructure.

• Demand for analytics engineers and data scientists did not become one-third of company personnel, but demand for the tools remains prevalent. Sarah thinks companies need data and analytics teams, just not armies of them.

• Sarah’s explicit correction: “That was something I got wrong.” Human-facing data catalogs struggled as the category, partly because catalog features inside Fivetran, dbt, Hex and Snowflake were “good enough” for humans. The missed opportunity may have been machine-facing metadata services—for agents, microservices and other systems—and governance rather than discoverability.

• GPU economics make loading efficiency material: if data cannot reach a GPU efficiently, the GPU sits idle and becomes a cost. Amplify-backed Spiral’s Vortex format targets that bottleneck.

2. Mega-seeds are obscuring the difference between capital needs and signaling

• “Crazy” in 2025 meant companies frequently raising upward of $100 million at seed with a long-term vision but no near-term roadmap. Sarah’s anxiety peaks when founders cannot explain the next six months yet demand a decision in seven days: that leaves little time to establish conviction or determine whether they need a partner who will work closely with them. Her preferred calculation is the resources needed for the next 12–24 months of milestones—headcount, compute and equipment—plus roughly a 20% buffer.

• Sarah admits making investments without a clear near-term roadmap when she knows the people well enough to believe they will figure it out. Some companies do need the money: Periodic, for example, needs to build a wet lab capable of supporting high-throughput biology, which is costly. Her concern is that other founders raise mainly because a large valuation makes recruiting easier; candidates are drawn to unicorn or near-unicorn status, with big labs offering prestige and money as the alternative.

• The recruiting pitch can turn less than 0.1% equity into a purported $10 million asset. The discussion also touched on loans or buybacks that provide employees a way to realize a dollar value. Sarah’s objection is that the valuation reflects little traded volume, and if the company spends its capital then exits below the amount raised, its team may receive nothing.

• The host’s Antithesis example showed how money can overwhelm substance: the announcement led with a $100 million seed associated with a Street firm. Sarah described Antithesis as deterministic simulation, and the host named Palanteer and WarpStream. Sarah agreed that funding currently sends a signal, but warned that pursuing “the most money at the highest valuation” is not a sound basis for choosing a job.

3. World models are investable before they are well-defined

• The host noted about three competing definitions of “world model.” Sarah says there is still substantial confusion about what the term means and therefore what such a system should be used for. She sees market potential in video models, autonomous driving and potentially coding, with the exact opportunity depending on the definition.

• The unresolved issue is transfer: a world model for game generation might not generalize to a factory or robotics. The host cited evidence from the General Intuition podcast that it could; Sarah said that is possible, but held the line that generality has not arrived yet.

4. Memory must graduate from stored facts to continuously learned skills

• Fast-growing AI applications often still show low retention and high churn. The host framed Cursor’s challenge as retaining users when Windsurf, cloud code, Cognition or another rival releases a feature. Sarah called Cursor Rules the “shittiest form of memory,” and agreed with the host that memory is poorly implemented across many surfaces. She also said consumers tolerate today’s weak implementations, so an early improvement is probably better than nothing.

• Sarah agreed that personalization or the consumerization of AI could be a key 2026 theme, but not only for consumers or prosumers. Enterprises adopting tools such as Devin or Augment also want their models to learn. The host argued that as AI’s magic becomes familiar, founders must return to ordinary SaaS disciplines such as K-factor and retention.

• Continual learning means absorbing skills from interactions and adapting as frameworks, languages and repositories change—not merely remembering preferences. But updating weights makes them stateful while inference is currently stateless, creating “fun, gnarly problems” around loading, unloading and caching. The host also noted that product management becomes harder when a product combines the user and memory: a bug may belong to memory or to the core system.

5. Real usage beats cloned RL environments, while applications direct research

• Sarah’s deliberately falsifiable call: “I’m actually okay to be wrong, but I think RL environments are just a fad.” She notes that labs pay seven or eight figures for them even though they could build them in-house, and compares this with their earlier willingness to pay for “piss-poor data annotation.” She allows that RL environments may create value in the short term.

• Sarah’s stronger point is that the best RL environment is the real world: a DoorDash clone is less compelling than logs and traces from DoorDash itself. The host pointed to Cursor’s use of real user activity to improve its coding agents and Tab. Task selection and rubrics still matter, but simply cloning an application is not very useful.

• Sarah’s favored companies connect research directly to an application unlock: Harvey has advanced retrieval in a way that benefits its product, Sierra focuses on rule following for customer support, and Sarah says she does not think Runway would have built models if it had not had to. Memory becomes compelling when it enables a dramatically better experience—or something previously impossible—not when treated as detached infrastructure.