Sarah Guo Predicts Large-Scale Post-Training for Domain AI Companies by 2026
Notable VC Sarah Guo predicts that the hunger for reasoning power in long-sequence agents will drive many domain-focused AI companies to engage in post-training.
She pointed out that the coding domain is just the starting point, and more vertical industries will turn to proprietary model post-training due to the demand for long-cycle agents in the future.
This trend reflects the acceleration of domain optimization paths beyond general foundational models driven by reasoning costs and performance demands.
Source: Public Information
ABAB AI Insight
Sarah Guo, as the founder of Conviction Fund, has long invested in early-stage AI infrastructure and application companies. She previously led multiple rounds of financing for Perplexity, Character.ai, and has publicly stated that she has seen several vertical AI companies in coding, legal, and financial sectors initiate privatized post-training in 2024-2025, indicating her judgment is based on observations from her actual investment portfolio.
In terms of capital pathways, venture capital is shifting from purely foundational models to a "foundational + domain post-training" dual structure. Domain companies are collecting proprietary data, building long-sequence task datasets, and fine-tuning open-source or closed-source large models (such as the Llama series or Claude) to achieve differentiated capabilities. The motivation is to capture high-margin application layer opportunities brought by the explosion of reasoning-intensive agents while reducing dependence on a single foundational model supplier.
Similar cases include the deep collaboration between legal tech company Harvey AI and Cohere in 2025, as well as medical AI company Abridge achieving long-sequence understanding of clinical documents through post-training. The current AI industry is transitioning from a "pre-training race" to a "post-training + agent deployment" phase.
Structural judgment: This essentially belongs to an industry chain reconstruction driven by technological substitution. The extremely high demand for context and tool usage capabilities in long-sequence agents is shifting pricing power from general pre-training labs to vertical companies that possess domain data. The mechanism is that post-training can achieve task specialization at a cost far lower than training from scratch, creating new entry barriers and profit pools.
ABAB News · Cognitive Law
The stronger the reasoning hunger, the more valuable vertical post-training becomes.
General is the entry point, domain is the harvesting field.
The first to eat crabs is never the last.