@FinanceYF5: Counterattack of the AI Application Layer 1/ Large model companies are being encroached upon from the other side. Cursor, Decagon, Harvey, Notion are all doing the same thing: moving from API to self-trained models. Not to save money, but to take back the flywheel.

X AI KOLs Following News

Summary

AI application layer companies such as Cursor, Decagon, Harvey, and Notion are shifting from using large model APIs to self-trained models. This trend aims to regain control of the data flywheel rather than merely saving costs.

🧵Counterattack of the AI Application Layer 1/🧭 Large model companies are being encroached upon from the other side. Cursor, Decagon, Harvey, Notion are all doing the same thing: moving from API to self-trained models. Not to save money, but to take back the flywheel.👇 https://t.co/HgPtoXuHM5
Original Article
View Cached Full Text

Cached at: 05/16/26, 09:17 AM

🧵The Counterattack of the AI Application Layer

1/🧙Large model companies are being eaten from below

Cursor, Decagon, Harvey, Notion are all doing the same thing: moving from API to self-trained models.

It’s not to save money—it’s to take back the flywheel. 👇 https://t.co/HgPtoXuHM5

The Counterattack of the AI Application Layer

1/ Large model companies are being reverse-sniped

Cursor, Decagon, Harvey, Notion are all doing the same thing: moving from API to self-trained models.

It’s not to save money—it’s to take back the flywheel.

2/ Cursor’s Composer is a signal

The base uses the open-source Kimi, the magic is in their own post-training.

The feedback signal for whether code completions are good or bad only exists inside Cursor—no one else can get it.

3/ Training your own model is now industry consensus

Decagon, Abridge, OpenEvidence, Hippocratic, Cognition, Lovable, Harvey, Gamma are all self-training.

Every company building long-horizon agents is leaving the frontier API.

4/ Sutton’s bitter lesson no longer applies

Pretraining scaling wins only when the target is fixed.

But “how should this doctor’s medical record be written” is a moving target—only RL against real-world outcomes can approach it, and those outcomes exist only inside the product.

5/ The economics of specialized models already win

At the same capability budget, a specialized open-source model matches or exceeds frontier models on in-distribution agentic tasks.

The longer the task and the more tool calls, the bigger the gap—unit economics are nearly an order of magnitude better.

6/ Big players can’t catch up—it’s an organizational problem

First, frontier labs are structured as “one model serving everyone.” Second, specialization requires “multiple models for segmented customers.” Third, OpenAI just deprecated their fine-tuning API, admitting this isn’t their core business.

7/ The moat has changed

After software costs collapse, the only surviving barrier is the training signal no one else can see.

Every user interaction generates data, every training run produces a better model, and the flywheel spins inside the product—big players are outside the flywheel.

That’s all, original author @oneill_c

If you like this topic:

  1. Follow me (@FinanceYF5)
  2. Like + retweet the first post below

GenAI Summit SF 2026: A massive conference with tens of thousands in the Bay Area!

Will fans exclusive ticket link (15% discount available): Exclusive code: WILL Exclusive links: Luma: https://luma.com/genaisummit26?coupon=WILL… Eventbrite: http://eventbrite.com/e/1985545163032/?discount=WILL…

https://x.com/FinanceYF5/status/2054120536675057736…

Fun interactive science app ideas | Part 4

Someone made a turbofan jet engine airflow simulator

UI design GPT Images 2

Code Gemini 3.1 Pro

More demos ↓

Similar Articles

@FinanceYF5: Meta's move is not just about cutting costs, but also about reshaping its internal architecture around AI infrastructure, foundation models, and AI commercialization. This means the company wants to allocate more human resources to building model training systems, developing the models themselves, and developing products that convert models into revenue.

X AI KOLs Following

Meta is reshaping its internal architecture around AI infrastructure, foundation models, and AI commercialization. It plans to allocate more human resources to building model training systems, model R&D, and product development, aiming to promote AI strategy implementation and increase revenue conversion.

@Phoenixyin13: This latest blockbuster paper from Meta FAIR aims to tell the AI industry an important bellwether: "Large model data is ushering in the era of intelligent scientists." In this paper, a 4B small model precisely refined by Autodata not only crushes the same-scale models trained with traditional synthetic data on legal reasoning tasks, but also...

X AI KOLs Timeline

Meta FAIR's latest paper proposes the Autodata method, which uses an intelligent data scientist Agent to autonomously generate and optimize high-quality data, enabling a 4B small model to defeat a 397B large model on legal reasoning tasks. This indicates that data quality can bridge the gap in parameter count, providing new insights for data pipelines and scaling.