Predictions for AI in 2025

Jiang Wu · 2025-01-01

*Originally published in Chinese on January 1, 2025.*

These are my predictions for AI in 2025. They are grouped by theme, and the "little dragons" in item 2 are the group of Chinese startups building foundation models.

Models and services become infrastructure

  1. Model APIs become infrastructure and are offered free, bundled with services such as storage and traffic.

How the AI industry reshuffles

  1. The little dragons split apart completely. Only two or three keep to the foundation model route, and the rest start plugging into service modules.
  2. Big companies' AI investment divides further, into product plus AI and AI productization.
  3. AI giants are classified by the compute reserves in their hands and no longer by revenue or user numbers.
  4. Chinese model vendors export in batches to countries other than the US and China.

Applications and product innovation

  1. At least a hundred products built by indie developers reach a scale of 100,000 daily active users.
  2. Image tools all become AI tools, with the basic module capabilities enhanced across the board.
  3. Video gets better in symmetry and consistency, and AI-made short dramas account for more than 10% of the supply on short-drama platforms.
  4. Phone operating systems become AI-driven and on-device models start to be used at scale. A phone's system assistant will replace some of today's inflated-metric AI assistants.
  5. AI's first real displacement in industry is the work of middle-level intelligence.
  6. Better multimodal capability pushes personalized education toward real deployment, and some large-scale revenue and fundraising will happen there.

New interactions and hardware

  1. Model capability drives new kinds of AI interaction, with breakthroughs beyond smart hardware in other specialized hardware, such as metrology, refrigeration, and special equipment.
  2. Smart glasses break out in the first half of next year, with an estimated 40 to 50 entrants. A first shakeout follows in the second half, leaving 5 to 10, and the one with the highest market share is H... (the sentence is cut off in the Chinese original).
  3. Another wave of competition begins over smart interaction and screens in the home, from smart speakers to smart mirrors.
  4. A home AI compute center may appear, as a dedicated NAS, a standalone server, or a router. I'm not sure which.
  5. Products built on highly structured knowledge bases start to be applied at scale, in areas like finance, legal, tax, and healthcare.
  6. Education becomes more finely divided by scenario. Beyond exam prep, vocational and general education get their first openings.

Personal data and AI assets

  1. Modularizing personal data assets becomes a business.
  2. Driven by AI, a person's stored assets will fall into three parts: hardware assets and cloud assets, with cloud assets divided into service-based and non-service-based.

AI assistants and the app ecosystem

  1. Apps of all kinds open interfaces to AI assistants. The user's decision center may be taken over by the operating system or by a standalone AI assistant.
  2. Physical services at the terminal extend upstream, while upstream AI decision services eat downward. The core connecting piece in the middle may change. Today the command center is the phone, and in the future it may be any hardware device.
  3. Several main directions for app startups: personal knowledge bases, personalized communities, vertical education, tool integration, and companion-style games.

Evaluating AI application value and startup opportunities

  1. The new standard for judging an application's added value is not token consumption. It is the cumulative absolute value of the tokens a service delivers.
  2. Between the API and the service, the chain has plenty of room for optimization, and startup opportunities in vertical applications and scenarios will come from there.

Home · All articles · Topics · 简体中文