Wei Wu

Wei Wu

Scientist and Engineer

Reflection of 2025 and Predictions for next year

2 minute read

Published:

Looking back at 2025, it was a year of striking progress and genuine breakthroughs in AI. At the same time, it was also a year filled with uncertainty—technological, economic, and societal.

The year opened with the release of DeepSeek R1, which for many marked the beginning of a more experience-driven era of AI. Reinforcement learning (RL) has moved beyond controlled benchmarks and is increasingly shaping how models learn from interaction, feedback, and long-horizon objectives. This shift represents a meaningful step in how AI systems adapt to complex, dynamic environments.

One of the core challenges in applying AI to real-world problems has always been reward definition. Whether in rule-based systems or neural networks, specifying good objectives is hard. Reinforcement learning—especially RLHF—demonstrated its impact earlier with ChatGPT, where preference learning and reward modeling made GPT-3.5 dramatically more usable than GPT-3 when it first appeared in 2020.

Still, most real-world domains resist clean reward signals. Mathematics and programming are exceptions. In math, answers can often be verified directly. In coding, correctness can be evaluated through test cases. It is therefore no surprise that these domains became early proving grounds for advanced AI systems.

In 2025, AI systems achieved historic results in competitive mathematics and competitive programming, reaching gold-medal-level performance in settings such as the IMO and ACM-ICPC, using natural language input and output. Beyond academic benchmarks, AI-assisted coding also saw massive real-world adoption. Tools like Cursor and Claude Code became widely used, changing how developers write, debug, and reason about software.

The obvious question now is: what comes after math and coding? The answer is far from clear—but 2026 promises to be an exciting year to find out.

Below are a few tentative predictions.

⸻

Startups

IPOs are likely to resume in 2026, with companies such as OpenAI and Anthropic among the first. At the same time, acqui-hires may become more common, making hiring increasingly difficult for early-stage AI startups as talent concentrates around a few dominant players.

⸻

AI Research

AI will continue to make steady progress in 2026. While breakthroughs in solving Clay Millennium Problems remain uncertain, we may see meaningful advances in formal reasoning, automated theorem proving, and hybrid systems that combine symbolic structure with learned representations.

⸻

Economy

AI-driven displacement will become more visible. While productivity gains are real, AI-related unemployment and job reshaping will become a more serious societal issue, forcing faster adaptation in education, labor markets, and policy.

⸻

You May Also Enjoy

Social Contract

5 minute read

Published:

The idea of a social contract is often introduced as a philosophical abstraction: an imagined agreement through which individuals consent to authority in exchange for security and order. Yet historically, social contracts are not ideals invented first and implemented later. They are ex post equilibria—practical arrangements that stabilize societies after major shifts in how value is created and distributed. Whenever a new production technology fundamentally alters this balance, the prevailing social contract comes under strain and is eventually redesigned.

Video Generation and SORA

less than 1 minute read

Published:

This is the begining of continous world modeling with deep learning. Scaling indeed works and opens up new possibilities.

Gen AI

1 minute read

Published:

I am going to post here about my thoughts on deep learning and AI, with regards to the recent development of large language models (LLMs) and world models, and the impact on the real world.