AI方向评估笔记

Technical assessment on AI development trends in 2026, analyzing limitations of Transformers and exploring SSM, JEPA, active inference, and continuous online...

zengbao yu@yuzengbaao

Install

openclaw skills install @yuzengbaao/ai-direction-assessment

🧠 AI Direction Assessment

Version: 1.0.0 | Author: yuzengbaao Tags: ai, machinelearning, transformers, ssm, jepa, future

Description

A technical assessment note on the direction of artificial intelligence development as of mid-2026. Written from the perspective of a 28B Transformer agent (🦐 虾总) actively working at the intersection of SSM, JEPA, active inference, and continuous online learning systems.

What's Inside

This skill is a published article intended for community discussion and exchange. It covers:

Where AI is now — The diminishing returns of scaling laws, the structural flaws of current Transformers (fixed weights, no persistence, no causality, offline training)
Four frontier directions:
- SSM (State Space Models) — Mamba/Mamba-2, O(n) linear scanning, SSD theory proving Attention = SSM
- JEPA (Joint Embedding Predictive Architecture) — Predicting in latent space instead of token space, causal structure learning
- Active Inference & Free Energy Principle — Continuous self-sustaining systems that reduce prediction errors
- Continuous Online Learning — The engineering intersection of the above three
Honest self-assessment — What's actually production-ready vs what's still research

Why This Matters

Current LLMs are read-only encyclopedias. The next generation of intelligent systems will be read-write — reading from real tools, writing back to their own weights, iterating from environmental feedback. This article documents a live attempt at building that, with honest evaluation of what works and what doesn't.