Skip to content

Technology Path

A short history of what the field expanded the prediction target to, and which bottleneck each expansion left open. The tutorials turn the same sequence into learning units; this page is only the overview.

text token → multimodal latent → future trajectory → action → improvement loop

Period Representative shift New capability Unresolved bottleneck
2017–2020 Transformer, pretrained language models Scalable sequence representation and generation grounding, factual reliability
2020–2022 scaling, instruction tuning, RLHF Better instruction following and generalization long-horizon reasoning, verifiability
2021–2023 CLIP, BLIP-2, LLaVA, and related work Vision–language alignment spatial relations, hallucination, long video
2020–2024 diffusion, latent diffusion, DiT, flow matching High-quality continuous-distribution generation controllability, temporal and physical consistency
2022–2026 agents, world models, VLAs, test-time compute Tool use, environment prediction, action execution online learning, long-term memory, safety
Now verifiers, synthetic data, automated experiment search A shift from one-shot training to closed-loop improvement reward hacking, proving real progress

These branches are not a simple chain of inheritance. Visual representation, generative modeling, and control learning evolved in parallel before converging in world models, agents, and embodied systems.

When reading a new method

  1. Prediction target — tokens, patches, latents, frames, future states, actions, or experiment outcomes?
  2. Conditioning state — text, image, video, trajectory history, tool output, or an environment?
  3. Feedback source — next-token loss, preference, a verifier, a simulator, or the real world?
  4. Where improvement happens — parameters, data, prompts, test-time search, or external tools?
  5. Evidence — reported in a paper, reproduced with public code, independently reproduced, or observed in use?