Loved the course!! What to do next to get updated to 2026

[za-world-xd]

New here, loved this course but it was made in 2022 what should i learn next to get updated to 2025/26 ai ml?
would be glad if someone could help

[yhay81]

The course is not obsolete. The parts that matter most — tensors, autograd, nn.Module, datasets/dataloaders, training/evaluation loops, transfer learning, experiment tracking, and deployment — are still the foundation. I would not restart from zero.

A practical 2025/26 bridge would be:

1. Re-run one course project on the current stable PyTorch, fixing any API/deprecation differences rather than only re-watching videos.
2. Learn the main PyTorch 2.x additions: torch.compile, automatic mixed precision, the profiler, and the current serialization/export path. The official PyTorch tutorials are the best source because they track the installed release.
3. Pick one specialization instead of trying to learn “all AI”:
   • vision: modern pretrained backbones, detection/segmentation, augmentation;
   • NLP/LLMs: Transformers, tokenization, fine-tuning/PEFT (LoRA), evaluation and inference;
   • generative vision: diffusion models;
   • ML engineering: reproducible data pipelines, testing, experiment tracking, serving and monitoring.
4. Build one end-to-end project with a dataset you care about: baseline → metrics → error analysis → improved model → small API/demo → README containing results and limitations.
5. Then read/replicate one recent paper or strong open-source implementation in that specialization. The course's paper-replication workflow is still exactly the right skill.

For an LLM-focused path, for example, I would do: current PyTorch refresher → Hugging Face Transformers/course → fine-tune a small open model with LoRA → create a held-out evaluation set → serve it locally and measure latency/memory. Avoid jumping straight to training a large model from scratch.

A good readiness check is whether you can explain and implement, without copying a notebook:

• a correct train/validation/test split;
• a training loop with mixed precision and checkpointing;
• appropriate metrics and error analysis;
• inference on unseen data;
• reproducible setup and a deployable demo.

If you can do those, the next step is projects and specialization, not another beginner course. Also, install commands change with hardware/CUDA versions, so use PyTorch's current Get Started selector rather than an old pinned command.