Trying automation tools
After more than five years away from day‑to‑day coding, curiosity about code‑automation pulled me back. In a policy‑reform class, the course pages looked like candidates for a lightweight plugin to surface definitions and references. An assistive browser extension took shape with automation as the co‑pilot—both to refresh skills and to watch how modern tooling changes development: what truly accelerates, and where human judgment still anchors decisions.
From deep learning to LLMs
The last time I was deep in the “AI” world (2016–18), the center of gravity was deep learning—CNNs for computer vision, RNNs/Seq2Seq, early attention, and libraries just getting real traction like TensorFlow and PyTorch. Word embeddings such as word2vec, GloVe and bag‑of‑words approaches were common pretraining stories.
Fast‑forward: transformers and scale changed the game. It’s fascinating to see today’s LLMs do in minutes what once took weeks of glue code and feature engineering. I still think architecture will need more breakthroughs for dependable, goal‑directed intelligence—more reinforcement‑learning‑like structure and designs less tightly coupled to raw data volume—but the trajectory is undeniable.
LLMs since 2019
A brief, non‑exhaustive list of notable LLM releases that shaped the ecosystem:
| Year | Model | Org | Notes |
|---|---|---|---|
| 2019 | GPT‑2 | OpenAI | Transformer LM at scale; staged release. |
| 2019 | T5 | Text‑to‑text framework; Unified transfer learning. | |
| 2020 | GPT‑3 | OpenAI | 175B params; few‑shot prompting mainstream. |
| 2021 | Gopher | DeepMind | Scaling studies on knowledge & reasoning. |
| 2022 | PaLM | Pathways; chain‑of‑thought prompting. | |
| 2022 | OPT | Meta | Open weights for research. |
| 2023 | GPT‑4 | OpenAI | Multimodal capability; stronger reliability. |
| 2023 | Claude 1/2 | Anthropic | Long context; constitutional AI. |
| 2023 | LLaMA | Meta | Efficient foundation weights; spurred open models. |
| 2023 | LLaMA‑2 | Meta | Commercially permissive license. |
| 2023 | Mistral / Mixtral | Mistral | Sparse MoE; strong open‑weights performance. |
| 2023 | Gemini | Multimodal suite (Nano/Pro/Ultra). | |
| 2024 | Llama 3 / 3.1 | Meta | Stronger instruction‑tuned open models. |
| 2024 | Qwen 2 / 2.5 | Alibaba | Competitive multilingual open weights. |
| 2024 | Claude 3 family | Anthropic | Opus/Sonnet/Haiku; safety‑aligned. |
Building the JDP plugin
The build journey leaned on code‑automation tools—primarily Cursor—to draft files, document changes, and cross‑check requirements. Documentation effort dropped noticeably because scaffolds and commit‑level notes were generated alongside edits. Packaging for the Chrome Web Store also became simpler; checklists and policy boilerplate were surfaced right when needed, making compliance and adherence to standards easier to satisfy without losing time in searches.
The resulting extension—Justice Definitions Project (JDP)—adds an instant definition pop‑up on right‑click and a side‑panel for repository‑wide search. For Indian law–facing communities—researchers, students and practitioners—faster access to expert‑curated definitions can reduce friction during reading and citation.
Try it on the Chrome Web Store and explore the open‑source repo on GitHub.
What automation helped—what it didn’t
Strengths
- Rapid scaffolding—UI shells, boilerplate, and docs are generated quickly.
- More time for the logic layer and data plumbing; less for rote tasks.
- Good for enumerating options, trade‑offs, and quick “what‑ifs”.
Pitfalls
- Direction drift: sometimes the assistant runs fast in a suboptimal path—needs course‑correction.
- Recursive loops: repeating fixes or re‑introducing removed code; supervision required.
- Resource hygiene: speed can accumulate tech‑debt; you still need strategy and constraints.
Useful acceleration, not abdication.
Engineering judgment remains central—defining intent, setting constraints, and choosing trade‑offs. Automation is excellent at momentum; you’re responsible for destination.
Acknowledgments
Prototype built as an experiment for the DAKSH team’s Justice Definitions Wiki Project. I got to know about DAKSH through the OpenTakshashila course on Judicial Policy and Reform.