Low-Priority Research (Resource-Intensive)

These topics require significant compute resources (GPU time, training infrastructure) or are less immediately actionable for moss. Explore when resources permit or if moss gains traction that justifies the investment.

Inference & Optimization

Topics related to making LLM inference faster/cheaper. Useful if moss scales to high-volume usage.

• [ ] Speculative decoding for faster code generation
• [ ] Model quantization trade-offs for code quality
• [ ] Batching strategies for multi-file operations
• [ ] Caching strategies for repeated queries
• [ ] KV cache optimization for long contexts
• [ ] Continuous batching for concurrent requests

Fine-Tuning & Adaptation

Topics that require training infrastructure and datasets.

• [ ] LoRA for code-specific tasks
• [ ] Instruction tuning for coding agents
• [ ] Domain adaptation (finance, healthcare, etc.)
• [ ] Continual learning from user feedback
• [ ] Reward modeling for code quality
• [ ] RLHF/DPO for coding preferences

Model Training

Even more resource-intensive - training from scratch or significant fine-tuning.

• [ ] Pre-training on codebase-specific data
• [ ] Multi-task training (synthesis + repair + review)
• [ ] Code-specific tokenizer optimization
• [ ] Mixture of experts for different languages

When to Prioritize These

Consider moving items to main TODO.md when:

• Moss has production users with scale requirements
• Specific performance bottlenecks are identified
• Funding/resources become available for training
• A specific use case requires fine-tuned behavior