Lucas Atkins is Arcee AI’s chief technology officer and head of research. His public work covers efficient model training, open model releases and the infrastructure needed for coding agents to work across long tasks.[1][5]
Start with Spectrum: Targeted Training on Signal to Noise Ratio →
Links
Arcee AI Research talkChief technology officer and head of research · Arcee AI
Career
Contributions
Choose which parts of a model to train
Atkins coauthored Spectrum, a post-training method that calculates a signal-to-noise measure for model modules, trains selected modules and freezes the rest. The authors compare it with full fine-tuning and QLoRA, reporting memory-efficiency and quality results in their experiments. The useful question is which parameters need to change for a task; the paper’s measurements do not establish the same savings for every model or training setup.[4]
Preserve the state a long-running agent needs
Atkins coauthored Arcee’s introduction of NAC, an open-source runtime for long engineering tasks. It separates work into structured episodes and persistent workstreams, coordinating parallel workers while retaining state across context windows. The design addresses a common failure in extended coding sessions: a growing transcript can bury earlier decisions, and compaction can remove details needed later. Persistent records make those decisions available for subsequent work.[5]
Tools, writing & talks
Writing, research & talks
Start here
A method for selectively training model modules while freezing the remaining parameters.
Read the module-selection procedure and comparisons before interpreting memory savings. Familiarity with fine-tuning helps; reproduce a small baseline before applying selective training to a larger model.
Co-author[4]
github.com