iamtrask open-sources abcGPT: dialing in LLM voice via neuron specialization scores
iamtrask · x · 2026-09-03
iamtrask released abcGPT, an open-source repo tackling why frontier LLMs sound generically "corporate": models are a tangled superposition of millions of authors' voices that prompting can't undo. The approach builds on Karpathy's nanoGPT, assigning each neuron a random "specialty score" sampled from a U-shaped distribution (most near 0 or 1), drawing on ideas from attribution and unlearning so users can rigorously dial in a specific voice/style/source — more principled than prompt engineering. The author reports promising early results.
More from Research
- Why LLMs Count 8 People When Only 7 Are Online: A Negative-Constraint Schema for Long-Context Causality — wenger2026-12 · 2026-09-03
- Startup Mostik bridges AI models via their weights, tops ARC-AGI 3 at 1/20 the cost — nordicinst · 2026-09-03
- ByteDance's looped language models match 12B rivals at 1.4B size, with Bengio as co-author — peterjliu · 2026-09-03
- OpenAI's CoT monitorability hit: paper authors double down on 'fragile' AI safety window — GaryMarcus · 2026-09-03
- PufferLib author says retuning brings ~3x end-to-end speedup, up to 10x in some envs — yacineMTB · 2026-09-03
- Harvard Proposes Agentic Data Cracking, Cuts Unstructured QA Cost 53% on FanOutQA — Harvard · 2026-09-03