csvoss: imagine steering LLMs to give 'good code' or 'no emojis' without begging in custom instructions
csvoss · x · 2026-10-13
csvoss argues that steering vectors remain unproductized despite their obvious potential: directly steering LLM outputs — "good code", "only sarcasm", "no emojis", "bad advice but funny" — without having to beg and cajole in custom instructions.
The point makes the practical value of interpretability research concrete: model behavior could be controlled like a dial, rather than through prompt engineering.
Related event: OpenAI researcher: steering vectors useful but not yet productized(3 posts)→
More from Research
- BenchCAD 2.0: best frontier model scores just 0.59/1.0 on professional CAD modeling — hanjie_chen · 2026-10-13
- Planning humanoid falls: IROS 2015 paper treats falling as dynamic programming over contacts — chris_j_paxton · 2026-10-13
- BlackboxNLP 2026 program goes live: EMNLP workshop in Budapest on Oct 29 — gsarti_ · 2026-10-13
- Label-free minibatch OT coupling beats conditional OT, hits 1.91 FID on ImageNet-256 — AlexanderTong7 · 2026-10-13
- 13B model trained only on pre-1931 text shows personas emerge from training data era — sethlazar · 2026-10-13
- EXL details GNN fraud detection system with 27-33% detection gains — AI Engineer · 2026-10-13