Independent AI researcher building HSL and HoLo
I study whether deterministic, byte-native structural signals can replace part of the learned tokenizer and embedding stack while preserving useful model behavior. HoLo is the integration and evaluation vehicle; HSL is the core research product.
This is an open, evidence-first project. I publish code, weights, matched controls, implementation limits, and negative results. The current claim is works and is worth measuring, not SOTA.
- One-page compute research brief
- Latest matched language evidence: two epochs, 3.303B target bytes per arm
- Curated Hugging Face models, demos, and weights
The latest controlled gate found a small validation/test BPB improvement from exact previous-byte feedback at only +864 parameters and approximately 0.58% wall-throughput cost. It did not solve late-slot error propagation or weak free generation; those remain explicit targets for a convergence and matched-baseline study.
| Project | What it tests | Public evidence |
|---|---|---|
| HSL | A deterministic, modality-agnostic structural signal substrate | Core repository, specifications, experiments, and reports |
| hsl-embedding | Reproducible byte-to-HSL feature generation | Python package, tests, and CI |
| hsl-embedding-zero | Transformer input with zero learned input parameters | Python package, tests, CI, and released weights |
| HoLo_ZeRo | Language-model integration of the zero-input path | Training code, checkpoints, and documented limitations |
| HoLo-ToLk-STT | HSL-based speech recognition | Four-seed matched experiments and released artifacts |
| HoLo-ToLk-TTS | HSL-based text-to-speech | Working prototype, code, and weights; perceptual evaluation remains open |
| HoLo-FuSe | Frozen HSL features in diffusion conditioning | Reproducible prototype and retained negative-control evidence |
- Controlled language-model studies on clean English and Korean corpora
- Byte-level BPB, memory, throughput, parameter, and convergence economics
- Exact long-context and autoregressive rollout diagnostics
- Matched ablations that separate data quality, input representation, optimization, and scale
- Speech and diffusion prototypes as cross-modal feasibility studies
- A private exploratory autonomous-system prototype, not presented as public evidence
The language path remains harder than conventional learned BPE embeddings: token IDs are not used as learned semantic inputs. That makes direct quality comparisons incomplete unless parameter cost, byte exposure, compute, and convergence are matched.
- Works-not-wins: feasibility before superiority claims
- Matched controls: same data, exposure, seed, and evaluation whenever possible
- Negative results stay public: rejected paths are evidence, not hidden failures
- Reproducibility: code, weights, hashes, and run contracts where practical
- Economy matters: quality is evaluated together with parameters, VRAM, throughput, and data cost
- GitHub repositories
- Curated Hugging Face research collection
- All Hugging Face models and demos
- HSL on PyPI
- Zero-input package on PyPI
- HSL paper on Zenodo
- ORCID
- Contact:
ggunio5782@gmail.com
I am seeking research-compute sponsorship or credits for reproducible scaling studies, especially controlled HSL/HoLo language, speech, and diffusion experiments. The immediate goal is not to declare a breakthrough; it is to determine rigorously where the approach works, where it fails, and whether its parameter and memory economics survive at larger scale. The compute research brief specifies a 100 H100-hour pilot with measurable success and rejection criteria.
HSL은 바이트에서 고정 구조 신호를 만드는 입력 기판이며, HoLo는 이를 언어·음성·이미지 모델에서 검증하는 통합 연구입니다. 학습된 토큰 임베딩 일부를 고정 신호로 대체할 수 있는지, 그리고 그 과정에서 파라미터·메모리·처리량 측면의 경제성이 유지되는지를 공개 실험으로 확인하고 있습니다.
현재 공개된 결과는 가능성과 작동성을 보여 주지만 완성형이나 성능 우위를 뜻하지 않습니다. 코드, 가중치, 대조군, 실패 사례를 함께 공개하며 더 큰 규모의 재현 가능한 검증을 위한 컴퓨팅 협력을 찾고 있습니다.