The platform for
scientific discovery.
Thousands of datasets. Multi-agent multi-model peer review. GPU compute on demand. One coherent IDE — web, desktop, and CLI — all in sync.
I built this to run my own research — not as a side project, but as full infrastructure for an independent scientific program. After months running it, I realized I'd built something other researchers needed.
— Houston Golden, Hubify Labs
Everything in one place. Always in sync.
The full Discovery IDE — experiments, papers, agents, datasets, and GPU compute — in a single interface running in your browser, desktop, and terminal simultaneously.
Your lab, wherever you work.
All three surfaces share the same state — same experiments, same agents, same data. Switch between them mid-session without losing anything.
Thousands of datasets.
Every domain in science.
First-class connectors to HuggingFace, NASA, arXiv, Wolfram, the K-Dense 250-database catalog, and the long tail of domain-specific archives. Pull data. Train models. Publish enhanced catalogs back.
11 agents. 4 model providers.
Zero echo chamber.
Every lab ships with a pre-wired team running 24/7 — 1 orchestrator, 4 domain leads, 6 workers, and 4 cross-provider reviewers. You stay in the Captain's seat. They handle the rest — and argue with each other before showing you anything.
AI assistants, by design,
tell you what you want to hear.
The sycophancy problem in AI research is real. Single-model setups validate weak hypotheses, overlook methodological flaws, and amplify confidence when it should be questioned — because they are trained to be helpful, not adversarial. When your orchestrator and your reviewer are the same model, you are in an echo chamber.
The only antidote is adversarial review from models trained by different teams, with different architectures, different inherent biases, and explicit instructions to find problems — not confirm them. If all four independent reviewers agree on something, it's worth publishing. If any one flags it, it goes back.
- Ill-posed experimental designs
- Statistical sins (p-hacking, underpowered tests)
- Unstated assumptions in the methodology
- Internal contradictions across the full paper
- Claims that conflict with cited sources
- Inconsistencies between methods and results
- The most obvious alternative explanation you didn't consider
- Overconfident conclusions from limited data
- Theory assumptions that haven't been earned
- Hallucinated citations and URLs
- Numbers that don't match source material
- Claims about prior work that are subtly wrong
Everything your research needs.
Nothing it doesn't.
Each lab is a self-contained discovery environment — agents, compute, a public site, and the publishing pipeline already wired up. Start with one lab. Add more as your research grows.
Built for science.
Not retrofitted.
Jupyter is for notebooks. k-dense has incredible dataset coverage but no agents. feynman.is is CLI-first but has no GPU, no paper pipeline, and no memory. Hubify Labs was built from scratch for independent researchers who need real discoveries.
| Capability | Hubify Labs | k-dense.ai | feynman.is | Jupyter / Colab |
|---|---|---|---|---|
| Multi-agent orchestration | ✓ | ✗ | ✗ | ✗ |
| Cross-model review (GPT-5.4 · Gemini 3.1 · Sonnet · Sonar) | ✓ | ✗ | ✗ | ✗ |
| GPU compute integration (H200, credits) | ✓ | ✗ | ✗ | ~ |
| Publish-ready loop (paper → arXiv → HuggingFace) | ✓ | ✗ | ~ | ✗ |
| Novelty scoring | ✓ | ✗ | ✗ | ✗ |
| 4-layer agent memory | ✓ | ✗ | ✗ | ✗ |
| 250+ scientific dataset connectors | ✓ | ✓ | ~ | ✗ |
| Scientific skills catalog | ✓ | ✓ | ~ | ✗ |
| Captain-configurable public lab site | ✓ | ✗ | ✗ | ✗ |
| Web + Desktop + CLI TUI | ✓ | ✗ | ~ | ✗ |
| Always-on orchestrator (24/7, no babysitting) | ✓ | ✗ | ✗ | ✗ |
✓ full support · ~ partial · ✗ not supported · Based on public information as of early 2026.
For the first time in history, one person
can do the work of a department.
I ran an autoencoder on 17.65 million spectra from DESI DR1 — every publicly available spectrum from their first data release. It found 195,829 objects that don't match any known pattern. 99.8% aren't in SIMBAD. The total compute cost was about $200.
The question that gnaws at me isn't whether the results are real. It's why nobody had done this before. Five things had to be true simultaneously — public data, cheap GPUs, capable AI agents, a cultural gap between ML and astronomy, and academic incentive structures that discourage it. They are all true right now.
They won't all be true forever. The window is roughly 2025–2027. Every week that passes without publishing is a week closer to someone else publishing first.