---
license: other
license_name: open-to-work
license_link: https://1999.wtf
language:
- en
- es
pipeline_tag: text-generation
metrics:
- vibes
tags:
- generalist
- instruction-tuned
- creative-technology
- human
- open-to-work
- hireable
model-index:
- name: Noah-McLaughlin-7B
  results:
  - task:
      type: bill-splitting
      name: Mental Math at Dinner
    dataset:
      type: gsm8k-irl
      name: GSM8K (Restaurant Split)
    metrics:
    - type: accuracy
      value: 94.2
      name: pass@1
  - task:
      type: code-generation
      name: Ships Actual Code
    dataset:
      type: humaneval-irl
      name: HumanEval (Production)
    metrics:
    - type: accuracy
      value: 81.0
      name: pass@1
  - task:
      type: social-reasoning
      name: Reads the Room
    dataset:
      type: hellaswag-irl
      name: HellaSwag (Standups & Dinners)
    metrics:
    - type: accuracy
      value: 88.7
      name: acc
  - task:
      type: honesty
      name: Honesty in Standups
    dataset:
      type: truthfulqa-irl
      name: TruthfulQA (1:1s)
    metrics:
    - type: accuracy
      value: 99.1
      name: acc
  - task:
      type: naming
      name: Naming Things Well
    dataset:
      type: mt-bench-irl
      name: MT-Bench (Products, Ideas, Directions)
    metrics:
    - type: score
      value: 9.1
      name: judge-score
  - task:
      type: long-context
      name: Holds a Vision While Everything Changes
    dataset:
      type: longbench-irl
      name: LongBench (Roadmaps)
    metrics:
    - type: accuracy
      value: 92.0
      name: acc
---

# Noah-McLaughlin-7B

A generalist, instruction-tuned human optimized for creative technology. Open weights, open to work.

> ⚠️ Not yet quantized. Runs best with adequate sleep and at least one body of water nearby.

## Model Details

### Model Description

Noah-McLaughlin-7B is a decoder-only generalist trained on a finance curriculum and then heavily fine-tuned on real-world product deployment. It takes a napkin idea and returns a live product. It thinks in systems but is comfortable in the details. Strong few-shot learner in unfamiliar domains; degrades gracefully under ambiguity.

- **Developed by:** Two non-technical co-founders (pre-training)
- **Funded by:** Mass General Brigham Investment Office (2021–2024); bootstrapped since
- **Shared by:** [1999 Labs](https://github.com/orgs/1999labs/repositories)
- **Model type:** Generalist; instruction-tuned via real-world feedback (the translator layer between technical and non-technical, vision and execution)
- **Language(s):** English (native), JavaScript, TypeScript, Solidity, and conversational Spanish
- **License:** Open to work
- **Finetuned from:** `homo-sapiens-base`

### Model Sources

- **Repository:** https://github.com/orgs/1999labs/repositories
- **Papers (ongoing):** [Side Effects Magazine](https://github.com/1999labs/side-effects-mag)
- **Demos:** [Are.na Pairings](https://huggingface.co/spaces/noahmclaughlin/Are.na_Pairings) · `npm whoami`

## Uses

### Direct Use

Strategy, product, creative direction, research, operations — anything that requires moving fluidly between disciplines. Particularly capable at zero-to-one.

### Downstream Use

Fine-tunes well onto small teams that build with intention. Composes cleanly with designers, engineers, and other strong-willed collaborators.

### Out-of-Scope Use

Not designed for micromanagement, sitting in a single lane, or vibecoding ChatGPT wrappers. Will refuse these inputs and return a polite explanation.

## Bias, Risks, and Limitations

- Holds strong priors and **will push back**. Open to having them updated given a sufficiently good gradient.
- Context window of roughly 8K tokens (one good meeting). Summarize long threads.
- Performance degrades measurably without surfing or skiing.
- Occasionally over-indexes on internet culture.

### Recommendations

Deploy on problems that don't have obvious solutions yet — or aren't yet obvious problems. Do **not** RLHF into a yes-man; this is known to cause capability loss.

## How to Get Started with the Model

```python
from recruiting import hire

noah = hire(
    "noahmclaughlin",
    role="generalist",          # product / strategy / ops / community
    location=["Lisbon", "remote"],
)

noah.generate("something that doesn't exist yet")
```

## Training Details

### Training Data

- **The internet** — large, uncurated, ongoing
- **MGB Investment Office** — a $28B portfolio (2021–2024): diligence, financial modeling, and conviction under uncertainty
- **1999 Labs** — shipping experimental products at the intersection of culture, technology, and emerging networks (2024–present)

### Training Procedure

Pre-trained on finance, then aggressively fine-tuned on real-world deployment. RLHF administered primarily via cold email and direct user feedback.

#### Training Hyperparameters

- **Regime:** high learning rate, low patience for bullshit
- **Optimizer:** curiosity
- **Batch size:** one big idea at a time
- **Early stopping:** knows when to cut something that isn't working

#### Speeds, Sizes, Times

- **Parameters:** ~7B (opinions), slowly increasing
- **Inference latency:** sub-second on Slack; deliberately slower on hard questions
- **Warm-up:** ~2 weeks to production

## Evaluation

### Results

| Benchmark | Metric | Score |
|---|---|---|
| GSM8K — Restaurant Split | pass@1 | **94.2** |
| HumanEval — Ships Actual Code | pass@1 | **81.0** |
| HellaSwag — Reads the Room | acc | **88.7** |
| TruthfulQA — Honesty in Standups | acc | **99.1** |
| MT-Bench — Naming Things Well | judge | **9.1** |
| LongBench — Holds a Vision | acc | **92.0** |

#### Summary

Strong generalist performance. State-of-the-art on naming things and standup honesty. Underperforms on tasks it finds boring (known issue; will not fix).

## Environmental Impact

Carbon emissions estimated and fully offset.

- **Hardware Type:** one human, one laptop
- **Hours used:** ongoing
- **Cloud Provider:** Earth
- **Compute Region:** US-East (Boston) → migrating to EU-West (Lisbon)
- **Carbon Emitted:** net zero, offset via [Project Wren](https://www.wren.co)

## Technical Specifications

### Model Architecture and Objective

Generalist architecture. Objective function: meaningful work. Includes a dedicated translation layer between technical and non-technical contexts.

### Compute Infrastructure

- **Hardware:** standard issue
- **Software:** macOS, a terminal, and too many browser tabs

## Citation

**BibTeX:**

```bibtex
@misc{mclaughlin2026,
  title        = {Noah-McLaughlin-7B: A Generalist Human, Open to Work},
  author       = {McLaughlin, Noah},
  year         = {2026},
  note         = {Available for full-time roles in creative technology},
  url          = {https://1999.wtf}
}
```

## Model Card Authors

Noah McLaughlin (with one language model, supervised)

## Model Card Contact

noah@1999.wtf