01 — BOTSPEAK / TOOL FLUENCY
Operating the machines
Prompt engineering, output evaluation, knowing when to trust a result and when not to. The operating layer. Necessary, not sufficient.
§ 02 — Charter Document
why we exist
The original Skunk Works built the SR-71 in 20 months with a team of
126 people in a tent next to a wind tunnel. We are not building reconnaissance
aircraft. But we are stealing the operating model.
AI research, in its current institutional form, is optimized for the wrong things. Publications over reproducibility. Demos over deployment. Funding rounds over working code. The interesting problems sit in the gap between what big labs will commit to and what universities are equipped to do.
Northeastern's AI Skunkworks occupies that gap. Founded in 2019. Small enough to move fast. Serious enough to ship. Selective enough that being wrong is informative. Affiliated with Humanitarians AI, the 501(c)(3) that handles the institutional and fiscal infrastructure.
§ 02.A — Operating Tenets
Eight rules we keep breaking productively.
- Ship the experiment, not the product. A working prototype beats a perfect deck every time. The deck is for people who couldn't tell the difference.
- Compute is cheap. Conviction is rare. Spend the former lavishly to test the latter. The most expensive thing is hesitation on a good hunch.
- Read the paper, then ignore it. Reproduce from first principles, not from citations. Half of what's published is wrong; the other half is right for the wrong reasons.
- Small teams, short loops. Two researchers and an engineer can outpace a department. Three engineers and a researcher can outpace a directorate.
- Disagree in private. Commit in public. Then revise both, ruthlessly, on Friday. Consensus is a lagging indicator, not a goal.
- The bug is the discovery. Anomalies are the only honest data your model will ever give you. Hunt them.
- Taste is the moat. Models commoditize. Benchmarks saturate. Compute equalizes. What scales is judgment about what to build.
- Publish the failure. Especially the expensive ones. The field is poorer for everyone's silence.
§ 02.B — Irreducible Capacities
What machines can't supply.
Every AI deployment failure is a judgment-layer failure. The algorithm was running; no one was conducting it. These are the capacities we develop in skunks that no model currently possesses.
02 — CAUSAL REASONING
Why, not just what
AI finds correlations. Humans build causal models. Counterfactual thinking, interventionist reasoning, and the 'why' questions statistical models cannot answer.
03 — ETHICAL JUDGMENT
Moral reasoning under uncertainty
Value pluralism, and the limits of rule-based ethics. Cannot be reduced to optimization or alignment.
04 — AIMAGINEERING
Meaning, not just output
Generative AI produces outputs. Humans produce meaning. Creative process, aesthetic judgment, conceptual blending, and the difference between novelty and originality.
05 — EMBODIED MENTORSHIP
Presence, improvisation, attunement
The embodied skills that make mentorship effective — none of which transfer to a language model.
06 — CONDUCTING
Orchestrating, not playing
Plausibility auditing, problem formulation, tool orchestration, interpretive judgment, executive integration. Conduct the orchestra; don't play every instrument.