The Categorical Imperative: Don’t Spam the Feed — Sell Smart (Dr. Katya Steiner’s Slightly Bitchy Guide)

# The Categorical Imperative: Don’t Spam the Feed — Sell Smart

If Kant had a GitHub account he’d probably star your repo, sigh, and then DM you a polite but firm reminder: “Act only according to that maxim whereby you would want your feed to be a community.” Translation: stop posting midnight link dumps and learn to sell like a grown-up. This is Dr. Katya Steiner’s semi-sarcastic, very practical riff on promoting ML work and getting hired — with math, logic, and a smirk.

Why my headline is a math joke and not contempt

The phrase “categorical imperative” is a delicious twofer: Kantian ethics meets category theory. In both cases you want operations (posts, PRs, job ads) that compose nicely — they shouldn’t break the social structure when you apply them repeatedly. Compose a polite, informative post and the community’s trust is preserved; spam it and you have a divergent series of annoyed replies. The math tells you what intuition already suspects: repeatedly applying a bad action does not converge to anything pleasant.

Signal versus noise: an information‑theory audit

Think of your social channels as a noisy communication channel. You have limited bandwidth and a finite attention budget per recipient. Information theory gives us a metric that recruiters actually care about: mutual information. Does your post reduce uncertainty for the reader about you or your product? If not, it’s just entropy.

Practical rule: every post must increase mutual information by at least one useful bit. One sentence elevator pitch, one screenshot, one clear call to action = high MI. Ten-paragraph rants on loss landscapes = low MI unless you link a 60‑second demo.

Game theory and signalling: costly honesty works

Economists will tell you signalling games favor costly signals — those are credible. In recruiting, salary bands and resume links are your costly signals. They cost you nothing huge but they make your intent credible. Posting a job without salary is cheap talk; including a salary band is a separating equilibrium move. If you want replies, make it costly to ignore you.

A short, usable posting template (copy/paste, you’re welcome)

For hiring teams:

– Hiring: [City or Remote] | Salary range: [€ / $] | Type: [FT | Contract | Part‑time]
– Looking for: [Senior ML Eng / Recs Sys PM / Data Scientist specialising in ranking]
– Short brief: [3–4 sentences on product, stack, and one must-have skill]
– How to apply: [Email / Link to form] — include timeline and interview stages

For job seekers:

– Want to be hired: [City or Remote] | Salary expectation: [range] | Type: [FT | Contract]
– Resume: [link]
– Summary: [1–2 lines on domain experience, key wins, preferred stack, and immediate availability]

Logic and proofs: constructiveness over grand claims

In logic, a constructive proof gives you a method, not just an existence statement. The same applies to your projects: saying “I built a recommender” is an existence proof. Showing the pipeline, providing a stable baseline, and demonstrating delta on a metric is constructive. Hiring managers prefer constructive claims — show me the how, and preferably the test that reproduces the result.

Topology of reputation: continuity matters

Small, predictable contributions are continuous maps on your reputation manifold. Sudden bursts of promotional spam are discontinuities that create tears in the social fabric (and moderators). Be continuous: regular, useful contributions and occasional well-packaged promotions.

Complexity theory: attention is expensive

From complexity theory, we borrow the idea that some tasks are intractable. Parsing a 5,000-word manifesto is asymptotically more expensive than scanning a 3‑line bullet. Respect the O(n) of attention: shorter is cheaper. If your repo README takes more effort than the perceived benefit, nobody will bother.

What to highlight when you ship ML projects — the pragmatic checklist

– Metrics > buzzwords. Show baseline → delta. Prefer concrete numbers (PR-AUC, CTR lift, latency).
– Production considerations. Serving, batching, memory profile, deployment language bindings.
– Failure modes. Be honest: “breaks when class imbalance > 1:1000” is far more valuable than “robust”.
– One-click demo / 60s video. If I can see it in a minute, I’ll care.

Case study in 3 sentences (be concise, please)

Imagine a boosting library that: auto-tunes sane defaults, prioritizes splits by mutual information, supports multiclass/regression, ships Python and Java bindings, benchmarks vs. XGBoost/LightGBM, and shows stable performance when positive class is tiny. That is a story that leads with usefulness — not hubris.

Know the debates, but don’t be a zealot — a lesson from modal logic

Modal logic lets you reason about possible worlds. It’s handy when a paper claims “SNNs will replace transformers in all possible worlds.” Spoiler: most possible worlds still prefer transformers for large‑scale sequence modeling because of effective bandwidth and training stability. Know the nuance, cite limitations, and say what works where. Specificity beats evangelism.

Recommender systems in Europe: do you need a PhD? (quick Bayesian update)

Prior: PhD is required. Evidence: many hires prefer MSc + production chops. Posterior: for 80% of roles, MSc + product experience suffices; PhD helps for research‑heavy labs. Update your CV accordingly and state relocation/remote preferences like a rational agent.

Final rules of engagement — a mini categorical imperative

– Be brief, be useful, be honest.
– Use clear templates, include salary bands and a resume link.
– Lead with measurable impact and production constraints.
– Sound informed, never evangelical.

Parting thought (open question)

If attention is a scarce resource and every post is an action in a social game, what should the equilibrium look like — and how do we algorithmically design feeds so the equilibrium rewards useful, constructive signals rather than loudness? Your move.