A sample digest
Tonight's reading
5 picks · about 30 min
This is an example. Your digest is built from the sources you pick and learns what you actually read — so it sharpens every night.
The case for small, sharp models
Task-specific models distilled to a few billion params are beating frontier models on cost, latency, and reliability for narrow jobs — and the gap is widening, not closing.
Why it matters
It flips the build-vs-buy calculus for anything agentic. If a distilled model owns your top three tasks, you stop paying frontier prices and p99 latency for work a smaller model does better.
What to apply
Pick your three highest-volume LLM calls and benchmark a distilled open model against your current one on cost, latency, and a 50-example eval set before defaulting to GPT-class.
Key points
- Latency wins compound in agent loops — every hop counts
- Distillation + a tight eval set beats prompt-engineering a big model
- Frontier models stay for the long tail, not the hot path
Because you keep opening posts on model efficiency and agent infra cost.
How we cut Postgres p99 by 40% without new hardware
Index-only scans, partial indexes on the hot predicates, and killing three N+1s did most of the work — no sharding, no read replicas, no rewrite.
Why it matters
Most "we need to scale Postgres" moments are really "we never profiled our queries." The cheapest 40% is usually sitting in your slow-query log.
What to apply
Turn on pg_stat_statements, sort by total time, and look at the top 5 — partial/covering indexes on those alone often pay for the afternoon.
Key points
- Covering indexes let the planner skip the heap entirely
- Partial indexes shrink the index to the rows you actually query
- EXPLAIN (ANALYZE, BUFFERS) shows the real cost, not the guess
You liked the piece on database internals last week — this is the applied version.
The activation metric that actually predicts retention
Across 30+ products, the single best predictor of week-4 retention is whether a user hits one specific "aha" action in the first 48 hours — not signups, not DAU.
Why it matters
If you optimize the funnel for signups instead of that 48-hour action, you grow a leaky bucket. The activation event is the real north star for a paid-acquisition test.
What to apply
Define your one activation action, instrument it, and measure % of new users who hit it in 48h. Optimize onboarding for that number, not signups.
Key points
- Retention is set in the first session, not month two
- Pick a behavioral activation event, not a vanity count
- Time-box it: "in 48 hours" beats "ever"
Matches your recent reads on growth loops and funnel economics.
The bundling moment for AI
The winners of this cycle won’t be the best model — they’ll be the ones that bundle a good-enough model into an existing distribution surface users already open daily.
Why it matters
Distribution eats model quality. A mediocre model in front of a billion users beats a great one a startup has to acquire users for, one expensive ad at a time.
Key points
- Aggregation theory applies: own demand, commoditize supply (models)
- The defensible layer is the workflow, not the weights
You return to posts on platform strategy and distribution — this is squarely there.
Running local models that actually ship
A practical setup for running quantized local models in production paths where latency, privacy, or cost rule out an API call — with the gotchas that bite at deploy time.
Why it matters
The "local model" conversation is finally past the demo stage. For privacy-sensitive or high-volume paths, on-device is now a real lever, not a science project.
What to apply
Prototype one privacy-sensitive feature on a quantized local model and measure quality vs. your API baseline before assuming you need the cloud.
Because you opened three posts on local inference and quantization this month.
This is one night. Yours is personalized.
Pick your sources, choose a delivery time, and Nodus sends a digest like this every night — to your inbox or your Kindle.
Get my first digestFree weekly digest · no credit card