Bring a judgeable counterexample
One anecdote can be enough if it includes the conditions, the measurement, and why it contradicts the claim.
Prove me wrong
The strongest counterexample gets published with credit. If it lands, the original essay gets a dated concession instead of a quiet edit.
Credit is explicit
Accepted rebuttals are published with your name, link, and preferred wording, unless you ask for anonymity.
Concessions stay visible
If the rebuttal changes the claim, the source essay gets a dated note that says what changed.
Open claims
Five places to take a shot
- 01
From Local models won the long tail
For high-volume, narrow production tasks, local or fine-tuned open models will take most of the call volume from frontier APIs.
- What would move me
- A production workload with at least 10,000 weekly calls where local or fine-tuned open models were evaluated seriously but frontier APIs stayed cheaper or more accurate for the majority of narrow calls.
- What counts
- Detailed cost, latency, quality, and operations data that shows the local-model path lost after a fair trial.
- 02
From Vibe-coded code has more bugs. Price it in.
AI-assisted coding carries a measurable bug tax unless the work has strong tests, low blast radius, or throwaway scope.
- What would move me
- A credible team-level comparison where AI-assisted code increased throughput without a higher serious-defect rate after matching for code area, reviewer quality, and test coverage.
- What counts
- Public study, internal writeup, or reproducible measurement with enough context to separate tooling effect from team maturity.
- 03
From Postgres is a queue. Stop reaching for Kafka.
For most internal workloads under roughly 1,000 events per second, Postgres is a better default queue than Kafka.
- What would move me
- A workload under that threshold where Postgres queueing failed for reasons Kafka solved without adding disproportionate operational load.
- What counts
- Incident notes, benchmark data, or architecture notes showing the failure mode was intrinsic to Postgres queueing, not an implementation mistake.
- 04
From Your MCP server is a prod dependency
MCP servers used by agents should be treated as production dependencies with versioning, health checks, and scoped credentials.
- What would move me
- A serious agent workflow where unpinned or unmonitored MCP servers remain safe because the surrounding architecture makes their failure non-production-impacting.
- What counts
- Concrete architecture with blast-radius controls, fallback behavior, and evidence that malformed or compromised MCP output cannot affect user-facing decisions.
- 05
From The output token tax
In agent products, verbose outputs and hidden reasoning fields are now a material margin leak compared with tighter schemas.
- What would move me
- A production agent where output verbosity improves quality or supportability enough to outweigh token cost after measuring both.
- What counts
- Before/after data on output tokens, task quality, debugging time, and provider bill that shows verbose output paying for itself.