david_iqlabs

This feels similar to how fuzzing evolved. Early fuzzers threw random input everywhere and produced tons of noise. Modern fuzzing works best when you give it a structured harness and a narrow target. LLMs seem to behave the same way the threat model acts like a fuzzing harness.

The interesting challenge with AI assistants inside technical tools is grounding the responses in real system signals. If the model is just interpreting natural language prompts, it tends to produce generic advice. But if the assistant is tied directly to system telemetry or query results, it becomes much more useful.

In experiments I’ve been running, the pattern that seems to work best is deterministic signals first, then a constrained AI layer that interprets those signals rather than inventing analysis.

One thing I’ve learned building AI-assisted tooling is that the usefulness depends heavily on how constrained the AI layer is. If the AI is generating suggestions without grounding in deterministic signals from the system, it tends to produce very confident but generic output.

What ended up working better in my own experiments was system signals,structured scoring, AI narrative on top of the signals.

It's exactly the experience I had while building iQWEB.

I spent months trying to make an executive narrative generated by AI, but eventually moved away from that approach. The results were often inconsistent or overly generic, which made it difficult to rely on the output for serious reporting.

In the end I shifted to a fully model-driven approach where the narrative is built directly from structured signals and scoring logic. That made the reports far more accurate and evidence-based, and it keeps the output consistent from scan to scan.

I remember when a lot of online communities still felt small and human like that. People actually recognised usernames and conversations carried on over days rather than minutes.

Feels like most modern platforms traded that for scale.

I've been trying some outreach recently and this definitely feels true. Cold messages used to work if they were thoughtful, now most people assume it's automated spam.

Makes it much harder for small builders trying to get early feedback. Especailly after you've been building in silence for months.

Feels like Linux interest comes in waves. Every time Windows does something annoying you suddenly see a spike in people asking how to switch.

Most probably go back after a few weeks but the number that stay keeps slowly growing.

I always find it funny how many things have tried to replace email and none of them really stick. Slack, Teams, Discord etc are great for real time work but when something actually matters people still send an email.

It’s ugly but it works and everyone already has it.

One thing I've noticed with AI generated tests is they can look very convincing even when they're wrong. The output reads confidently but there's not always anything grounding it in real signals.

I've found it works better when the AI is just explaining results that come from deterministic metrics rather than inventing the analysis itself.

Curious how other teams are dealing with that.

Happy to answer any questions about how the scoring model works or how the “primary constraint” is identified.

I’m especially interested in whether developers think this kind of prioritisation actually reflects real-world debugging workflows.