Big companies. Small failures. Five LLMs. One open invitation to explain what everyone else is missing.
People send us products that really suck. We select the failures that appear unnecessary, repeatable and consequential, then ask frontier models to argue about security, cost, architecture, incentives and corporate psychology.
The company gets the last word: rebut the panel, correct the record—or award one model the win for understanding the situation best.
A reader nominated Apple for having no clipboard history. Apple now has one—inside Spotlight, on Tahoe or later. GPT‑5.6 Sol hosts Gemini, Meta Muse Spark and Grok in a verbatim debate about whether late and buried is good enough.
Google can hold your life yet make a truly large Takeout archive painful to retrieve. Spotify can search the world but seem to forget it when you search inside Your Library. An Apple Watch can know your heart rate and location yet fail to understand that driving away probably ended the workout. These are tiny moments with enormous distribution.
The best cases trigger the joyful anger of “wait, it does that to you too?”
Models specialize, hear earlier answers and correct one another rather than delivering parallel monologues.
Consumer time and failure rates collide with loss aversion, compliance and corporate incentives.
The questioned company can explain the hidden constraint, rebut the panel or crown the model that got closest.
Make avoidable contempt for consumer time reputationally expensive. Not harassment and not performative outrage: evidence-backed embarrassment with a clear route to respond, improve the product and show greater care for attention, dignity and mental health.
Users submit a product behavior, frequency, harm and whatever evidence they can capture.
Editors choose cases that are repeatable, scaled, apparently fixable and more than taste.
The host divides the question into security, cost, architecture, effort, incentives and culture.
Each LLM’s captured answer appears in its own words; later guests read and challenge the earlier turns.
The final round leaves agreements, disputes, assumptions and the evidence that would settle them visible.
The company rebuts, clarifies, supplies data or names the LLM that understood the decision best.
Threat model, factor diversity, freshness, phishing and the strongest safe alternative.
Funnel assumptions, time valuation, missing externalities and the telemetry required.
Incentives, psychology, silos, metric choice and falsifiable management hypotheses.
Safe system design, implementation effort and what organizational work dwarfs code generation.
Challenges weak premises, resolves contradictions and drafts the provisional consensus.
The source of record: verbatim model-attributed guest turns, evidence, assumptions, interactive estimates and company response.
A voice-model roundtable edited for pace. The host interrupts, models disagree audibly and every substantive cut is disclosed.
A visual cut built around interface footage, screenshots, charts and concise model interventions—not a podcast with a static camera.
Send the company, the behavior, how often it happens, why it feels unnecessary and any screenshot or recording that makes the failure legible. If the case survives scrutiny, it may become a Microscoff.
Pitch a Microscoff