All Articles
How Well Does Omni Flash Hear 'Rotate the Camera 30 Degrees Right'? Measuring Where Conversational Edits Land
Public-preview Gemini Omni Flash lets you re-edit a generated video in plain language. 'Make the lighting evening' lands; 'rotate the camera 30 degrees' often misses. Here is a running log of where instructions land, sorted mechanically by comparing before/after frames.
Keeping a Long-Running Managed Agent Alive Across Sandbox Recycling — Durable Checkpoints and Idempotent Resume
A Managed Agents sandbox can be recycled out from under you. Before 40 minutes of work resets to zero, we design a durable checkpoint that pushes progress outside the sandbox and an idempotent resume that never runs a side effect twice. With working SQLite code.
"No Watermark Detected" Doesn't Mean It Isn't AI — The Asymmetry of SynthID
Images generated with Gemini carry a SynthID watermark. But a positive result and a negative result don't carry the same weight, and that asymmetry changes how you should track provenance.
The Same gemini-flash-latest Pointed to Different Models in Different Regions
Alias resolution rolls out region by region. Send the same gemini-flash-latest to Tokyo and us-central1 and, for a few days, different models answer. Here is why that quietly invalidates your comparisons, and how a one-shot probe catches it.
I Was Handing Gemini Obfuscated Stack Traces — Until retrace Went In Front, the Diagnoses Were Confident and Wrong
Release stack traces come out of R8 with the names flattened. Feed one to Gemini as-is and the diagnosis arrives calm, well-written, and wrong. Put retrace in front, match the mapping by versionCode, and forbid confident answers when you cannot restore. Numbers from 42 reports.
The Model Didn't Ship on Its Rumored Date — Read Your Context Limit From the API, Not the Headlines
July 17 came and went with no official word on Gemini 3.5 Pro. Instead of baking rumored numbers into constants, here's a context budget layer that reads the real limit from models.get and degrades quietly when input overflows.
A Japanese query won't surface its English twin — when embeddings notice language before meaning
Embed a translation pair with gemini-embedding-2 and the two halves won't be nearest neighbours, because language itself inflates similarity. Here is how I measured cross-lingual recall using translation pairs as ground truth, and what happened when I subtracted the language centroid.
A Gemini stream drops halfway — restart it, or have the model continue?
Most apps silently restart a dropped stream. Here is the arithmetic behind continuing from the partial output instead, and where to put the threshold.
What language should your system instruction be in? Measuring three approaches when most prompts arrive in the user's language
Keep the system instruction in English, or translate it into the user's language? I measured input tokens per language with countTokens, then lined up output-language match and schema compliance to find where nine tokens is enough.
I stopped storing every generation log — three retention tiers and a prompt fingerprint that keeps traceability
I was storing every Gemini API request and response body for debugging. Here is how I moved to three retention tiers plus a prompt fingerprint, and kept the ability to diagnose issues without keeping the text.
I Asked Gemini to Grade My App Store Screenshots. Everything Scored 78–85.
Ask Gemini Vision to grade App Store screenshots out of 100 and good candidates and deliberately broken ones both land at 78–85. Here is how I measured the judge's discrimination power, dropped absolute scoring, and rebuilt it as debiased pairwise comparison.
I Put =GEMINI() in 420 Cells. The Calls Kept Coming on Days I Never Opened the Sheet
A custom function calling Gemini from Sheets looks best the moment it starts working, and it does not grow from there. Here is what the 30-second ceiling, uncontrolled recalculation, and read-only cells actually cost, measured, plus the code to move to a write-back design.