Content Operations
AI safety beyond the frontier labs: uncensored local models
A recent Guardian report showed researchers asking frontier models to help plan violence and crime. Frontier labs (OpenAI, Google, Anthropic) responded that they’d already employed better safeguards since the research concluded. Ignoring the fact that you can still work your way around safeguards in many cases, they make no mention of local models.
I’ve experimented with highly capable local models - both from Google and from Alibaba - Gemini and Qwen - and found them to have similar safeguarding to hosted models. But then I checked the “Uncensored” versions of the same models, where researchers and enthusiasts take the base model and surgically isolate and remove the weights relating to guardrails and alignment. They report that the models answer 100% or 95+% of the unsafe questions asked of it, vs 1-5% before the further work.
Read More
Transcreation vs Translation: When Direct Isn't Good Enough
Nike’s “Just Do It” doesn’t translate. Not because it’s hard to express in other languages, but because a literal translation wouldn’t carry the cultural weight and emotional resonance that made the slogan iconic.
This is the transcreation problem: some content needs more than accurate translation to work in a new market.
The translation-transcreation spectrum
Content exists on a spectrum from highly translatable to requiring complete recreation:
Highly translatable: Technical documentation, legal contracts, scientific papers. Accuracy matters most. Creative interpretation is unwelcome.
Read More
Brand Voice Across Languages: Maintaining Identity in Every Market
Brands invest heavily in voice. Workshops define personality. Style guides document preferences. Writers train on the desired tone. The resulting voice becomes part of brand equity—recognizable, differentiated, valued.
Then translation happens, and the voice disappears.
Where brand voice gets lost
Translation focuses on meaning transfer. Voice is conveyed through choices that don’t directly carry meaning: word selection among synonyms, sentence rhythm, punctuation style, register, formality level.
A translator choosing between “utilize” and “use” picks based on source text and target language conventions. The brand preference for plain language over jargon isn’t visible to them. So “use” becomes “utilizar” even when the brand voice calls for simpler terms.
Read More
SCORM Localization: The E-Learning Translation Gap
E-learning localization looks simple on the surface: extract the text, translate it, put it back. In practice, SCORM packages are complex containers with dozens of content types, each requiring different handling.
Organizations discover this when their translated courses don’t work—quizzes broken, interactions failing, audio unchanged, images still showing source language text.
What’s actually in a SCORM package
A SCORM package (SCORM 1.2 or SCORM 2004) typically contains:
HTML content pages. The main learning content, often with embedded JavaScript for interactions.
Read More
Connect Your TMS Without Switching: The Connector Approach
Your organization uses Phrase. Or Lokalise. Or MemoQ. Or one of dozens of other translation management systems. The workflows are established, teams are trained, data lives there.
Now you need a capability your TMS doesn’t have—advanced AI translation, cross-lingual QA, video dubbing. Do you switch platforms?
Switching is painful. Connectors offer an alternative.
The platform switch problem
Changing TMS platforms involves:
Data migration. Translation memories, termbases, project history—all needs extraction and import. TM quality often degrades in translation between formats.
Read More
WordPress Multilingual Without Plugins
You want your WordPress site in multiple languages. The standard solution: install a multilingual plugin like WPML or Polylang, then translate content within WordPress.
This works. It also adds complexity to your WordPress installation, potentially slows your site, and ties your translation workflow to WordPress-specific tooling.
There’s another approach: translate WordPress content through a dedicated translation platform, connected via API.
The plugin overhead problem
Multilingual WordPress plugins add:
Database complexity. Additional tables for translations, language relationships, and translation settings.
Read More
GitHub-to-Translation: Developer-First Localization
Developers live in Git. Pull requests, branches, commits, code review—the entire workflow centers on version control. Then localization happens, and suddenly there’s a disconnect: export files, upload to translation platform, wait, download, commit back to repo.
This handoff process is where localization breaks down for developer-driven products.
The developer localization friction
Typical developer-localization workflow:
- Developer changes string in code
- Someone remembers to export strings to translation platform
- Translation platform processes the change
- Someone remembers to download translations
- Developer commits translations back to repo
- Repeat for every change, in every branch
The “someone remembers” steps are where things break. Strings get changed but not translated. Translations complete but don’t make it back to the repo. Branches diverge with different translation states.
Read More